Merge pull request #226 from nils-asmussen/fix/Gauss

Fix compiling of MSource::Gauss for single precision

.gitignore

@@ -83,6 +83,7 @@ ltmain.sh
 .Trashes
 ehthumbs.db
 Thumbs.db
+.dirstamp
 
 # build directory #
 ###################
@@ -93,14 +94,12 @@ build*/*
 *.xcodeproj/*
 build.sh
 .vscode
+*.code-workspace
 
 # Eigen source #
 ################
-lib/Eigen/*
-
-# FFTW source #
-################
-lib/fftw/*
+Grid/Eigen
+Eigen/*
 
 # libtool macros #
 ##################
@@ -111,15 +110,8 @@ m4/libtool.m4
 ################
 gh-pages/
 
-# Buck files #
-##############
-.buck*
-buck-out
-BUCK
-make-bin-BUCK.sh
-
 # generated sources #
 #####################
-lib/qcd/spin/gamma-gen/*.h
-lib/qcd/spin/gamma-gen/*.cc
-
+Grid/qcd/spin/gamma-gen/*.h
+Grid/qcd/spin/gamma-gen/*.cc
+Grid/util/Version.h

.travis.yml

@@ -9,68 +9,11 @@ matrix:
     - os:        osx
       osx_image: xcode8.3
       compiler: clang
-    - compiler: gcc
-      dist: trusty
-      sudo: required
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-4.9
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      env: VERSION=-4.9
-    - compiler: gcc
-      dist: trusty
-      sudo: required
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-5
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      env: VERSION=-5
-    - compiler: clang
-      dist: trusty
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-4.8
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      env: CLANG_LINK=http://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
-    - compiler: clang
-      dist: trusty
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-4.8
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      env: CLANG_LINK=http://llvm.org/releases/3.7.0/clang+llvm-3.7.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
+      env: PREC=single
+    - os:        osx
+      osx_image: xcode8.3
+      compiler: clang
+      env: PREC=double
       
 before_install:
     - export GRIDDIR=`pwd`
@@ -78,9 +21,11 @@ before_install:
     - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export PATH="${GRIDDIR}/clang/bin:${PATH}"; fi
     - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc openssl; fi
     
 install:
+    - export CWD=`pwd`
+    - echo $CWD
     - export CC=$CC$VERSION
     - export CXX=$CXX$VERSION
     - echo $PATH
@@ -93,22 +38,24 @@ install:
     - which $CXX
     - $CXX --version
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export EXTRACONF='--with-openssl=/usr/local/opt/openssl'; fi
     
 script:
     - ./bootstrap.sh
     - mkdir build
     - cd build
-    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none
+    - mkdir lime
+    - cd lime
+    - mkdir build
+    - cd build
+    - wget http://usqcd-software.github.io/downloads/c-lime/lime-1.3.2.tar.gz
+    - tar xf lime-1.3.2.tar.gz
+    - cd lime-1.3.2
+    - ./configure --prefix=$CWD/build/lime/install
+    - make -j4
+    - make install
+    - cd $CWD/build
+    - ../configure --enable-precision=$PREC --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
     - make -j4 
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
-    - echo make clean
-    - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
-    - make -j4
-    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - make check
-    - echo make clean
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto ; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then make -j4; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
-
-

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/Grid.h

@@ -41,7 +41,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #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>
 
 #endif

Grid/GridCore.h

@@ -48,6 +48,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/serialisation/Serialisation.h>
 #include <Grid/threads/Threads.h>
 #include <Grid/util/Util.h>
+#include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 
 #include <Grid/cartesian/Cartesian.h>    
 #include <Grid/tensors/Tensors.h>      

Grid/GridStd.h

@@ -7,6 +7,7 @@
 #include <cassert>
 #include <complex>
 #include <vector>
+#include <string>
 #include <iostream>
 #include <iomanip>
 #include <random>
@@ -18,6 +19,7 @@
 #include <ctime>
 #include <sys/time.h>
 #include <chrono>
+#include <zlib.h>
 
 ///////////////////
 // Grid config

Grid/Grid_Eigen_Dense.h

@@ -1,4 +1,9 @@
 #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"

Grid/Makefile.am

@@ -0,0 +1,63 @@
+extra_sources=
+extra_headers=
+
+if BUILD_COMMS_MPI3
+  extra_sources+=communicator/Communicator_mpi3.cc
+  extra_sources+=communicator/Communicator_base.cc
+  extra_sources+=communicator/SharedMemoryMPI.cc
+  extra_sources+=communicator/SharedMemory.cc
+endif
+
+if BUILD_COMMS_NONE
+  extra_sources+=communicator/Communicator_none.cc
+  extra_sources+=communicator/Communicator_base.cc
+  extra_sources+=communicator/SharedMemoryNone.cc
+  extra_sources+=communicator/SharedMemory.cc
+endif
+
+if BUILD_HDF5
+  extra_sources+=serialisation/Hdf5IO.cc 
+  extra_headers+=serialisation/Hdf5IO.h
+  extra_headers+=serialisation/Hdf5Type.h
+endif
+
+all: version-cache
+
+version-cache:
+	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
+		a="uncommited changes";\
+	else\
+		a="clean";\
+	fi;\
+	echo "`git log -n 1 --format=format:"#define GITHASH \\"%H:%d $$a\\"%n" HEAD`" > vertmp;\
+	if [ -e version-cache ]; then\
+		d=`diff vertmp version-cache`;\
+		if [ "$${d}" != "" ]; then\
+			mv vertmp version-cache;\
+			rm -f Version.h;\
+		fi;\
+	else\
+		mv vertmp version-cache;\
+		rm -f Version.h;\
+	fi;\
+	rm -f vertmp
+
+Version.h:
+	cp version-cache Version.h
+
+.PHONY: version-cache
+
+#
+# Libraries
+#
+include Make.inc
+include Eigen.inc
+
+lib_LIBRARIES = libGrid.a
+
+CCFILES += $(extra_sources)
+HFILES  += $(extra_headers) Config.h Version.h
+
+libGrid_a_SOURCES              = $(CCFILES)
+libGrid_adir                   = $(includedir)/Grid
+nobase_dist_pkginclude_HEADERS = $(HFILES) $(eigen_files) $(eigen_unsupp_files)

Grid/algorithms/Algorithms.h

@@ -1,6 +1,6 @@
     /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/Algorithms.h
 
@@ -37,37 +37,27 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/approx/Chebyshev.h>
 #include <Grid/algorithms/approx/Remez.h>
 #include <Grid/algorithms/approx/MultiShiftFunction.h>
+#include <Grid/algorithms/approx/Forecast.h>
 
+#include <Grid/algorithms/iterative/Deflation.h>
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
-
-// Lanczos support
-//#include <Grid/algorithms/iterative/MatrixUtils.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>
 
-// Eigen/lanczos
-// EigCg
-// MCR
-// Pcg
-// Multishift CG
-// Hdcg
-// GCR
-// etc..
-
-// integrator/Leapfrog
-// integrator/Omelyan
-// integrator/ForceGradient
-
-// montecarlo/hmc
-// montecarlo/rhmc
-// montecarlo/metropolis
-// etc...
-
-
 #endif

Grid/algorithms/CoarsenedMatrix.h

@@ -103,29 +103,32 @@ namespace Grid {
     GridBase *CoarseGrid;
     GridBase *FineGrid;
     std::vector<Lattice<Fobj> > subspace;
+    int checkerboard;
 
-    Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid) : 
-      CoarseGrid(_CoarseGrid),
+  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
+    CoarseGrid(_CoarseGrid),
       FineGrid(_FineGrid),
-      subspace(nbasis,_FineGrid)
+      subspace(nbasis,_FineGrid),
+      checkerboard(_checkerboard)
 	{
 	};
   
     void Orthogonalise(void){
       CoarseScalar InnerProd(CoarseGrid); 
+      std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
       blockOrthogonalise(InnerProd,subspace);
+      std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
+      blockOrthogonalise(InnerProd,subspace);
+      //      std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
+      //      CheckOrthogonal();
     } 
     void CheckOrthogonal(void){
       CoarseVector iProj(CoarseGrid); 
       CoarseVector eProj(CoarseGrid); 
-      Lattice<CComplex> pokey(CoarseGrid);
-
-      
       for(int i=0;i<nbasis;i++){
 	blockProject(iProj,subspace[i],subspace);
-
 	eProj=zero; 
-	for(int ss=0;ss<CoarseGrid->oSites();ss++){
+	parallel_for(int ss=0;ss<CoarseGrid->oSites();ss++){
 	  eProj._odata[ss](i)=CComplex(1.0);
 	}
 	eProj=eProj - iProj;
@@ -137,6 +140,7 @@ namespace Grid {
       blockProject(CoarseVec,FineVec,subspace);
     }
     void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+      FineVec.checkerboard = subspace[0].checkerboard;
       blockPromote(CoarseVec,FineVec,subspace);
     }
     void CreateSubspaceRandom(GridParallelRNG &RNG){
@@ -147,6 +151,7 @@ namespace Grid {
       Orthogonalise();
     }
 
+    /*
     virtual void CreateSubspaceLanczos(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) 
     {
       // Run a Lanczos with sloppy convergence
@@ -195,7 +200,7 @@ namespace Grid {
 	  std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
 	}
     }
-
+    */
     virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
 
       RealD scale;
@@ -206,6 +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;
@@ -290,13 +296,58 @@ namespace Grid {
       return norm2(out);
     };
 
-    RealD Mdag (const CoarseVector &in, CoarseVector &out){ 
-      return M(in,out);
+    RealD Mdag (const CoarseVector &in, CoarseVector &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);
     };
 
-    // 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){};
+    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;
+      }();
+
+      parallel_for(int ss=0;ss<Grid()->oSites();ss++){
+        siteVector res = zero;
+        siteVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        SE=Stencil.GetEntry(ptype,point,ss);
+
+        if(SE->_is_local&&SE->_permute) {
+          permute(nbr,in._odata[SE->_offset],ptype);
+        } else if(SE->_is_local) {
+          nbr = in._odata[SE->_offset];
+        } else {
+          nbr = Stencil.CommBuf()[SE->_offset];
+        }
+
+        res = res + A[point]._odata[ss]*nbr;
+
+        vstream(out._odata[ss],res);
+      }
+    };
+
+    void Mdiag(const CoarseVector &in, CoarseVector &out){
+      Mdir(in, out, 0, 0); // use the self coupling (= last) point of the stencil
+    };
 
     CoarsenedMatrix(GridCartesian &CoarseGrid) 	: 
 
@@ -412,7 +463,7 @@ namespace Grid {
       std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
 #endif
       //      ForceHermitian();
-      AssertHermitian();
+      // AssertHermitian();
       // ForceDiagonal();
     }
     void ForceDiagonal(void) {

Grid/algorithms/FFT.h

@@ -230,6 +230,7 @@ namespace Grid {
       // Barrel shift and collect global pencil
       std::vector<int> lcoor(Nd), gcoor(Nd);
       result = source;
+      int pc = processor_coor[dim];
       for(int p=0;p<processors[dim];p++) {
         PARALLEL_REGION
         {
@@ -240,7 +241,8 @@ namespace Grid {
           for(int idx=0;idx<sgrid->lSites();idx++) {
             sgrid->LocalIndexToLocalCoor(idx,cbuf);
             peekLocalSite(s,result,cbuf);
-            cbuf[dim]+=p*L;
+	    cbuf[dim]+=((pc+p) % processors[dim])*L;
+	    //            cbuf[dim]+=p*L;
             pokeLocalSite(s,pgbuf,cbuf);
           }
         }
@@ -278,7 +280,6 @@ namespace Grid {
       flops+= flops_call*NN;
       
       // writing out result
-      int pc = processor_coor[dim];
       PARALLEL_REGION
       {
         std::vector<int> clbuf(Nd), cgbuf(Nd);

Grid/algorithms/LinearOperator.h

@@ -51,7 +51,7 @@ namespace Grid {
 
       virtual void Op     (const Field &in, Field &out) = 0; // Abstract base
       virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
-      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2) = 0;
       virtual void HermOp(const Field &in, Field &out)=0;
     };
 
@@ -162,15 +162,10 @@ namespace Grid {
 	_Mat.M(in,out);
       }
       void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	ComplexD dot;
-
 	_Mat.M(in,out);
 	
-	dot= innerProduct(in,out);
-	n1=real(dot);
-
-	dot = innerProduct(out,out);
-	n2=real(dot);
+	ComplexD dot= innerProduct(in,out); n1=real(dot);
+	n2=norm2(out);
       }
       void HermOp(const Field &in, Field &out){
 	_Mat.M(in,out);
@@ -183,19 +178,21 @@ namespace Grid {
     //////////////////////////////////////////////////////////
 
     template<class Field>
-      class SchurOperatorBase :  public LinearOperatorBase<Field> {
+    class SchurOperatorBase :  public LinearOperatorBase<Field> {
     public:
       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);
+      Field tmp(in._grid);
+      tmp.checkerboard = in.checkerboard;
 	ni=Mpc(in,tmp);
 	no=MpcDag(tmp,out);
       }
-      void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+      out.checkerboard = in.checkerboard;
 	MpcDagMpc(in,out,n1,n2);
       }
-      void HermOp(const Field &in, Field &out){
+      virtual void HermOp(const Field &in, Field &out){
 	RealD n1,n2;
 	HermOpAndNorm(in,out,n1,n2);
       }
@@ -212,22 +209,22 @@ namespace Grid {
       void OpDir  (const Field &in, Field &out,int dir,int disp) {
 	assert(0);
       }
-
     };
     template<class Matrix,class Field>
-      class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
-    protected:
-      Matrix &_Mat;
+    class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
     public:
+      Matrix &_Mat;
       SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
       virtual  RealD Mpc      (const Field &in, Field &out) {
-	Field tmp(in._grid);
-//	std::cout <<"grid pointers: in._grid="<< in._grid << " out._grid=" << out._grid << "  _Mat.Grid=" << _Mat.Grid() << " _Mat.RedBlackGrid=" << _Mat.RedBlackGrid() << std::endl;
+      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;
 	_Mat.Mooee(in,out);
 	return axpy_norm(out,-1.0,tmp,out);
       }
@@ -270,7 +267,6 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
-
     template<class Matrix,class Field>
       class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
     protected:
@@ -299,6 +295,82 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    // Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
+    // Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo ) Moo^-1 phi=eta ; psi = Moo^-1 phi
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
+    template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    //  Staggered use
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Matrix,class Field>
+      class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
+    protected:
+      Matrix &_Mat;
+      Field tmp;
+      RealD mass;
+      double tMpc;
+      double tIP;
+      double tMeo;
+      double taxpby_norm;
+      uint64_t ncall;
+    public:
+      void Report(void)
+      {
+	std::cout << GridLogMessage << " HermOpAndNorm.Mpc "<< tMpc/ncall<<" usec "<<std::endl;
+	std::cout << GridLogMessage << " HermOpAndNorm.IP  "<< tIP /ncall<<" usec "<<std::endl;
+	std::cout << GridLogMessage << " Mpc.MeoMoe        "<< tMeo/ncall<<" usec "<<std::endl;
+	std::cout << GridLogMessage << " Mpc.axpby_norm    "<< taxpby_norm/ncall<<" usec "<<std::endl;
+      }
+      SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
+      { 
+	assert( _Mat.isTrivialEE() );
+	mass = _Mat.Mass();
+	tMpc=0;
+	tIP =0;
+        tMeo=0;
+        taxpby_norm=0;
+	ncall=0;
+      }
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+	ncall++;
+	tMpc-=usecond();
+	n2 = Mpc(in,out);
+	tMpc+=usecond();
+	tIP-=usecond();
+	ComplexD dot= innerProduct(in,out);
+	tIP+=usecond();
+	n1 = real(dot);
+      }
+      virtual void HermOp(const Field &in, Field &out){
+	ncall++;
+	tMpc-=usecond();
+	_Mat.Meooe(in,out);
+	_Mat.Meooe(out,tmp);
+	tMpc+=usecond();
+	taxpby_norm-=usecond();
+	axpby(out,-1.0,mass*mass,tmp,in);
+	taxpby_norm+=usecond();
+      }
+      virtual  RealD Mpc      (const Field &in, Field &out) {
+	tMeo-=usecond();
+	_Mat.Meooe(in,out);
+	_Mat.Meooe(out,tmp);
+	tMeo+=usecond();
+	taxpby_norm-=usecond();
+	RealD nn=axpby_norm(out,-1.0,mass*mass,tmp,in);
+	taxpby_norm+=usecond();
+	return nn;
+      }
+      virtual  RealD MpcDag   (const Field &in, Field &out){
+	return Mpc(in,out);
+      }
+      virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+	assert(0);// Never need with staggered
+      }
+    };
+    template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
 
 
     /////////////////////////////////////////////////////////////
@@ -307,6 +379,12 @@ 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 {
@@ -314,6 +392,14 @@ namespace Grid {
       virtual void operator() (const Field &in, Field &out) = 0;
     };
 
+    template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
+    public:
+      void operator() (const Field &in, Field &out){
+	out = in;
+      };
+    };
+
+
     /////////////////////////////////////////////////////////////
     // Base classes for Multishift solvers for operators
     /////////////////////////////////////////////////////////////
@@ -336,6 +422,64 @@ namespace Grid {
      };
     */
 
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hermitian operator Linear function and operator function
+  ////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Field>
+    class HermOpOperatorFunction : public OperatorFunction<Field> {
+      void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
+	Linop.HermOp(in,out);
+      };
+    };
+
+    template<typename Field>
+      class PlainHermOp : public LinearFunction<Field> {
+    public:
+      LinearOperatorBase<Field> &_Linop;
+      
+      PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
+      {}
+      
+      void operator()(const Field& in, Field& out) {
+	_Linop.HermOp(in,out);
+      }
+    };
+
+    template<typename Field>
+    class FunctionHermOp : public LinearFunction<Field> {
+    public:
+      OperatorFunction<Field>   & _poly;
+      LinearOperatorBase<Field> &_Linop;
+      
+      FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop) 
+	: _poly(poly), _Linop(linop) {};
+      
+      void operator()(const Field& in, Field& out) {
+	_poly(_Linop,in,out);
+      }
+    };
+
+  template<class Field>
+  class Polynomial : public OperatorFunction<Field> {
+  private:
+    std::vector<RealD> Coeffs;
+  public:
+    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);
+      AtoN = in;
+      out = AtoN*Coeffs[0];
+      for(int n=1;n<Coeffs.size();n++){
+	Mtmp = AtoN;
+	Linop.HermOp(Mtmp,AtoN);
+	out=out+AtoN*Coeffs[n];
+      }
+    };
+  };
 
 }
 

Grid/algorithms/SparseMatrix.h

@@ -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;

Grid/algorithms/approx/Chebyshev.h

@@ -8,6 +8,7 @@
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Christoph Lehner <clehner@bnl.gov>
 
     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
@@ -33,41 +34,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  // Simple general polynomial with user supplied coefficients
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field>
-  class HermOpOperatorFunction : public OperatorFunction<Field> {
-    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-      Linop.HermOp(in,out);
-    };
-  };
-
-  template<class Field>
-  class Polynomial : public OperatorFunction<Field> {
-  private:
-    std::vector<RealD> Coeffs;
-  public:
-    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);
-      AtoN = in;
-      out = AtoN*Coeffs[0];
-//            std::cout <<"Poly in " <<norm2(in)<<" size "<< Coeffs.size()<<std::endl;
-//            std::cout <<"Coeffs[0]= "<<Coeffs[0]<< " 0 " <<norm2(out)<<std::endl;
-      for(int n=1;n<Coeffs.size();n++){
-	Mtmp = AtoN;
-	Linop.HermOp(Mtmp,AtoN);
-	out=out+AtoN*Coeffs[n];
-//            std::cout <<"Coeffs "<<n<<"= "<< Coeffs[n]<< " 0 " <<std::endl;
-//		std::cout << n<<" " <<norm2(out)<<std::endl;
-      }
-    };
-  };
+struct ChebyParams : Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
+				  RealD, alpha,  
+				  RealD, beta,   
+				  int, Npoly);
+};
 
   ////////////////////////////////////////////////////////////////////////////////////////////
   // Generic Chebyshev approximations
@@ -82,8 +54,10 @@ namespace Grid {
 
   public:
     void csv(std::ostream &out){
-	RealD diff = hi-lo;
-      for (RealD x=lo-0.2*diff; x<hi+0.2*diff; x+=(hi-lo)/1000) {
+      RealD diff = hi-lo;
+      RealD delta = (hi-lo)*1.0e-9;
+      for (RealD x=lo; x<hi; x+=delta) {
+	delta*=1.1;
 	RealD f = approx(x);
 	out<< x<<" "<<f<<std::endl;
       }
@@ -99,6 +73,7 @@ namespace Grid {
     };
 
     Chebyshev(){};
+    Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
     Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
     Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};
 
@@ -193,6 +168,47 @@ namespace Grid {
       return sum;
     };
 
+    RealD approxD(RealD x)
+    {
+      RealD Un;
+      RealD Unm;
+      RealD Unp;
+      
+      RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
+      
+      RealD U0=1;
+      RealD U1=2*y;
+      
+      RealD sum;
+      sum = Coeffs[1]*U0;
+      sum+= Coeffs[2]*U1*2.0;
+      
+      Un =U1;
+      Unm=U0;
+      for(int i=2;i<order-1;i++){
+	Unp=2*y*Un-Unm;
+	Unm=Un;
+	Un =Unp;
+	sum+= Un*Coeffs[i+1]*(i+1.0);
+      }
+      return sum/(0.5*(hi-lo));
+    };
+    
+    RealD approxInv(RealD z, RealD x0, int maxiter, RealD resid) {
+      RealD x = x0;
+      RealD eps;
+      
+      int i;
+      for (i=0;i<maxiter;i++) {
+	eps = approx(x) - z;
+	if (fabs(eps / z) < resid)
+	  return x;
+	x = x - eps / approxD(x);
+      }
+      
+      return std::numeric_limits<double>::quiet_NaN();
+    }
+    
     // Implement the required interface
     void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 

Grid/algorithms/approx/Forecast.h

@@ -0,0 +1,152 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/approx/Forecast.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: David Murphy <dmurphy@phys.columbia.edu>
+
+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 INCLUDED_FORECAST_H
+#define INCLUDED_FORECAST_H
+
+namespace Grid {
+
+  // Abstract base class.
+  // Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
+  // and returns a forecasted solution to the system D*psi = phi (psi).
+  template<class Matrix, class Field>
+  class Forecast
+  {
+    public:
+      virtual Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
+  };
+
+  // Implementation of Brower et al.'s chronological inverter (arXiv:hep-lat/9509012),
+  // used to forecast solutions across poles of the EOFA heatbath.
+  //
+  // Modified from CPS (cps_pp/src/util/dirac_op/d_op_base/comsrc/minresext.C)
+  template<class Matrix, class Field>
+  class ChronoForecast : public Forecast<Matrix,Field>
+  {
+    public:
+      Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& prev_solns)
+      {
+        int degree = prev_solns.size();
+        Field chi(phi); // forecasted solution
+
+        // Trivial cases
+        if(degree == 0){ chi = zero; return chi; }
+        else if(degree == 1){ return prev_solns[0]; }
+
+        RealD dot;
+        ComplexD xp;
+        Field r(phi); // residual
+        Field Mv(phi);
+        std::vector<Field> v(prev_solns); // orthonormalized previous solutions
+        std::vector<Field> MdagMv(degree,phi);
+
+        // Array to hold the matrix elements
+        std::vector<std::vector<ComplexD>> G(degree, std::vector<ComplexD>(degree));
+
+        // Solution and source vectors
+        std::vector<ComplexD> a(degree);
+        std::vector<ComplexD> b(degree);
+
+        // Orthonormalize the vector basis
+        for(int i=0; i<degree; i++){
+          v[i] *= 1.0/std::sqrt(norm2(v[i]));
+          for(int j=i+1; j<degree; j++){ v[j] -= innerProduct(v[i],v[j]) * v[i]; }
+        }
+
+        // Perform sparse matrix multiplication and construct rhs
+        for(int i=0; i<degree; i++){
+          b[i] = innerProduct(v[i],phi);
+          Mat.M(v[i],Mv);
+          Mat.Mdag(Mv,MdagMv[i]);
+          G[i][i] = innerProduct(v[i],MdagMv[i]);
+        }
+
+        // Construct the matrix
+        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]);
+        }}
+
+        // Gauss-Jordan elimination with partial pivoting
+        for(int i=0; i<degree; i++){
+
+          // 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; } }
+          if(k != i){
+            xp = b[k];
+            b[k] = b[i];
+            b[i] = xp;
+            for(int j=0; j<degree; j++){
+              xp = G[k][j];
+              G[k][j] = G[i][j];
+              G[i][j] = xp;
+            }
+          }
+
+          // Convert matrix to upper triangular form
+          for(int j=i+1; j<degree; j++){
+            xp = G[j][i]/G[i][i];
+            b[j] -= xp * b[i];
+            for(int k=0; k<degree; k++){ G[j][k] -= xp*G[i][k]; }
+          }
+        }
+
+        // Use Gaussian elimination to solve equations and calculate initial guess
+        chi = zero;
+        r = phi;
+        for(int i=degree-1; i>=0; i--){
+          a[i] = 0.0;
+          for(int j=i+1; j<degree; j++){ a[i] += G[i][j] * a[j]; }
+          a[i] = (b[i]-a[i])/G[i][i];
+          chi += a[i]*v[i];
+          r -= a[i]*MdagMv[i];
+        }
+
+        RealD true_r(0.0);
+        ComplexD tmp;
+        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;
+          true_r += std::sqrt(tmp.real());
+        }
+
+        RealD error = std::sqrt(norm2(r)/norm2(phi));
+        std::cout << GridLogMessage << "ChronoForecast: |res|/|src| = " << error << std::endl;
+
+        return chi;
+      };
+  };
+
+}
+
+#endif

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -0,0 +1,698 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/BlockConjugateGradient.h
+
+Copyright (C) 2017
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_BLOCK_CONJUGATE_GRADIENT_H
+#define GRID_BLOCK_CONJUGATE_GRADIENT_H
+
+
+namespace Grid {
+
+enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
+
+//////////////////////////////////////////////////////////////////////////
+// Block conjugate gradient. Dimension zero should be the block direction
+//////////////////////////////////////////////////////////////////////////
+template <class Field>
+class BlockConjugateGradient : public OperatorFunction<Field> {
+ public:
+
+  typedef typename Field::scalar_type scomplex;
+
+  int blockDim ;
+  int Nblock;
+
+  BlockCGtype CGtype;
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
+  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),PrintInterval(100)
+  {};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Thin QR factorisation (google it)
+////////////////////////////////////////////////////////////////////////////////////////////////////
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  //Dimensions
+  // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
+  //
+  // Rdag R = m_rr = Herm = L L^dag        <-- Cholesky decomposition (LLT routine in Eigen)
+  //
+  //   Q  C = R => Q = R C^{-1}
+  //
+  // Want  Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock} 
+  //
+  // Set C = L^{dag}, and then Q^dag Q = ident 
+  //
+  // Checks:
+  // 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());
+
+  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
+
+  C    = L.adjoint();
+  Cinv = C.inverse();
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Q = R C^{-1}
+  //
+  // Q_j  = R_i Cinv(i,j) 
+  //
+  // NB maddMatrix conventions are Right multiplication X[j] a[j,i] already
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  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
+////////////////////////////////////////////////////////////////////////////////////////////////////
+void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
+{
+  if ( CGtype == BlockCGrQ ) {
+    BlockCGrQsolve(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:
+//--------------------------
+// X is guess/Solution
+// B is RHS
+// Solve A X_i = B_i    ;        i refers to Nblock index
+////////////////////////////////////////////////////////////////////////////
+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];
+/* FAKE */
+  Nblock=8;
+  std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
+
+  X.checkerboard = B.checkerboard;
+  conformable(X, B);
+
+  Field tmp(B);
+  Field Q(B);
+  Field D(B);
+  Field Z(B);
+  Field AD(B);
+
+  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);
+
+  sliceNorm(ssq,B,Orthog);
+  RealD sssum=0;
+  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+
+  sliceNorm(residuals,B,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  sliceNorm(residuals,X,Orthog);
+  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<<"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;  
+
+  ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
+  D=Q;
+
+  std::cout << GridLogMessage<<"BlockCGrQ 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();
+    Linop.HermOp(D, Z);      
+    MatrixTimer.Stop();
+
+    //4. M  = [D^dag Z]^{-1}
+    sliceInnerTimer.Start();
+    sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
+    sliceInnerTimer.Stop();
+    m_M       = m_DZ.inverse();
+    
+    //5. X  = X + D MC
+    m_tmp     = m_M * m_C;
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(X,m_tmp, D,X,Orthog);     
+    sliceMaddTimer.Stop();
+
+    //6. QS = Q - ZM
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(tmp,m_M,Z,Q,Orthog,-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();
+    sliceMaddMatrix(D,m_tmp,D,Q,Orthog);
+    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 << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
+	      <<" ave "<<std::sqrt(rrsum/sssum) << " max "<< 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;
+
+      Linop.HermOp(X, AD);
+      AD = AD-B;
+      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AD)/norm2(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;
+    }
+
+  }
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) 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];
+
+  std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
+
+  Psi.checkerboard = Src.checkerboard;
+  conformable(Psi, Src);
+
+  Field P(Src);
+  Field AP(Src);
+  Field R(Src);
+  
+  std::vector<ComplexD> v_pAp(Nblock);
+  std::vector<RealD> v_rr (Nblock);
+  std::vector<RealD> v_rr_inv(Nblock);
+  std::vector<RealD> v_alpha(Nblock);
+  std::vector<RealD> v_beta(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);
+
+  R = Src - AP;  
+  P = R;
+  sliceNorm(v_rr,R,Orthog);
+
+  GridStopWatch sliceInnerTimer;
+  GridStopWatch sliceMaddTimer;
+  GridStopWatch sliceNormTimer;
+  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(v_rr[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();
+    sliceInnerProductVector(v_pAp,P,AP,Orthog);
+    sliceInnerTimer.Stop();
+    for(int b=0;b<Nblock;b++){
+      v_alpha[b] = v_rr[b]/real(v_pAp[b]);
+    }
+
+    // Psi, R update
+    sliceMaddTimer.Start();
+    sliceMaddVector(Psi,v_alpha, P,Psi,Orthog);     // add alpha *  P to psi
+    sliceMaddVector(R  ,v_alpha,AP,  R,Orthog,-1.0);// sub alpha * AP to resid
+    sliceMaddTimer.Stop();
+
+    // Beta
+    for(int b=0;b<Nblock;b++){
+      v_rr_inv[b] = 1.0/v_rr[b];
+    }
+    sliceNormTimer.Start();
+    sliceNorm(v_rr,R,Orthog);
+    sliceNormTimer.Stop();
+    for(int b=0;b<Nblock;b++){
+      v_beta[b] = v_rr_inv[b] *v_rr[b];
+    }
+
+    // Search update
+    sliceMaddTimer.Start();
+    sliceMaddVector(P,v_beta,P,R,Orthog);
+    sliceMaddTimer.Stop();
+
+    /*********************
+     * convergence monitor
+     *********************
+     */
+    RealD max_resid=0;
+    for(int b=0;b<Nblock;b++){
+      RealD rr = v_rr[b]/ssq[b];
+      if ( rr > max_resid ) max_resid = rr;
+    }
+    
+    if ( max_resid < Tolerance*Tolerance ) { 
+
+      SolverTimer.Stop();
+
+      std::cout << GridLogMessage<<"MultiRHS solver converged in " <<k<<" iterations"<<std::endl;
+      for(int b=0;b<Nblock;b++){
+	std::cout << GridLogMessage<< "\t\tBlock "<<b<<" computed resid "<< std::sqrt(v_rr[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 << "\tTrue 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 << "\tNorm       " << sliceNormTimer.Elapsed() <<std::endl;
+      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
+
+
+      IterationsToComplete = k;
+      return;
+    }
+
+  }
+  std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
+
+  if (ErrorOnNoConverge) assert(0);
+  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] + (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] += (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;
+    }
+
+  }
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
+
+  if (ErrorOnNoConverge) assert(0);
+  IterationsToComplete = k;
+}
+
+
+
+};
+
+}
+#endif

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,244 @@
+/*************************************************************************************
+
+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:
+  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<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> 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));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, w, i);
+
+      qrUpdate(i);
+
+      cp = std::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 - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (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] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD 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]     = std::conj(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] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + v[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/ConjugateGradient.h

@@ -52,8 +52,9 @@ class ConjugateGradient : public OperatorFunction<Field> {
         MaxIterations(maxit),
         ErrorOnNoConverge(err_on_no_conv){};
 
-  void operator()(LinearOperatorBase<Field> &Linop, const Field &src,
-                  Field &psi) {
+  void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
+
+
     psi.checkerboard = src.checkerboard;
     conformable(psi, src);
 
@@ -69,7 +70,6 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
     
     Linop.HermOpAndNorm(psi, mmp, d, b);
-    
 
     r = src - mmp;
     p = r;
@@ -78,24 +78,29 @@ class ConjugateGradient : public OperatorFunction<Field> {
     cp = a;
     ssq = norm2(src);
 
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: guess " << guess << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   src " << ssq << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:    mp " << d << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   mmp " << b << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:  cp,r " << cp << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:     p " << a << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:     p " << a << std::endl;
 
     RealD rsq = Tolerance * Tolerance * ssq;
 
     // Check if guess is really REALLY good :)
     if (cp <= rsq) {
+      std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
+      IterationsToComplete = 0;	
       return;
     }
 
-    std::cout << GridLogIterative << std::setprecision(4)
+    std::cout << GridLogIterative << std::setprecision(8)
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
     GridStopWatch LinalgTimer;
+    GridStopWatch InnerTimer;
+    GridStopWatch AxpyNormTimer;
+    GridStopWatch LinearCombTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
 
@@ -105,29 +110,32 @@ class ConjugateGradient : public OperatorFunction<Field> {
       c = cp;
 
       MatrixTimer.Start();
-      Linop.HermOpAndNorm(p, mmp, d, qq);
+      Linop.HermOp(p, mmp);
       MatrixTimer.Stop();
 
       LinalgTimer.Start();
-      //  RealD    qqck = norm2(mmp);
-      //  ComplexD dck  = innerProduct(p,mmp);
 
+      InnerTimer.Start();
+      ComplexD dc  = innerProduct(p,mmp);
+      InnerTimer.Stop();
+      d = dc.real();
       a = c / d;
-      b_pred = a * (a * qq - d) / c;
 
+      AxpyNormTimer.Start();
       cp = axpy_norm(r, -a, mmp, r);
+      AxpyNormTimer.Stop();
       b = cp / c;
 
-      // Fuse these loops ; should be really easy
-      psi = a * p + psi;
-      p = p * b + r;
-
+      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]);
+      }
+      LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
       std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
-                << " residual " << cp << " target " << rsq << std::endl;
-      std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << "  b = "<< b << std::endl;
-      std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << "  c = "<< c << std::endl;
+                << " residual^2 " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -148,6 +156,9 @@ class ConjugateGradient : public OperatorFunction<Field> {
 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 
@@ -156,8 +167,7 @@ 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;

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -30,8 +30,11 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
 
 namespace 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 +53,12 @@ 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){ };
@@ -149,6 +157,8 @@ namespace Grid {
     }
   };
 
+
+
 }
 
 #endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -43,6 +43,7 @@ namespace Grid {
 public:                                                
     RealD   Tolerance;
     Integer MaxIterations;
+    Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
     int verbose;
     MultiShiftFunction shifts;
 
@@ -163,7 +164,16 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   for(int s=0;s<nshift;s++) {
     axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
   }
-  
+ 
+  ///////////////////////////////////////
+  // Timers
+  ///////////////////////////////////////
+  GridStopWatch AXPYTimer;
+  GridStopWatch ShiftTimer;
+  GridStopWatch QRTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch SolverTimer;
+  SolverTimer.Start();
   
   // Iteration loop
   int k;
@@ -171,7 +181,9 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   for (k=1;k<=MaxIterations;k++){
     
     a = c /cp;
+    AXPYTimer.Start();
     axpy(p,a,p,r);
+    AXPYTimer.Stop();
     
     // Note to self - direction ps is iterated seperately
     // for each shift. Does not appear to have any scope
@@ -180,6 +192,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     // However SAME r is used. Could load "r" and update
     // ALL ps[s]. 2/3 Bandwidth saving
     // New Kernel: Load r, vector of coeffs, vector of pointers ps
+    AXPYTimer.Start();
     for(int s=0;s<nshift;s++){
       if ( ! converged[s] ) { 
 	if (s==0){
@@ -190,22 +203,34 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	}
       }
     }
+    AXPYTimer.Stop();
     
     cp=c;
+    MatrixTimer.Start();  
+    //Linop.HermOpAndNorm(p,mmp,d,qq); // d is used
+    // The below is faster on KNL
+    Linop.HermOp(p,mmp); 
+    d=real(innerProduct(p,mmp));
     
-    Linop.HermOpAndNorm(p,mmp,d,qq);
+    MatrixTimer.Stop();  
+
+    AXPYTimer.Start();
     axpy(mmp,mass[0],p,mmp);
+    AXPYTimer.Stop();
     RealD rn = norm2(p);
     d += rn*mass[0];
     
     bp=b;
     b=-cp/d;
     
+    AXPYTimer.Start();
     c=axpy_norm(r,b,mmp,r);
+    AXPYTimer.Stop();
 
     // Toggle the recurrence history
     bs[0] = b;
     iz = 1-iz;
+    ShiftTimer.Start();
     for(int s=1;s<nshift;s++){
       if((!converged[s])){
 	RealD z0 = z[s][1-iz];
@@ -215,6 +240,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
       }
     }
+    ShiftTimer.Stop();
     
     for(int s=0;s<nshift;s++){
       int ss = s;
@@ -257,6 +283,9 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     
     if ( all_converged ){
 
+    SolverTimer.Stop();
+
+
       std::cout<<GridLogMessage<< "CGMultiShift: All shifts have converged iteration "<<k<<std::endl;
       std::cout<<GridLogMessage<< "CGMultiShift: Checking solutions"<<std::endl;
       
@@ -269,8 +298,19 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	RealD cn = norm2(src);
 	std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
       }
+
+      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
+      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+
+      IterationsToComplete = k;	
+
       return;
     }
+
+   
   }
   // ugly hack
   std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -0,0 +1,256 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientReliableUpdate.h
+
+    Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@phys.columbia.edu>
+
+    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_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
+#define GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_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 ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
+  public:
+    bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+    // Defaults true.
+    RealD Tolerance;
+    Integer MaxIterations;
+    Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+    Integer ReliableUpdatesPerformed;
+
+    bool DoFinalCleanup; //Final DP cleanup, defaults to true
+    Integer IterationsToCleanup; //Final DP cleanup step iterations
+    
+    LinearOperatorBase<FieldF> &Linop_f;
+    LinearOperatorBase<FieldD> &Linop_d;
+    GridBase* SinglePrecGrid;
+    RealD Delta; //reliable update parameter
+
+    //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
+    LinearOperatorBase<FieldF> *Linop_fallback;
+    RealD fallback_transition_tol;
+
+    
+    ConjugateGradientReliableUpdate(RealD tol, Integer maxit, RealD _delta, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d, bool err_on_no_conv = true)
+      : Tolerance(tol),
+        MaxIterations(maxit),
+	Delta(_delta),
+	Linop_f(_Linop_f),
+	Linop_d(_Linop_d),
+	SinglePrecGrid(_sp_grid),
+        ErrorOnNoConverge(err_on_no_conv),
+	DoFinalCleanup(true),
+	Linop_fallback(NULL)
+    {};
+
+    void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
+      Linop_fallback = &_Linop_fallback;
+      fallback_transition_tol = _fallback_transition_tol;      
+    }
+    
+    void operator()(const FieldD &src, FieldD &psi) {
+      LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
+      bool using_fallback = false;
+      
+      psi.checkerboard = src.checkerboard;
+      conformable(psi, src);
+
+      RealD cp, c, a, d, b, ssq, qq, b_pred;
+
+      FieldD p(src);
+      FieldD mmp(src);
+      FieldD r(src);
+
+      // Initial residual computation & set up
+      RealD guess = norm2(psi);
+      assert(std::isnan(guess) == 0);
+    
+      Linop_d.HermOpAndNorm(psi, mmp, d, b);
+    
+      r = src - mmp;
+      p = r;
+
+      a = norm2(p);
+      cp = a;
+      ssq = norm2(src);
+
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: guess " << guess << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:   src " << ssq << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:    mp " << d << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:   mmp " << b << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:  cp,r " << cp << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:     p " << a << std::endl;
+
+      RealD rsq = Tolerance * Tolerance * ssq;
+
+      // 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;
+	return;
+      }
+
+      //Single prec initialization
+      FieldF r_f(SinglePrecGrid);
+      r_f.checkerboard = r.checkerboard;
+      precisionChange(r_f, r);
+
+      FieldF psi_f(r_f);
+      psi_f = zero;
+
+      FieldF p_f(r_f);
+      FieldF mmp_f(r_f);
+
+      RealD MaxResidSinceLastRelUp = cp; //initial residual    
+    
+      std::cout << GridLogIterative << std::setprecision(4)
+		<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
+
+      GridStopWatch LinalgTimer;
+      GridStopWatch MatrixTimer;
+      GridStopWatch SolverTimer;
+
+      SolverTimer.Start();
+      int k = 0;
+      int l = 0;
+    
+      for (k = 1; k <= MaxIterations; k++) {
+	c = cp;
+
+	MatrixTimer.Start();
+	Linop_f_use->HermOpAndNorm(p_f, mmp_f, d, qq);
+	MatrixTimer.Stop();
+
+	LinalgTimer.Start();
+
+	a = c / d;
+	b_pred = a * (a * qq - d) / c;
+
+	cp = axpy_norm(r_f, -a, mmp_f, r_f);
+	b = cp / c;
+
+	// Fuse these loops ; should be really easy
+	psi_f = a * p_f + psi_f;
+	//p_f = p_f * b + r_f;
+
+	LinalgTimer.Stop();
+
+	std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: Iteration " << k
+		  << " residual " << cp << " target " << rsq << std::endl;
+	std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << "  b = "<< b << std::endl;
+	std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << "  c = "<< c << std::endl;
+
+	if(cp > MaxResidSinceLastRelUp){
+	  std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: updating MaxResidSinceLastRelUp : " << MaxResidSinceLastRelUp << " -> " << cp << std::endl;
+	  MaxResidSinceLastRelUp = cp;
+	}
+	  
+	// Stopping condition
+	if (cp <= rsq) {
+	  //Although not written in the paper, I assume that I have to add on the final solution
+	  precisionChange(mmp, psi_f);
+	  psi = psi + mmp;
+	
+	
+	  SolverTimer.Stop();
+	  Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	  p = mmp - src;
+
+	  RealD srcnorm = sqrt(norm2(src));
+	  RealD resnorm = 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 << "\tTrue residual " << true_residual<<std::endl;
+	  std::cout << GridLogMessage << "\tTarget " << Tolerance << 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 << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+
+	  IterationsToComplete = k;	
+	  ReliableUpdatesPerformed = l;
+	  
+	  if(DoFinalCleanup){
+	    //Do a final CG to cleanup
+	    std::cout << GridLogMessage << "ConjugateGradientReliableUpdate performing final cleanup.\n";
+	    ConjugateGradient<FieldD> CG(Tolerance,MaxIterations);
+	    CG.ErrorOnNoConverge = ErrorOnNoConverge;
+	    CG(Linop_d,src,psi);
+	    IterationsToCleanup = CG.IterationsToComplete;
+	  }
+	  else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
+	  return;
+	}
+	else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
+		    << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
+	  precisionChange(mmp, psi_f);
+	  psi = psi + mmp;
+
+	  Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	  r = src - mmp;
+
+	  psi_f = zero;
+	  precisionChange(r_f, r);
+	  cp = norm2(r);
+	  MaxResidSinceLastRelUp = cp;
+
+	  b = cp/c;
+	  
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate new residual " << cp << std::endl;
+	  
+	  l = l+1;
+	}
+
+	p_f = p_f * b + r_f; //update search vector after reliable update appears to help convergence
+
+	if(!using_fallback && Linop_fallback != NULL && cp < fallback_transition_tol){
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate switching to fallback linear operator on iteration " << k << " at residual " << cp << std::endl;
+	  Linop_f_use = Linop_fallback;
+	  using_fallback = true;
+	}
+
+	
+      }
+      std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
+		<< std::endl;
+      
+      if (ErrorOnNoConverge) assert(0);
+      IterationsToComplete = k;
+      ReliableUpdatesPerformed = l;      
+    }    
+  };
+
+
+};
+
+
+
+#endif

Grid/algorithms/iterative/Deflation.h

@@ -0,0 +1,108 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.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_DEFLATION_H
+#define GRID_DEFLATION_H
+
+namespace Grid { 
+
+template<class Field>
+class ZeroGuesser: public LinearFunction<Field> {
+public:
+  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:
+  virtual void operator()(const Field &src, Field &guess) { guess = src; };
+};
+
+////////////////////////////////
+// Fine grid deflation
+////////////////////////////////
+template<class Field>
+class DeflatedGuesser: public LinearFunction<Field> {
+private:
+  const std::vector<Field> &evec;
+  const std::vector<RealD> &eval;
+
+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;
+    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;
+  }
+};
+
+template<class FineField, class CoarseField>
+class LocalCoherenceDeflatedGuesser: public LinearFunction<FineField> {
+private:
+  const std::vector<FineField>   &subspace;
+  const std::vector<CoarseField> &evec_coarse;
+  const std::vector<RealD>       &eval_coarse;
+public:
+  
+  LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
+				const std::vector<CoarseField> &_evec_coarse,
+				const std::vector<RealD>       &_eval_coarse)
+    : subspace(_subspace), 
+      evec_coarse(_evec_coarse), 
+      eval_coarse(_eval_coarse)  
+  {
+  }
+  
+  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;
+    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;
+  };
+};
+
+
+
+}
+#endif

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,256 @@
+/*************************************************************************************
+
+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:
+  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<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> 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 = std::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 - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (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] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD 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]     = std::conj(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] - H(k, i) * y[k];
+      y[i] = y[i] / 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 +1,254 @@
+/*************************************************************************************
+
+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:
+  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<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> 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 = std::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 - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (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] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD 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]     = std::conj(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] - H(k, i) * y[k];
+      y[i] = y[i] / 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 +1,242 @@
+/*************************************************************************************
+
+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:
+  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<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> 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 = std::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 - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (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] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD 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]     = std::conj(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] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + v[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -0,0 +1,842 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+Author: Christoph Lehner <clehner@bnl.gov>
+
+    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_BIRL_H
+#define GRID_BIRL_H
+
+#include <string.h> //memset
+//#include <zlib.h>
+#include <sys/stat.h>
+
+namespace Grid { 
+
+  ////////////////////////////////////////////////////////
+  // Move following 100 LOC to lattice/Lattice_basis.h
+  ////////////////////////////////////////////////////////
+template<class Field>
+void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k) 
+{
+  for(int j=0; j<k; ++j){
+    auto ip = innerProduct(basis[j],w);
+    w = w - ip*basis[j];
+  }
+}
+
+template<class Field>
+void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm) 
+{
+  typedef typename Field::vector_object vobj;
+  GridBase* grid = basis[0]._grid;
+      
+  parallel_region
+  {
+
+    std::vector < vobj , commAllocator<vobj> > B(Nm); // Thread private
+       
+    parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
+      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];
+	}
+      }
+      for(int j=j0; j<j1; ++j){
+	  basis[j]._odata[ss] = B[j];
+      }
+    }
+  }
+}
+
+// Extract a single rotated vector
+template<class Field>
+void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
+{
+  typedef typename Field::vector_object vobj;
+  GridBase* grid = basis[0]._grid;
+
+  result.checkerboard = basis[0].checkerboard;
+  parallel_for(int ss=0;ss < grid->oSites();ss++){
+    vobj B = zero;
+    for(int k=k0; k<k1; ++k){
+      B +=Qt(j,k) * basis[k]._odata[ss];
+    }
+    result._odata[ss] = B;
+  }
+}
+
+template<class Field>
+void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx) 
+{
+  int vlen = idx.size();
+
+  assert(vlen>=1);
+  assert(vlen<=sort_vals.size());
+  assert(vlen<=_v.size());
+
+  for (size_t i=0;i<vlen;i++) {
+
+    if (idx[i] != i) {
+
+      //////////////////////////////////////
+      // idx[i] is a table of desired sources giving a permutation.
+      // Swap v[i] with v[idx[i]].
+      // Find  j>i for which _vnew[j] = _vold[i],
+      // track the move idx[j] => idx[i]
+      // track the move idx[i] => i
+      //////////////////////////////////////
+      size_t j;
+      for (j=i;j<idx.size();j++)
+	if (idx[j]==i)
+	  break;
+
+      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
+      std::swap(sort_vals[i],sort_vals[idx[i]]);
+
+      idx[j] = idx[i];
+      idx[i] = i;
+    }
+  }
+}
+
+inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals) 
+{
+  std::vector<int> idx(sort_vals.size());
+  std::iota(idx.begin(), idx.end(), 0);
+
+  // sort indexes based on comparing values in v
+  std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
+    return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
+  });
+  return idx;
+}
+
+template<class Field>
+void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse) 
+{
+  std::vector<int> idx = basisSortGetIndex(sort_vals);
+  if (reverse)
+    std::reverse(idx.begin(), idx.end());
+  
+  basisReorderInPlace(_v,sort_vals,idx);
+}
+
+/////////////////////////////////////////////////////////////
+// Implicitly restarted lanczos
+/////////////////////////////////////////////////////////////
+template<class Field> class ImplicitlyRestartedLanczosTester 
+{
+ public:
+  virtual int TestConvergence(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
+  virtual int ReconstructEval(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
+};
+
+enum IRLdiagonalisation { 
+  IRLdiagonaliseWithDSTEGR,
+  IRLdiagonaliseWithQR,
+  IRLdiagonaliseWithEigen
+};
+
+template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public ImplicitlyRestartedLanczosTester<Field>
+{
+ public:
+
+  LinearFunction<Field>       &_HermOp;
+  ImplicitlyRestartedLanczosHermOpTester(LinearFunction<Field> &HermOp) : _HermOp(HermOp)  {  };
+  int ReconstructEval(int j,RealD resid,Field &B, RealD &eval,RealD evalMaxApprox)
+  {
+    return TestConvergence(j,resid,B,eval,evalMaxApprox);
+  }
+  int TestConvergence(int j,RealD eresid,Field &B, RealD &eval,RealD evalMaxApprox)
+  {
+    Field v(B);
+    RealD eval_poly = eval;
+    // Apply operator
+    _HermOp(B,v);
+
+    RealD vnum = real(innerProduct(B,v)); // HermOp.
+    RealD vden = norm2(B);
+    RealD vv0  = norm2(v);
+    eval   = vnum/vden;
+    v -= eval*B;
+
+    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+
+    int conv=0;
+    if( (vv<eresid*eresid) ) conv = 1;
+
+    return conv;
+  }
+};
+
+template<class Field> 
+class ImplicitlyRestartedLanczos {
+ private:
+  const RealD small = 1.0e-8;
+  int MaxIter;
+  int MinRestart; // Minimum number of restarts; only check for convergence after
+  int Nstop;   // Number of evecs checked for convergence
+  int Nk;      // Number of converged sought
+  //  int Np;      // Np -- Number of spare vecs in krylov space //  == Nm - Nk
+  int Nm;      // Nm -- total number of vectors
+  IRLdiagonalisation diagonalisation;
+  int orth_period;
+    
+  RealD OrthoTime;
+  RealD eresid, betastp;
+  ////////////////////////////////
+  // Embedded objects
+  ////////////////////////////////
+  LinearFunction<Field>       &_PolyOp;
+  LinearFunction<Field>       &_HermOp;
+  ImplicitlyRestartedLanczosTester<Field> &_Tester;
+  // Default tester provided (we need a ref to something in default case)
+  ImplicitlyRestartedLanczosHermOpTester<Field> SimpleTester;
+  /////////////////////////
+  // Constructor
+  /////////////////////////
+  
+public:       
+
+  //////////////////////////////////////////////////////////////////
+  // PAB:
+  //////////////////////////////////////////////////////////////////
+  // Too many options  & knobs. 
+  // Eliminate:
+  //   orth_period
+  //   betastp
+  //   MinRestart
+  //
+  // Do we really need orth_period
+  // What is the theoretical basis & guarantees of betastp ?
+  // Nstop=Nk viable?
+  // MinRestart avoidable with new convergence test?
+  // Could cut to PolyOp, HermOp, Tester, Nk, Nm, resid, maxiter (+diagonalisation)
+  // HermOp could be eliminated if we dropped the Power method for max eval.
+  // -- also: The eval, eval2, eval2_copy stuff is still unnecessarily unclear
+  //////////////////////////////////////////////////////////////////
+ ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
+			    LinearFunction<Field> & HermOp,
+			    ImplicitlyRestartedLanczosTester<Field> & Tester,
+			    int _Nstop, // sought vecs
+			    int _Nk, // sought vecs
+			    int _Nm, // spare vecs
+			    RealD _eresid, // resid in lmdue deficit 
+			    int _MaxIter, // Max iterations
+			    RealD _betastp=0.0, // if beta(k) < betastp: converged
+			    int _MinRestart=1, int _orth_period = 1,
+			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
+    SimpleTester(HermOp), _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(Tester),
+    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
+    eresid(_eresid),      betastp(_betastp),
+    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
+
+    ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
+			       LinearFunction<Field> & HermOp,
+			       int _Nstop, // sought vecs
+			       int _Nk, // sought vecs
+			       int _Nm, // spare vecs
+			       RealD _eresid, // resid in lmdue deficit 
+			       int _MaxIter, // Max iterations
+			       RealD _betastp=0.0, // if beta(k) < betastp: converged
+			       int _MinRestart=1, int _orth_period = 1,
+			       IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
+    SimpleTester(HermOp),  _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(SimpleTester),
+    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
+    eresid(_eresid),      betastp(_betastp),
+    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
+
+  ////////////////////////////////
+  // Helpers
+  ////////////////////////////////
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  void orthogonalize(Field& w, std::vector<Field>& evec,int k)
+  {
+    OrthoTime-=usecond()/1e6;
+    basisOrthogonalize(evec,w,k);
+    normalise(w);
+    OrthoTime+=usecond()/1e6;
+  }
+
+/* Rudy Arthur's thesis pp.137
+------------------------
+Require: M > K P = M − K †
+Compute the factorization AVM = VM HM + fM eM 
+repeat
+  Q=I
+  for i = 1,...,P do
+    QiRi =HM −θiI Q = QQi
+    H M = Q †i H M Q i
+  end for
+  βK =HM(K+1,K) σK =Q(M,K)
+  r=vK+1βK +rσK
+  VK =VM(1:M)Q(1:M,1:K)
+  HK =HM(1:K,1:K)
+  →AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
+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);
+    
+    GridLogIRL.TimingMode(1);
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<" -- seek   Nk    = " << Nk    <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- total  Nm    = " << Nm    <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- size of eval = " << eval.size() << std::endl;
+    std::cout << GridLogIRL <<" -- size of evec = " << evec.size() << std::endl;
+    if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
+      std::cout << GridLogIRL << "Diagonalisation is DSTEGR "<<std::endl;
+    } else if ( diagonalisation == IRLdiagonaliseWithQR ) { 
+      std::cout << GridLogIRL << "Diagonalisation is QR "<<std::endl;
+    }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
+      std::cout << GridLogIRL << "Diagonalisation is Eigen "<<std::endl;
+    }
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+	
+    assert(Nm <= evec.size() && Nm <= eval.size());
+    
+    // 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_IRL_MEVAPP_ = 50;
+      for (int i=0;i<_MAX_ITER_IRL_MEVAPP_;i++) {
+	normalise(src_n);
+	_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) < 0.05)
+	  i=_MAX_ITER_IRL_MEVAPP_;
+	evalMaxApprox = na;
+	std::cout << GridLogIRL << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+	src_n = tmp;
+      }
+    }
+	
+    std::vector<RealD> lme(Nm);  
+    std::vector<RealD> lme2(Nm);
+    std::vector<RealD> eval2(Nm);
+    std::vector<RealD> eval2_copy(Nm);
+    Eigen::MatrixXd Qt = Eigen::MatrixXd::Zero(Nm,Nm);
+
+    Field f(grid);
+    Field v(grid);
+    int k1 = 1;
+    int k2 = Nk;
+    RealD beta_k;
+
+    Nconv = 0;
+  
+    // Set initial vector
+    evec[0] = src;
+    normalise(evec[0]);
+	
+    // Initial Nk steps
+    OrthoTime=0.;
+    for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
+    std::cout<<GridLogIRL <<"Initial "<< Nk <<"steps done "<<std::endl;
+    std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+
+    //////////////////////////////////
+    // Restarting loop begins
+    //////////////////////////////////
+    int iter;
+    for(iter = 0; iter<MaxIter; ++iter){
+      
+      OrthoTime=0.;
+
+      std::cout<< GridLogMessage <<" **********************"<< std::endl;
+      std::cout<< GridLogMessage <<" Restart iteration = "<< iter << std::endl;
+      std::cout<< GridLogMessage <<" **********************"<< std::endl;
+
+      std::cout<<GridLogIRL <<" running "<<Nm-Nk <<" steps: "<<std::endl;
+      for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
+      f *= lme[Nm-1];
+
+      std::cout<<GridLogIRL <<" "<<Nm-Nk <<" steps done "<<std::endl;
+      std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+	  
+      //////////////////////////////////
+      // getting eigenvalues
+      //////////////////////////////////
+      for(int k=0; k<Nm; ++k){
+	eval2[k] = eval[k+k1-1];
+	lme2[k] = lme[k+k1-1];
+      }
+      Qt = Eigen::MatrixXd::Identity(Nm,Nm);
+      diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
+      std::cout<<GridLogIRL <<" diagonalized "<<std::endl;
+
+      //////////////////////////////////
+      // sorting
+      //////////////////////////////////
+      eval2_copy = eval2;
+      std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
+      std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
+      const int chunk=8;
+      for(int io=0; io<k2;io+=chunk){
+	std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
+	for(int ii=0;ii<chunk;ii++){
+	  if ( (io+ii)<k2 )
+	    std::cout<< " "<< std::setw(12)<< eval2[io+ii];
+	}
+	std::cout << std::endl;
+      }
+
+      //////////////////////////////////
+      // Implicitly shifted QR transformations
+      //////////////////////////////////
+      Qt = Eigen::MatrixXd::Identity(Nm,Nm);
+      for(int ip=k2; ip<Nm; ++ip){ 
+	QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
+      }
+      std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
+
+      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;
+      
+      ////////////////////////////////////////////////////
+      // Compressed vector f and beta(k2)
+      ////////////////////////////////////////////////////
+      f *= Qt(k2-1,Nm-1);
+      f += lme[k2-1] * evec[k2];
+      beta_k = norm2(f);
+      beta_k = sqrt(beta_k);
+      std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
+	  
+      RealD betar = 1.0/beta_k;
+      evec[k2] = betar * f;
+      lme[k2-1] = beta_k;
+	  
+      ////////////////////////////////////////////////////
+      // Convergence test
+      ////////////////////////////////////////////////////
+      for(int k=0; k<Nm; ++k){    
+	eval2[k] = eval[k];
+	lme2[k] = lme[k];
+      }
+      Qt = Eigen::MatrixXd::Identity(Nm,Nm);
+      diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
+      std::cout<<GridLogIRL <<" Diagonalized "<<std::endl;
+	  
+      Nconv = 0;
+      if (iter >= MinRestart) {
+
+	std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
+
+	Field B(grid); B.checkerboard = evec[0].checkerboard;
+
+	//  power of two search pattern;  not every evalue in eval2 is assessed.
+	int allconv =1;
+	for(int jj = 1; jj<=Nstop; jj*=2){
+	  int j = Nstop-jj;
+	  RealD e = eval2_copy[j]; // Discard the evalue
+	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
+	  if( !_Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) {
+	    allconv=0;
+	  }
+	}
+	// Do evec[0] for good measure
+	{ 
+	  int j=0;
+	  RealD e = eval2_copy[0]; 
+	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
+	  if( !_Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) allconv=0;
+	}
+	if ( allconv ) Nconv = Nstop;
+
+	// test if we converged, if so, terminate
+	std::cout<<GridLogIRL<<" #modes converged: >= "<<Nconv<<"/"<<Nstop<<std::endl;
+	//	if( Nconv>=Nstop || beta_k < betastp){
+	if( Nconv>=Nstop){
+	  goto converged;
+	}
+	  
+      } else {
+	std::cout << GridLogIRL << "iter < MinRestart: do not yet test for convergence\n";
+      } // end of iter loop
+    }
+
+    std::cout<<GridLogError<<"\n NOT converged.\n";
+    abort();
+	
+  converged:
+    {
+      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;
+      //////////////////////////////////////////////////////////////////////
+      // Full final convergence test; unconditionally applied
+      //////////////////////////////////////////////////////////////////////
+      for(int j = 0; j<=Nk; j++){
+	B=evec[j];
+	if( _Tester.ReconstructEval(j,eresid,B,eval2[j],evalMaxApprox) ) {
+	  Nconv++;
+	}
+      }
+
+      if ( Nconv < Nstop )
+	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
+
+      eval=eval2;
+      
+      //Keep only converged
+      eval.resize(Nconv);// Nstop?
+      evec.resize(Nconv,grid);// Nstop?
+      basisSortInPlace(evec,eval,reverse);
+      
+    }
+       
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL << " -- Iterations  = "<< iter   << "\n";
+    std::cout << GridLogIRL << " -- beta(k)     = "<< beta_k << "\n";
+    std::cout << GridLogIRL << " -- Nconv       = "<< Nconv  << "\n";
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+  }
+
+ private:
+/* Saad PP. 195
+1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
+2. For k = 1,2,...,m Do:
+3. wk:=Avk−βkv_{k−1}      
+4. αk:=(wk,vk)       // 
+5. wk:=wk−αkvk       // wk orthog vk 
+6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+7. vk+1 := wk/βk+1
+8. EndDo
+ */
+  void step(std::vector<RealD>& lmd,
+	    std::vector<RealD>& lme, 
+	    std::vector<Field>& evec,
+	    Field& w,int Nm,int k)
+  {
+    const RealD tiny = 1.0e-20;
+    assert( k< Nm );
+
+    GridStopWatch gsw_op,gsw_o;
+
+    Field& evec_k = evec[k];
+
+    _PolyOp(evec_k,w);    std::cout<<GridLogIRL << "PolyOp" <<std::endl;
+
+    if(k>0) w -= lme[k-1] * evec[k-1];
+
+    ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
+    RealD     alph = real(zalph);
+
+    w = w - alph * evec_k;// 5. wk:=wk−αkvk
+
+    RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+    // 7. vk+1 := wk/βk+1
+
+    lmd[k] = alph;
+    lme[k] = beta;
+
+    if (k>0 && k % orth_period == 0) {
+      orthogonalize(w,evec,k); // orthonormalise
+      std::cout<<GridLogIRL << "Orthogonalised " <<std::endl;
+    }
+
+    if(k < Nm-1) evec[k+1] = w;
+
+    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+    if ( beta < tiny ) 
+      std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
+  }
+
+  void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
+			 int Nk, int Nm,  
+			 Eigen::MatrixXd & Qt, // Nm x Nm
+			 GridBase *grid)
+  {
+    Eigen::MatrixXd TriDiag = Eigen::MatrixXd::Zero(Nk,Nk);
+
+    for(int i=0;i<Nk;i++)   TriDiag(i,i)   = lmd[i];
+    for(int i=0;i<Nk-1;i++) TriDiag(i,i+1) = lme[i];
+    for(int i=0;i<Nk-1;i++) TriDiag(i+1,i) = lme[i];
+    
+    Eigen::SelfAdjointEigenSolver<Eigen::MatrixXd> eigensolver(TriDiag);
+
+    for (int i = 0; i < Nk; i++) {
+      lmd[Nk-1-i] = eigensolver.eigenvalues()(i);
+    }
+    for (int i = 0; i < Nk; i++) {
+      for (int j = 0; j < Nk; j++) {
+	Qt(Nk-1-i,j) = eigensolver.eigenvectors()(j,i);
+      }
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // File could end here if settle on Eigen ??? !!!
+  ///////////////////////////////////////////////////////////////////////////
+  void QR_decomp(std::vector<RealD>& lmd,   // Nm 
+		 std::vector<RealD>& lme,   // Nm 
+		 int Nk, int Nm,            // Nk, Nm
+		 Eigen::MatrixXd& Qt,       // Nm x Nm matrix
+		 RealD Dsh, int kmin, int kmax)
+  {
+    int k = kmin-1;
+    RealD x;
+    
+    RealD Fden = 1.0/hypot(lmd[k]-Dsh,lme[k]);
+    RealD c = ( lmd[k] -Dsh) *Fden;
+    RealD s = -lme[k] *Fden;
+      
+    RealD tmpa1 = lmd[k];
+    RealD tmpa2 = lmd[k+1];
+    RealD tmpb  = lme[k];
+
+    lmd[k]   = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
+    lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
+    lme[k]   = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
+    x        =-s*lme[k+1];
+    lme[k+1] = c*lme[k+1];
+      
+    for(int i=0; i<Nk; ++i){
+      RealD Qtmp1 = Qt(k,i);
+      RealD Qtmp2 = Qt(k+1,i);
+      Qt(k,i)  = c*Qtmp1 - s*Qtmp2;
+      Qt(k+1,i)= s*Qtmp1 + c*Qtmp2; 
+    }
+
+    // Givens transformations
+    for(int k = kmin; k < kmax-1; ++k){
+      
+      RealD Fden = 1.0/hypot(x,lme[k-1]);
+      RealD c = lme[k-1]*Fden;
+      RealD s = - x*Fden;
+	
+      RealD tmpa1 = lmd[k];
+      RealD tmpa2 = lmd[k+1];
+      RealD tmpb  = lme[k];
+
+      lmd[k]   = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
+      lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
+      lme[k]   = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
+      lme[k-1] = c*lme[k-1] -s*x;
+
+      if(k != kmax-2){
+	x = -s*lme[k+1];
+	lme[k+1] = c*lme[k+1];
+      }
+
+      for(int i=0; i<Nk; ++i){
+	RealD Qtmp1 = Qt(k,i);
+	RealD Qtmp2 = Qt(k+1,i);
+	Qt(k,i)     = c*Qtmp1 -s*Qtmp2;
+	Qt(k+1,i)   = s*Qtmp1 +c*Qtmp2;
+      }
+    }
+  }
+
+  void diagonalize(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
+		   int Nk, int Nm,   
+		   Eigen::MatrixXd & Qt,
+		   GridBase *grid)
+  {
+    Qt = Eigen::MatrixXd::Identity(Nm,Nm);
+    if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
+      diagonalize_lapack(lmd,lme,Nk,Nm,Qt,grid);
+    } else if ( diagonalisation == IRLdiagonaliseWithQR ) { 
+      diagonalize_QR(lmd,lme,Nk,Nm,Qt,grid);
+    }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
+      diagonalize_Eigen(lmd,lme,Nk,Nm,Qt,grid);
+    } else { 
+      assert(0);
+    }
+  }
+
+#ifdef USE_LAPACK
+void LAPACK_dstegr(char *jobz, char *range, int *n, double *d, double *e,
+                   double *vl, double *vu, int *il, int *iu, double *abstol,
+                   int *m, double *w, double *z, int *ldz, int *isuppz,
+                   double *work, int *lwork, int *iwork, int *liwork,
+                   int *info);
+#endif
+
+void diagonalize_lapack(std::vector<RealD>& lmd,
+			std::vector<RealD>& lme, 
+			int Nk, int Nm,  
+			Eigen::MatrixXd& Qt,
+			GridBase *grid)
+{
+#ifdef USE_LAPACK
+  const int size = Nm;
+  int NN = Nk;
+  double evals_tmp[NN];
+  double evec_tmp[NN][NN];
+  memset(evec_tmp[0],0,sizeof(double)*NN*NN);
+  double DD[NN];
+  double EE[NN];
+  for (int i = 0; i< NN; i++) {
+    for (int j = i - 1; j <= i + 1; j++) {
+      if ( j < NN && j >= 0 ) {
+	if (i==j) DD[i] = lmd[i];
+	if (i==j) evals_tmp[i] = lmd[i];
+	if (j==(i-1)) EE[j] = lme[j];
+      }
+    }
+  }
+  int evals_found;
+  int lwork = ( (18*NN) > (1+4*NN+NN*NN)? (18*NN):(1+4*NN+NN*NN)) ;
+  int liwork =  3+NN*10 ;
+  int iwork[liwork];
+  double work[lwork];
+  int isuppz[2*NN];
+  char jobz = 'V'; // calculate evals & evecs
+  char range = 'I'; // calculate all evals
+  //    char range = 'A'; // calculate all evals
+  char uplo = 'U'; // refer to upper half of original matrix
+  char compz = 'I'; // Compute eigenvectors of tridiagonal matrix
+  int ifail[NN];
+  int info;
+  int total = grid->_Nprocessors;
+  int node  = grid->_processor;
+  int interval = (NN/total)+1;
+  double vl = 0.0, vu = 0.0;
+  int il = interval*node+1 , iu = interval*(node+1);
+  if (iu > NN)  iu=NN;
+  double tol = 0.0;
+  if (1) {
+    memset(evals_tmp,0,sizeof(double)*NN);
+    if ( il <= NN){
+      LAPACK_dstegr(&jobz, &range, &NN,
+		    (double*)DD, (double*)EE,
+		    &vl, &vu, &il, &iu, // these four are ignored if second parameteris 'A'
+		    &tol, // tolerance
+		    &evals_found, evals_tmp, (double*)evec_tmp, &NN,
+		    isuppz,
+		    work, &lwork, iwork, &liwork,
+		    &info);
+      for (int i = iu-1; i>= il-1; i--){
+	evals_tmp[i] = evals_tmp[i - (il-1)];
+	if (il>1) evals_tmp[i-(il-1)]=0.;
+	for (int j = 0; j< NN; j++){
+	  evec_tmp[i][j] = evec_tmp[i - (il-1)][j];
+	  if (il>1) evec_tmp[i-(il-1)][j]=0.;
+	}
+      }
+    }
+    {
+      grid->GlobalSumVector(evals_tmp,NN);
+      grid->GlobalSumVector((double*)evec_tmp,NN*NN);
+    }
+  } 
+  // Safer to sort instead of just reversing it, 
+  // but the document of the routine says evals are sorted in increasing order. 
+  // qr gives evals in decreasing order.
+  for(int i=0;i<NN;i++){
+    lmd [NN-1-i]=evals_tmp[i];
+    for(int j=0;j<NN;j++){
+      Qt((NN-1-i),j)=evec_tmp[i][j];
+    }
+  }
+#else 
+  assert(0);
+#endif
+}
+
+void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
+		    int Nk, int Nm,   
+		    Eigen::MatrixXd & Qt,
+		    GridBase *grid)
+{
+  int QRiter = 100*Nm;
+  int kmin = 1;
+  int kmax = Nk;
+  
+  // (this should be more sophisticated)
+  for(int iter=0; iter<QRiter; ++iter){
+    
+    // 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 Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
+    // (Dsh: shift)
+    
+    // transformation
+    QR_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
+    
+    // Convergence criterion (redef of kmin and kamx)
+    for(int j=kmax-1; j>= kmin; --j){
+      RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
+      if(fabs(lme[j-1])+dds > dds){
+	kmax = j+1;
+	goto continued;
+      }
+    }
+    QRiter = iter;
+    return;
+    
+  continued:
+    for(int j=0; j<kmax-1; ++j){
+      RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
+      if(fabs(lme[j])+dds > dds){
+	kmin = j+1;
+	break;
+      }
+    }
+  }
+  std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<QRiter<<"\n";
+  abort();
+}
+};
+}
+#endif

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -0,0 +1,406 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/LocalCoherenceLanczos.h
+
+    Copyright (C) 2015
+
+Author: Christoph Lehner <clehner@bnl.gov>
+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 */
+#ifndef GRID_LOCAL_COHERENCE_IRL_H
+#define GRID_LOCAL_COHERENCE_IRL_H
+
+namespace Grid { 
+
+
+struct LanczosParams : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
+				  ChebyParams, Cheby,/*Chebyshev*/
+				  int, Nstop,    /*Vecs in Lanczos must converge Nstop < Nk < Nm*/
+				  int, Nk,       /*Vecs in Lanczos seek converge*/
+				  int, Nm,       /*Total vecs in Lanczos include restart*/
+				  RealD, resid,  /*residual*/
+ 				  int, MaxIt, 
+				  RealD, betastp,  /* ? */
+				  int, MinRes);    // Must restart
+};
+
+struct LocalCoherenceLanczosParams : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
+				  bool, saveEvecs,
+				  bool, doFine,
+				  bool, doFineRead,
+				  bool, doCoarse,
+	       			  bool, doCoarseRead,
+				  LanczosParams, FineParams,
+				  LanczosParams, CoarseParams,
+				  ChebyParams,   Smoother,
+				  RealD        , coarse_relax_tol,
+				  std::vector<int>, blockSize,
+				  std::string, config,
+				  std::vector < std::complex<double>  >, omega,
+				  RealD, mass,
+				  RealD, M5);
+};
+
+// Duplicate functionality; ProjectedFunctionHermOp could be used with the trivial function
+template<class Fobj,class CComplex,int nbasis>
+class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+  LinearOperatorBase<FineField> &_Linop;
+  std::vector<FineField>        &subspace;
+
+  ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) : 
+    _Linop(linop), subspace(_subspace)
+  {  
+    assert(subspace.size() >0);
+  };
+
+  void operator()(const CoarseField& in, CoarseField& out) {
+    GridBase *FineGrid = subspace[0]._grid;    
+    int   checkerboard = subspace[0].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;
+    blockProject(out,fout,subspace);     std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+
+  OperatorFunction<FineField>   & _poly;
+  LinearOperatorBase<FineField> &_Linop;
+  std::vector<FineField>        &subspace;
+
+  ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,
+			  LinearOperatorBase<FineField>& linop, 
+			  std::vector<FineField> & _subspace) :
+    _poly(poly),
+    _Linop(linop),
+    subspace(_subspace)
+  {  };
+
+  void operator()(const CoarseField& in, CoarseField& out) {
+    
+    GridBase *FineGrid = subspace[0]._grid;    
+    int   checkerboard = subspace[0].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;
+    blockProject(out,fout,subspace);           std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanczosTester<Lattice<iVector<CComplex,nbasis > > >
+{
+ public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+  LinearFunction<CoarseField> & _Poly;
+  OperatorFunction<FineField>   & _smoother;
+  LinearOperatorBase<FineField> &_Linop;
+  RealD                          _coarse_relax_tol;
+  std::vector<FineField>        &_subspace;
+  
+  ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
+					   OperatorFunction<FineField>   &smoother,
+					   LinearOperatorBase<FineField> &Linop,
+					   std::vector<FineField>        &subspace,
+					   RealD coarse_relax_tol=5.0e3) 
+    : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
+      _coarse_relax_tol(coarse_relax_tol)  
+  {    };
+
+  int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+  {
+    CoarseField v(B);
+    RealD eval_poly = eval;
+
+    // Apply operator
+    _Poly(B,v);
+
+    RealD vnum = real(innerProduct(B,v)); // HermOp.
+    RealD vden = norm2(B);
+    RealD vv0  = norm2(v);
+    eval   = vnum/vden;
+    v -= eval*B;
+
+    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+
+    int conv=0;
+    if( (vv<eresid*eresid) ) conv = 1;
+    return conv;
+  }
+  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;
+
+    blockPromote(B,fv,_subspace);  
+    
+    _smoother(_Linop,fv,fB); 
+
+    RealD eval_poly = eval;
+    _Linop.HermOp(fB,fv);
+
+    RealD vnum = real(innerProduct(fB,fv)); // HermOp.
+    RealD vden = norm2(fB);
+    RealD vv0  = norm2(fv);
+    eval   = vnum/vden;
+    fv -= eval*fB;
+    RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
+    if( (vv<eresid*eresid) ) return 1;
+    return 0;
+  }
+};
+
+////////////////////////////////////////////
+// Make serializable Lanczos params
+////////////////////////////////////////////
+template<class Fobj,class CComplex,int nbasis>
+class LocalCoherenceLanczos 
+{
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CComplex>                   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<Fobj>                       FineField;
+
+protected:
+  GridBase *_CoarseGrid;
+  GridBase *_FineGrid;
+  int _checkerboard;
+  LinearOperatorBase<FineField>                 & _FineOp;
+  
+  std::vector<RealD>                              &evals_fine;
+  std::vector<RealD>                              &evals_coarse; 
+  std::vector<FineField>                          &subspace;
+  std::vector<CoarseField>                        &evec_coarse;
+
+private:
+  std::vector<RealD>                              _evals_fine;
+  std::vector<RealD>                              _evals_coarse; 
+  std::vector<FineField>                          _subspace;
+  std::vector<CoarseField>                        _evec_coarse;
+
+public:
+
+  LocalCoherenceLanczos(GridBase *FineGrid,
+			GridBase *CoarseGrid,
+			LinearOperatorBase<FineField> &FineOp,
+			int checkerboard) :
+    _CoarseGrid(CoarseGrid),
+    _FineGrid(FineGrid),
+    _FineOp(FineOp),
+    _checkerboard(checkerboard),
+    evals_fine  (_evals_fine),
+    evals_coarse(_evals_coarse),
+    subspace    (_subspace),
+    evec_coarse(_evec_coarse)
+  {
+    evals_fine.resize(0);
+    evals_coarse.resize(0);
+  };
+  //////////////////////////////////////////////////////////////////////////
+  // Alternate constructore, external storage for use by Hadrons module
+  //////////////////////////////////////////////////////////////////////////
+  LocalCoherenceLanczos(GridBase *FineGrid,
+			GridBase *CoarseGrid,
+			LinearOperatorBase<FineField> &FineOp,
+			int checkerboard,
+			std::vector<FineField>   &ext_subspace,
+			std::vector<CoarseField> &ext_coarse,
+			std::vector<RealD>       &ext_eval_fine,
+			std::vector<RealD>       &ext_eval_coarse
+			) :
+    _CoarseGrid(CoarseGrid),
+    _FineGrid(FineGrid),
+    _FineOp(FineOp),
+    _checkerboard(checkerboard),
+    evals_fine  (ext_eval_fine), 
+    evals_coarse(ext_eval_coarse),
+    subspace    (ext_subspace),
+    evec_coarse (ext_coarse)
+  {
+    evals_fine.resize(0);
+    evals_coarse.resize(0);
+  };
+
+  void Orthogonalise(void ) {
+    CoarseScalar InnerProd(_CoarseGrid);
+    std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
+    blockOrthogonalise(InnerProd,subspace);
+    std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
+    blockOrthogonalise(InnerProd,subspace);
+  };
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = ::sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+  /*
+  void fakeFine(void)
+  {
+    int Nk = nbasis;
+    subspace.resize(Nk,_FineGrid);
+    subspace[0]=1.0;
+    subspace[0].checkerboard=_checkerboard;
+    normalise(subspace[0]);
+    PlainHermOp<FineField>    Op(_FineOp);
+    for(int k=1;k<Nk;k++){
+      subspace[k].checkerboard=_checkerboard;
+      Op(subspace[k-1],subspace[k]);
+      normalise(subspace[k]);
+    }
+  }
+  */
+
+  void testFine(RealD resid) 
+  {
+    assert(evals_fine.size() == nbasis);
+    assert(subspace.size() == nbasis);
+    PlainHermOp<FineField>    Op(_FineOp);
+    ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
+    for(int k=0;k<nbasis;k++){
+      assert(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
+    }
+  }
+
+  void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
+  {
+    assert(evals_fine.size() == nbasis);
+    assert(subspace.size() == nbasis);
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    Chebyshev<FineField>                          ChebySmooth(cheby_smooth);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,subspace);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
+
+    for(int k=0;k<evec_coarse.size();k++){
+      if ( k < nbasis ) { 
+	assert(ChebySmoothTester.ReconstructEval(k,resid,evec_coarse[k],evals_coarse[k],1.0)==1);
+      } else { 
+	assert(ChebySmoothTester.ReconstructEval(k,resid*relax,evec_coarse[k],evals_coarse[k],1.0)==1);
+      }
+    }
+  }
+
+  void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
+		RealD MaxIt, RealD betastp, int MinRes)
+  {
+    assert(nbasis<=Nm);
+    Chebyshev<FineField>      Cheby(cheby_parms);
+    FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
+    PlainHermOp<FineField>    Op(_FineOp);
+
+    evals_fine.resize(Nm);
+    subspace.resize(Nm,_FineGrid);
+
+    ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
+
+    FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
+
+    int Nconv;
+    IRL.calc(evals_fine,subspace,src,Nconv,false);
+    
+    // Shrink down to number saved
+    assert(Nstop>=nbasis);
+    assert(Nconv>=nbasis);
+    evals_fine.resize(nbasis);
+    subspace.resize(nbasis,_FineGrid);
+  }
+  void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
+		  int Nstop, int Nk, int Nm,RealD resid, 
+		  RealD MaxIt, RealD betastp, int MinRes)
+  {
+    Chebyshev<FineField>                          Cheby(cheby_op);
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
+
+    evals_coarse.resize(Nm);
+    evec_coarse.resize(Nm,_CoarseGrid);
+
+    CoarseField src(_CoarseGrid);     src=1.0; 
+
+    ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
+    int Nconv=0;
+    IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
+    assert(Nconv>=Nstop);
+    evals_coarse.resize(Nstop);
+    evec_coarse.resize (Nstop,_CoarseGrid);
+    for (int i=0;i<Nstop;i++){
+      std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl;
+    }
+  }
+};
+
+}
+#endif

Grid/algorithms/iterative/MinimalResidual.h

@@ -0,0 +1,156 @@
+/*************************************************************************************
+
+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:
+  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);
+
+    Complex a, c;
+    Real    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:    mp " << d << 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 +1,273 @@
+/*************************************************************************************
+
+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:
+  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<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> 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 = std::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 - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (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] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD 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]     = std::conj(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] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#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/PrecGeneralisedConjugateResidual.h

@@ -139,8 +139,11 @@ namespace Grid {
       MatTimer.Start();
       Linop.HermOpAndNorm(psi,Az,zAz,zAAz); 
       MatTimer.Stop();
+
+      LinalgTimer.Start();
       r=src-Az;
-      
+      LinalgTimer.Stop();
+
       /////////////////////
       // p = Prec(r)
       /////////////////////
@@ -152,8 +155,10 @@ 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 +171,14 @@ 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 +188,14 @@ 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);  
+	cp = axpy_norm(r,-a,q[peri_k],r);
+        LinalgTimer.Stop();
 
 	if((k==nstep-1)||(cp<rsq)){
 	  return cp;
@@ -202,6 +211,8 @@ 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; 
 
@@ -219,9 +230,9 @@ namespace Grid {
 
 	}
 	qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
-
-
+        LinalgTimer.Stop();
       }
+
       assert(0); // never reached
       return cp;
     }

Grid/algorithms/iterative/SchurRedBlack.h

@@ -0,0 +1,486 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/SchurRedBlack.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_SCHUR_RED_BLACK_H
+#define GRID_SCHUR_RED_BLACK_H
+
+
+  /*
+   * Red black Schur decomposition
+   *
+   *  M = (Mee Meo) =  (1             0 )   (Mee   0               )  (1 Mee^{-1} Meo)
+   *      (Moe Moo)    (Moe Mee^-1    1 )   (0   Moo-Moe Mee^-1 Meo)  (0   1         )
+   *                =         L                     D                     U
+   *
+   * L^-1 = (1              0 )
+   *        (-MoeMee^{-1}   1 )   
+   * L^{dag} = ( 1       Mee^{-dag} Moe^{dag} )
+   *           ( 0       1                    )
+   * L^{-d}  = ( 1      -Mee^{-dag} Moe^{dag} )
+   *           ( 0       1                    )
+   *
+   * U^-1 = (1   -Mee^{-1} Meo)
+   *        (0    1           )
+   * U^{dag} = ( 1                 0)
+   *           (Meo^dag Mee^{-dag} 1)
+   * U^{-dag} = (  1                 0)
+   *            (-Meo^dag Mee^{-dag} 1)
+   ***********************
+   *     M psi = eta
+   ***********************
+   *Odd
+   * i)                 D_oo psi_o =  L^{-1}  eta_o
+   *                        eta_o' = (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * Wilson:
+   *      (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1}  eta_o
+   * Stag:
+   *      D_oo psi_o = L^{-1}  eta =    (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * L^-1 eta_o= (1              0 ) (e
+   *             (-MoeMee^{-1}   1 )   
+   *
+   *Even
+   * ii)  Mee psi_e + Meo psi_o = src_e
+   *
+   *   => sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+   *
+   * 
+   * TODO: Other options:
+   * 
+   * a) change checkerboards for Schur e<->o
+   *
+   * Left precon by Moo^-1
+   * b) Doo^{dag} M_oo^-dag Moo^-1 Doo psi_0 =  (D_oo)^dag M_oo^-dag Moo^-1 L^{-1}  eta_o
+   *                              eta_o'     = (D_oo)^dag  M_oo^-dag Moo^-1 (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * Right precon by Moo^-1
+   * c) M_oo^-dag Doo^{dag} Doo Moo^-1 phi_0 = M_oo^-dag (D_oo)^dag L^{-1}  eta_o
+   *                              eta_o'     = M_oo^-dag (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
+   *                              psi_o = M_oo^-1 phi_o
+   * TODO: Deflation 
+   */
+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
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  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:
+
+    SchurRedBlackBase(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+        const bool _solnAsInitGuess = false)  :
+    _HermitianRBSolver(HermitianRBSolver),
+    useSolnAsInitGuess(_solnAsInitGuess)
+    { 
+      CBfactorise = 0;
+      subtractGuess(initSubGuess);
+    };
+    void subtractGuess(const bool initSubGuess)
+    {
+      subGuess = initSubGuess;
+    }
+    bool isSubtractGuess(void)
+    {
+      return subGuess;
+    }
+
+    /////////////////////////////////////////////////////////////
+    // Shared code
+    /////////////////////////////////////////////////////////////
+    void operator() (Matrix & _Matrix,const Field &in, Field &out){
+      ZeroGuesser<Field> guess;
+      (*this)(_Matrix,in,out,guess);
+    }
+    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
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field resid(fgrid);
+      Field src_o(grid);
+      Field src_e(grid);
+      Field sol_o(grid);
+
+      ////////////////////////////////////////////////
+      // RedBlack source
+      ////////////////////////////////////////////////
+      RedBlackSource(_Matrix,in,src_e,src_o);
+
+      ////////////////////////////////
+      // Construct the guess
+      ////////////////////////////////
+      if(useSolnAsInitGuess) {
+        pickCheckerboard(Odd, sol_o, out);
+      } else {
+        guess(src_o,sol_o);
+      }
+
+      Field  guess_save(grid);
+      guess_save = sol_o;
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      RedBlackSolve(_Matrix,src_o,sol_o);
+
+      ////////////////////////////////////////////////
+      // Fionn A2A boolean behavioural control
+      ////////////////////////////////////////////////
+      if (subGuess)      sol_o= sol_o-guess_save;
+
+      ///////////////////////////////////////////////////
+      // RedBlack solution needs the even source
+      ///////////////////////////////////////////////////
+      RedBlackSolution(_Matrix,sol_o,src_e,out);
+
+      // Verify the unprec residual
+      if ( ! subGuess ) {
+        _Matrix.M(out,resid); 
+        resid = resid-in;
+        RealD ns = norm2(in);
+        RealD nr = norm2(resid);
+
+        std::cout<<GridLogMessage << "SchurRedBlackBase solver true unprec resid "<< std::sqrt(nr/ns) << std::endl;
+      } else {
+        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>;
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Site diagonal has Mooee on it.
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagMooeeSolve : public SchurRedBlackBase<Field> {
+  public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
+
+    SchurRedBlackDiagMooeeSolve(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 = 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   tmp(grid);
+      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_i = src_e-tmp;               assert(  src_e_i.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e_i,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)
+    {
+      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); 
+    }
+  };
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // 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 : public SchurRedBlackBase<Field> {
+  public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+      const bool _solnAsInitGuess = false)  
+    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
+
+    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   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
+      _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   sol_o_i(grid);
+      Field   tmp(grid);
+      Field   sol_e(grid);
+
+      ////////////////////////////////////////////////
+      // 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_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(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

@@ -0,0 +1,125 @@
+#include <Grid/GridCore.h>
+#include <fcntl.h>
+
+namespace Grid {
+
+MemoryStats *MemoryProfiler::stats = nullptr;
+bool         MemoryProfiler::debug = false;
+
+int PointerCache::victim;
+
+PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
+
+void *PointerCache::Insert(void *ptr,size_t bytes) {
+
+  if (bytes < 4096 ) return ptr;
+
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  void * ret = NULL;
+  int v = -1;
+
+  for(int e=0;e<Ncache;e++) {
+    if ( Entries[e].valid==0 ) {
+      v=e; 
+      break;
+    }
+  }
+
+  if ( v==-1 ) {
+    v=victim;
+    victim = (victim+1)%Ncache;
+  }
+
+  if ( Entries[v].valid ) {
+    ret = Entries[v].address;
+    Entries[v].valid = 0;
+    Entries[v].address = NULL;
+    Entries[v].bytes = 0;
+  }
+
+  Entries[v].address=ptr;
+  Entries[v].bytes  =bytes;
+  Entries[v].valid  =1;
+
+  return ret;
+}
+
+void *PointerCache::Lookup(size_t bytes) {
+
+ if (bytes < 4096 ) return NULL;
+
+#ifdef _OPENMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  for(int e=0;e<Ncache;e++){
+    if ( Entries[e].valid && ( Entries[e].bytes == bytes ) ) {
+      Entries[e].valid = 0;
+      return Entries[e].address;
+    }
+  }
+  return NULL;
+}
+
+
+void check_huge_pages(void *Buf,uint64_t BYTES)
+{
+#ifdef __linux__
+  int fd = open("/proc/self/pagemap", O_RDONLY);
+  assert(fd >= 0);
+  const int page_size = 4096;
+  uint64_t virt_pfn = (uint64_t)Buf / page_size;
+  off_t offset = sizeof(uint64_t) * virt_pfn;
+  uint64_t npages = (BYTES + page_size-1) / page_size;
+  uint64_t pagedata[npages];
+  uint64_t ret = lseek(fd, offset, SEEK_SET);
+  assert(ret == offset);
+  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
+  assert(ret == sizeof(uint64_t) * npages);
+  int nhugepages = npages / 512;
+  int n4ktotal, nnothuge;
+  n4ktotal = 0;
+  nnothuge = 0;
+  for (int i = 0; i < nhugepages; ++i) {
+    uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
+    for (int j = 0; j < 512; ++j) {
+      uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
+      ++n4ktotal;
+      if (pageaddr != baseaddr + j * page_size)
+	++nnothuge;
+      }
+  }
+  int rank = CartesianCommunicator::RankWorld();
+  printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
+#endif
+}
+
+std::string sizeString(const size_t bytes)
+{
+  constexpr unsigned int bufSize = 256;
+  const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
+  char                   buf[256];
+  size_t                 s     = 0;
+  double                 count = bytes;
+  
+  while (count >= 1024 && s < 7)
+  {
+      s++;
+      count /= 1024;
+  }
+  if (count - floor(count) == 0.0)
+  {
+      snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
+  }
+  else
+  {
+      snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
+  }
+  
+  return std::string(buf);
+}
+
+}

Grid/allocator/AlignedAllocator.h

@@ -63,6 +63,66 @@ namespace Grid {
     static void *Lookup(size_t bytes) ;
 
   };
+  
+  std::string sizeString(size_t bytes);
+
+  struct MemoryStats
+  {
+    size_t totalAllocated{0}, maxAllocated{0}, 
+           currentlyAllocated{0}, totalFreed{0};
+  };
+    
+  class MemoryProfiler
+  {
+  public:
+    static MemoryStats *stats;
+    static bool        debug;
+  };
+
+  #define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
+  #define profilerDebugPrint \
+  if (MemoryProfiler::stats)\
+  {\
+    auto s = MemoryProfiler::stats;\
+    std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
+              << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
+              << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
+              << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
+              << std::endl;\
+  }
+
+  #define profilerAllocate(bytes)\
+  if (MemoryProfiler::stats)\
+  {\
+    auto s = MemoryProfiler::stats;\
+    s->totalAllocated     += (bytes);\
+    s->currentlyAllocated += (bytes);\
+    s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated);\
+  }\
+  if (MemoryProfiler::debug)\
+  {\
+    std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl;\
+    profilerDebugPrint;\
+  }
+
+  #define profilerFree(bytes)\
+  if (MemoryProfiler::stats)\
+  {\
+    auto s = MemoryProfiler::stats;\
+    s->totalFreed         += (bytes);\
+    s->currentlyAllocated -= (bytes);\
+  }\
+  if (MemoryProfiler::debug)\
+  {\
+    std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl;\
+    profilerDebugPrint;\
+  }
+
+  void check_huge_pages(void *Buf,uint64_t BYTES);
 
 ////////////////////////////////////////////////////////////////////
 // A lattice of something, but assume the something is SIMDized.
@@ -90,20 +150,39 @@ public:
   pointer allocate(size_type __n, const void* _p= 0)
   { 
     size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
 
     _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
-    
-#ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,128);
-#else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(128,bytes);
-#endif
+    //    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 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;
+    // 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){
+      cp[n]=0;
+    }
     return ptr;
   }
 
   void deallocate(pointer __p, size_type __n) { 
     size_type bytes = __n * sizeof(_Tp);
+
+    profilerFree(bytes);
+
     pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
 
 #ifdef HAVE_MM_MALLOC_H
@@ -154,10 +233,13 @@ public:
 #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(__n*sizeof(_Tp),64);
+    _Tp *ptr = (_Tp *) shmem_align(bytes,64);
 #else
-    _Tp *ptr = (_Tp *) shmem_align(64,__n*sizeof(_Tp));
+    _Tp *ptr = (_Tp *) shmem_align(64,bytes);
 #endif
 #ifdef PARANOID_SYMMETRIC_HEAP
     static void * bcast;
@@ -175,20 +257,39 @@ public:
 #endif 
     return ptr;
   }
-  void deallocate(pointer __p, size_type) { 
+  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(__n*sizeof(_Tp),128);
+    _Tp * ptr = (_Tp *) _mm_malloc(bytes, GRID_ALLOC_ALIGN);
 #else
-    _Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
+    _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) { 
+  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

Grid/cartesian/Cartesian_base.h

@@ -44,11 +44,20 @@ namespace Grid{
   class GridBase : public CartesianCommunicator , public GridThread {
 
 public:
-
+    int dummy;
     // 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,
+	     const CartesianCommunicator &parent,
+	     int &split_rank) 
+      : CartesianCommunicator(processor_grid,parent,split_rank) {};
+    GridBase(const std::vector<int> & processor_grid,
+	     const CartesianCommunicator &parent) 
+      : CartesianCommunicator(processor_grid,parent,dummy) {};
+
+    virtual ~GridBase() = default;
 
 
     // Physics Grid information.
@@ -63,13 +72,14 @@ public:
     int _isites;
     int _fsites;                  // _isites*_osites = product(dimensions).
     int _gsites;
-    std::vector<int> _slice_block;   // subslice information
+    std::vector<int> _slice_block;// subslice information
     std::vector<int> _slice_stride;
     std::vector<int> _slice_nblock;
 
-    // Might need these at some point
-    //    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
+    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
+
+    bool _isCheckerBoarded; 
 
 public:
 
@@ -176,6 +186,7 @@ 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;};
@@ -186,17 +197,18 @@ public:
     ////////////////////////////////////////////////////////////////
 
     void show_decomposition(){
-      std::cout << GridLogMessage << "Full Dimensions    : " << _fdimensions << std::endl;
-      std::cout << GridLogMessage << "Global Dimensions  : " << _gdimensions << std::endl;
-      std::cout << GridLogMessage << "Local Dimensions   : " << _ldimensions << std::endl;
-      std::cout << GridLogMessage << "Reduced Dimensions : " << _rdimensions << std::endl;
-      std::cout << GridLogMessage << "Outer strides      : " << _ostride << std::endl;
-      std::cout << GridLogMessage << "Inner strides      : " << _istride << std::endl;
-      std::cout << GridLogMessage << "iSites             : " << _isites << std::endl;
-      std::cout << GridLogMessage << "oSites             : " << _osites << std::endl;
-      std::cout << GridLogMessage << "lSites             : " << lSites() << std::endl;        
-      std::cout << GridLogMessage << "gSites             : " << gSites() << std::endl;
-      std::cout << GridLogMessage << "Nd                 : " << _ndimension << std::endl;             
+      std::cout << GridLogMessage << "\tFull Dimensions    : " << _fdimensions << std::endl;
+      std::cout << GridLogMessage << "\tSIMD layout        : " << _simd_layout << std::endl;
+      std::cout << GridLogMessage << "\tGlobal Dimensions  : " << _gdimensions << std::endl;
+      std::cout << GridLogMessage << "\tLocal Dimensions   : " << _ldimensions << std::endl;
+      std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
+      std::cout << GridLogMessage << "\tOuter strides      : " << _ostride << std::endl;
+      std::cout << GridLogMessage << "\tInner strides      : " << _istride << std::endl;
+      std::cout << GridLogMessage << "\tiSites             : " << _isites << std::endl;
+      std::cout << GridLogMessage << "\toSites             : " << _osites << std::endl;
+      std::cout << GridLogMessage << "\tlSites             : " << lSites() << std::endl;        
+      std::cout << GridLogMessage << "\tgSites             : " << gSites() << std::endl;
+      std::cout << GridLogMessage << "\tNd                 : " << _ndimension << std::endl;             
     } 
 
     ////////////////////////////////////////////////////////////////
@@ -210,9 +222,6 @@ public:
       assert(lidx<lSites());
       Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
     }
-
-
-
     void GlobalCoorToGlobalIndex(const std::vector<int> & gcoor,int & gidx){
       gidx=0;
       int mult=1;

Grid/cartesian/Cartesian_full.h

@@ -0,0 +1,174 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/cartesian/Cartesian_full.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_CARTESIAN_FULL_H
+#define GRID_CARTESIAN_FULL_H
+
+namespace Grid{
+    
+/////////////////////////////////////////////////////////////////////////////////////////
+// Grid Support.
+/////////////////////////////////////////////////////////////////////////////////////////
+
+
+class GridCartesian: public GridBase {
+
+public:
+    int dummy;
+    virtual int  CheckerBoardFromOindexTable (int Oindex) {
+      return 0;
+    }
+    virtual int  CheckerBoardFromOindex (int Oindex)
+    {
+      return 0;
+    }
+    virtual int CheckerBoarded(int dim){
+      return 0;
+    }
+    virtual int CheckerBoard(const std::vector<int> &site){
+        return 0;
+    }
+    virtual int CheckerBoardDestination(int cb,int shift,int dim){
+        return 0;
+    }
+    virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift, int ocb){
+      return shift;
+    }
+    virtual int CheckerBoardShift(int source_cb,int dim,int shift, int osite){
+      return shift;
+    }
+    /////////////////////////////////////////////////////////////////////////
+    // 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,
+		  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,
+		  const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
+    {
+      Init(dimensions,simd_layout,processor_grid);
+    }
+    /////////////////////////////////////////////////////////////////////////
+    // 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)
+    {
+      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)
+    {
+      ///////////////////////
+      // Grid information
+      ///////////////////////
+      _isCheckerBoarded = false;
+      _ndimension = dimensions.size();
+
+      _fdimensions.resize(_ndimension);
+      _gdimensions.resize(_ndimension);
+      _ldimensions.resize(_ndimension);
+      _rdimensions.resize(_ndimension);
+      _simd_layout.resize(_ndimension);
+      _lstart.resize(_ndimension);
+      _lend.resize(_ndimension);
+
+      _ostride.resize(_ndimension);
+      _istride.resize(_ndimension);
+
+      _fsites = _gsites = _osites = _isites = 1;
+
+      for (int d = 0; d < _ndimension; d++)
+      {
+        _fdimensions[d] = dimensions[d];   // Global dimensions
+        _gdimensions[d] = _fdimensions[d]; // Global dimensions
+        _simd_layout[d] = simd_layout[d];
+        _fsites = _fsites * _fdimensions[d];
+        _gsites = _gsites * _gdimensions[d];
+
+        // Use a reduced simd grid
+        _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
+        //std::cout << _ldimensions[d] << "  " << _gdimensions[d] << "  " << _processors[d] << std::endl;
+        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+
+        _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
+        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+
+        _lstart[d] = _processor_coor[d] * _ldimensions[d];
+        _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
+        _osites *= _rdimensions[d];
+        _isites *= _simd_layout[d];
+
+        // Addressing support
+        if (d == 0)
+        {
+          _ostride[d] = 1;
+          _istride[d] = 1;
+        }
+        else
+        {
+          _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
+          _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
+        }
+      }
+
+      ///////////////////////
+      // subplane information
+      ///////////////////////
+      _slice_block.resize(_ndimension);
+      _slice_stride.resize(_ndimension);
+      _slice_nblock.resize(_ndimension);
+
+      int block = 1;
+      int nblock = 1;
+      for (int d = 0; d < _ndimension; d++)
+        nblock *= _rdimensions[d];
+
+      for (int d = 0; d < _ndimension; d++)
+      {
+        nblock /= _rdimensions[d];
+        _slice_block[d] = block;
+        _slice_stride[d] = _ostride[d] * _rdimensions[d];
+        _slice_nblock[d] = nblock;
+        block = block * _rdimensions[d];
+      }
+    };
+
+};
+}
+#endif

Grid/cartesian/Cartesian_red_black.h

@@ -112,151 +112,209 @@ public:
       }
     };
 
-    GridRedBlackCartesian(const GridBase *base) : GridRedBlackCartesian(base->_fdimensions,base->_simd_layout,base->_processors)  {};
+    ////////////////////////////////////////////////////////////
+    // Create Redblack from original grid; require full grid pointer ?
+    ////////////////////////////////////////////////////////////
+    GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
+    {
+      int dims = base->_ndimension;
+      std::vector<int> checker_dim_mask(dims,1);
+      int checker_dim = 0;
+      Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
+    };
 
-    GridRedBlackCartesian(const std::vector<int> &dimensions,
+    ////////////////////////////////////////////////////////////
+    // Create redblack from original grid, with non-trivial checker dim mask
+    ////////////////////////////////////////////////////////////
+    GridRedBlackCartesian(const GridBase *base,
+			  const std::vector<int> &checker_dim_mask,
+			  int checker_dim
+			  ) :  GridBase(base->_processors,*base) 
+    {
+      Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim)  ;
+    }
+
+    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) 
+			  ) :  GridBase(processor_grid,*base) 
     {
       Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
     }
-    GridRedBlackCartesian(const std::vector<int> &dimensions,
+
+    ////////////////////////////////////////////////////////////
+    // 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) 
+			  const std::vector<int> &processor_grid) : GridBase(processor_grid,*base) 
     {
       std::vector<int> checker_dim_mask(dimensions.size(),1);
-      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,0);
+      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,
-	      int checker_dim)
+              const std::vector<int> &simd_layout,
+              const std::vector<int> &processor_grid,
+              const std::vector<int> &checker_dim_mask,
+              int checker_dim)
     {
-    ///////////////////////
-    // Grid information
-    ///////////////////////
+
+      _isCheckerBoarded = true;
       _checker_dim = checker_dim;
-      assert(checker_dim_mask[checker_dim]==1);
+      assert(checker_dim_mask[checker_dim] == 1);
       _ndimension = dimensions.size();
-      assert(checker_dim_mask.size()==_ndimension);
-      assert(processor_grid.size()==_ndimension);
-      assert(simd_layout.size()==_ndimension);
-      
+      assert(checker_dim_mask.size() == _ndimension);
+      assert(processor_grid.size() == _ndimension);
+      assert(simd_layout.size() == _ndimension);
+
       _fdimensions.resize(_ndimension);
       _gdimensions.resize(_ndimension);
       _ldimensions.resize(_ndimension);
       _rdimensions.resize(_ndimension);
       _simd_layout.resize(_ndimension);
-      
+      _lstart.resize(_ndimension);
+      _lend.resize(_ndimension);
+
       _ostride.resize(_ndimension);
       _istride.resize(_ndimension);
-      
+
       _fsites = _gsites = _osites = _isites = 1;
-	
-      _checker_dim_mask=checker_dim_mask;
 
-      for(int d=0;d<_ndimension;d++){
-	_fdimensions[d] = dimensions[d];
-	_gdimensions[d] = _fdimensions[d];
-	_fsites = _fsites * _fdimensions[d];
-	_gsites = _gsites * _gdimensions[d];
-        
-	if (d==_checker_dim) {
-	  _gdimensions[d] = _gdimensions[d]/2; // Remove a checkerboard
-	}
-	_ldimensions[d] = _gdimensions[d]/_processors[d];
+      _checker_dim_mask = checker_dim_mask;
 
-	// Use a reduced simd grid
-	_simd_layout[d] = simd_layout[d];
-	_rdimensions[d]= _ldimensions[d]/_simd_layout[d];
-	assert(_rdimensions[d]>0);
+      for (int d = 0; d < _ndimension; d++)
+      {
+        _fdimensions[d] = dimensions[d];
+        _gdimensions[d] = _fdimensions[d];
+        _fsites = _fsites * _fdimensions[d];
+        _gsites = _gsites * _gdimensions[d];
 
-	// all elements of a simd vector must have same checkerboard.
-	// If Ls vectorised, this must still be the case; e.g. dwf rb5d
-	if ( _simd_layout[d]>1 ) {
-	  if ( checker_dim_mask[d] ) { 
-	    assert( (_rdimensions[d]&0x1) == 0 );
-	  }
-	}
+        if (d == _checker_dim)
+        {
+          assert((_gdimensions[d] & 0x1) == 0);
+          _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
+	  _gsites /= 2;
+        }
+        _ldimensions[d] = _gdimensions[d] / _processors[d];
+        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        _lstart[d] = _processor_coor[d] * _ldimensions[d];
+        _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
 
-	_osites *= _rdimensions[d];
-	_isites *= _simd_layout[d];
-        
-	// Addressing support
-	if ( d==0 ) {
-	  _ostride[d] = 1;
-	  _istride[d] = 1;
-	} else {
-	  _ostride[d] = _ostride[d-1]*_rdimensions[d-1];
-	  _istride[d] = _istride[d-1]*_simd_layout[d-1];
-	}
+        // Use a reduced simd grid
+        _simd_layout[d] = simd_layout[d];
+        _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
+        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        assert(_rdimensions[d] > 0);
 
+        // all elements of a simd vector must have same checkerboard.
+        // If Ls vectorised, this must still be the case; e.g. dwf rb5d
+        if (_simd_layout[d] > 1)
+        {
+          if (checker_dim_mask[d])
+          {
+            assert((_rdimensions[d] & 0x1) == 0);
+          }
+        }
 
+        _osites *= _rdimensions[d];
+        _isites *= _simd_layout[d];
+
+        // Addressing support
+        if (d == 0)
+        {
+          _ostride[d] = 1;
+          _istride[d] = 1;
+        }
+        else
+        {
+          _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
+          _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
+        }
       }
-            
+
       ////////////////////////////////////////////////////////////////////////////////////////////
       // subplane information
       ////////////////////////////////////////////////////////////////////////////////////////////
       _slice_block.resize(_ndimension);
       _slice_stride.resize(_ndimension);
       _slice_nblock.resize(_ndimension);
-        
-      int block =1;
-      int nblock=1;
-      for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
-      
-      for(int d=0;d<_ndimension;d++){
-	nblock/=_rdimensions[d];
-	_slice_block[d] =block;
-	_slice_stride[d]=_ostride[d]*_rdimensions[d];
-	_slice_nblock[d]=nblock;
-	block = block*_rdimensions[d];
+
+      int block = 1;
+      int nblock = 1;
+      for (int d = 0; d < _ndimension; d++)
+        nblock *= _rdimensions[d];
+
+      for (int d = 0; d < _ndimension; d++)
+      {
+        nblock /= _rdimensions[d];
+        _slice_block[d] = block;
+        _slice_stride[d] = _ostride[d] * _rdimensions[d];
+        _slice_nblock[d] = nblock;
+        block = block * _rdimensions[d];
       }
 
       ////////////////////////////////////////////////
       // Create a checkerboard lookup table
       ////////////////////////////////////////////////
       int rvol = 1;
-      for(int d=0;d<_ndimension;d++){
-	rvol=rvol * _rdimensions[d];
+      for (int d = 0; d < _ndimension; d++)
+      {
+        rvol = rvol * _rdimensions[d];
       }
       _checker_board.resize(rvol);
-      for(int osite=0;osite<_osites;osite++){
-	_checker_board[osite] = CheckerBoardFromOindex (osite);
+      for (int osite = 0; osite < _osites; osite++)
+      {
+        _checker_board[osite] = CheckerBoardFromOindex(osite);
       }
-      
     };
-protected:
+
+  protected:
     virtual int oIndex(std::vector<int> &coor)
     {
-      int idx=0;
-      for(int d=0;d<_ndimension;d++) {
-	if( d==_checker_dim ) {
-	  idx+=_ostride[d]*((coor[d]/2)%_rdimensions[d]);
-	} else {
-	  idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
-	}
+      int idx = 0;
+      for (int d = 0; d < _ndimension; d++)
+      {
+        if (d == _checker_dim)
+        {
+          idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
+        }
+        else
+        {
+          idx += _ostride[d] * (coor[d] % _rdimensions[d]);
+        }
       }
       return idx;
     };
-        
+
     virtual int iIndex(std::vector<int> &lcoor)
     {
-        int idx=0;
-        for(int d=0;d<_ndimension;d++) {
-	  if( d==_checker_dim ) {
-	    idx+=_istride[d]*(lcoor[d]/(2*_rdimensions[d]));
-	  } else { 
-	    idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
-	  }
-	}
-        return idx;
+      int idx = 0;
+      for (int d = 0; d < _ndimension; d++)
+      {
+        if (d == _checker_dim)
+        {
+          idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
+        }
+        else
+        {
+          idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
+        }
+      }
+      return idx;
     }
 };
-
 }
 #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/communicator/SharedMemory.h>
 #include <Grid/communicator/Communicator_base.h>
 
 #endif

Grid/communicator/Communicator_base.cc

@@ -26,40 +26,24 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/GridCore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+#include <sys/mman.h>
 
 namespace Grid {
 
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////
-void *              CartesianCommunicator::ShmCommBuf;
-uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 128*1024*1024; 
-CartesianCommunicator::CommunicatorPolicy_t  CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
-
-/////////////////////////////////
-// Alloc, free shmem region
-/////////////////////////////////
-void *CartesianCommunicator::ShmBufferMalloc(size_t bytes){
-  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
-  void *ptr = (void *)heap_top;
-  heap_top  += bytes;
-  heap_bytes+= bytes;
-  if (heap_bytes >= MAX_MPI_SHM_BYTES) {
-    std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
-    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
-    std::cout<< " Current value is " << (MAX_MPI_SHM_BYTES/(1024*1024)) <<std::endl;
-    assert(heap_bytes<MAX_MPI_SHM_BYTES);
-  }
-  return ptr;
-}
-void CartesianCommunicator::ShmBufferFreeAll(void) { 
-  heap_top  =(size_t)ShmBufferSelf();
-  heap_bytes=0;
-}
+CartesianCommunicator::CommunicatorPolicy_t  
+CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
+int CartesianCommunicator::nCommThreads = -1;
 
 /////////////////////////////////
 // Grid information queries
 /////////////////////////////////
+int                      CartesianCommunicator::Dimensions(void)        { return _ndimension; };
 int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
 int                      CartesianCommunicator::BossRank(void)          { return 0; };
 int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
@@ -87,43 +71,6 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);
 }
-
-#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPI3L)
-
-int                      CartesianCommunicator::NodeCount(void)    { return ProcessorCount();};
-
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						       void *xmit,
-						       int xmit_to_rank,
-						       void *recv,
-						       int recv_from_rank,
-						       int bytes)
-{
-  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
-  return 2.0*bytes;
-}
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
-{
-  SendToRecvFromComplete(waitall);
-}
-void CartesianCommunicator::StencilBarrier(void){};
-
-commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
-
-void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
-
-void *CartesianCommunicator::ShmBuffer(int rank) {
-  return NULL;
-}
-void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) { 
-  return NULL;
-}
-void CartesianCommunicator::ShmInitGeneric(void){
-  ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
-  ShmCommBuf=(void *)&ShmBufStorageVector[0];
-}
-
-#endif
   
 }
 

Grid/communicator/Communicator_base.h

@@ -32,114 +32,57 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 // Processor layout information
 ///////////////////////////////////
-#ifdef GRID_COMMS_MPI
-#include <mpi.h>
-#endif
-#ifdef GRID_COMMS_MPI3
-#include <mpi.h>
-#endif
-#ifdef GRID_COMMS_MPI3L
-#include <mpi.h>
-#endif
-#ifdef GRID_COMMS_SHMEM
-#include <mpp/shmem.h>
-#endif
+#include <Grid/communicator/SharedMemory.h>
 
 namespace Grid {
 
-class CartesianCommunicator {
-  public:    
+class CartesianCommunicator : public SharedMemory {
 
-  // 65536 ranks per node adequate for now
-  // 128MB shared memory for comms enought for 48^4 local vol comms
-  // Give external control (command line override?) of this
+public:    
 
-  static const int      MAXLOG2RANKSPERNODE = 16;            
-  static uint64_t MAX_MPI_SHM_BYTES;
+  ////////////////////////////////////////////
+  // Policies
+  ////////////////////////////////////////////
+  enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
+  static CommunicatorPolicy_t CommunicatorPolicy;
+  static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
+  static int       nCommThreads;
 
+  ////////////////////////////////////////////
   // 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.
   int              _processor;       // linear processor rank
   std::vector<int> _processor_coor;  // linear processor coordinate
-  unsigned long _ndimension;
-
-#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPI3L)
-  static MPI_Comm communicator_world;
-         MPI_Comm communicator;
-  typedef MPI_Request CommsRequest_t;
-#else 
-  typedef int CommsRequest_t;
-#endif
-
-  ////////////////////////////////////////////////////////////////////
-  // Helper functionality for SHM Windows common to all other impls
-  ////////////////////////////////////////////////////////////////////
-  // Longer term; drop this in favour of a master / slave model with 
-  // cartesian communicator on a subset of ranks, slave ranks controlled
-  // by group leader with data xfer via shared memory
-  ////////////////////////////////////////////////////////////////////
-#ifdef GRID_COMMS_MPI3
-
-  static int ShmRank;
-  static int ShmSize;
-  static int GroupRank;
-  static int GroupSize;
-  static int WorldRank;
-  static int WorldSize;
-
-  std::vector<int>  WorldDims;
-  std::vector<int>  GroupDims;
-  std::vector<int>  ShmDims;
-  
-  std::vector<int> GroupCoor;
-  std::vector<int> ShmCoor;
-  std::vector<int> WorldCoor;
-
-  static std::vector<int> GroupRanks; 
-  static std::vector<int> MyGroup;
-  static int ShmSetup;
-  static MPI_Win ShmWindow; 
-  static MPI_Comm ShmComm;
-  
-  std::vector<int>  LexicographicToWorldRank;
-  
-  static std::vector<void *> ShmCommBufs;
-
-#else 
-  static void ShmInitGeneric(void);
-  static commVector<uint8_t> ShmBufStorageVector;
-#endif 
-
-  /////////////////////////////////
-  // Grid information and queries
-  // Implemented in Communicator_base.C
-  /////////////////////////////////
-  static void * ShmCommBuf;
-
-  // Isend/Irecv/Wait, or Sendrecv blocking
-  enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
-  static CommunicatorPolicy_t CommunicatorPolicy;
-  static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
-
-  size_t heap_top;
-  size_t heap_bytes;
-
-  void *ShmBufferSelf(void);
-  void *ShmBuffer(int rank);
-  void *ShmBufferTranslate(int rank,void * local_p);
-  void *ShmBufferMalloc(size_t bytes);
-  void ShmBufferFreeAll(void) ;
+  unsigned long    _ndimension;
+  static Grid_MPI_Comm      communicator_world;
+  Grid_MPI_Comm             communicator;
+  std::vector<Grid_MPI_Comm> communicator_halo;
   
   ////////////////////////////////////////////////
   // Must call in Grid startup
   ////////////////////////////////////////////////
   static void Init(int *argc, char ***argv);
-  
+
   ////////////////////////////////////////////////
-  // Constructor of any given grid
+  // 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);
+  virtual ~CartesianCommunicator();
+
+ private:
+
+  ////////////////////////////////////////////////
+  // 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);
+
+ public:
+
   
   ////////////////////////////////////////////////////////////////////////////////////////
   // Wraps MPI_Cart routines, or implements equivalent on other impls
@@ -148,13 +91,13 @@ class CartesianCommunicator {
   int  RankFromProcessorCoor(std::vector<int> &coor);
   void ProcessorCoorFromRank(int rank,std::vector<int> &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)     ;
   int                      ProcessorCount(void)    ;
-  int                      NodeCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -175,6 +118,8 @@ class CartesianCommunicator {
   void GlobalSumVector(ComplexF *c,int N);
   void GlobalSum(ComplexD &c);
   void GlobalSumVector(ComplexD *c,int N);
+  void GlobalXOR(uint32_t &);
+  void GlobalXOR(uint64_t &);
   
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
@@ -207,14 +152,21 @@ class CartesianCommunicator {
   
   void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
 
+  double StencilSendToRecvFrom(void *xmit,
+			       int xmit_to_rank,
+			       void *recv,
+			       int recv_from_rank,
+			       int bytes,int dir);
+
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-				  void *xmit,
-				  int xmit_to_rank,
-				  void *recv,
-				  int recv_from_rank,
-				  int bytes);
+				    void *xmit,
+				    int xmit_to_rank,
+				    void *recv,
+				    int recv_from_rank,
+				    int bytes,int dir);
   
-  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+  
+  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
   void StencilBarrier(void);
 
   ////////////////////////////////////////////////////////////
@@ -226,6 +178,23 @@ class CartesianCommunicator {
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
   void Broadcast(int root,void* data, int bytes);
+
+  ////////////////////////////////////////////////////////////
+  // All2All down one dimension
+  ////////////////////////////////////////////////////////////
+  template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
+    assert(dim>=0);
+    assert(dim<_ndimension);
+    assert(in.size()==out.size());
+    int numnode = _processors[dim];
+    uint64_t bytes=sizeof(T);
+    uint64_t words=in.size()/numnode;
+    assert(numnode * words == in.size());
+    assert(words < (1ULL<<31));
+    AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
+  }
+  void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);
+  void AllToAll(void  *in,void *out,uint64_t words         ,uint64_t bytes);
   
   template<class obj> void Broadcast(int root,obj &data)
     {

Grid/communicator/Communicator_mpi3.cc

@@ -0,0 +1,509 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/Communicator_mpi.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/GridCore.h>
+#include <Grid/communicator/SharedMemory.h>
+
+namespace Grid {
+
+Grid_MPI_Comm       CartesianCommunicator::communicator_world;
+
+////////////////////////////////////////////
+// First initialise of comms system
+////////////////////////////////////////////
+void CartesianCommunicator::Init(int *argc, char ***argv) 
+{
+
+  int flag;
+  int provided;
+
+  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) )
+      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();
+}
+
+///////////////////////////////////////////////////////////////////////////
+// Use cartesian communicators now even in MPI3
+///////////////////////////////////////////////////////////////////////////
+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 rank;
+  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
+  assert(ierr==0);
+  return rank;
+}
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
+{
+  coor.resize(_ndimension);
+  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
+  assert(ierr==0);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Initialises from communicator_world
+////////////////////////////////////////////////////////////////////////////////////////////////////////
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
+{
+  MPI_Comm optimal_comm;
+  ////////////////////////////////////////////////////
+  // Remap using the shared memory optimising routine
+  // The remap creates a comm which must be freed
+  ////////////////////////////////////////////////////
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
+  InitFromMPICommunicator(processors,optimal_comm);
+  SetCommunicator(optimal_comm);
+  ///////////////////////////////////////////////////
+  // Free the temp communicator
+  ///////////////////////////////////////////////////
+  MPI_Comm_free(&optimal_comm);
+}
+
+//////////////////////////////////
+// Try to subdivide communicator
+//////////////////////////////////
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)    
+{
+  _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);
+
+  // Can make 5d grid from 4d etc...
+  int pad = _ndimension-parent_ndimension;
+  for(int d=0;d<parent_ndimension;d++){
+    parent_processor_coor[pad+d]=parent._processor_coor[d];
+    parent_processors    [pad+d]=parent._processors[d];
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // split the communicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  //  int Nparent = parent._processors ; 
+  int Nparent;
+  MPI_Comm_size(parent.communicator,&Nparent);
+
+  int childsize=1;
+  for(int d=0;d<processors.size();d++) {
+    childsize *= processors[d];
+  }
+  int Nchild = Nparent/childsize;
+  assert (childsize * Nchild == Nparent);
+
+  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
+
+  for(int d=0;d<_ndimension;d++){
+    ccoor[d] = parent_processor_coor[d] % processors[d];
+    scoor[d] = parent_processor_coor[d] / processors[d];
+    ssize[d] = parent_processors[d]     / processors[d];
+  }
+
+  // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
+  int crank;  
+  // Mpi uses the reverse Lexico convention to us; so reversed routines called
+  Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
+  Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
+
+  MPI_Comm comm_split;
+  if ( Nchild > 1 ) { 
+
+    if(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
+    ////////////////////////////////////////////////////////////////
+    int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
+    assert(ierr==0);
+
+  } else {
+    srank = 0;
+    int ierr = MPI_Comm_dup (parent.communicator,&comm_split);
+    assert(ierr==0);
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Set up from the new split communicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  InitFromMPICommunicator(processors,comm_split);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take the right SHM buffers
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  SetCommunicator(comm_split);
+  
+  ///////////////////////////////////////////////
+  // Free the temp communicator 
+  ///////////////////////////////////////////////
+  MPI_Comm_free(&comm_split);
+
+  if(0){ 
+    std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
+    for(int d=0;d<processors.size();d++){
+      std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
+    }
+  }
+  for(int d=0;d<processors.size();d++){
+    assert(_processor_coor[d] == ccoor[d] );
+  }
+}
+
+void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
+{
+  ////////////////////////////////////////////////////
+  // Creates communicator, and the communicator_halo
+  ////////////////////////////////////////////////////
+  _ndimension = processors.size();
+  _processor_coor.resize(_ndimension);
+
+  /////////////////////////////////
+  // Count the requested nodes
+  /////////////////////////////////
+  _Nprocessors=1;
+  _processors = processors;
+  for(int i=0;i<_ndimension;i++){
+    _Nprocessors*=_processors[i];
+  }
+
+  std::vector<int> 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]);
+
+  if ( 0 && (communicator_base != communicator_world) ) {
+    std::cout << "InitFromMPICommunicator Cartesian communicator created with a non-world communicator"<<std::endl;
+    std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";
+    for(int d=0;d<_processors.size();d++){
+      std::cout << _processor_coor[d]<<" ";
+    }
+    std::cout << std::endl;
+  }
+
+  int Size;
+  MPI_Comm_size(communicator,&Size);
+
+  communicator_halo.resize (2*_ndimension);
+  for(int i=0;i<_ndimension*2;i++){
+    MPI_Comm_dup(communicator,&communicator_halo[i]);
+  }
+  assert(Size==_Nprocessors);
+}
+
+CartesianCommunicator::~CartesianCommunicator()
+{
+  int MPI_is_finalised;
+  MPI_Finalized(&MPI_is_finalised);
+  if (communicator && !MPI_is_finalised) {
+    MPI_Comm_free(&communicator);
+    for(int i=0;i<communicator_halo.size();i++){
+      MPI_Comm_free(&communicator_halo[i]);
+    }
+  }  
+}
+void CartesianCommunicator::GlobalSum(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(float &f){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(float *f,int N)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(double *d,int N)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFrom(void *xmit,
+					   int dest,
+					   void *recv,
+					   int from,
+					   int bytes)
+{
+  std::vector<CommsRequest_t> reqs(0);
+  //    unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  //    unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  //    xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
+  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
+  SendToRecvFromComplete(reqs);
+  //    rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+  //    printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
+}
+void CartesianCommunicator::SendRecvPacket(void *xmit,
+					   void *recv,
+					   int sender,
+					   int receiver,
+					   int bytes)
+{
+  MPI_Status stat;
+  assert(sender != receiver);
+  int tag = sender;
+  if ( _processor == sender ) {
+    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
+  }
+  if ( _processor == receiver ) { 
+    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
+  }
+}
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes)
+{
+  int myrank = _processor;
+  int ierr;
+
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
+
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    
+    assert(ierr==0);
+    list.push_back(xrq);
+    list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
+  }
+}
+
+double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
+						     int dest,
+						     void *recv,
+						     int from,
+						     int bytes,int dir)
+{
+  std::vector<CommsRequest_t> list;
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
+  StencilSendToRecvFromComplete(list,dir);
+  return offbytes;
+}
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,
+							 int dest,
+							 void *recv,
+							 int from,
+							 int bytes,int dir)
+{
+  int ncomm  =communicator_halo.size(); 
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+
+  if ( gfrom ==MPI_UNDEFINED) {
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
+    assert(ierr==0);
+    list.push_back(rrq);
+    off_node_bytes+=bytes;
+  }
+
+  if ( gdest == MPI_UNDEFINED ) {
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
+    assert(ierr==0);
+    list.push_back(xrq);
+    off_node_bytes+=bytes;
+  }
+
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
+    this->StencilSendToRecvFromComplete(list,dir);
+  }
+
+  return off_node_bytes;
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
+{
+  SendToRecvFromComplete(waitall);
+}
+void CartesianCommunicator::StencilBarrier(void)
+{
+  MPI_Barrier  (ShmComm);
+}
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  int nreq=list.size();
+
+  if (nreq==0) return;
+
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+  list.resize(0);
+}
+void CartesianCommunicator::Barrier(void)
+{
+  int ierr = MPI_Barrier(communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+{
+  int ierr=MPI_Bcast(data,
+		     bytes,
+		     MPI_BYTE,
+		     root,
+		     communicator);
+  assert(ierr==0);
+}
+int CartesianCommunicator::RankWorld(void){ 
+  int r; 
+  MPI_Comm_rank(communicator_world,&r);
+  return r;
+}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+  int ierr= MPI_Bcast(data,
+		      bytes,
+		      MPI_BYTE,
+		      root,
+		      communicator_world);
+  assert(ierr==0);
+}
+
+void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  std::vector<int> row(_ndimension,1);
+  assert(dim>=0 && dim<_ndimension);
+
+  //  Split the communicator
+  row[dim] = _processors[dim];
+
+  int me;
+  CartesianCommunicator Comm(row,*this,me);
+  Comm.AllToAll(in,out,words,bytes);
+}
+void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  // MPI is a pain and uses "int" arguments
+  // 64*64*64*128*16 == 500Million elements of data.
+  // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
+  // (Turns up on 32^3 x 64 Gparity too)
+  MPI_Datatype object;
+  int iwords; 
+  int ibytes;
+  iwords = words;
+  ibytes = bytes;
+  assert(words == iwords); // safe to cast to int ?
+  assert(bytes == ibytes); // safe to cast to int ?
+  MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
+  MPI_Type_commit(&object);
+  MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
+  MPI_Type_free(&object);
+}
+
+
+
+}
+

Grid/communicator/Communicator_none.cc

@@ -32,16 +32,27 @@ namespace Grid {
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////////////////////////////////////////
+Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
-  ShmInitGeneric();
+  GlobalSharedMemory::Init(communicator_world);
+  GlobalSharedMemory::SharedMemoryAllocate(
+		   GlobalSharedMemory::MAX_MPI_SHM_BYTES,
+		   GlobalSharedMemory::Hugepages);
+}
+
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank) 
+  : CartesianCommunicator(processors) 
+{
+  srank=0;
+  SetCommunicator(communicator_world);
 }
 
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _processors = processors;
-  _ndimension = processors.size();
+  _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
   
   // Require 1^N processor grid for fake
@@ -51,14 +62,19 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
     assert(_processors[d]==1);
     _processor_coor[d] = 0;
   }
+  SetCommunicator(communicator_world);
 }
 
+CartesianCommunicator::~CartesianCommunicator(){}
+
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}
 void CartesianCommunicator::GlobalSum(uint32_t &){}
 void CartesianCommunicator::GlobalSum(uint64_t &){}
 void CartesianCommunicator::GlobalSumVector(double *,int N){}
+void CartesianCommunicator::GlobalXOR(uint32_t &){}
+void CartesianCommunicator::GlobalXOR(uint64_t &){}
 
 void CartesianCommunicator::SendRecvPacket(void *xmit,
 					   void *recv,
@@ -93,6 +109,14 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
 {
   assert(0);
 }
+void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  bcopy(in,out,bytes*words);
+}
+void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  bcopy(in,out,bytes*words);
+}
 
 int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
@@ -106,6 +130,36 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   dest=0;
 }
 
+double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
+						     int xmit_to_rank,
+						     void *recv,
+						     int recv_from_rank,
+						     int bytes, int dir)
+{
+  std::vector<CommsRequest_t> list;
+  // Discard the "dir"
+  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  SendToRecvFromComplete(list);
+  return 2.0*bytes;
+}
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,
+							 int xmit_to_rank,
+							 void *recv,
+							 int recv_from_rank,
+							 int bytes, int dir)
+{
+  // Discard the "dir"
+  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  return 2.0*bytes;
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
+{
+  SendToRecvFromComplete(waitall);
+}
+
+void CartesianCommunicator::StencilBarrier(void){};
+
 
 }
 

Grid/communicator/SharedMemory.cc

@@ -0,0 +1,92 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/GridCore.h>
+
+namespace Grid { 
+
+// static data
+
+uint64_t            GlobalSharedMemory::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
+int                 GlobalSharedMemory::Hugepages = 0;
+int                 GlobalSharedMemory::_ShmSetup;
+int                 GlobalSharedMemory::_ShmAlloc;
+uint64_t            GlobalSharedMemory::_ShmAllocBytes;
+
+std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
+
+Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm;
+int                 GlobalSharedMemory::WorldShmRank;
+int                 GlobalSharedMemory::WorldShmSize;
+std::vector<int>    GlobalSharedMemory::WorldShmRanks;
+
+Grid_MPI_Comm       GlobalSharedMemory::WorldComm;
+int                 GlobalSharedMemory::WorldSize;
+int                 GlobalSharedMemory::WorldRank;
+
+int                 GlobalSharedMemory::WorldNodes;
+int                 GlobalSharedMemory::WorldNode;
+
+void GlobalSharedMemory::SharedMemoryFree(void)
+{
+  assert(_ShmAlloc);
+  assert(_ShmAllocBytes>0);
+  for(int r=0;r<WorldShmSize;r++){
+    munmap(WorldShmCommBufs[r],_ShmAllocBytes);
+  }
+  _ShmAlloc = 0;
+  _ShmAllocBytes = 0;
+}
+/////////////////////////////////
+// Alloc, free shmem region
+/////////////////////////////////
+void *SharedMemory::ShmBufferMalloc(size_t bytes){
+  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
+  void *ptr = (void *)heap_top;
+  heap_top  += bytes;
+  heap_bytes+= bytes;
+  if (heap_bytes >= heap_size) {
+    std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
+    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
+    std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
+    assert(heap_bytes<heap_size);
+  }
+  return ptr;
+}
+void SharedMemory::ShmBufferFreeAll(void) { 
+  heap_top  =(size_t)ShmBufferSelf();
+  heap_bytes=0;
+}
+void *SharedMemory::ShmBufferSelf(void)
+{
+  return ShmCommBufs[ShmRank];
+}
+
+
+
+}

Grid/communicator/SharedMemory.h

@@ -0,0 +1,167 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+
+// 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>
+
+#if defined (GRID_COMMS_MPI3) 
+#include <mpi.h>
+#endif 
+#include <semaphore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+#include <sys/types.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+#include <sys/mman.h>
+#include <zlib.h>
+#ifdef HAVE_NUMAIF_H
+#include <numaif.h>
+#endif
+
+namespace Grid {
+
+#if defined (GRID_COMMS_MPI3) 
+  typedef MPI_Comm    Grid_MPI_Comm;
+  typedef MPI_Request CommsRequest_t;
+#else 
+  typedef int CommsRequest_t;
+  typedef int Grid_MPI_Comm;
+#endif
+
+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:
+  static int      ShmSetup(void)      { return _ShmSetup; }
+  static int      ShmAlloc(void)      { return _ShmAlloc; }
+  static uint64_t ShmAllocBytes(void) { return _ShmAllocBytes; }
+  static uint64_t      MAX_MPI_SHM_BYTES;
+  static int           Hugepages;
+
+  static std::vector<void *> WorldShmCommBufs;
+
+  static Grid_MPI_Comm WorldComm;
+  static int           WorldRank;
+  static int           WorldSize;
+
+  static Grid_MPI_Comm WorldShmComm;
+  static int           WorldShmRank;
+  static int           WorldShmSize;
+
+  static int           WorldNodes;
+  static int           WorldNode;
+
+  static std::vector<int>  WorldShmRanks;
+
+  //////////////////////////////////////////////////////////////////////////////////////
+  // 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 OptimalCommunicatorHypercube(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicatorSharedMemory(const std::vector<int> &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);
+
+};
+
+//////////////////////////////
+// one per communicator
+//////////////////////////////
+class SharedMemory 
+{
+ private:
+  static const int     MAXLOG2RANKSPERNODE = 16;            
+
+  size_t heap_top;
+  size_t heap_bytes;
+  size_t heap_size;
+
+ protected:
+
+  Grid_MPI_Comm    ShmComm; // for barriers
+  int    ShmRank; 
+  int    ShmSize;
+  std::vector<void *> ShmCommBufs;
+  std::vector<int>    ShmRanks;// Mapping comm ranks to Shm ranks
+
+ public:
+  SharedMemory() {};
+  ~SharedMemory();
+  ///////////////////////////////////////////////////////////////////////////////////////
+  // set the buffers & sizes
+  ///////////////////////////////////////////////////////////////////////////////////////
+  void SetCommunicator(Grid_MPI_Comm comm);
+
+  ////////////////////////////////////////////////////////////////////////
+  // For this instance ; disjoint buffer sets between splits if split grid
+  ////////////////////////////////////////////////////////////////////////
+  void ShmBarrier(void); 
+
+  ///////////////////////////////////////////////////
+  // Call on any instance
+  ///////////////////////////////////////////////////
+  void SharedMemoryTest(void);
+  void *ShmBufferSelf(void);
+  void *ShmBuffer    (int rank);
+  void *ShmBufferTranslate(int rank,void * local_p);
+  void *ShmBufferMalloc(size_t bytes);
+  void  ShmBufferFreeAll(void) ;
+  
+  //////////////////////////////////////////////////////////////////////////
+  // Make info on Nodes & ranks and Shared memory available
+  //////////////////////////////////////////////////////////////////////////
+  int NodeCount(void) { return GlobalSharedMemory::WorldNodes;};
+  int RankCount(void) { return GlobalSharedMemory::WorldSize;};
+
+};
+
+}

Grid/communicator/SharedMemoryMPI.cc

@@ -0,0 +1,667 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/GridCore.h>
+#include <pwd.h>
+
+namespace Grid { 
+
+/*Construct from an MPI communicator*/
+void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
+{
+  assert(_ShmSetup==0);
+  WorldComm = comm;
+  MPI_Comm_rank(WorldComm,&WorldRank);
+  MPI_Comm_size(WorldComm,&WorldSize);
+  // WorldComm, WorldSize, WorldRank
+
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory
+  /////////////////////////////////////////////////////////////////////
+  MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
+  MPI_Comm_rank(WorldShmComm     ,&WorldShmRank);
+  MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
+  // WorldShmComm, WorldShmSize, WorldShmRank
+
+  // WorldNodes
+  WorldNodes = WorldSize/WorldShmSize;
+  assert( (WorldNodes * WorldShmSize) == WorldSize );
+
+  // FIXME: Check all WorldShmSize are the same ?
+
+  /////////////////////////////////////////////////////////////////////
+  // find world ranks in our SHM group (i.e. which ranks are on our node)
+  /////////////////////////////////////////////////////////////////////
+  MPI_Group WorldGroup, ShmGroup;
+  MPI_Comm_group (WorldComm, &WorldGroup); 
+  MPI_Comm_group (WorldShmComm, &ShmGroup);
+
+  std::vector<int> world_ranks(WorldSize);   for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
+
+  WorldShmRanks.resize(WorldSize); 
+  MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &WorldShmRanks[0]); 
+
+  ///////////////////////////////////////////////////////////////////
+  // Identify who is in my group and nominate the leader
+  ///////////////////////////////////////////////////////////////////
+  int g=0;
+  std::vector<int> MyGroup;
+  MyGroup.resize(WorldShmSize);
+  for(int rank=0;rank<WorldSize;rank++){
+    if(WorldShmRanks[rank]!=MPI_UNDEFINED){
+      assert(g<WorldShmSize);
+      MyGroup[g++] = rank;
+    }
+  }
+  
+  std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
+  int myleader = MyGroup[0];
+  
+  std::vector<int> leaders_1hot(WorldSize,0);
+  std::vector<int> leaders_group(WorldNodes,0);
+  leaders_1hot [ myleader ] = 1;
+    
+  ///////////////////////////////////////////////////////////////////
+  // global sum leaders over comm world
+  ///////////////////////////////////////////////////////////////////
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,WorldComm);
+  assert(ierr==0);
+
+  ///////////////////////////////////////////////////////////////////
+  // find the group leaders world rank
+  ///////////////////////////////////////////////////////////////////
+  int group=0;
+  for(int l=0;l<WorldSize;l++){
+    if(leaders_1hot[l]){
+      leaders_group[group++] = l;
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////
+  // Identify the node of the group in which I (and my leader) live
+  ///////////////////////////////////////////////////////////////////
+  WorldNode=-1;
+  for(int g=0;g<WorldNodes;g++){
+    if (myleader == leaders_group[g]){
+      WorldNode=g;
+    }
+  }
+  assert(WorldNode!=-1);
+  _ShmSetup=1;
+}
+// Gray encode support 
+int BinaryToGray (int  binary) {
+  int gray = (binary>>1)^binary;
+  return gray;
+}
+int Log2Size(int TwoToPower,int MAXLOG2)
+{
+  int log2size = -1;
+  for(int i=0;i<=MAXLOG2;i++){
+    if ( (0x1<<i) == TwoToPower ) {
+      log2size = i;
+      break;
+    }
+  }
+  return log2size;
+}
+void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &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) OptimalCommunicatorHypercube(processors,optimal_comm);
+  else         OptimalCommunicatorSharedMemory(processors,optimal_comm);
+}
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
+  assert(log2size != -1);
+
+  ////////////////////////////////////////////////////////////////
+  // Identify the hypercube coordinate of this node using hostname
+  ////////////////////////////////////////////////////////////////
+  // n runs 0...7 9...16 18...25 27...34     (8*4)  5 bits
+  // i runs 0..7                                    3 bits
+  // r runs 0..3                                    2 bits
+  // 2^10 = 1024 nodes
+  const int maxhdim = 10; 
+  std::vector<int> HyperCubeCoords(maxhdim,0);
+  std::vector<int> RootHyperCubeCoords(maxhdim,0);
+  int R;
+  int I;
+  int N;
+  const int namelen = _POSIX_HOST_NAME_MAX;
+  char name[namelen];
+
+  // Parse ICE-XA hostname to get hypercube location
+  gethostname(name,namelen);
+  int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
+  assert(nscan==3);
+
+  int nlo = N%9;
+  int nhi = N/9;
+  uint32_t hypercoor = (R<<8)|(I<<5)|(nhi<<3)|nlo ;
+  uint32_t rootcoor  = hypercoor;
+
+  //////////////////////////////////////////////////////////////////
+  // Print debug info
+  //////////////////////////////////////////////////////////////////
+  for(int d=0;d<maxhdim;d++){
+    HyperCubeCoords[d] = (hypercoor>>d)&0x1;
+  }
+
+  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;
+
+  //////////////////////////////////////////////////////////////////
+  // broadcast node 0's base coordinate for this partition.
+  //////////////////////////////////////////////////////////////////
+  MPI_Bcast(&rootcoor, sizeof(rootcoor), MPI_BYTE, 0, WorldComm); 
+  hypercoor=hypercoor-rootcoor;
+  assert(hypercoor<WorldSize);
+  assert(hypercoor>=0);
+
+  //////////////////////////////////////
+  // Printing
+  //////////////////////////////////////
+  for(int d=0;d<maxhdim;d++){
+    HyperCubeCoords[d] = (hypercoor>>d)&0x1;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Identify subblock of ranks on node spreading across dims
+  // 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);
+  std::vector<int> HyperCoor(ndimension);
+  int dim = 0;
+  for(int l2=0;l2<log2size;l2++){
+    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
+    ShmDims[dim]*=2;
+    dim=(dim+1)%ndimension;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Establish torus of processes and nodes with sub-blockings
+  ////////////////////////////////////////////////////////////////
+  for(int d=0;d<ndimension;d++){
+    NodeDims[d] = WorldDims[d]/ShmDims[d];
+  }
+  ////////////////////////////////////////////////////////////////
+  // Map Hcube according to physical lattice 
+  // must partition. Loop over dims and find out who would join.
+  ////////////////////////////////////////////////////////////////
+  int hcoor = hypercoor;
+  for(int d=0;d<ndimension;d++){
+     int bits = Log2Size(NodeDims[d],MAXLOG2RANKSPERNODE);
+     int msk  = (0x1<<bits)-1;
+     HyperCoor[d]=hcoor & msk;  
+     HyperCoor[d]=BinaryToGray(HyperCoor[d]); // Space filling curve magic
+     hcoor = hcoor >> bits;
+  } 
+  ////////////////////////////////////////////////////////////////
+  // Check processor counts match
+  ////////////////////////////////////////////////////////////////
+  int Nprocessors=1;
+  for(int i=0;i<ndimension;i++){
+    Nprocessors*=processors[i];
+  }
+  assert(WorldSize==Nprocessors);
+
+  ////////////////////////////////////////////////////////////////
+  // Establish mapping between lexico physics coord and WorldRank
+  ////////////////////////////////////////////////////////////////
+  int rank;
+
+  Lexicographic::CoorFromIndexReversed(NodeCoor,WorldNode   ,NodeDims);
+
+  for(int d=0;d<ndimension;d++) NodeCoor[d]=HyperCoor[d];
+
+  Lexicographic::CoorFromIndexReversed(ShmCoor ,WorldShmRank,ShmDims);
+  for(int d=0;d<ndimension;d++) WorldCoor[d] = NodeCoor[d]*ShmDims[d]+ShmCoor[d];
+  Lexicographic::IndexFromCoorReversed(WorldCoor,rank,WorldDims);
+
+  /////////////////////////////////////////////////////////////////
+  // Build the new communicator
+  /////////////////////////////////////////////////////////////////
+  int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
+  assert(ierr==0);
+}
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
+  assert(log2size != -1);
+
+  ////////////////////////////////////////////////////////////////
+  // Identify subblock of ranks on node spreading across dims
+  // 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);
+  int dim = 0;
+  for(int l2=0;l2<log2size;l2++){
+    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
+    ShmDims[dim]*=2;
+    dim=(dim+1)%ndimension;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Establish torus of processes and nodes with sub-blockings
+  ////////////////////////////////////////////////////////////////
+  for(int d=0;d<ndimension;d++){
+    NodeDims[d] = WorldDims[d]/ShmDims[d];
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Check processor counts match
+  ////////////////////////////////////////////////////////////////
+  int Nprocessors=1;
+  for(int i=0;i<ndimension;i++){
+    Nprocessors*=processors[i];
+  }
+  assert(WorldSize==Nprocessors);
+
+  ////////////////////////////////////////////////////////////////
+  // Establish mapping between lexico physics coord and WorldRank
+  ////////////////////////////////////////////////////////////////
+  int rank;
+
+  Lexicographic::CoorFromIndexReversed(NodeCoor,WorldNode   ,NodeDims);
+  Lexicographic::CoorFromIndexReversed(ShmCoor ,WorldShmRank,ShmDims);
+  for(int d=0;d<ndimension;d++) WorldCoor[d] = NodeCoor[d]*ShmDims[d]+ShmCoor[d];
+  Lexicographic::IndexFromCoorReversed(WorldCoor,rank,WorldDims);
+
+  /////////////////////////////////////////////////////////////////
+  // Build the new communicator
+  /////////////////////////////////////////////////////////////////
+  int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
+  assert(ierr==0);
+}
+////////////////////////////////////////////////////////////////////////////////////////////
+// SHMGET
+////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMGET
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  std::cout << "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+  std::vector<int> shmids(WorldShmSize);
+
+  if ( WorldShmRank == 0 ) {
+    for(int r=0;r<WorldShmSize;r++){
+      size_t size = bytes;
+      key_t key   = IPC_PRIVATE;
+      int flags = IPC_CREAT | SHM_R | SHM_W;
+#ifdef SHM_HUGETLB
+      if (Hugepages) flags|=SHM_HUGETLB;
+#endif
+      if ((shmids[r]= shmget(key,size, flags)) ==-1) {
+        int errsv = errno;
+        printf("Errno %d\n",errsv);
+        printf("key   %d\n",key);
+        printf("size  %ld\n",size);
+        printf("flags %d\n",flags);
+        perror("shmget");
+        exit(1);
+      }
+    }
+  }
+  MPI_Barrier(WorldShmComm);
+  MPI_Bcast(&shmids[0],WorldShmSize*sizeof(int),MPI_BYTE,0,WorldShmComm);
+  MPI_Barrier(WorldShmComm);
+
+  for(int r=0;r<WorldShmSize;r++){
+    WorldShmCommBufs[r] = (uint64_t *)shmat(shmids[r], NULL,0);
+    if (WorldShmCommBufs[r] == (uint64_t *)-1) {
+      perror("Shared memory attach failure");
+      shmctl(shmids[r], IPC_RMID, NULL);
+      exit(2);
+    }
+  }
+  MPI_Barrier(WorldShmComm);
+  ///////////////////////////////////
+  // Mark for clean up
+  ///////////////////////////////////
+  for(int r=0;r<WorldShmSize;r++){
+    shmctl(shmids[r], IPC_RMID,(struct shmid_ds *)NULL);
+  }
+  MPI_Barrier(WorldShmComm);
+
+  _ShmAlloc=1;
+  _ShmAllocBytes  = bytes;
+}
+#endif
+ 
+////////////////////////////////////////////////////////////////////////////////////////////
+// Hugetlbfs mapping intended
+////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMMMAP
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hugetlbfs and others map filesystems as mappable huge pages
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  char shm_name [NAME_MAX];
+  for(int r=0;r<WorldShmSize;r++){
+    
+    sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",WorldNode,r);
+    int fd=open(shm_name,O_RDWR|O_CREAT,0666);
+    if ( fd == -1) { 
+      printf("open %s failed\n",shm_name);
+      perror("open hugetlbfs");
+      exit(0);
+    }
+    int mmap_flag = MAP_SHARED ;
+#ifdef MAP_POPULATE    
+    mmap_flag|=MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+    if ( flags ) mmap_flag |= MAP_HUGETLB;
+#endif
+    void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
+    if ( ptr == (void *)MAP_FAILED ) {    
+      printf("mmap %s failed\n",shm_name);
+      perror("failed mmap");      assert(0);    
+    }
+    assert(((uint64_t)ptr&0x3F)==0);
+    close(fd);
+    WorldShmCommBufs[r] =ptr;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+  }
+  _ShmAlloc=1;
+  _ShmAllocBytes  = bytes;
+};
+#endif // MMAP
+
+#ifdef GRID_MPI3_SHM_NONE
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hugetlbf and others map filesystems as mappable huge pages
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  char shm_name [NAME_MAX];
+  assert(WorldShmSize == 1);
+  for(int r=0;r<WorldShmSize;r++){
+    
+    int fd=-1;
+    int mmap_flag = MAP_SHARED |MAP_ANONYMOUS ;
+#ifdef MAP_POPULATE    
+    mmap_flag|=MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+    if ( flags ) mmap_flag |= MAP_HUGETLB;
+#endif
+    void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
+    if ( ptr == (void *)MAP_FAILED ) {    
+      printf("mmap %s failed\n",shm_name);
+      perror("failed mmap");      assert(0);    
+    }
+    assert(((uint64_t)ptr&0x3F)==0);
+    close(fd);
+    WorldShmCommBufs[r] =ptr;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+  }
+  _ShmAlloc=1;
+  _ShmAllocBytes  = bytes;
+};
+#endif // MMAP
+
+#ifdef GRID_MPI3_SHMOPEN
+////////////////////////////////////////////////////////////////////////////////////////////
+// POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
+// tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for 
+// the posix shm virtual file system
+////////////////////////////////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{ 
+  std::cout << "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0); 
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+
+  char shm_name [NAME_MAX];
+  if ( WorldShmRank == 0 ) {
+    for(int r=0;r<WorldShmSize;r++){
+	
+      size_t size = bytes;
+      
+      struct passwd *pw = getpwuid (getuid());
+      sprintf(shm_name,"/Grid_%s_mpi3_shm_%d_%d",pw->pw_name,WorldNode,r);
+      
+      shm_unlink(shm_name);
+      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
+      if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
+      ftruncate(fd, size);
+	
+      int mmap_flag = MAP_SHARED;
+#ifdef MAP_POPULATE 
+      mmap_flag |= MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+      if (flags) mmap_flag |= MAP_HUGETLB;
+#endif
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
+      
+      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
+      if ( ptr == (void * )MAP_FAILED ) {       
+	perror("failed mmap");     
+	assert(0);    
+      }
+      assert(((uint64_t)ptr&0x3F)==0);
+      
+      WorldShmCommBufs[r] =ptr;
+      close(fd);
+    }
+  }
+
+  MPI_Barrier(WorldShmComm);
+  
+  if ( WorldShmRank != 0 ) { 
+    for(int r=0;r<WorldShmSize;r++){
+
+      size_t size = bytes ;
+      
+      struct passwd *pw = getpwuid (getuid());
+      sprintf(shm_name,"/Grid_%s_mpi3_shm_%d_%d",pw->pw_name,WorldNode,r);
+      
+      int fd=shm_open(shm_name,O_RDWR,0666);
+      if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
+      
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      WorldShmCommBufs[r] =ptr;
+
+      close(fd);
+    }
+  }
+  _ShmAlloc=1;
+  _ShmAllocBytes = bytes;
+}
+#endif
+
+
+
+
+  ////////////////////////////////////////////////////////
+  // Global shared functionality finished
+  // Now move to per communicator functionality
+  ////////////////////////////////////////////////////////
+void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
+{
+  int rank, size;
+  MPI_Comm_rank(comm,&rank);
+  MPI_Comm_size(comm,&size);
+  ShmRanks.resize(size);
+
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory
+  /////////////////////////////////////////////////////////////////////
+  MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
+  MPI_Comm_rank(ShmComm     ,&ShmRank);
+  MPI_Comm_size(ShmComm     ,&ShmSize);
+  ShmCommBufs.resize(ShmSize);
+
+  //////////////////////////////////////////////////////////////////////
+  // Map ShmRank to WorldShmRank and use the right buffer
+  //////////////////////////////////////////////////////////////////////
+  assert (GlobalSharedMemory::ShmAlloc()==1);
+  heap_size = GlobalSharedMemory::ShmAllocBytes();
+  for(int r=0;r<ShmSize;r++){
+
+    uint32_t wsr = (r==ShmRank) ? GlobalSharedMemory::WorldShmRank : 0 ;
+
+    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();
+
+  /////////////////////////////////////////////////////////////////////
+  // find comm ranks in our SHM group (i.e. which ranks are on our node)
+  /////////////////////////////////////////////////////////////////////
+  MPI_Group FullGroup, ShmGroup;
+  MPI_Comm_group (comm   , &FullGroup); 
+  MPI_Comm_group (ShmComm, &ShmGroup);
+
+  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]); 
+}
+//////////////////////////////////////////////////////////////////
+// On node barrier
+//////////////////////////////////////////////////////////////////
+void SharedMemory::ShmBarrier(void)
+{
+  MPI_Barrier  (ShmComm);
+}
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Test the shared memory is working
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+void SharedMemory::SharedMemoryTest(void)
+{
+  ShmBarrier();
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+      uint64_t * check = (uint64_t *) ShmCommBufs[r];
+      check[0] = GlobalSharedMemory::WorldNode;
+      check[1] = r;
+      check[2] = 0x5A5A5A;
+    }
+  }
+  ShmBarrier();
+  for(int r=0;r<ShmSize;r++){
+    uint64_t * check = (uint64_t *) ShmCommBufs[r];
+    
+    assert(check[0]==GlobalSharedMemory::WorldNode);
+    assert(check[1]==r);
+    assert(check[2]==0x5A5A5A);
+    
+  }
+  ShmBarrier();
+}
+
+void *SharedMemory::ShmBuffer(int rank)
+{
+  int gpeer = ShmRanks[rank];
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    return ShmCommBufs[gpeer];
+  }
+}
+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){
+    return NULL;
+  } else { 
+    uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
+    uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
+    return (void *) remote;
+  }
+}
+SharedMemory::~SharedMemory()
+{
+  int MPI_is_finalised;  MPI_Finalized(&MPI_is_finalised);
+  if ( !MPI_is_finalised ) { 
+    MPI_Comm_free(&ShmComm);
+  }
+};
+
+}

Grid/communicator/SharedMemoryNone.cc

@@ -0,0 +1,128 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/GridCore.h>
+
+namespace Grid { 
+
+/*Construct from an MPI communicator*/
+void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
+{
+  assert(_ShmSetup==0);
+  WorldComm = 0;
+  WorldRank = 0;
+  WorldSize = 1;
+  WorldShmComm = 0 ;
+  WorldShmRank = 0 ;
+  WorldShmSize = 1 ;
+  WorldNodes   = 1 ;
+  WorldNode    = 0 ;
+  WorldShmRanks.resize(WorldSize); WorldShmRanks[0] = 0;
+  WorldShmCommBufs.resize(1);
+  _ShmSetup=1;
+}
+
+void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+  optimal_comm = WorldComm;
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////
+// Hugetlbfs mapping intended, use anonymous mmap
+////////////////////////////////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  void * ShmCommBuf ; 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+  int mmap_flag =0;
+#ifdef MAP_ANONYMOUS
+  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
+#endif
+#ifdef MAP_ANON
+  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
+#endif
+#ifdef MAP_HUGETLB
+  if ( flags ) mmap_flag |= MAP_HUGETLB;
+#endif
+  ShmCommBuf =(void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag, -1, 0); 
+  if (ShmCommBuf == (void *)MAP_FAILED) {
+    perror("mmap failed ");
+    exit(EXIT_FAILURE);  
+  }
+#ifdef MADV_HUGEPAGE
+  if (!Hugepages ) madvise(ShmCommBuf,bytes,MADV_HUGEPAGE);
+#endif
+  bzero(ShmCommBuf,bytes);
+  WorldShmCommBufs[0] = ShmCommBuf;
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+};
+
+  ////////////////////////////////////////////////////////
+  // Global shared functionality finished
+  // Now move to per communicator functionality
+  ////////////////////////////////////////////////////////
+void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
+{
+  assert(GlobalSharedMemory::ShmAlloc()==1);
+  ShmRanks.resize(1);
+  ShmCommBufs.resize(1);
+  ShmRanks[0] = 0;
+  ShmRank     = 0;
+  ShmSize     = 1;
+  //////////////////////////////////////////////////////////////////////
+  // Map ShmRank to WorldShmRank and use the right buffer
+  //////////////////////////////////////////////////////////////////////
+  ShmCommBufs[0] = GlobalSharedMemory::WorldShmCommBufs[0];
+  heap_size      = GlobalSharedMemory::ShmAllocBytes();
+  ShmBufferFreeAll();
+  return;
+}
+//////////////////////////////////////////////////////////////////
+// On node barrier
+//////////////////////////////////////////////////////////////////
+void SharedMemory::ShmBarrier(void){ return ; }
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Test the shared memory is working
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+void SharedMemory::SharedMemoryTest(void) { return; }
+
+void *SharedMemory::ShmBuffer(int rank)
+{
+  return NULL;
+}
+void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
+{
+  return NULL;
+}
+SharedMemory::~SharedMemory()
+{};
+
+}

Grid/cshift/Cshift.h

@@ -42,7 +42,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/cshift/Cshift_mpi.h>
 #endif 
 
-#ifdef GRID_COMMS_MPI3L
+#ifdef GRID_COMMS_MPIT
 #include <Grid/cshift/Cshift_mpi.h>
 #endif 
 

Grid/cshift/Cshift_common.h

@@ -45,31 +45,33 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
   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);
 
   int stride=rhs._grid->_slice_stride[dimension];
   if ( cbmask == 0x3 ) { 
-    parallel_for_nest2(int n=0;n<e1;n++){
+    for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o  = n*stride;
 	int bo = n*e2;
-	buffer[off+bo+b]=rhs._odata[so+o+b];
+	table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
       }
     }
   } else { 
      int bo=0;
-     std::vector<std::pair<int,int> > table;
      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);
 	 if ( ocb &cbmask ) {
-	   table.push_back(std::pair<int,int> (bo++,o+b));
+	   table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
 	 }
        }
      }
-     parallel_for(int i=0;i<table.size();i++){
-       buffer[off+table[i].first]=rhs._odata[so+table[i].second];
-     }
+  }
+  parallel_for(int i=0;i<ent;i++){
+    buffer[table[i].first]=rhs._odata[table[i].second];
   }
 }
 
@@ -140,31 +142,35 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
   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;
+
   if ( cbmask ==0x3 ) {
-    parallel_for_nest2(int n=0;n<e1;n++){
+
+    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];
-	rhs._odata[so+o+b]=buffer[bo+b];
+	table[ent++] = std::pair<int,int>(so+o+b,bo+b);
       }
     }
+
   } else { 
-    std::vector<std::pair<int,int> > table;
     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
 	if ( ocb & cbmask ) {
-	  table.push_back(std::pair<int,int> (so+o+b,bo++));
+	  table[ent++]=std::pair<int,int> (so+o+b,bo++);
 	}
       }
     }
-    parallel_for(int i=0;i<table.size();i++){
-       //       std::cout << "Rcv"<< table[i].first << " " << table[i].second << " " <<buffer[table[i].second]<<std::endl;
-       rhs._odata[table[i].first]=buffer[table[i].second];
-     }
+  }
+
+  parallel_for(int i=0;i<ent;i++){
+    rhs._odata[table[i].first]=buffer[table[i].second];
   }
 }
 
@@ -228,29 +234,32 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int e1=rhs._grid->_slice_nblock[dimension]; // clearly loop invariant for icpc
   int e2=rhs._grid->_slice_block[dimension];
   int stride = rhs._grid->_slice_stride[dimension];
+  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+  int ent=0;
+
   if(cbmask == 0x3 ){
-    parallel_for_nest2(int n=0;n<e1;n++){
+    for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
- 
         int o =n*stride+b;
-  	//lhs._odata[lo+o]=rhs._odata[ro+o];
-	vstream(lhs._odata[lo+o],rhs._odata[ro+o]);
+	table[ent++] = std::pair<int,int>(lo+o,ro+o);
       }
     }
   } else { 
-    parallel_for_nest2(int n=0;n<e1;n++){
+    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);
         if ( ocb&cbmask ) {
-  	//lhs._odata[lo+o]=rhs._odata[ro+o];
-	  vstream(lhs._odata[lo+o],rhs._odata[ro+o]);
+	  table[ent++] = std::pair<int,int>(lo+o,ro+o);
 	}
       }
     }
   }
-  
+
+  parallel_for(int i=0;i<ent;i++){
+    lhs._odata[table[i].first]=rhs._odata[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)
@@ -269,16 +278,28 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int e2=rhs._grid->_slice_block [dimension];
   int stride = rhs._grid->_slice_stride[dimension];
 
-  parallel_for_nest2(int n=0;n<e1;n++){
-  for(int b=0;b<e2;b++){
+  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++){
+      int o  =n*stride;
+      table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+    }}
+  } else {
+    for(int n=0;n<e1;n++){
+    for(int b=0;b<e2;b++){
       int o  =n*stride;
       int ocb=1<<lhs._grid->CheckerBoardFromOindex(o+b);
-      if ( ocb&cbmask ) {
-	permute(lhs._odata[lo+o+b],rhs._odata[ro+o+b],permute_type);
-      }
+      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);
+  }
 }
 
 //////////////////////////////////////////////////////
@@ -291,6 +312,8 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
   sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
   sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
 
+  double t_local;
+  
   if ( sshift[0] == sshift[1] ) {
     Cshift_local(ret,rhs,dimension,shift,0x3);
   } else {
@@ -299,7 +322,7 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
   }
 }
 
-template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
+template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
   GridBase *grid = rhs._grid;
   int fd = grid->_fdimensions[dimension];
@@ -325,11 +348,7 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
 
     int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
     int sx     = (x+sshift)%rd;
-
-    // FIXME : This must change where we have a 
-    // Rotate slice.
     
-    // Document how this works ; why didn't I do this when I first wrote it...
     // wrap is whether sshift > rd.
     //  num is sshift mod rd.
     // 
@@ -365,10 +384,8 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
 
     if ( permute_slice ) Copy_plane_permute(ret,rhs,dimension,x,sx,cbmask,permute_type_dist);
     else                 Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-
   
   }
-  return ret;
 }
 }
 #endif

Grid/cshift/Cshift_mpi.h

@@ -54,13 +54,13 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
 
 
   if ( !comm_dim ) {
-    //    std::cout << "Cshift_local" <<std::endl;
+    //std::cout << "CSHIFT: Cshift_local" <<std::endl;
     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
   } else if ( splice_dim ) {
-    //    std::cout << "Cshift_comms_simd" <<std::endl;
+    //std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift);
   } else {
-    //    std::cout << "Cshift_comms" <<std::endl;
+    //std::cout << "CSHIFT: Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift);
   }
   return ret;
@@ -91,9 +91,12 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
   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] ) {
+    //std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
   } else {
+    //std::cout << "Two pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
@@ -175,6 +178,10 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int simd_layout     = grid->_simd_layout[dimension];
   int comm_dim        = grid->_processors[dimension] >1 ;
 
+  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
+  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
+  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
+
   assert(comm_dim==1);
   assert(simd_layout==2);
   assert(shift>=0);

Grid/lattice/Lattice_arith.h

@@ -244,19 +244,11 @@ namespace Grid {
 
   template<class sobj,class vobj> strong_inline
   RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
-    conformable(ret,x);
-    conformable(x,y);
-    axpy(ret,a,x,y);
-    return norm2(ret);
+    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){
-    ret.checkerboard = x.checkerboard;
-    conformable(ret,x);
-    conformable(x,y);
-    axpby(ret,a,b,x,y);
-    return norm2(ret); // FIXME implement parallel norm in ss loop
+    return axpby_norm_fast(ret,a,b,x,y);
   }
 
 }

Grid/lattice/Lattice_base.h

@@ -85,7 +85,7 @@ class LatticeTrinaryExpression :public std::pair<Op,std::tuple<T1,T2,T3> >, publ
 
 void inline conformable(GridBase *lhs,GridBase *rhs)
 {
-  assert(lhs == rhs);
+  assert((lhs == rhs) && " conformable check pointers mismatch ");
 }
 
 template<class vobj>
@@ -256,9 +256,42 @@ public:
       _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;
     
   void reset(GridBase* grid) {
@@ -277,15 +310,6 @@ public:
     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;
-  }
   
   // *=,+=,-= operators inherit behvour from correspond */+/- operation
   template<class T> strong_inline Lattice<vobj> &operator *=(const T &r) {

Grid/lattice/Lattice_comparison_utils.h

@@ -179,7 +179,7 @@ namespace Grid {
       return ret;
     }
 
-#define DECLARE_RELATIONAL(op,functor) \
+#define DECLARE_RELATIONAL_EQ(op,functor) \
   template<class vsimd,IfSimd<vsimd> = 0>\
     inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
     {\
@@ -198,11 +198,6 @@ namespace Grid {
       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 iScalar<vsimd> &rhs)\
-    {									\
-      return lhs._internal op rhs._internal;				\
-    }									\
   template<class vsimd>\
     inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
     {									\
@@ -212,14 +207,21 @@ namespace Grid {
     inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
     {									\
       return lhs op rhs._internal;					\
-    }									
+    }									\
 
+#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)\
+    {									\
+      return lhs._internal op rhs._internal;				\
+    }									
 
 DECLARE_RELATIONAL(<,slt);
 DECLARE_RELATIONAL(<=,sle);
 DECLARE_RELATIONAL(>,sgt);
 DECLARE_RELATIONAL(>=,sge);
-DECLARE_RELATIONAL(==,seq);
+DECLARE_RELATIONAL_EQ(==,seq);
 DECLARE_RELATIONAL(!=,sne);
 
 #undef DECLARE_RELATIONAL

Grid/lattice/Lattice_coordinate.h

@@ -52,23 +52,5 @@ namespace Grid {
       }
     };
 
-    // LatticeCoordinate();
-    // FIXME for debug; deprecate this; made obscelete by 
-    template<class vobj> void lex_sites(Lattice<vobj> &l){
-      Real *v_ptr = (Real *)&l._odata[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

Grid/lattice/Lattice_reduction.h

@@ -1,4 +1,4 @@
- /*************************************************************************************
+/*************************************************************************************
     Grid physics library, www.github.com/paboyle/Grid 
     Source file: ./lib/lattice/Lattice_reduction.h
     Copyright (C) 2015
@@ -33,41 +33,94 @@ namespace Grid {
   // Deterministic Reduction operations
   ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
-  ComplexD nrm = innerProduct(arg,arg);
+  auto nrm = innerProduct(arg,arg);
   return std::real(nrm); 
 }
 
 // Double inner product
 template<class vobj>
-inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) 
+inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
 {
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
-  scalar_type  nrm;
-  
   GridBase *grid = left._grid;
-  
-  std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
-  
+  const int pad = 8;
+
+  ComplexD  inner;
+  Vector<ComplexD> sumarray(grid->SumArraySize()*pad);
+
   parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
     int nwork, mywork, myoff;
     GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
     
-    decltype(innerProductD(left._odata[0],right._odata[0])) vnrm=zero; // private to thread; sub summation
+    decltype(innerProductD(left._odata[0],right._odata[0])) vinner=zero; // private to thread; sub summation
     for(int ss=myoff;ss<mywork+myoff; ss++){
-      vnrm = vnrm + innerProductD(left._odata[ss],right._odata[ss]);
+      vinner = vinner + innerProductD(left._odata[ss],right._odata[ss]);
     }
-    sumarray[thr]=TensorRemove(vnrm) ;
+    // 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);
   }
   
-  vector_type vvnrm; vvnrm=zero;  // sum across threads
+  inner=0.0;
   for(int i=0;i<grid->SumArraySize();i++){
-    vvnrm = vvnrm+sumarray[i];
+    inner = inner+sumarray[i*pad];
   } 
-  nrm = Reduce(vvnrm);// sum across simd
-  right._grid->GlobalSum(nrm);
-  return nrm;
+  right._grid->GlobalSum(inner);
+  return inner;
 }
+
+/////////////////////////
+// Fast axpby_norm
+// z = a x + b y
+// return norm z
+/////////////////////////
+template<class sobj,class vobj> strong_inline RealD 
+axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y) 
+{
+  sobj one(1.0);
+  return axpby_norm_fast(z,a,one,x,y);
+}
+
+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;
+  conformable(z,x);
+  conformable(x,y);
+
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_typeD vector_type;
+  RealD  nrm;
+  
+  GridBase *grid = x._grid;
+  
+  Vector<RealD> sumarray(grid->SumArraySize()*pad);
+  
+  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
+    int nwork, mywork, myoff;
+    GridThread::GetWork(x._grid->oSites(),thr,mywork,myoff);
+    
+    // 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)))) ;
+  }
+  
+  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);
+  return nrm; 
+}
+
  
 template<class Op,class T1>
 inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
@@ -221,6 +274,115 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   }
 }
 
+template<class vobj>
+static void mySliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)
+{
+  // std::cout << GridLogMessage << "Start mySliceInnerProductVector" << std::endl;
+
+  typedef typename vobj::scalar_type scalar_type;
+  std::vector<scalar_type> lsSum;
+  localSliceInnerProductVector(result, lhs, rhs, lsSum, orthogdim);
+  globalSliceInnerProductVector(result, lhs, lsSum, orthogdim);
+  // std::cout << GridLogMessage << "End mySliceInnerProductVector" << std::endl;
+}
+
+template <class vobj>
+static void localSliceInnerProductVector(std::vector<ComplexD> &result, const Lattice<vobj> &lhs, const Lattice<vobj> &rhs, std::vector<typename vobj::scalar_type> &lsSum, int orthogdim)
+{
+  // 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;
+  assert(grid!=NULL);
+  conformable(grid,rhs._grid);
+
+  const int    Nd = grid->_ndimension;
+  const int Nsimd = grid->Nsimd();
+
+  assert(orthogdim >= 0);
+  assert(orthogdim < Nd);
+
+  int fd=grid->_fdimensions[orthogdim];
+  int ld=grid->_ldimensions[orthogdim];
+  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
+  lsSum.resize(ld,scalar_type(0.0));                    // sum across these down to scalars
+  std::vector<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;
+  }
+
+  int e1=    grid->_slice_nblock[orthogdim];
+  int e2=    grid->_slice_block [orthogdim];
+  int stride=grid->_slice_stride[orthogdim];
+  // 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++)
+  {
+
+    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]));
+        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);
+  for(int rt=0;rt<rd;rt++){
+
+    iScalar<vector_type> temp; 
+    temp._internal = lvSum[rt];
+    extract(temp,extracted);
+
+    for(int idx=0;idx<Nsimd;idx++){
+
+      grid->iCoorFromIindex(icoor,idx);
+
+      int ldx =rt+icoor[orthogdim]*rd;
+
+      lsSum[ldx]=lsSum[ldx]+extracted[idx]._internal;
+
+    }
+  }
+  // std::cout << GridLogMessage << "End sum over simd lanes" << std::endl;
+}
+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;
+  int fd = result.size();
+  int ld = lsSum.size();
+  // sum over nodes.
+  std::vector<scalar_type> gsum;
+  gsum.resize(fd, scalar_type(0.0));
+  // std::cout << GridLogMessage << "Start of gsum[t] creation:" << std::endl;
+  for(int t=0;t<fd;t++){
+    int pt = t/ld; // processor plane
+    int lt = t%ld;
+    if ( pt == grid->_processor_coor[orthogdim] ) {
+      gsum[t]=lsSum[lt];
+    }
+  }
+  // std::cout << GridLogMessage << "End of gsum[t] creation:" << std::endl;
+  // std::cout << GridLogMessage << "Start of GlobalSumVector:" << std::endl;
+  grid->GlobalSumVector(&gsum[0], fd);
+  // std::cout << GridLogMessage << "End of GlobalSumVector:" << std::endl;
+
+  result = gsum;
+}
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {
@@ -328,6 +490,8 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   typedef typename vobj::vector_type vector_type;
   typedef typename vobj::tensor_reduced tensor_reduced;
   
+  scalar_type zscale(scale);
+
   GridBase *grid  = X._grid;
 
   int Nsimd  =grid->Nsimd();
@@ -353,7 +517,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
       grid->iCoorFromIindex(icoor,l);
       int ldx =r+icoor[orthogdim]*rd;
       scalar_type *as =(scalar_type *)&av;
-      as[l] = scalar_type(a[ldx])*scale;
+      as[l] = scalar_type(a[ldx])*zscale;
     }
 
     tensor_reduced at; at=av;
@@ -367,71 +531,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   }
 };
 
-
 /*
-template<class vobj>
-static void sliceMaddVectorSlow (Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
-			     int Orthog,RealD scale=1.0) 
-{    
-  // FIXME: Implementation is slow
-  // Best base the linear combination by constructing a 
-  // set of vectors of size grid->_rdimensions[Orthog].
-  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);
-  // If we based this on Cshift it would work for spread out
-  // but it would be even slower
-  for(int i=0;i<Nblock;i++){
-    ExtractSlice(Rslice,Y,i,Orthog);
-    ExtractSlice(Xslice,X,i,Orthog);
-    Rslice = Rslice + Xslice*(scale*a[i]);
-    InsertSlice(Rslice,R,i,Orthog);
-  }
-};
-template<class vobj>
-static void sliceInnerProductVectorSlow( std::vector<ComplexD> & vec, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
-  {
-    // FIXME: Implementation is slow
-    // Look at localInnerProduct implementation,
-    // and do inside a site loop with block strided iterators
-    typedef typename vobj::scalar_object sobj;
-    typedef typename vobj::scalar_type scalar_type;
-    typedef typename vobj::vector_type vector_type;
-    typedef typename vobj::tensor_reduced scalar;
-    typedef typename scalar::scalar_object  scomplex;
-  
-    int Nblock = lhs._grid->GlobalDimensions()[Orthog];
-    vec.resize(Nblock);
-    std::vector<scomplex> sip(Nblock);
-    Lattice<scalar> IP(lhs._grid); 
-    IP=localInnerProduct(lhs,rhs);
-    sliceSum(IP,sip,Orthog);
-  
-    for(int ss=0;ss<Nblock;ss++){
-      vec[ss] = TensorRemove(sip[ss]);
-    }
-  }
-*/
-
-//////////////////////////////////////////////////////////////////////////////////////////
-// FIXME: Implementation is slow
-// If we based this on Cshift it would work for spread out
-// but it would be even slower
-//
-// Repeated extract slice is inefficient
-//
-// Best base the linear combination by constructing a 
-// set of vectors of size grid->_rdimensions[Orthog].
-//////////////////////////////////////////////////////////////////////////////////////////
-
 inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
 {
   int NN    = BlockSolverGrid->_ndimension;
@@ -450,7 +550,7 @@ inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Or
   }
   return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
 }
-
+*/
 
 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) 
@@ -460,57 +560,169 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   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);
-  
-  for(int i=0;i<Nblock;i++){
-    ExtractSlice(Rslice,Y,i,Orthog);
-    for(int j=0;j<Nblock;j++){
-      ExtractSlice(Xslice,X,j,Orthog);
-      Rslice = Rslice + Xslice*(scale*aa(j,i));
-    }
-    InsertSlice(Rslice,R,i,Orthog);
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+
+  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];
+#pragma omp parallel 
+  {
+    std::vector<vobj> s_x(Nblock);
+
+#pragma omp for collapse(2)
+    for(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[o+i*ostride];
+      }
+
+      vobj dot;
+      for(int i=0;i<Nblock;i++){
+	dot = Y[o+i*ostride];
+	for(int j=0;j<Nblock;j++){
+	  dot = dot + s_x[j]*(scale*aa(j,i));
+	}
+	R[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;
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+  int nh =  FullGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  int nl=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];
+#pragma omp parallel 
+  {
+    std::vector<vobj> s_x(Nblock);
+
+#pragma omp for collapse(2)
+    for(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[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[o+i*ostride]=dot;
+      }
+    }}
+  }
+
+};
+
+
 template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
-  // FIXME: Implementation is slow
-  // Not sure of best solution.. think about it
   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);
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
   
   int Nblock = FullGrid->GlobalDimensions()[Orthog];
   
-  Lattice<vobj> Lslice(SliceGrid);
-  Lattice<vobj> Rslice(SliceGrid);
+  //  Lattice<vobj> Lslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
   
   mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  
-  for(int i=0;i<Nblock;i++){
-    ExtractSlice(Lslice,lhs,i,Orthog);
-    for(int j=0;j<Nblock;j++){
-      ExtractSlice(Rslice,rhs,j,Orthog);
-      mat(i,j) = innerProduct(Lslice,Rslice);
-    }
+
+  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;
+
+#pragma omp parallel 
+  {
+    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++){
+    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];
+      }
+
+      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);
+      }}
+    }}
+#pragma omp critical
+    {
+      mat += mat_thread;
+    }  
   }
-#undef FORCE_DIAG
-#ifdef FORCE_DIAG
+
   for(int i=0;i<Nblock;i++){
-    for(int j=0;j<Nblock;j++){
-      if ( i != j ) mat(i,j)=0.0;
-    }
-  }
-#endif
+  for(int j=0;j<Nblock;j++){
+    ComplexD sum = mat(i,j);
+    FullGrid->GlobalSum(sum);
+    mat(i,j)=sum;
+  }}
+
   return;
 }
 

Grid/lattice/Lattice_rng.h

@@ -77,9 +77,6 @@ namespace Grid {
 
   
 // merge of April 11 2017
-//<<<<<<< HEAD
-
-
   // this function is necessary for the LS vectorised field
   inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
   {
@@ -91,7 +88,6 @@ namespace Grid {
     // all further divisions are local
     for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
     for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
-    
 
     // then divide the number of local sites
     // check that the total number of sims agree, meanse the iSites are the same
@@ -102,27 +98,6 @@ namespace Grid {
 
     return fine->lSites() / coarse->lSites();
   }
-
-  /*
-  // Wrap seed_seq to give common interface with random_device
-  class fixedSeed {
-  public:
-    typedef std::seed_seq::result_type result_type;
-    std::seed_seq src;
-    
-    fixedSeed(const std::vector<int> &seeds) : src(seeds.begin(),seeds.end()) {};
-
-    result_type operator () (void){
-      std::vector<result_type> list(1);
-      src.generate(list.begin(),list.end());
-      return list[0];
-    }
-
-  };
-
-=======
->>>>>>> develop
-  */
   
   // real scalars are one component
   template<class scalar,class distribution,class generator> 
@@ -171,7 +146,7 @@ namespace Grid {
     // support for parallel init
     ///////////////////////
 #ifdef RNG_FAST_DISCARD
-    static void Skip(RngEngine &eng)
+    static void Skip(RngEngine &eng,uint64_t site)
     {
       /////////////////////////////////////////////////////////////////////////////////////
       // Skip by 2^40 elements between successive lattice sites
@@ -183,9 +158,21 @@ namespace Grid {
       // tens of seconds per trajectory so this is clean in all reasonable cases,
       // and margin of safety is orders of magnitude.
       // We could hack Sitmo to skip in the higher order words of state if necessary
+      //
+      // Replace with 2^30 ; avoid problem on large volumes
+      //
       /////////////////////////////////////////////////////////////////////////////////////
-      uint64_t skip = 0x1; skip = skip<<40;
+      //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
+      const int shift = 30;
+
+      uint64_t skip = site;
+
+      skip = skip<<shift;
+
+      assert((skip >> shift)==site); // check for overflow
+
       eng.discard(skip);
+      //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
     } 
 #endif
     static RngEngine Reseed(RngEngine &eng)
@@ -264,7 +251,7 @@ namespace Grid {
 
       dist[0].reset();
       for(int idx=0;idx<words;idx++){
-  fillScalar(buf[idx],dist[0],_generators[0]);
+	fillScalar(buf[idx],dist[0],_generators[0]);
       }
 
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
@@ -296,7 +283,7 @@ namespace Grid {
       RealF *pointer=(RealF *)&l;
       dist[0].reset();
       for(int i=0;i<2*vComplexF::Nsimd();i++){
-  fillScalar(pointer[i],dist[0],_generators[0]);
+	fillScalar(pointer[i],dist[0],_generators[0]);
       }
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
@@ -304,7 +291,7 @@ namespace Grid {
       RealD *pointer=(RealD *)&l;
       dist[0].reset();
       for(int i=0;i<2*vComplexD::Nsimd();i++){
-  fillScalar(pointer[i],dist[0],_generators[0]);
+	fillScalar(pointer[i],dist[0],_generators[0]);
       }
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
@@ -312,7 +299,7 @@ namespace Grid {
       RealF *pointer=(RealF *)&l;
       dist[0].reset();
       for(int i=0;i<vRealF::Nsimd();i++){
-  fillScalar(pointer[i],dist[0],_generators[0]);
+	fillScalar(pointer[i],dist[0],_generators[0]);
       }
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
@@ -330,6 +317,19 @@ namespace Grid {
       std::seed_seq src(seeds.begin(),seeds.end());
       Seed(src,0);
     }
+
+    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 serial RNG with unique string '" 
+                << s << "'" << std::endl;
+      std::cout << GridLogMessage << "Seed SHA256: " << sha.str() << std::endl;
+      SeedFixedIntegers(seeds);
+    }
   };
 
   class GridParallelRNG : public GridRNGbase {
@@ -390,6 +390,14 @@ namespace Grid {
       _time_counter += usecond()- inner_time_counter;
     };
 
+    void SeedUniqueString(const std::string &s){
+      std::vector<int> seeds;
+      seeds = GridChecksum::sha256_seeds(s);
+      std::cout << GridLogMessage << "Intialising parallel RNG with unique string '" 
+                << s << "'" << std::endl;
+      std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
+      SeedFixedIntegers(seeds);
+    }
     void SeedFixedIntegers(const std::vector<int> &seeds){
 
       // Everyone generates the same seed_seq based on input seeds
@@ -407,15 +415,14 @@ namespace Grid {
       // MT implementation does not implement fast discard even though
       // in principle this is possible
       ////////////////////////////////////////////////
-      std::vector<int> gcoor;
-      int rank,o_idx,i_idx;
 
       // Everybody loops over global volume.
-      for(int gidx=0;gidx<_grid->_gsites;gidx++){
-
-	Skip(master_engine); // Skip to next RNG sequence
+      parallel_for(int gidx=0;gidx<_grid->_gsites;gidx++){
 
 	// Where is it?
+	int rank,o_idx,i_idx;
+	std::vector<int> gcoor;
+
 	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
 
@@ -423,6 +430,7 @@ namespace Grid {
 	if( rank == _grid->ThisRank() ){
 	  int l_idx=generator_idx(o_idx,i_idx);
 	  _generators[l_idx] = master_engine;
+	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
 
       }

Grid/lattice/Lattice_transfer.h

@@ -50,26 +50,22 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
   ////////////////////////////////////////////////////////////////////////////////////////////
   template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
     half.checkerboard = cb;
-    int ssh=0;
-    //parallel_for
-    for(int ss=0;ss<full._grid->oSites();ss++){
-      std::vector<int> coor;
+
+    parallel_for(int ss=0;ss<full._grid->oSites();ss++){
       int cbos;
-      
+      std::vector<int> 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];
-	ssh++;
       }
     }
   }
   template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
     int cb = half.checkerboard;
-    int ssh=0;
-    //parallel_for
-    for(int ss=0;ss<full._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<full._grid->oSites();ss++){
       std::vector<int> coor;
       int cbos;
 
@@ -77,8 +73,8 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
       cbos=half._grid->CheckerBoard(coor);
       
       if (cbos==cb) {
+	int ssh=half._grid->oIndex(coor);
 	full._odata[ss]=half._odata[ssh];
-	ssh++;
       }
     }
   }
@@ -109,8 +105,8 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
 
   coarseData=zero;
 
-  // Loop with a cache friendly loop ordering
-  for(int sf=0;sf<fine->oSites();sf++){
+  // Loop over coars parallel, and then loop over fine associated with coarse.
+  parallel_for(int sf=0;sf<fine->oSites();sf++){
 
     int sc;
     std::vector<int> coor_c(_ndimension);
@@ -119,8 +115,9 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     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
     for(int i=0;i<nbasis;i++) {
-      
+
       coarseData._odata[sc](i)=coarseData._odata[sc](i)
 	+ innerProduct(Basis[i]._odata[sf],fineData._odata[sf]);
 
@@ -139,6 +136,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
   GridBase * coarse= coarseA._grid;
 
   fineZ.checkerboard=fineX.checkerboard;
+  assert(fineX.checkerboard==fineY.checkerboard);
   subdivides(coarse,fine); // require they map
   conformable(fineX,fineY);
   conformable(fineX,fineZ);
@@ -180,9 +178,10 @@ template<class vobj,class CComplex>
   GridBase *coarse(CoarseInner._grid);
   GridBase *fine  (fineX._grid);
 
-  Lattice<dotp> fine_inner(fine);
+  Lattice<dotp> fine_inner(fine); fine_inner.checkerboard = fineX.checkerboard;
   Lattice<dotp> coarse_inner(coarse);
 
+  // Precision promotion?
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
   parallel_for(int ss=0;ss<coarse->oSites();ss++){
@@ -193,7 +192,7 @@ 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;
+  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);
@@ -216,19 +215,25 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
 
+  // Turn this around to loop threaded over sc and interior loop 
+  // over sf would thread better
   coarseData=zero;
-  for(int sf=0;sf<fine->oSites();sf++){
-    
+  parallel_region {
+
     int sc;
     std::vector<int> coor_c(_ndimension);
     std::vector<int> 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);
-
-    coarseData._odata[sc]=coarseData._odata[sc]+fineData._odata[sf];
+    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
+    
+      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];
 
+    }
   }
   return;
 }
@@ -238,7 +243,7 @@ inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vob
 {
   GridBase * fine = unpicked._grid;
 
-  Lattice<vobj> zz(fine);
+  Lattice<vobj> zz(fine); zz.checkerboard = unpicked.checkerboard;
   Lattice<iScalar<vInteger> > fcoor(fine);
 
   zz = zero;
@@ -303,20 +308,21 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   }
 
   // Loop with a cache friendly loop ordering
-  for(int sf=0;sf<fine->oSites();sf++){
-
+  parallel_region {
     int sc;
     std::vector<int> coor_c(_ndimension);
     std::vector<int> 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];
+    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
 
+      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];
+      }
     }
   }
   return;
@@ -458,8 +464,10 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[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
@@ -479,7 +487,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
 
 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;
 
@@ -493,8 +501,10 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slic
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[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
@@ -551,7 +561,10 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
 //Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
 template<typename vobj, typename sobj>
-typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in){
+typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type 
+unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
+{
+
   typedef typename vobj::vector_type vtype;
   
   GridBase* in_grid = in._grid;
@@ -591,10 +604,152 @@ typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>
   }
 }
 
+template<typename vobj, typename sobj>
+typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type 
+unvectorizeToRevLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
+{
+
+  typedef typename vobj::vector_type vtype;
+  
+  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);
+      
+  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
+    //Assemble vector of pointers to output elements
+    std::vector<sobj*> out_ptrs(in_nsimd);
+
+    std::vector<int> in_ocoor(ndim);
+    in_grid->oCoorFromOindex(in_ocoor, in_oidx);
+
+    std::vector<int> lcoor(in_grid->Nd());
+      
+    for(int lane=0; lane < in_nsimd; lane++){
+      for(int mu=0;mu<ndim;mu++)
+	lcoor[mu] = in_ocoor[mu] + in_grid->_rdimensions[mu]*in_icoor[lane][mu];
+
+      int lex;
+      Lexicographic::IndexFromCoorReversed(lcoor, lex, in_grid->_ldimensions);
+      out_ptrs[lane] = &out[lex];
+    }
+    
+    //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
+template<typename vobj, typename sobj>
+typename std::enable_if<isSIMDvectorized<vobj>::value 
+                    && !isSIMDvectorized<sobj>::value, void>::type 
+vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
+{
+
+  typedef typename vobj::vector_type vtype;
+  
+  GridBase* grid = out._grid;
+  assert(in.size()==grid->lSites());
+  
+  int ndim     = grid->Nd();
+  int nsimd    = vtype::Nsimd();
+
+  std::vector<std::vector<int> > 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
+    //Assemble vector of pointers to output elements
+    std::vector<sobj*> ptrs(nsimd);
+
+    std::vector<int> ocoor(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++){
+	lcoor[mu] = ocoor[mu] + grid->_rdimensions[mu]*icoor[lane][mu];
+      }
+
+      int lex;
+      Lexicographic::IndexFromCoor(lcoor, lex, grid->_ldimensions);
+      ptrs[lane] = &in[lex];
+    }
+    
+    //pack from those ptrs
+    vobj vecobj;
+    merge1(vecobj, ptrs, 0);
+    out._odata[oidx] = vecobj; 
+  }
+}
+
+template<typename vobj, typename sobj>
+typename std::enable_if<isSIMDvectorized<vobj>::value 
+                    && !isSIMDvectorized<sobj>::value, void>::type 
+vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
+{
+
+  typedef typename vobj::vector_type vtype;
+  
+  GridBase* grid = out._grid;
+  assert(in.size()==grid->lSites());
+  
+  int ndim     = grid->Nd();
+  int nsimd    = vtype::Nsimd();
+
+  std::vector<std::vector<int> > 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
+    //Assemble vector of pointers to output elements
+    std::vector<sobj*> ptrs(nsimd);
+
+    std::vector<int> ocoor(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++){
+	lcoor[mu] = ocoor[mu] + grid->_rdimensions[mu]*icoor[lane][mu];
+      }
+
+      int lex;
+      Lexicographic::IndexFromCoorReversed(lcoor, lex, grid->_ldimensions);
+      ptrs[lane] = &in[lex];
+    }
+    
+    //pack from those ptrs
+    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;
@@ -615,7 +770,7 @@ 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(int out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
+  parallel_for(uint64_t out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
     std::vector<int> out_ocoor(ndim);
     out_grid->oCoorFromOindex(out_ocoor, out_oidx);
 
@@ -633,6 +788,302 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
     merge(out._odata[out_oidx], ptrs, 0);
   }
 }
- 
+
+////////////////////////////////////////////////////////////////////////////////
+// Communicate between grids
+////////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+// SIMPLE CASE:
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+//
+// Mesh of nodes (2x2) ; subdivide to  1x1 subdivisions
+//
+// Lex ord:   
+//          N0 va0 vb0 vc0 vd0       N1 va1 vb1 vc1 vd1  
+//          N2 va2 vb2 vc2 vd2       N3 va3 vb3 vc3 vd3 
+//
+// Ratio = full[dim] / split[dim]
+//
+// For each dimension do an all to all; get Nvec -> Nvec / ratio
+//                                          Ldim -> Ldim * ratio
+//                                          LocalVol -> LocalVol * ratio
+// full AllToAll(0)
+//          N0 va0 vb0 va1 vb1       N1 vc0 vd0 vc1 vd1   
+//          N2 va2 vb2 va3 vb3       N3 vc2 vd2 vc3 vd3 
+//
+// REARRANGE
+//          N0 va01 vb01      N1 vc01 vd01
+//          N2 va23 vb23      N3 vc23 vd23
+//
+// full AllToAll(1)           // Not what is wanted. FIXME
+//          N0 va01 va23      N1 vc01 vc23 
+//          N2 vb01 vb23      N3 vd01 vd23
+// 
+// REARRANGE
+//          N0 va0123      N1 vc0123
+//          N2 vb0123      N3 vd0123
+//
+// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
+// NB: Easiest to programme if keep in lex order.
+/*
+ *  Let chunk = (fvol*nvec)/sP be size of a chunk.         ( Divide lexico vol * nvec into fP/sP = M chunks )
+ *  
+ *  2nd A2A (over sP nodes; subdivide the fP into sP chunks of M)
+ * 
+ *     node 0     1st chunk of node 0M..(1M-1); 2nd chunk of node 0M..(1M-1)..   data chunk x M x sP = fL / sP * M * sP = fL * M growth
+ *     node 1     1st chunk of node 1M..(2M-1); 2nd chunk of node 1M..(2M-1)..
+ *     node 2     1st chunk of node 2M..(3M-1); 2nd chunk of node 2M..(3M-1)..
+ *     node 3     1st chunk of node 3M..(3M-1); 2nd chunk of node 2M..(3M-1)..
+ *  etc...
+ */
+template<class Vobj>
+void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  typedef typename Vobj::scalar_object Sobj;
+
+  int full_vecs   = full.size();
+
+  assert(full_vecs>=1);
+
+  GridBase * full_grid = full[0]._grid;
+  GridBase *split_grid = split._grid;
+
+  int       ndim  = full_grid->_ndimension;
+  int  full_nproc = full_grid->_Nprocessors;
+  int split_nproc =split_grid->_Nprocessors;
+
+  ////////////////////////////////
+  // Checkerboard management
+  ////////////////////////////////
+  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);
+    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]);
+    }
+  }
+
+  int   nvector   =full_nproc/split_nproc; 
+  assert(nvector*split_nproc==full_nproc);
+  assert(nvector == full_vecs);
+
+  std::vector<int> ratio(ndim);
+  for(int d=0;d<ndim;d++){
+    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  }
+
+  uint64_t lsites = full_grid->lSites();
+  uint64_t     sz = lsites * nvector;
+  std::vector<Sobj> tmpdata(sz);
+  std::vector<Sobj> alldata(sz);
+  std::vector<Sobj> scalardata(lsites); 
+
+  for(int v=0;v<nvector;v++){
+    unvectorizeToLexOrdArray(scalardata,full[v]);    
+    parallel_for(int site=0;site<lsites;site++){
+      alldata[v*lsites+site] = scalardata[site];
+    }
+  }
+
+  int nvec = nvector; // Counts down to 1 as we collapse dims
+  std::vector<int> ldims = full_grid->_ldimensions;
+
+  for(int d=ndim-1;d>=0;d--){
+
+    if ( ratio[d] != 1 ) {
+
+      full_grid ->AllToAll(d,alldata,tmpdata);
+      if ( split_grid->_processors[d] > 1 ) {
+	alldata=tmpdata;
+	split_grid->AllToAll(d,alldata,tmpdata);
+      }
+
+      auto rdims = ldims; 
+      auto     M = ratio[d];
+      auto rsites= lsites*M;// increases rsites by M
+      nvec      /= M;       // Reduce nvec by subdivision factor
+      rdims[d]  *= M;       // increase local dim by same factor
+
+      int sP =   split_grid->_processors[d];
+      int fP =    full_grid->_processors[d];
+
+      int fvol   = lsites;
+      
+      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);
+	for(int m=0;m<M;m++){
+	  for(int s=0;s<sP;s++){
+	    
+	    // addressing; use lexico
+	    int lex_r;
+	    uint64_t lex_c        = c+chunk*m+chunk*M*s;
+	    uint64_t lex_fvol_vec = c+chunk*s;
+	    uint64_t lex_fvol     = lex_fvol_vec%fvol;
+	    uint64_t lex_vec      = lex_fvol_vec/fvol;
+
+	    // which node sets an adder to the coordinate
+	    Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);	  
+	    coor[d] += m*ldims[d];
+	    Lexicographic::IndexFromCoor(coor, lex_r, rdims);	  
+	    lex_r += lex_vec * rsites;
+
+	    // LexicoFind coordinate & vector number within split lattice
+	    alldata[lex_r] = tmpdata[lex_c];
+
+	  }
+	}
+      }
+      ldims[d]*= ratio[d];
+      lsites  *= ratio[d];
+
+    }
+  }
+  vectorizeFromLexOrdArray(alldata,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);
+  for(int n=0;n<nvector;n++){
+    full_v[n] = full;
+  }
+  Grid_split(full_v,split);
+}
+
+template<class Vobj>
+void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  typedef typename Vobj::scalar_object Sobj;
+
+  int full_vecs   = full.size();
+
+  assert(full_vecs>=1);
+
+  GridBase * full_grid = full[0]._grid;
+  GridBase *split_grid = split._grid;
+
+  int       ndim  = full_grid->_ndimension;
+  int  full_nproc = full_grid->_Nprocessors;
+  int split_nproc =split_grid->_Nprocessors;
+
+  ////////////////////////////////
+  // Checkerboard management
+  ////////////////////////////////
+  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);
+    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]);
+    }
+  }
+
+  int   nvector   =full_nproc/split_nproc; 
+  assert(nvector*split_nproc==full_nproc);
+  assert(nvector == full_vecs);
+
+  std::vector<int> ratio(ndim);
+  for(int d=0;d<ndim;d++){
+    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  }
+
+  uint64_t lsites = full_grid->lSites();
+  uint64_t     sz = lsites * nvector;
+  std::vector<Sobj> tmpdata(sz);
+  std::vector<Sobj> alldata(sz);
+  std::vector<Sobj> scalardata(lsites); 
+
+  unvectorizeToLexOrdArray(alldata,split);    
+
+  /////////////////////////////////////////////////////////////////
+  // Start from split grid and work towards full grid
+  /////////////////////////////////////////////////////////////////
+
+  int nvec = 1;
+  uint64_t rsites        = split_grid->lSites();
+  std::vector<int> rdims = split_grid->_ldimensions;
+
+  for(int d=0;d<ndim;d++){
+
+    if ( ratio[d] != 1 ) {
+
+      auto     M = ratio[d];
+
+      int sP =   split_grid->_processors[d];
+      int fP =    full_grid->_processors[d];
+      
+      auto ldims = rdims;  ldims[d]  /= M;  // Decrease local dims by same factor
+      auto lsites= rsites/M;                // Decreases rsites by M
+      
+      int fvol   = lsites;
+      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);
+	  for(int m=0;m<M;m++){
+	    for(int s=0;s<sP;s++){
+
+	      // addressing; use lexico
+	      int lex_r;
+	      uint64_t lex_c = c+chunk*m+chunk*M*s;
+	      uint64_t lex_fvol_vec = c+chunk*s;
+	      uint64_t lex_fvol     = lex_fvol_vec%fvol;
+	      uint64_t lex_vec      = lex_fvol_vec/fvol;
+	      
+	      // which node sets an adder to the coordinate
+	      Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);	  
+	      coor[d] += m*ldims[d];
+	      Lexicographic::IndexFromCoor(coor, lex_r, rdims);	  
+	      lex_r += lex_vec * rsites;
+	      
+	      // LexicoFind coordinate & vector number within split lattice
+	      tmpdata[lex_c] = alldata[lex_r];
+	    }
+	  }
+	}
+      }
+
+      if ( split_grid->_processors[d] > 1 ) {
+	split_grid->AllToAll(d,tmpdata,alldata);
+	tmpdata=alldata;
+      }
+      full_grid ->AllToAll(d,tmpdata,alldata);
+      rdims[d]/= M;
+      rsites  /= M;
+      nvec    *= M;       // Increase nvec by subdivision factor
+    }
+  }
+
+  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());
+      scalardata[site] = alldata[v*lsites+site];
+    }
+    vectorizeFromLexOrdArray(scalardata,full[v]);    
+  }
+}
+
 }
 #endif

Grid/lattice/Lattice_unary.h

@@ -62,14 +62,20 @@ namespace Grid {
     return ret;
   }
 
-  template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs, ComplexD alpha, Integer Nexp = DEFAULT_MAT_EXP){
+  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);
     }
+
     return ret;
+
+    
+    
+
+    
   }
 
 

Grid/log/Log.cc

@@ -50,7 +50,7 @@ namespace Grid {
     return (status==0) ? res.get() : name ;
   }
   
-GridStopWatch Logger::StopWatch;
+GridStopWatch Logger::GlobalStopWatch;
 int Logger::timestamp;
 std::ostream Logger::devnull(0);
 
@@ -59,13 +59,16 @@ void GridLogTimestamp(int on){
 }
 
 Colours GridLogColours(0);
-GridLogger GridLogError(1, "Error", GridLogColours, "RED");
+GridLogger GridLogMG     (1, "MG"    , GridLogColours, "NORMAL");
+GridLogger GridLogIRL    (1, "IRL"   , GridLogColours, "NORMAL");
+GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
+GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
-GridLogger GridLogDebug(1, "Debug", GridLogColours, "PURPLE");
+GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
-GridLogger GridLogIterative(1, "Iterative", GridLogColours, "BLUE");
-GridLogger GridLogIntegrator(1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
+GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(0);
@@ -74,19 +77,18 @@ 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++) {
-    if (logstreams[i] == std::string("Error")) GridLogError.Active(1);
-    if (logstreams[i] == std::string("Warning")) GridLogWarning.Active(1);
-    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("Integrator")) GridLogIntegrator.Active(1);
-    if (logstreams[i] == std::string("Colours")) GridLogColours.Active(1);
+    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
+    if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
+    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("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }
 
@@ -95,7 +97,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
 ////////////////////////////////////////////////////////////
 void Grid_quiesce_nodes(void) {
   int me = 0;
-#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPI3L)
+#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
   MPI_Comm_rank(MPI_COMM_WORLD, &me);
 #endif
 #ifdef GRID_COMMS_SHMEM

Grid/log/Log.h

@@ -85,12 +85,16 @@ class Logger {
 protected:
   Colours &Painter;
   int active;
+  int timing_mode;
+  int topWidth{-1}, chanWidth{-1};
   static int timestamp;
   std::string name, topName;
   std::string COLOUR;
 
 public:
-  static GridStopWatch StopWatch;
+  static GridStopWatch GlobalStopWatch;
+  GridStopWatch         LocalStopWatch;
+  GridStopWatch *StopWatch;
   static std::ostream devnull;
 
   std::string background() {return Painter.colour["NORMAL"];}
@@ -101,22 +105,52 @@ public:
     name(nm),
     topName(topNm),
     Painter(col_class),
-    COLOUR(col) {} ;
+    timing_mode(0),
+    COLOUR(col) 
+    {
+      StopWatch = & GlobalStopWatch;
+    };
   
   void Active(int on) {active = on;};
   int  isActive(void) {return active;};
   static void Timestamp(int on) {timestamp = on;};
-  
+  void Reset(void) { 
+    StopWatch->Reset(); 
+    StopWatch->Start(); 
+  }
+  void TimingMode(int on) { 
+    timing_mode = on; 
+    if(on) { 
+      StopWatch = &LocalStopWatch;
+      Reset(); 
+    }
+  }
+  void setTopWidth(const int w) {topWidth = w;}
+  void setChanWidth(const int w) {chanWidth = w;}
+
   friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
     if ( log.active ) {
-      stream << log.background()<< std::setw(8) << std::left << log.topName << log.background()<< " : ";
-      stream << log.colour() << std::setw(10) << std::left << log.name << log.background() << " : ";
+      stream << log.background()<<  std::left;
+      if (log.topWidth > 0)
+      {
+        stream << std::setw(log.topWidth);
+      }
+      stream << log.topName << log.background()<< " : ";
+      stream << log.colour() <<  std::left;
+      if (log.chanWidth > 0)
+      {
+        stream << std::setw(log.chanWidth);
+      }
+      stream << log.name << log.background() << " : ";
       if ( log.timestamp ) {
-	StopWatch.Stop();
-	GridTime now = StopWatch.Elapsed();
-	StopWatch.Start();
-	stream << log.evidence()<< now << log.background() << " : " ;
+	log.StopWatch->Stop();
+	GridTime now = log.StopWatch->Elapsed();
+	
+	if ( log.timing_mode==1 ) log.StopWatch->Reset();
+	log.StopWatch->Start();
+	stream << log.evidence()
+	       << now	       << log.background() << " : " ;
       }
       stream << log.colour();
       return stream;
@@ -135,6 +169,9 @@ public:
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
 
+extern GridLogger GridLogMG;
+extern GridLogger GridLogIRL;
+extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;

Grid/parallelIO/BinaryIO.cc

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

Grid/parallelIO/BinaryIO.h

@@ -0,0 +1,760 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/parallelIO/BinaryIO.h
+
+    Copyright (C) 2015
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu<guido.cossu@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_BINARY_IO_H
+#define GRID_BINARY_IO_H
+
+#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT) 
+#define USE_MPI_IO
+#else
+#undef  USE_MPI_IO
+#endif
+
+#ifdef HAVE_ENDIAN_H
+#include <endian.h>
+#endif
+
+#include <arpa/inet.h>
+#include <algorithm>
+
+namespace Grid { 
+
+
+/////////////////////////////////////////////////////////////////////////////////
+// Byte reversal garbage
+/////////////////////////////////////////////////////////////////////////////////
+inline uint32_t byte_reverse32(uint32_t f) { 
+      f = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
+      return f;
+}
+inline uint64_t byte_reverse64(uint64_t f) { 
+  uint64_t g;
+  g = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
+  g = g << 32;
+  f = f >> 32;
+  g|= ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
+  return g;
+}
+
+#if BYTE_ORDER == BIG_ENDIAN 
+inline uint64_t Grid_ntohll(uint64_t A) { return A; }
+#else
+inline uint64_t Grid_ntohll(uint64_t A) { 
+  return byte_reverse64(A);
+}
+#endif
+
+// A little helper
+inline void removeWhitespace(std::string &key)
+{
+  key.erase(std::remove_if(key.begin(), key.end(), ::isspace),key.end());
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Static class holding the parallel IO code
+// Could just use a namespace
+///////////////////////////////////////////////////////////////////////////////////////////////////
+class BinaryIO {
+ public:
+  static int latticeWriteMaxRetry;
+
+  /////////////////////////////////////////////////////////////////////////////
+  // more byte manipulation helpers
+  /////////////////////////////////////////////////////////////////////////////
+
+  template<class vobj> static inline void Uint32Checksum(Lattice<vobj> &lat,uint32_t &nersc_csum)
+  {
+    typedef typename vobj::scalar_object sobj;
+
+    GridBase *grid = lat._grid;
+    uint64_t lsites = grid->lSites();
+
+    std::vector<sobj> scalardata(lsites); 
+    unvectorizeToLexOrdArray(scalardata,lat);    
+
+    NerscChecksum(grid,scalardata,nersc_csum);
+  }
+
+  template <class fobj>
+  static inline void NerscChecksum(GridBase *grid, std::vector<fobj> &fbuf, uint32_t &nersc_csum)
+  {
+    const uint64_t size32 = sizeof(fobj) / sizeof(uint32_t);
+
+    uint64_t lsites = grid->lSites();
+    if (fbuf.size() == 1)
+    {
+      lsites = 1;
+    }
+
+PARALLEL_REGION
+    {
+      uint32_t nersc_csum_thr = 0;
+
+PARALLEL_FOR_LOOP_INTERN
+      for (uint64_t local_site = 0; local_site < lsites; local_site++)
+      {
+        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
+      {
+        nersc_csum += nersc_csum_thr;
+      }
+    }
+  }
+
+  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();
+
+PARALLEL_REGION
+    { 
+      std::vector<int> 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++){
+
+	uint32_t * site_buf = (uint32_t *)&fbuf[local_site];
+
+	/* 
+	 * Scidac csum  is rather more heavyweight
+	 * FIXME -- 128^3 x 256 x 16 will overflow.
+	 */
+	
+	int global_site;
+
+	Lexicographic::CoorFromIndex(coor,local_site,local_vol);
+
+	for(int d=0;d<nd;d++) {
+	  coor[d] = coor[d]+local_start[d];
+	}
+
+	Lexicographic::IndexFromCoor(coor,global_site,global_vol);
+
+	uint32_t gsite29   = global_site%29;
+	uint32_t gsite31   = global_site%31;
+	
+	site_crc = crc32(0,(unsigned char *)site_buf,sizeof(fobj));
+	//	std::cout << "Site "<<local_site << " crc "<<std::hex<<site_crc<<std::dec<<std::endl;
+	//	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
+      {
+	scidac_csuma^= scidac_csuma_thr;
+	scidac_csumb^= scidac_csumb_thr;
+      }
+    }
+  }
+
+  // Network is big endian
+  static inline void htobe32_v(void *file_object,uint32_t bytes){ be32toh_v(file_object,bytes);} 
+  static inline void htobe64_v(void *file_object,uint32_t bytes){ be64toh_v(file_object,bytes);} 
+  static inline void htole32_v(void *file_object,uint32_t bytes){ le32toh_v(file_object,bytes);} 
+  static inline void htole64_v(void *file_object,uint32_t bytes){ le64toh_v(file_object,bytes);} 
+
+  static inline void be32toh_v(void *file_object,uint64_t bytes)
+  {
+    uint32_t * f = (uint32_t *)file_object;
+    uint64_t count = bytes/sizeof(uint32_t);
+    parallel_for(uint64_t i=0;i<count;i++){  
+      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;
+
+    uint64_t count = bytes/sizeof(uint32_t);
+    parallel_for(uint64_t i=0;i<count;i++){  
+      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
+  static inline void be64toh_v(void *file_object,uint64_t bytes)
+  {
+    uint64_t * f = (uint64_t *)file_object;
+    uint64_t count = bytes/sizeof(uint64_t);
+    parallel_for(uint64_t i=0;i<count;i++){  
+      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 count = bytes/sizeof(uint64_t);
+    parallel_for(uint64_t i=0;i<count;i++){  
+      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) ; 
+      g = g << 32;
+      f = f >> 32;
+      g|= ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
+      fp[i] = Grid_ntohll(g);
+    }
+  }
+  /////////////////////////////////////////////////////////////////////////////
+  // Real action:
+  // Read or Write distributed lexico array of ANY object to a specific location in file 
+  //////////////////////////////////////////////////////////////////////////////////////
+
+  static const int BINARYIO_MASTER_APPEND = 0x10;
+  static const int BINARYIO_UNORDERED     = 0x08;
+  static const int BINARYIO_LEXICOGRAPHIC = 0x04;
+  static const int BINARYIO_READ          = 0x02;
+  static const int BINARYIO_WRITE         = 0x01;
+
+  template<class word,class fobj>
+  static inline void IOobject(word w,
+			      GridBase *grid,
+			      std::vector<fobj> &iodata,
+			      std::string file,
+			      uint64_t& offset,
+			      const std::string &format, int control,
+			      uint32_t &nersc_csum,
+			      uint32_t &scidac_csuma,
+			      uint32_t &scidac_csumb)
+  {
+    grid->Barrier();
+    GridStopWatch timer; 
+    GridStopWatch bstimer;
+    
+    nersc_csum=0;
+    scidac_csuma=0;
+    scidac_csumb=0;
+
+    int ndim                 = grid->Dimensions();
+    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();
+
+    std::vector<int> lStart(ndim);
+    std::vector<int> gStart(ndim);
+
+    // Flatten the file
+    uint64_t lsites = grid->lSites();
+    if ( control & BINARYIO_MASTER_APPEND )  {
+      assert(iodata.size()==1);
+    } else {
+      assert(lsites==iodata.size());
+    }
+    for(int d=0;d<ndim;d++){
+      gStart[d] = lLattice[d]*pcoor[d];
+      lStart[d] = 0;
+    }
+
+#ifdef USE_MPI_IO
+    std::vector<int> distribs(ndim,MPI_DISTRIBUTE_BLOCK);
+    std::vector<int> dargs   (ndim,MPI_DISTRIBUTE_DFLT_DARG);
+    MPI_Datatype mpiObject;
+    MPI_Datatype fileArray;
+    MPI_Datatype localArray;
+    MPI_Datatype mpiword;
+    MPI_Offset disp = offset;
+    MPI_File fh ;
+    MPI_Status status;
+    int numword;
+
+    if ( sizeof( word ) == sizeof(float ) ) {
+      numword = sizeof(fobj)/sizeof(float);
+      mpiword = MPI_FLOAT;
+    } else {
+      numword = sizeof(fobj)/sizeof(double);
+      mpiword = MPI_DOUBLE;
+    }
+
+    //////////////////////////////////////////////////////////////////////////////
+    // Sobj in MPI phrasing
+    //////////////////////////////////////////////////////////////////////////////
+    int ierr;
+    ierr = MPI_Type_contiguous(numword,mpiword,&mpiObject);    assert(ierr==0);
+    ierr = MPI_Type_commit(&mpiObject);
+
+    //////////////////////////////////////////////////////////////////////////////
+    // File global array data type
+    //////////////////////////////////////////////////////////////////////////////
+    ierr=MPI_Type_create_subarray(ndim,&gLattice[0],&lLattice[0],&gStart[0],MPI_ORDER_FORTRAN, mpiObject,&fileArray);    assert(ierr==0);
+    ierr=MPI_Type_commit(&fileArray);    assert(ierr==0);
+
+    //////////////////////////////////////////////////////////////////////////////
+    // local lattice array
+    //////////////////////////////////////////////////////////////////////////////
+    ierr=MPI_Type_create_subarray(ndim,&lLattice[0],&lLattice[0],&lStart[0],MPI_ORDER_FORTRAN, mpiObject,&localArray);    assert(ierr==0);
+    ierr=MPI_Type_commit(&localArray);    assert(ierr==0);
+#endif
+
+    //////////////////////////////////////////////////////////////////////////////
+    // Byte order
+    //////////////////////////////////////////////////////////////////////////////
+    int ieee32big = (format == std::string("IEEE32BIG"));
+    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
+    //////////////////////////////////////////////////////////////////////////////
+    if ( control & BINARYIO_READ ) { 
+
+      timer.Start();
+
+      if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
+#ifdef USE_MPI_IO
+	std::cout<< GridLogMessage<<"IOobject: MPI read I/O "<< file<< std::endl;
+	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);    assert(ierr==0);
+	ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);    assert(ierr==0);
+	ierr=MPI_File_read_all(fh, &iodata[0], 1, localArray, &status);    assert(ierr==0);
+	MPI_File_close(&fh);
+	MPI_Type_free(&fileArray);
+	MPI_Type_free(&localArray);
+#else 
+	assert(0);
+#endif
+      } else {
+	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
+                  << 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)
+        {
+          fin.seekg(-sizeof(fobj), fin.end);
+        }
+        else
+        {
+          fin.seekg(offset + myrank * lsites * sizeof(fobj));
+        }
+        fin.read((char *)&iodata[0], iodata.size() * sizeof(fobj));
+        assert(fin.fail() == 0);
+        fin.close();
+      }
+      timer.Stop();
+
+      grid->Barrier();
+
+      bstimer.Start();
+      ScidacChecksum(grid,iodata,scidac_csuma,scidac_csumb);
+      if (ieee32big) be32toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      if (ieee32)    le32toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      if (ieee64big) be64toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      if (ieee64)    le64toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      NerscChecksum(grid,iodata,nersc_csum);
+      bstimer.Stop();
+    }
+    
+    if ( control & BINARYIO_WRITE ) { 
+
+      bstimer.Start();
+      NerscChecksum(grid,iodata,nersc_csum);
+      if (ieee32big) htobe32_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      if (ieee32)    htole32_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      if (ieee64big) htobe64_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      if (ieee64)    htole64_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
+      ScidacChecksum(grid,iodata,scidac_csuma,scidac_csumb);
+      bstimer.Stop();
+
+      grid->Barrier();
+
+      timer.Start();
+      if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
+#ifdef USE_MPI_IO
+        std::cout << GridLogMessage <<"IOobject: MPI write I/O " << file << std::endl;
+        ierr = MPI_File_open(grid->communicator, (char *)file.c_str(), MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh);
+	//        std::cout << GridLogMessage << "Checking for errors" << std::endl;
+        if (ierr != MPI_SUCCESS)
+        {
+          char error_string[BUFSIZ];
+          int length_of_error_string, error_class;
+
+          MPI_Error_class(ierr, &error_class);
+          MPI_Error_string(error_class, error_string, &length_of_error_string);
+          fprintf(stderr, "%3d: %s\n", myrank, error_string);
+          MPI_Error_string(ierr, error_string, &length_of_error_string);
+          fprintf(stderr, "%3d: %s\n", myrank, error_string);
+          MPI_Abort(MPI_COMM_WORLD, 1); //assert(ierr == 0);
+        }
+
+        std::cout << GridLogDebug << "MPI write I/O set view " << file << std::endl;
+        ierr = MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);
+        assert(ierr == 0);
+
+        std::cout << GridLogDebug << "MPI write I/O write all " << file << std::endl;
+        ierr = MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);
+        assert(ierr == 0);
+
+        MPI_Offset os;
+        MPI_File_get_position(fh, &os);
+        MPI_File_get_byte_offset(fh, os, &disp);
+        offset = disp;
+
+
+        MPI_File_close(&fh);
+        MPI_Type_free(&fileArray);
+        MPI_Type_free(&localArray);
+#else 
+	assert(0);
+#endif
+      } else { 
+
+        std::cout << GridLogMessage << "IOobject: C++ write I/O " << file << " : "
+                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
+        
+	std::ofstream fout; 
+	fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
+	try {
+	  if (offset) { // Must already exist and contain data
+	    fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
+	  } else {     // Allow create
+	    fout.open(file,std::ios::binary|std::ios::out);
+	  }
+	} catch (const std::fstream::failure& exc) {
+	  std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
+	  std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
+	  //	  std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
+#ifdef USE_MPI_IO
+	  MPI_Abort(MPI_COMM_WORLD,1);
+#else
+	  exit(1);
+#endif
+	}
+	
+	if ( control & BINARYIO_MASTER_APPEND )  {
+	  try {
+	    fout.seekp(0,fout.end);
+	  } catch (const std::fstream::failure& exc) {
+	    std::cout << "Exception in seeking file end " << file << std::endl;
+	  }
+	} else {
+	  try { 
+	    fout.seekp(offset+myrank*lsites*sizeof(fobj));
+	  } catch (const std::fstream::failure& exc) {
+	    std::cout << "Exception in seeking file " << file <<" offset "<< offset << std::endl;
+	  }
+	}
+
+	try {
+	  fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));//assert( fout.fail()==0);
+	}
+	catch (const std::fstream::failure& exc) {
+	  std::cout << "Exception in writing file " << file << std::endl;
+	  std::cout << GridLogError << "Exception description: "<< exc.what() << std::endl;
+#ifdef USE_MPI_IO
+	  MPI_Abort(MPI_COMM_WORLD,1);
+#else
+	  exit(1);
+#endif
+	}
+  offset  = fout.tellp();
+	fout.close();
+      }
+      timer.Stop();
+    }
+    
+    std::cout<<GridLogMessage<<"IOobject: ";
+    if ( control & BINARYIO_READ) std::cout << " read  ";
+    else                          std::cout << " write ";
+    uint64_t bytes = sizeof(fobj)*iodata.size()*nrank;
+    std::cout<< bytes <<" bytes in "<<timer.Elapsed() <<" "
+	     << (double)bytes/ (double)timer.useconds() <<" MB/s "<<std::endl;
+
+    std::cout<<GridLogMessage<<"IOobject: endian and checksum overhead "<<bstimer.Elapsed()  <<std::endl;
+
+    //////////////////////////////////////////////////////////////////////////////
+    // Safety check
+    //////////////////////////////////////////////////////////////////////////////
+    // if the data size is 1 we do not want to sum over the MPI ranks
+    if (iodata.size() != 1){
+      grid->Barrier();
+      grid->GlobalSum(nersc_csum);
+      grid->GlobalXOR(scidac_csuma);
+      grid->GlobalXOR(scidac_csumb);
+      grid->Barrier();
+    }
+  }
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Read a Lattice of object
+  //////////////////////////////////////////////////////////////////////////////////////
+  template<class vobj,class fobj,class munger>
+  static inline void readLatticeObject(Lattice<vobj> &Umu,
+				       std::string file,
+				       munger munge,
+				       uint64_t offset,
+				       const std::string &format,
+				       uint32_t &nersc_csum,
+				       uint32_t &scidac_csuma,
+				       uint32_t &scidac_csumb)
+  {
+    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();
+
+    std::vector<sobj> scalardata(lsites); 
+    std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
+    
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+
+    GridStopWatch timer; 
+    timer.Start();
+
+    parallel_for(uint64_t x=0;x<lsites;x++) munge(iodata[x], scalardata[x]);
+
+    vectorizeFromLexOrdArray(scalardata,Umu);    
+    grid->Barrier();
+
+    timer.Stop();
+    std::cout<<GridLogMessage<<"readLatticeObject: vectorize overhead "<<timer.Elapsed()  <<std::endl;
+  }
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Write a Lattice of object
+  //////////////////////////////////////////////////////////////////////////////////////
+  template<class vobj,class fobj,class munger>
+    static inline void writeLatticeObject(Lattice<vobj> &Umu,
+					  std::string file,
+					  munger munge,
+					  uint64_t offset,
+					  const std::string &format,
+					  uint32_t &nersc_csum,
+					  uint32_t &scidac_csuma,
+					  uint32_t &scidac_csumb)
+  {
+    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(), 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
+
+    //////////////////////////////////////////////////////////////////////////////
+    // Munge [ .e.g 3rd row recon ]
+    //////////////////////////////////////////////////////////////////////////////
+    GridStopWatch timer; timer.Start();
+    unvectorizeToLexOrdArray(scalardata,Umu);    
+
+    parallel_for(uint64_t x=0;x<lsites;x++) 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;
+          parallel_for(uint64_t x=0;x<lsites;x++) 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;
+  }
+  
+  /////////////////////////////////////////////////////////////////////////////
+  // Read a RNG;  use IOobject and lexico map to an array of state 
+  //////////////////////////////////////////////////////////////////////////////////////
+  static inline void readRNG(GridSerialRNG &serial,
+			     GridParallelRNG &parallel,
+			     std::string file,
+			     uint64_t offset,
+			     uint32_t &nersc_csum,
+			     uint32_t &scidac_csuma,
+			     uint32_t &scidac_csumb)
+  {
+    typedef typename GridSerialRNG::RngStateType RngStateType;
+    const int RngStateCount = GridSerialRNG::RngStateCount;
+    typedef std::array<RngStateType,RngStateCount> RNGstate;
+    typedef RngStateType word;    word w=0;
+
+    std::string format = "IEEE32BIG";
+
+    GridBase *grid = parallel._grid;
+    uint64_t gsites = grid->gSites();
+    uint64_t lsites = grid->lSites();
+
+    uint32_t nersc_csum_tmp   = 0;
+    uint32_t scidac_csuma_tmp = 0;
+    uint32_t scidac_csumb_tmp = 0;
+
+    GridStopWatch timer;
+
+    std::cout << GridLogMessage << "RNG read I/O on file " << file << std::endl;
+
+    std::vector<RNGstate> iodata(lsites);
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+
+    timer.Start();
+    parallel_for(uint64_t lidx=0;lidx<lsites;lidx++){
+      std::vector<RngStateType> tmp(RngStateCount);
+      std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
+      parallel.SetState(tmp,lidx);
+    }
+    timer.Stop();
+
+    iodata.resize(1);
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_MASTER_APPEND,
+	     nersc_csum_tmp,scidac_csuma_tmp,scidac_csumb_tmp);
+
+    {
+      std::vector<RngStateType> tmp(RngStateCount);
+      std::copy(iodata[0].begin(),iodata[0].end(),tmp.begin());
+      serial.SetState(tmp,0);
+    }
+
+    nersc_csum   = nersc_csum   + nersc_csum_tmp;
+    scidac_csuma = scidac_csuma ^ scidac_csuma_tmp;
+    scidac_csumb = scidac_csumb ^ scidac_csumb_tmp;
+
+    std::cout << GridLogMessage << "RNG file nersc_checksum   " << std::hex << nersc_csum << std::dec << std::endl;
+    std::cout << GridLogMessage << "RNG file scidac_checksuma " << std::hex << scidac_csuma << std::dec << std::endl;
+    std::cout << GridLogMessage << "RNG file scidac_checksumb " << std::hex << scidac_csumb << std::dec << std::endl;
+
+    std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
+  }
+  /////////////////////////////////////////////////////////////////////////////
+  // Write a RNG; lexico map to an array of state and use IOobject
+  //////////////////////////////////////////////////////////////////////////////////////
+  static inline void writeRNG(GridSerialRNG &serial,
+			      GridParallelRNG &parallel,
+			      std::string file,
+			      uint64_t offset,
+			      uint32_t &nersc_csum,
+			      uint32_t &scidac_csuma,
+			      uint32_t &scidac_csumb)
+  {
+    typedef typename GridSerialRNG::RngStateType RngStateType;
+    typedef RngStateType word; word w=0;
+    const int RngStateCount = GridSerialRNG::RngStateCount;
+    typedef std::array<RngStateType,RngStateCount> RNGstate;
+
+    GridBase *grid = parallel._grid;
+    uint64_t gsites = grid->gSites();
+    uint64_t lsites = grid->lSites();
+
+    uint32_t nersc_csum_tmp;
+    uint32_t scidac_csuma_tmp;
+    uint32_t scidac_csumb_tmp;
+
+    GridStopWatch timer;
+    std::string format = "IEEE32BIG";
+
+    std::cout << GridLogMessage << "RNG write I/O on file " << file << std::endl;
+
+    timer.Start();
+    std::vector<RNGstate> iodata(lsites);
+    parallel_for(uint64_t lidx=0;lidx<lsites;lidx++){
+      std::vector<RngStateType> tmp(RngStateCount);
+      parallel.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,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+    iodata.resize(1);
+    {
+      std::vector<RngStateType> tmp(RngStateCount);
+      serial.GetState(tmp,0);
+      std::copy(tmp.begin(),tmp.end(),iodata[0].begin());
+    }
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_MASTER_APPEND,
+	     nersc_csum_tmp,scidac_csuma_tmp,scidac_csumb_tmp);
+
+    nersc_csum   = nersc_csum   + nersc_csum_tmp;
+    scidac_csuma = scidac_csuma ^ scidac_csuma_tmp;
+    scidac_csumb = scidac_csumb ^ scidac_csumb_tmp;
+    
+    std::cout << GridLogMessage << "RNG file checksum " << std::hex << nersc_csum    << std::dec << std::endl;
+    std::cout << GridLogMessage << "RNG file checksuma " << std::hex << scidac_csuma << std::dec << std::endl;
+    std::cout << GridLogMessage << "RNG file checksumb " << std::hex << scidac_csumb << std::dec << std::endl;
+    std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
+  }
+};
+
+}
+#endif

Grid/parallelIO/IldgIO.h

@@ -0,0 +1,938 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/parallelIO/IldgIO.h
+
+Copyright (C) 2015
+
+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_ILDG_IO_H
+#define GRID_ILDG_IO_H
+
+#ifdef HAVE_LIME
+#include <algorithm>
+#include <fstream>
+#include <iomanip>
+#include <iostream>
+#include <map>
+
+#include <pwd.h>
+#include <sys/utsname.h>
+#include <unistd.h>
+
+//C-Lime is a must have for this functionality
+extern "C" {  
+#include "lime.h"
+}
+
+namespace Grid {
+namespace QCD {
+
+#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
+  /////////////////////////////////
+ template<class word> inline std::string ScidacWordMnemonic(void){ return std::string("unknown"); }
+ template<> inline std::string ScidacWordMnemonic<double>  (void){ return std::string("D"); }
+ template<> inline std::string ScidacWordMnemonic<float>   (void){ return std::string("F"); }
+ template<> inline std::string ScidacWordMnemonic< int32_t>(void){ return std::string("I32_t"); }
+ template<> inline std::string ScidacWordMnemonic<uint32_t>(void){ return std::string("U32_t"); }
+ template<> inline std::string ScidacWordMnemonic< int64_t>(void){ return std::string("I64_t"); }
+ template<> inline std::string ScidacWordMnemonic<uint64_t>(void){ return std::string("U64_t"); }
+
+  /////////////////////////////////////////
+  // Encode a generic tensor as a string
+  /////////////////////////////////////////
+ template<class vobj> std::string ScidacRecordTypeString(int &colors, int &spins, int & typesize,int &datacount) { 
+
+   typedef typename getPrecision<vobj>::real_scalar_type stype;
+
+   int _ColourN       = indexRank<ColourIndex,vobj>();
+   int _ColourScalar  =  isScalar<ColourIndex,vobj>();
+   int _ColourVector  =  isVector<ColourIndex,vobj>();
+   int _ColourMatrix  =  isMatrix<ColourIndex,vobj>();
+
+   int _SpinN       = indexRank<SpinIndex,vobj>();
+   int _SpinScalar  =  isScalar<SpinIndex,vobj>();
+   int _SpinVector  =  isVector<SpinIndex,vobj>();
+   int _SpinMatrix  =  isMatrix<SpinIndex,vobj>();
+
+   int _LorentzN       = indexRank<LorentzIndex,vobj>();
+   int _LorentzScalar  =  isScalar<LorentzIndex,vobj>();
+   int _LorentzVector  =  isVector<LorentzIndex,vobj>();
+   int _LorentzMatrix  =  isMatrix<LorentzIndex,vobj>();
+
+   std::stringstream stream;
+
+   stream << "GRID_";
+   stream << ScidacWordMnemonic<stype>();
+
+   if ( _LorentzVector )   stream << "_LorentzVector"<<_LorentzN;
+   if ( _LorentzMatrix )   stream << "_LorentzMatrix"<<_LorentzN;
+
+   if ( _SpinVector )   stream << "_SpinVector"<<_SpinN;
+   if ( _SpinMatrix )   stream << "_SpinMatrix"<<_SpinN;
+
+   if ( _ColourVector )   stream << "_ColourVector"<<_ColourN;
+   if ( _ColourMatrix )   stream << "_ColourMatrix"<<_ColourN;
+
+   if ( _ColourScalar && _LorentzScalar && _SpinScalar )   stream << "_Complex";
+
+
+   typesize = sizeof(typename vobj::scalar_type);
+
+   if ( _ColourMatrix ) typesize*= _ColourN*_ColourN;
+   else                 typesize*= _ColourN;
+
+   if ( _SpinMatrix )   typesize*= _SpinN*_SpinN;
+   else                 typesize*= _SpinN;
+
+   colors    = _ColourN;
+   spins     = _SpinN;
+   datacount = _LorentzN;
+
+   return stream.str();
+ }
+ 
+ template<class vobj> std::string ScidacRecordTypeString(Lattice<vobj> & lat,int &colors, int &spins, int & typesize,int &datacount) { 
+   return ScidacRecordTypeString<vobj>(colors,spins,typesize,datacount);
+ };
+
+
+ ////////////////////////////////////////////////////////////
+ // Helper to fill out metadata
+ ////////////////////////////////////////////////////////////
+ template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
+					  FieldMetaData &header,
+					  scidacRecord & _scidacRecord,
+					  scidacFile   & _scidacFile) 
+ {
+   typedef typename getPrecision<vobj>::real_scalar_type stype;
+
+   /////////////////////////////////////
+   // Pull Grid's metadata
+   /////////////////////////////////////
+   PrepareMetaData(field,header);
+
+   /////////////////////////////////////
+   // Scidac Private File structure
+   /////////////////////////////////////
+   _scidacFile              = scidacFile(field._grid);
+
+   /////////////////////////////////////
+   // Scidac Private Record structure
+   /////////////////////////////////////
+   scidacRecord sr;
+   sr.datatype   = ScidacRecordTypeString(field,sr.colors,sr.spins,sr.typesize,sr.datacount);
+   sr.date       = header.creation_date;
+   sr.precision  = ScidacWordMnemonic<stype>();
+   sr.recordtype = GRID_IO_FIELD;
+
+   _scidacRecord = sr;
+
+   //   std::cout << GridLogMessage << "Build SciDAC datatype " <<sr.datatype<<std::endl;
+ }
+ 
+ ///////////////////////////////////////////////////////
+ // Scidac checksum
+ ///////////////////////////////////////////////////////
+ static int scidacChecksumVerify(scidacChecksum &scidacChecksum_,uint32_t scidac_csuma,uint32_t scidac_csumb)
+ {
+   uint32_t scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
+   uint32_t scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
+   if ( scidac_csuma !=scidac_checksuma) return 0;
+   if ( scidac_csumb !=scidac_checksumb) return 0;
+   return 1;
+ }
+
+////////////////////////////////////////////////////////////////////////////////////
+// Lime, ILDG and Scidac I/O classes
+////////////////////////////////////////////////////////////////////////////////////
+class GridLimeReader : public BinaryIO {
+ public:
+   ///////////////////////////////////////////////////
+   // FIXME: format for RNG? Now just binary out instead
+   ///////////////////////////////////////////////////
+
+   FILE       *File;
+   LimeReader *LimeR;
+   std::string filename;
+
+   /////////////////////////////////////////////
+   // Open the file
+   /////////////////////////////////////////////
+   void open(const std::string &_filename) 
+   {
+     filename= _filename;
+     File = fopen(filename.c_str(), "r");
+     if (File == nullptr)
+     {
+       std::cerr << "cannot open file '" << filename << "'" << std::endl;
+       abort();
+     }
+     LimeR = limeCreateReader(File);
+   }
+   /////////////////////////////////////////////
+   // Close the file
+   /////////////////////////////////////////////
+   void close(void){
+     fclose(File);
+     //     limeDestroyReader(LimeR);
+   }
+
+  ////////////////////////////////////////////
+  // Read a generic lattice field and verify checksum
+  ////////////////////////////////////////////
+  template<class vobj>
+  void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
+  {
+    typedef typename vobj::scalar_object sobj;
+    scidacChecksum scidacChecksum_;
+    FieldNormMetaData  FieldNormMetaData_;
+    uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+
+    std::string format = getFormatString<vobj>();
+
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+
+      uint64_t file_bytes =limeReaderBytes(LimeR);
+
+      //      std::cout << GridLogMessage << limeReaderType(LimeR) << " "<< file_bytes <<" bytes "<<std::endl;
+      //      std::cout << GridLogMessage<< " readLimeObject seeking "<<  record_name <<" found record :" <<limeReaderType(LimeR) <<std::endl;
+
+      if ( !strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
+
+	//	std::cout << GridLogMessage<< " readLimeLatticeBinaryObject matches ! " <<std::endl;
+
+	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 Payload expected " <<PayloadSize<<std::endl;
+	//	std::cout << "R file size " <<file_bytes <<std::endl;
+
+	assert(PayloadSize == file_bytes);// Must match or user error
+
+	uint64_t offset= ftello(File);
+	//	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
+	/////////////////////////////////////////////
+	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
+  ////////////////////////////////////////////
+  void readLimeObject(std::string &xmlstring,std::string record_name)
+  {
+    // should this be a do while; can we miss a first record??
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+
+      //      std::cout << GridLogMessage<< " readLimeObject seeking "<< record_name <<" found record :" <<limeReaderType(LimeR) <<std::endl;
+      uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
+
+      if ( !strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
+
+	//	std::cout << GridLogMessage<< " readLimeObject matches ! " << record_name <<std::endl;
+	std::vector<char> xmlc(nbytes+1,'\0');
+	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
+	//	std::cout << GridLogMessage<< " readLimeObject matches XML " << &xmlc[0] <<std::endl;
+
+	xmlstring = std::string(&xmlc[0]);
+	return;
+      }
+
+    }  
+    assert(0);
+  }
+
+  template<class serialisable_object>
+  void readLimeObject(serialisable_object &object,std::string object_name,std::string record_name)
+  {
+    std::string xmlstring;
+
+    readLimeObject(xmlstring, record_name);
+    XmlReader RD(xmlstring, true, "");
+    read(RD,object_name,object);
+  }
+};
+
+class GridLimeWriter : public BinaryIO 
+{
+ public:
+
+   ///////////////////////////////////////////////////
+   // FIXME: format for RNG? Now just binary out instead
+   // FIXME: collective calls or not ?
+   //      : must know if I am the I/O boss
+   ///////////////////////////////////////////////////
+   FILE       *File;
+   LimeWriter *LimeW;
+   std::string filename;
+   bool        boss_node;
+   GridLimeWriter( bool isboss = true) {
+     boss_node = isboss;
+   }
+   void open(const std::string &_filename) { 
+     filename= _filename;
+     if ( boss_node ) {
+       File = fopen(filename.c_str(), "w");
+       LimeW = limeCreateWriter(File); assert(LimeW != NULL );
+     }
+   }
+   /////////////////////////////////////////////
+   // Close the file
+   /////////////////////////////////////////////
+   void close(void) {
+     if ( boss_node ) {
+       fclose(File);
+     }
+     //  limeDestroyWriter(LimeW);
+   }
+  ///////////////////////////////////////////////////////
+  // Lime utility functions
+  ///////////////////////////////////////////////////////
+  int createLimeRecordHeader(std::string message, int MB, int ME, size_t PayloadSize)
+  {
+    if ( boss_node ) {
+      LimeRecordHeader *h;
+      h = limeCreateHeader(MB, ME, const_cast<char *>(message.c_str()), PayloadSize);
+      assert(limeWriteRecordHeader(h, LimeW) >= 0);
+      limeDestroyHeader(h);
+    }
+    return LIME_SUCCESS;
+  }
+  ////////////////////////////////////////////
+  // Write a generic serialisable object
+  ////////////////////////////////////////////
+  void writeLimeObject(int MB,int ME,XmlWriter &writer,std::string object_name,std::string record_name)
+  {
+    if ( boss_node ) {
+      std::string xmlstring = writer.docString();
+
+      //    std::cout << "WriteLimeObject" << record_name <<std::endl;
+      uint64_t nbytes = xmlstring.size();
+      //    std::cout << " xmlstring "<< nbytes<< " " << xmlstring <<std::endl;
+      int err;
+      LimeRecordHeader *h = limeCreateHeader(MB, ME,const_cast<char *>(record_name.c_str()), nbytes); 
+      assert(h!= NULL);
+      
+      err=limeWriteRecordHeader(h, LimeW);                    assert(err>=0);
+      err=limeWriteRecordData(&xmlstring[0], &nbytes, LimeW); assert(err>=0);
+      err=limeWriterCloseRecord(LimeW);                       assert(err>=0);
+      limeDestroyHeader(h);
+    }
+  }
+
+  template<class serialisable_object>
+  void writeLimeObject(int MB,int ME,serialisable_object &object,std::string object_name,std::string record_name, const unsigned int scientificPrec = 0)
+  {
+    XmlWriter WR("","");
+
+    if (scientificPrec)
+    {
+      WR.scientificFormat(true);
+      WR.setPrecision(scientificPrec);
+    }
+    write(WR,object_name,object);
+    writeLimeObject(MB, ME, WR, object_name, record_name);
+  }
+  ////////////////////////////////////////////////////
+  // Write a generic lattice field and csum
+  // This routine is Collectively called by all nodes
+  // in communicator used by the field._grid
+  ////////////////////////////////////////////////////
+  template<class vobj>
+  void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
+  {
+    ////////////////////////////////////////////////////////////////////
+    // NB: FILE and iostream are jointly writing disjoint sequences in the
+    // the same file through different file handles (integer units).
+    // 
+    // These are both buffered, so why I think this code is right is as follows.
+    //
+    // i)  write record header to FILE *File, telegraphing the size; flush
+    // ii) ftello reads the offset from FILE *File . 
+    // iii) iostream / MPI Open independently seek this offset. Write sequence direct to disk.
+    //      Closes iostream and flushes.
+    // iv) fseek on FILE * to end of this disjoint section.
+    //  v) Continue writing scidac record.
+    ////////////////////////////////////////////////////////////////////
+    
+    GridBase *grid = field._grid;
+    assert(boss_node == field._grid->IsBoss() );
+
+    FieldNormMetaData FNMD; FNMD.norm2 = norm2(field);
+
+    ////////////////////////////////////////////
+    // Create record header
+    ////////////////////////////////////////////
+    typedef typename vobj::scalar_object sobj;
+    int err;
+    uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+    uint64_t PayloadSize = sizeof(sobj) * grid->_gsites;
+    if ( boss_node ) {
+      createLimeRecordHeader(record_name, 0, 0, PayloadSize);
+      fflush(File);
+    }
+    
+    //    std::cout << "W sizeof(sobj)"      <<sizeof(sobj)<<std::endl;
+    //    std::cout << "W Gsites "           <<field._grid->_gsites<<std::endl;
+    //    std::cout << "W Payload expected " <<PayloadSize<<std::endl;
+
+    ////////////////////////////////////////////////
+    // Check all nodes agree on file position
+    ////////////////////////////////////////////////
+    uint64_t offset1;
+    if ( boss_node ) {
+      offset1 = ftello(File);    
+    }
+    grid->Broadcast(0,(void *)&offset1,sizeof(offset1));
+
+    ///////////////////////////////////////////
+    // The above is collective. Write by other means into the binary record
+    ///////////////////////////////////////////
+    std::string format = getFormatString<vobj>();
+    BinarySimpleMunger<sobj,sobj> munge;
+    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb);
+
+    ///////////////////////////////////////////
+    // Wind forward and close the record
+    ///////////////////////////////////////////
+    if ( boss_node ) {
+      fseek(File,0,SEEK_END);             
+      uint64_t offset2 = ftello(File);     //    std::cout << " now at offset "<<offset2 << std::endl;
+      assert( (offset2-offset1) == PayloadSize);
+    }
+
+    /////////////////////////////////////////////////////////////
+    // Check MPI-2 I/O did what we expect to file
+    /////////////////////////////////////////////////////////////
+
+    if ( boss_node ) { 
+      err=limeWriterCloseRecord(LimeW);  assert(err>=0);
+    }
+    ////////////////////////////////////////
+    // Write checksum element, propagaing forward from the BinaryIO
+    // Always pair a checksum with a binary object, and close message
+    ////////////////////////////////////////
+    scidacChecksum checksum;
+    std::stringstream streama; streama << std::hex << scidac_csuma;
+    std::stringstream streamb; streamb << std::hex << scidac_csumb;
+    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));
+    }
+  }
+};
+
+class ScidacWriter : public GridLimeWriter {
+ public:
+
+  ScidacWriter(bool isboss =true ) : GridLimeWriter(isboss)  { };
+
+  template<class SerialisableUserFile>
+  void writeScidacFileRecord(GridBase *grid,SerialisableUserFile &_userFile)
+  {
+    scidacFile    _scidacFile(grid);
+    if ( this->boss_node ) {
+      writeLimeObject(1,0,_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
+      writeLimeObject(0,1,_userFile,_userFile.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
+    }
+  }
+  ////////////////////////////////////////////////
+  // Write generic lattice field in scidac format
+  ////////////////////////////////////////////////
+  template <class vobj, class userRecord>
+  void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
+                              const unsigned int recordScientificPrec = 0) 
+  {
+    GridBase * grid = field._grid;
+
+    ////////////////////////////////////////
+    // fill the Grid header
+    ////////////////////////////////////////
+    FieldMetaData header;
+    scidacRecord  _scidacRecord;
+    scidacFile    _scidacFile;
+
+    ScidacMetaData(field,header,_scidacRecord,_scidacFile);
+
+    //////////////////////////////////////////////
+    // Fill the Lime file record by record
+    //////////////////////////////////////////////
+    if ( this->boss_node ) {
+      writeLimeObject(1,0,header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
+      writeLimeObject(0,0,_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML), recordScientificPrec);
+      writeLimeObject(0,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
+    }
+    // Collective call
+    writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
+  }
+};
+
+
+class ScidacReader : public GridLimeReader {
+ public:
+
+   template<class SerialisableUserFile>
+   void readScidacFileRecord(GridBase *grid,SerialisableUserFile &_userFile)
+   {
+     scidacFile    _scidacFile(grid);
+     readLimeObject(_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
+     readLimeObject(_userFile,_userFile.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
+   }
+  ////////////////////////////////////////////////
+  // Write generic lattice field in scidac format
+  ////////////////////////////////////////////////
+  template <class vobj, class userRecord>
+  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord) 
+  {
+    typedef typename vobj::scalar_object sobj;
+    GridBase * grid = field._grid;
+
+    ////////////////////////////////////////
+    // fill the Grid header
+    ////////////////////////////////////////
+    FieldMetaData header;
+    scidacRecord  _scidacRecord;
+    scidacFile    _scidacFile;
+
+    //////////////////////////////////////////////
+    // Fill the Lime file record by record
+    //////////////////////////////////////////////
+    readLimeObject(header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
+    readLimeObject(_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
+    readLimeObject(_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
+    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));
+  }
+  void skipPastBinaryRecord(void) {
+    std::string rec_name(ILDG_BINARY_DATA);
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+      if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
+	return;
+      }
+    }    
+  }
+  void skipPastObjectRecord(std::string rec_name) {
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+      if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+	return;
+      }
+    }
+  }
+  void skipScidacFieldRecord() {
+    skipPastObjectRecord(std::string(GRID_FORMAT));
+    skipPastObjectRecord(std::string(SCIDAC_RECORD_XML));
+    skipPastObjectRecord(std::string(SCIDAC_PRIVATE_RECORD_XML));
+    skipPastBinaryRecord();
+  }
+};
+
+
+class IldgWriter : public ScidacWriter {
+ public:
+  
+  IldgWriter(bool isboss) : ScidacWriter(isboss) {};
+
+  ///////////////////////////////////
+  // A little helper
+  ///////////////////////////////////
+  void writeLimeIldgLFN(std::string &LFN)
+  {
+    uint64_t PayloadSize = LFN.size();
+    int err;
+    createLimeRecordHeader(ILDG_DATA_LFN, 0 , 0, PayloadSize);
+    err=limeWriteRecordData(const_cast<char*>(LFN.c_str()), &PayloadSize,LimeW); assert(err>=0);
+    err=limeWriterCloseRecord(LimeW); assert(err>=0);
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Special ILDG operations ; gauge configs only.
+  // Don't require scidac records EXCEPT checksum
+  // Use Grid MetaData object if present.
+  ////////////////////////////////////////////////////////////////
+  template <class vsimd>
+  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
+  {
+    GridBase * grid = Umu._grid;
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef typename vobj::scalar_object sobj;
+
+    ////////////////////////////////////////
+    // fill the Grid header
+    ////////////////////////////////////////
+    FieldMetaData header;
+    scidacRecord  _scidacRecord;
+    scidacFile    _scidacFile;
+
+    ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
+
+    std::string format = header.floating_point;
+    header.ensemble_id    = description;
+    header.ensemble_label = description;
+    header.sequence_number = sequence;
+    header.ildg_lfn = LFN;
+
+    assert ( (format == std::string("IEEE32BIG"))  
+           ||(format == std::string("IEEE64BIG")) );
+
+    //////////////////////////////////////////////////////
+    // Fill ILDG header data struct
+    //////////////////////////////////////////////////////
+    ildgFormat ildgfmt ;
+    ildgfmt.field     = std::string("su3gauge");
+
+    if ( format == std::string("IEEE32BIG") ) { 
+      ildgfmt.precision = 32;
+    } else { 
+      ildgfmt.precision = 64;
+    }
+    ildgfmt.version = 1.0;
+    ildgfmt.lx = header.dimension[0];
+    ildgfmt.ly = header.dimension[1];
+    ildgfmt.lz = header.dimension[2];
+    ildgfmt.lt = header.dimension[3];
+    assert(header.nd==4);
+    assert(header.nd==header.dimension.size());
+
+    //////////////////////////////////////////////////////////////////////////////
+    // Field norm tests
+    //////////////////////////////////////////////////////////////////////////////
+    FieldNormMetaData FieldNormMetaData_;
+    FieldNormMetaData_.norm2 = norm2(Umu);
+
+    //////////////////////////////////////////////////////////////////////////////
+    // Fill the USQCD info field
+    //////////////////////////////////////////////////////////////////////////////
+    usqcdInfo info;
+    info.version=1.0;
+    info.plaq   = header.plaquette;
+    info.linktr = header.link_trace;
+
+    //    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));
+    writeLimeObject(0,0,info,info.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
+    writeLimeObject(0,0,ildgfmt,std::string("ildgFormat")   ,std::string(ILDG_FORMAT)); // rec
+    writeLimeIldgLFN(header.ildg_lfn);                                                 // rec
+    writeLimeLatticeBinaryObject(Umu,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
+    //    limeDestroyWriter(LimeW);
+  }
+};
+
+class IldgReader : public GridLimeReader {
+ public:
+
+  ////////////////////////////////////////////////////////////////
+  // Read either Grid/SciDAC/ILDG configuration
+  // Don't require scidac records EXCEPT checksum
+  // Use Grid MetaData object if present.
+  // Else use ILDG MetaData object if present.
+  // Else use SciDAC MetaData object if present.
+  ////////////////////////////////////////////////////////////////
+  template <class vsimd>
+  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
+
+    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef typename GaugeField::vector_object  vobj;
+    typedef typename vobj::scalar_object sobj;
+
+    typedef LorentzColourMatrixF fobj;
+    typedef LorentzColourMatrixD dobj;
+
+    GridBase *grid = Umu._grid;
+
+    std::vector<int> dims = Umu._grid->FullDimensions();
+
+    assert(dims.size()==4);
+
+    // Metadata holders
+    ildgFormat     ildgFormat_    ;
+    std::string    ildgLFN_       ;
+    scidacChecksum scidacChecksum_; 
+    usqcdInfo      usqcdInfo_     ;
+    FieldNormMetaData FieldNormMetaData_;
+
+    // track what we read from file
+    int found_ildgFormat    =0;
+    int found_ildgLFN       =0;
+    int found_scidacChecksum=0;
+    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;
+
+    // Binary format
+    std::string format;
+
+    //////////////////////////////////////////////////////////////////////////
+    // Loop over all records
+    // -- Order is poorly guaranteed except ILDG header preceeds binary section.
+    // -- Run like an event loop.
+    // -- Impose trust hierarchy. Grid takes precedence & look for ILDG, and failing
+    //    that Scidac. 
+    // -- Insist on Scidac checksum record.
+    //////////////////////////////////////////////////////////////////////////
+
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+
+      uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
+      
+      //////////////////////////////////////////////////////////////////
+      // If not BINARY_DATA read a string and parse
+      //////////////////////////////////////////////////////////////////
+      if ( strncmp(limeReaderType(LimeR), ILDG_BINARY_DATA,strlen(ILDG_BINARY_DATA) )  ) {
+	
+	// Copy out the string
+	std::vector<char> xmlc(nbytes+1,'\0');
+	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
+	//	std::cout << GridLogMessage<< "Non binary record :" <<limeReaderType(LimeR) <<std::endl; //<<"\n"<<(&xmlc[0])<<std::endl;
+
+	//////////////////////////////////
+	// ILDG format record
+
+	std::string xmlstring(&xmlc[0]);
+	if ( !strncmp(limeReaderType(LimeR), ILDG_FORMAT,strlen(ILDG_FORMAT)) ) { 
+
+	  XmlReader RD(xmlstring, true, "");
+	  read(RD,"ildgFormat",ildgFormat_);
+
+	  if ( ildgFormat_.precision == 64 ) format = std::string("IEEE64BIG");
+	  if ( ildgFormat_.precision == 32 ) format = std::string("IEEE32BIG");
+
+	  assert( ildgFormat_.lx == dims[0]);
+	  assert( ildgFormat_.ly == dims[1]);
+	  assert( ildgFormat_.lz == dims[2]);
+	  assert( ildgFormat_.lt == dims[3]);
+
+	  found_ildgFormat = 1;
+	}
+
+	if ( !strncmp(limeReaderType(LimeR), ILDG_DATA_LFN,strlen(ILDG_DATA_LFN)) ) {
+	  FieldMetaData_.ildg_lfn = xmlstring;
+	  found_ildgLFN = 1;
+	}
+
+	if ( !strncmp(limeReaderType(LimeR), GRID_FORMAT,strlen(ILDG_FORMAT)) ) { 
+
+	  XmlReader RD(xmlstring, true, "");
+	  read(RD,"FieldMetaData",FieldMetaData_);
+
+	  format = FieldMetaData_.floating_point;
+
+	  assert(FieldMetaData_.dimension[0] == dims[0]);
+	  assert(FieldMetaData_.dimension[1] == dims[1]);
+	  assert(FieldMetaData_.dimension[2] == dims[2]);
+	  assert(FieldMetaData_.dimension[3] == dims[3]);
+
+	  found_FieldMetaData = 1;
+	}
+
+	if ( !strncmp(limeReaderType(LimeR), SCIDAC_RECORD_XML,strlen(SCIDAC_RECORD_XML)) ) { 
+	  // is it a USQCD info field
+	  if ( xmlstring.find(std::string("usqcdInfo")) != std::string::npos ) { 
+	    //	    std::cout << GridLogMessage<<"...found a usqcdInfo field"<<std::endl;
+	    XmlReader RD(xmlstring, true, "");
+	    read(RD,"usqcdInfo",usqcdInfo_);
+	    found_usqcdInfo = 1;
+	  }
+	}
+
+	if ( !strncmp(limeReaderType(LimeR), SCIDAC_CHECKSUM,strlen(SCIDAC_CHECKSUM)) ) { 
+	  XmlReader RD(xmlstring, true, "");
+	  read(RD,"scidacChecksum",scidacChecksum_);
+	  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;
+	uint64_t offset= ftello(File);
+	if ( format == std::string("IEEE64BIG") ) {
+	  GaugeSimpleMunger<dobj, sobj> munge;
+	  BinaryIO::readLatticeObject< vobj, dobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+	} else { 
+	  GaugeSimpleMunger<fobj, sobj> munge;
+	  BinaryIO::readLatticeObject< vobj, fobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+	}
+
+	found_ildgBinary = 1;
+      }
+
+    }
+
+    //////////////////////////////////////////////////////
+    // Minimally must find binary segment and checksum
+    // Since this is an ILDG reader require ILDG format
+    //////////////////////////////////////////////////////
+    assert(found_ildgBinary);
+    assert(found_ildgFormat);
+    assert(found_scidacChecksum);
+
+    // Must find something with the lattice dimensions
+    assert(found_FieldMetaData||found_ildgFormat);
+
+    if ( found_FieldMetaData ) {
+
+      std::cout << GridLogMessage<<"Grid MetaData was record found: configuration was probably written by Grid ! Yay ! "<<std::endl;
+
+    } else { 
+
+      assert(found_ildgFormat);
+      assert ( ildgFormat_.field == std::string("su3gauge") );
+
+      ///////////////////////////////////////////////////////////////////////////////////////
+      // Populate our Grid metadata as best we can
+      ///////////////////////////////////////////////////////////////////////////////////////
+
+      std::ostringstream vers; vers << ildgFormat_.version;
+      FieldMetaData_.hdr_version = vers.str();
+      FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
+
+      FieldMetaData_.nd=4;
+      FieldMetaData_.dimension.resize(4);
+
+      FieldMetaData_.dimension[0] = ildgFormat_.lx ;
+      FieldMetaData_.dimension[1] = ildgFormat_.ly ;
+      FieldMetaData_.dimension[2] = ildgFormat_.lz ;
+      FieldMetaData_.dimension[3] = ildgFormat_.lt ;
+
+      if ( found_usqcdInfo ) { 
+	FieldMetaData_.plaquette = usqcdInfo_.plaq;
+	FieldMetaData_.link_trace= usqcdInfo_.linktr;
+	std::cout << GridLogMessage <<"This configuration was probably written by USQCD "<<std::endl;
+	std::cout << GridLogMessage <<"USQCD xml record Plaquette : "<<FieldMetaData_.plaquette<<std::endl;
+	std::cout << GridLogMessage <<"USQCD xml record LinkTrace : "<<FieldMetaData_.link_trace<<std::endl;
+      } else { 
+	FieldMetaData_.plaquette = 0.0;
+	FieldMetaData_.link_trace= 0.0;
+	std::cout << GridLogWarning << "This configuration is unsafe with no plaquette records that can verify it !!! "<<std::endl;
+      }
+    }
+
+    ////////////////////////////////////////////////////////////
+    // 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);
+      scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb);
+      assert( scidac_csuma ==FieldMetaData_.scidac_checksuma);
+      assert( scidac_csumb ==FieldMetaData_.scidac_checksumb);
+      std::cout << GridLogMessage<<"SciDAC checksums match " << std::endl;
+    } else { 
+      std::cout << GridLogWarning<<"SciDAC checksums not found. This is unsafe. " << std::endl;
+      assert(0); // Can I insist always checksum ?
+    }
+
+    if ( found_FieldMetaData || found_usqcdInfo ) {
+      FieldMetaData checker;
+      GaugeStatistics(Umu,checker);
+      assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
+      assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
+      std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;
+    }
+  }
+ };
+
+}}
+
+//HAVE_LIME
+#endif
+
+#endif

Grid/parallelIO/IldgIOtypes.h

@@ -0,0 +1,237 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/parallelIO/IldgIO.h
+
+Copyright (C) 2015
+
+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_ILDGTYPES_IO_H
+#define GRID_ILDGTYPES_IO_H
+
+#ifdef HAVE_LIME
+extern "C" { // for linkage
+#include "lime.h"
+}
+
+namespace Grid {
+
+/////////////////////////////////////////////////////////////////////////////////
+// Data representation of records that enter ILDG and SciDac formats
+/////////////////////////////////////////////////////////////////////////////////
+
+#define GRID_FORMAT      "grid-format"
+#define ILDG_FORMAT      "ildg-format"
+#define ILDG_BINARY_DATA "ildg-binary-data"
+#define ILDG_DATA_LFN    "ildg-data-lfn"
+#define SCIDAC_CHECKSUM           "scidac-checksum"
+#define SCIDAC_PRIVATE_FILE_XML   "scidac-private-file-xml"
+#define SCIDAC_FILE_XML           "scidac-file-xml"
+#define SCIDAC_PRIVATE_RECORD_XML "scidac-private-record-xml"
+#define SCIDAC_RECORD_XML         "scidac-record-xml"
+#define SCIDAC_BINARY_DATA        "scidac-binary-data"
+// Unused SCIDAC records names; could move to support this functionality
+#define SCIDAC_SITELIST           "scidac-sitelist"
+
+  ////////////////////////////////////////////////////////////
+  const int GRID_IO_SINGLEFILE = 0; // hardcode lift from QIO compat
+  const int GRID_IO_MULTIFILE  = 1; // hardcode lift from QIO compat
+  const int GRID_IO_FIELD      = 0; // hardcode lift from QIO compat
+  const int GRID_IO_GLOBAL     = 1; // hardcode lift from QIO compat
+  ////////////////////////////////////////////////////////////
+
+/////////////////////////////////////////////////////////////////////////////////
+// QIO uses mandatory "private" records fixed format
+// Private is in principle "opaque" however it can't be changed now because that would break existing 
+// file compatability, so should be correct to assume the undocumented but defacto file structure.
+/////////////////////////////////////////////////////////////////////////////////
+
+struct emptyUserRecord : Serializable { 
+  GRID_SERIALIZABLE_CLASS_MEMBERS(emptyUserRecord,int,dummy);
+  emptyUserRecord() { dummy=0; };
+};
+
+////////////////////////
+// Scidac private file xml
+// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile>
+////////////////////////
+struct scidacFile : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(scidacFile,
+                                  double, version,
+                                  int, spacetime,
+				  std::string, dims, // must convert to int
+                                  int, volfmt);
+
+  std::vector<int> getDimensions(void) { 
+    std::stringstream stream(dims);
+    std::vector<int> dimensions;
+    int n;
+    while(stream >> n){
+      dimensions.push_back(n);
+    }
+    return dimensions;
+  }
+
+  void setDimensions(std::vector<int> dimensions) { 
+    char delimiter = ' ';
+    std::stringstream stream;
+    for(int i=0;i<dimensions.size();i++){ 
+      stream << dimensions[i];
+      if ( i != dimensions.size()-1) { 
+	stream << delimiter <<std::endl;
+      }
+    }
+    dims = stream.str();
+  }
+
+  // Constructor provides Grid
+  scidacFile() =default; // default constructor
+  scidacFile(GridBase * grid){
+    version      = 1.0;
+    spacetime    = grid->_ndimension;
+    setDimensions(grid->FullDimensions()); 
+    volfmt       = GRID_IO_SINGLEFILE;
+  }
+
+};
+
+///////////////////////////////////////////////////////////////////////
+// scidac-private-record-xml : example
+// <scidacRecord>
+// <version>1.1</version><date>Tue Jul 26 21:14:44 2011 UTC</date><recordtype>0</recordtype>
+// <datatype>QDP_D3_ColorMatrix</datatype><precision>D</precision><colors>3</colors><spins>4</spins>
+// <typesize>144</typesize><datacount>4</datacount>
+// </scidacRecord>
+///////////////////////////////////////////////////////////////////////
+
+struct scidacRecord : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(scidacRecord,
+                                  double, version,
+                                  std::string, date,
+				  int, recordtype,
+				  std::string, datatype,
+				  std::string, precision,
+				  int, colors,
+				  int, spins,
+				  int, typesize,
+				  int, datacount);
+
+  scidacRecord()
+  : version(1.0), recordtype(0), colors(0), spins(0), typesize(0), datacount(0)
+  {}
+};
+
+////////////////////////
+// ILDG format
+////////////////////////
+struct ildgFormat : Serializable {
+public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(ildgFormat,
+				  double, version,
+				  std::string, field,
+				  int, precision,
+				  int, lx,
+				  int, ly,
+				  int, lz,
+				  int, lt);
+  ildgFormat() { version=1.0; };
+};
+////////////////////////
+// USQCD info
+////////////////////////
+struct usqcdInfo : Serializable { 
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdInfo,
+				  double, version,
+				  double, plaq,
+				  double, linktr,
+				  std::string, info);
+  usqcdInfo() { 
+    version=1.0; 
+  };
+};
+////////////////////////
+// Scidac Checksum
+////////////////////////
+struct scidacChecksum : Serializable { 
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(scidacChecksum,
+				  double, version,
+				  std::string, suma,
+				  std::string, sumb);
+  scidacChecksum() { 
+    version=1.0; 
+  };
+};
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Type:           scidac-file-xml         <title>MILC ILDG archival gauge configuration</title>
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Type:           
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+////////////////////////
+// Scidac private file xml 
+// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile> 
+////////////////////////                                                                                                                                                                              
+
+#if 0
+////////////////////////////////////////////////////////////////////////////////////////
+// From http://www.physics.utah.edu/~detar/scidac/qio_2p3.pdf
+////////////////////////////////////////////////////////////////////////////////////////
+struct usqcdPropFile : Serializable { 
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropFile,
+				  double, version,
+				  std::string, type,
+				  std::string, info);
+  usqcdPropFile() { 
+    version=1.0; 
+  };
+};
+struct usqcdSourceInfo : Serializable { 
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdSourceInfo,
+				  double, version,
+				  std::string, info);
+  usqcdSourceInfo() { 
+    version=1.0; 
+  };
+};
+struct usqcdPropInfo : Serializable { 
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropInfo,
+				  double, version,
+				  int, spin,
+				  int, color,
+				  std::string, info);
+  usqcdPropInfo() { 
+    version=1.0; 
+  };
+};
+#endif
+
+}
+#endif
+#endif

Grid/parallelIO/MetaData.h

@@ -0,0 +1,331 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/parallelIO/NerscIO.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 */
+
+#include <algorithm>
+#include <iostream>
+#include <iomanip>
+#include <fstream>
+#include <map>
+#include <unistd.h>
+#include <sys/utsname.h>
+#include <pwd.h>
+
+namespace Grid {
+
+  ///////////////////////////////////////////////////////
+  // Precision mapping
+  ///////////////////////////////////////////////////////
+  template<class vobj> static std::string getFormatString (void)
+  {
+    std::string format;
+    typedef typename getPrecision<vobj>::real_scalar_type stype;
+    if ( sizeof(stype) == sizeof(float) ) {
+      format = std::string("IEEE32BIG");
+    }
+    if ( sizeof(stype) == sizeof(double) ) {
+      format = std::string("IEEE64BIG");
+    }
+    return format;
+  }
+  ////////////////////////////////////////////////////////////////////////////////
+  // header specification/interpretation
+  ////////////////////////////////////////////////////////////////////////////////
+    class FieldNormMetaData : Serializable {
+    public:
+      GRID_SERIALIZABLE_CLASS_MEMBERS(FieldNormMetaData, double, norm2);
+    };
+    class FieldMetaData : Serializable {
+    public:
+
+      GRID_SERIALIZABLE_CLASS_MEMBERS(FieldMetaData,
+				      int, nd,
+				      std::vector<int>, dimension,
+				      std::vector<std::string>, boundary,
+				      int, data_start,
+				      std::string, hdr_version,
+				      std::string, storage_format,
+				      double, link_trace,
+				      double, plaquette,
+				      uint32_t, checksum,
+				      uint32_t, scidac_checksuma,
+				      uint32_t, scidac_checksumb,
+				      unsigned int, sequence_number,
+				      std::string, data_type,
+				      std::string, ensemble_id,
+				      std::string, ensemble_label,
+				      std::string, ildg_lfn,
+				      std::string, creator,
+				      std::string, creator_hardware,
+				      std::string, creation_date,
+				      std::string, archive_date,
+				      std::string, floating_point);
+      // WARNING: non-initialised values might lead to twisted parallel IO
+      // issues, std::string are fine because they initliase to size 0
+      // as per C++ standard.
+      FieldMetaData(void) 
+      : nd(4), dimension(4,0), boundary(4, ""), data_start(0),
+      link_trace(0.), plaquette(0.), checksum(0),
+      scidac_checksuma(0), scidac_checksumb(0), sequence_number(0)
+      {}
+    };
+
+  namespace QCD {
+
+    using namespace Grid;
+
+
+    //////////////////////////////////////////////////////////////////////
+    // Bit and Physical Checksumming and QA of data
+    //////////////////////////////////////////////////////////////////////
+    inline void GridMetaData(GridBase *grid,FieldMetaData &header)
+    {
+      int nd = grid->_ndimension;
+      header.nd = nd;
+      header.dimension.resize(nd);
+      header.boundary.resize(nd);
+      header.data_start = 0;
+      for(int d=0;d<nd;d++) {
+	header.dimension[d] = grid->_fdimensions[d];
+      }
+      for(int d=0;d<nd;d++) {
+	header.boundary[d] = std::string("PERIODIC");
+      }
+    }
+
+    inline void MachineCharacteristics(FieldMetaData &header)
+    {
+      // Who
+      struct passwd *pw = getpwuid (getuid());
+      if (pw) header.creator = std::string(pw->pw_name); 
+
+      // When
+      std::time_t t = std::time(nullptr);
+      std::tm tm_ = *std::localtime(&t);
+      std::ostringstream oss; 
+      //      oss << std::put_time(&tm_, "%c %Z");
+      header.creation_date = oss.str();
+      header.archive_date  = header.creation_date;
+
+      // What
+      struct utsname name;  uname(&name);
+      header.creator_hardware = std::string(name.nodename)+"-";
+      header.creator_hardware+= std::string(name.machine)+"-";
+      header.creator_hardware+= std::string(name.sysname)+"-";
+      header.creator_hardware+= std::string(name.release);
+    }
+
+#define dump_meta_data(field, s)					\
+      s << "BEGIN_HEADER"      << std::endl;				\
+      s << "HDR_VERSION = "    << field.hdr_version    << std::endl;	\
+      s << "DATATYPE = "       << field.data_type      << std::endl;	\
+      s << "STORAGE_FORMAT = " << field.storage_format << std::endl;	\
+      for(int i=0;i<4;i++){						\
+	s << "DIMENSION_" << i+1 << " = " << field.dimension[i] << std::endl ; \
+      }									\
+      s << "LINK_TRACE = " << std::setprecision(10) << field.link_trace << std::endl; \
+      s << "PLAQUETTE  = " << std::setprecision(10) << field.plaquette  << std::endl; \
+      for(int i=0;i<4;i++){						\
+	s << "BOUNDARY_"<<i+1<<" = " << field.boundary[i] << std::endl;	\
+      }									\
+									\
+      s << "CHECKSUM = "<< std::hex << std::setw(10) << field.checksum << std::dec<<std::endl; \
+      s << "SCIDAC_CHECKSUMA = "<< std::hex << std::setw(10) << field.scidac_checksuma << std::dec<<std::endl; \
+      s << "SCIDAC_CHECKSUMB = "<< std::hex << std::setw(10) << field.scidac_checksumb << std::dec<<std::endl; \
+      s << "ENSEMBLE_ID = "     << field.ensemble_id      << std::endl;	\
+      s << "ENSEMBLE_LABEL = "  << field.ensemble_label   << std::endl;	\
+      s << "SEQUENCE_NUMBER = " << field.sequence_number  << std::endl;	\
+      s << "CREATOR = "         << field.creator          << std::endl;	\
+      s << "CREATOR_HARDWARE = "<< field.creator_hardware << std::endl;	\
+      s << "CREATION_DATE = "   << field.creation_date    << std::endl;	\
+      s << "ARCHIVE_DATE = "    << field.archive_date     << std::endl;	\
+      s << "FLOATING_POINT = "  << field.floating_point   << std::endl;	\
+      s << "END_HEADER"         << std::endl;
+
+template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMetaData &header)
+{
+  GridBase *grid = field._grid;
+  std::string format = getFormatString<vobj>();
+   header.floating_point = format;
+   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
+   GridMetaData(grid,header); 
+   MachineCharacteristics(header);
+ }
+ 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);
+ }
+ 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);
+ }
+ template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header)
+ {
+   
+   GridBase *grid = field._grid;
+   std::string format = getFormatString<vLorentzColourMatrixF>();
+   header.floating_point = format;
+   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
+   GridMetaData(grid,header); 
+   GaugeStatistics(field,header);
+   MachineCharacteristics(header);
+ }
+ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
+ {
+   GridBase *grid = field._grid;
+   std::string format = getFormatString<vLorentzColourMatrixD>();
+   header.floating_point = format;
+   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
+   GridMetaData(grid,header); 
+   GaugeStatistics(field,header);
+   MachineCharacteristics(header);
+ }
+
+    //////////////////////////////////////////////////////////////////////
+    // Utilities ; these are QCD aware
+    //////////////////////////////////////////////////////////////////////
+    inline void reconstruct3(LorentzColourMatrix & cm)
+    {
+      const int x=0;
+      const int y=1;
+      const int z=2;
+      for(int mu=0;mu<Nd;mu++){
+	cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
+	cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
+	cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
+      }
+    }
+
+    ////////////////////////////////////////////////////////////////////////////////
+    // Some data types for intermediate storage
+    ////////////////////////////////////////////////////////////////////////////////
+    template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
+
+    typedef iLorentzColour2x3<Complex>  LorentzColour2x3;
+    typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
+    typedef iLorentzColour2x3<ComplexD> LorentzColour2x3D;
+
+/////////////////////////////////////////////////////////////////////////////////
+// Simple classes for precision conversion
+/////////////////////////////////////////////////////////////////////////////////
+template <class fobj, class sobj>
+struct BinarySimpleUnmunger {
+  typedef typename getPrecision<fobj>::real_scalar_type fobj_stype;
+  typedef typename getPrecision<sobj>::real_scalar_type sobj_stype;
+  
+  void operator()(sobj &in, fobj &out) {
+    // take word by word and transform accoding to the status
+    fobj_stype *out_buffer = (fobj_stype *)&out;
+    sobj_stype *in_buffer = (sobj_stype *)&in;
+    size_t fobj_words = sizeof(out) / sizeof(fobj_stype);
+    size_t sobj_words = sizeof(in) / sizeof(sobj_stype);
+    assert(fobj_words == sobj_words);
+    
+    for (unsigned int word = 0; word < sobj_words; word++)
+      out_buffer[word] = in_buffer[word];  // type conversion on the fly
+    
+  }
+};
+
+template <class fobj, class sobj>
+struct BinarySimpleMunger {
+  typedef typename getPrecision<fobj>::real_scalar_type fobj_stype;
+  typedef typename getPrecision<sobj>::real_scalar_type sobj_stype;
+
+  void operator()(fobj &in, sobj &out) {
+    // take word by word and transform accoding to the status
+    fobj_stype *in_buffer = (fobj_stype *)&in;
+    sobj_stype *out_buffer = (sobj_stype *)&out;
+    size_t fobj_words = sizeof(in) / sizeof(fobj_stype);
+    size_t sobj_words = sizeof(out) / sizeof(sobj_stype);
+    assert(fobj_words == sobj_words);
+    
+    for (unsigned int word = 0; word < sobj_words; word++)
+      out_buffer[word] = in_buffer[word];  // type conversion on the fly
+    
+  }
+};
+
+
+    template<class fobj,class sobj>
+    struct GaugeSimpleMunger{
+      void operator()(fobj &in, sobj &out) {
+        for (int mu = 0; mu < Nd; mu++) {
+          for (int i = 0; i < Nc; i++) {
+          for (int j = 0; j < Nc; j++) {
+	    out(mu)()(i, j) = in(mu)()(i, j);
+	  }}
+        }
+      };
+    };
+
+    template <class fobj, class sobj>
+    struct GaugeSimpleUnmunger {
+
+      void operator()(sobj &in, fobj &out) {
+        for (int mu = 0; mu < Nd; mu++) {
+          for (int i = 0; i < Nc; i++) {
+          for (int j = 0; j < Nc; j++) {
+	    out(mu)()(i, j) = in(mu)()(i, j);
+	  }}
+        }
+      };
+    };
+
+    template<class fobj,class sobj>
+    struct Gauge3x2munger{
+      void operator() (fobj &in,sobj &out){
+	for(int mu=0;mu<Nd;mu++){
+	  for(int i=0;i<2;i++){
+	  for(int j=0;j<3;j++){
+	    out(mu)()(i,j) = in(mu)(i)(j);
+	  }}
+	}
+	reconstruct3(out);
+      }
+    };
+
+    template<class fobj,class sobj>
+    struct Gauge3x2unmunger{
+      void operator() (sobj &in,fobj &out){
+	for(int mu=0;mu<Nd;mu++){
+	  for(int i=0;i<2;i++){
+	  for(int j=0;j<3;j++){
+	    out(mu)(i)(j) = in(mu)()(i,j);
+	  }}
+	}
+      }
+    };
+  }
+
+
+}

Grid/parallelIO/NerscIO.h

@@ -0,0 +1,363 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/parallelIO/NerscIO.h
+
+    Copyright (C) 2015
+
+    Author: Matt Spraggs <matthew.spraggs@gmail.com>
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_NERSC_IO_H
+#define GRID_NERSC_IO_H
+
+namespace Grid {
+  namespace QCD {
+
+    using namespace Grid;
+
+    ////////////////////////////////////////////////////////////////////////////////
+    // Write and read from fstream; comput header offset for payload
+    ////////////////////////////////////////////////////////////////////////////////
+    class NerscIO : public BinaryIO { 
+    public:
+
+      static inline void truncate(std::string file){
+	std::ofstream fout(file,std::ios::out);
+      }
+  
+      static inline unsigned int writeHeader(FieldMetaData &field,std::string file)
+      {
+      std::ofstream fout(file,std::ios::out|std::ios::in);
+      fout.seekp(0,std::ios::beg);
+      dump_meta_data(field, fout);
+      field.data_start = fout.tellp();
+      return field.data_start;
+    }
+
+      // 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;
+
+      //////////////////////////////////////////////////
+      // read the header
+      //////////////////////////////////////////////////
+      std::ifstream fin(file);
+
+      getline(fin,line); // read one line and insist is 
+
+      removeWhitespace(line);
+      std::cout << GridLogMessage << "* " << line << std::endl;
+
+      assert(line==std::string("BEGIN_HEADER"));
+
+      do {
+      getline(fin,line); // read one line
+      std::cout << GridLogMessage << "* "<<line<< std::endl;
+      int eq = line.find("=");
+      if(eq >0) {
+      std::string key=line.substr(0,eq);
+      std::string val=line.substr(eq+1);
+      removeWhitespace(key);
+      removeWhitespace(val);
+      
+      header[key] = val;
+    }
+    } while( line.find("END_HEADER") == std::string::npos );
+
+      field.data_start = fin.tellg();
+
+      //////////////////////////////////////////////////
+      // chomp the values
+      //////////////////////////////////////////////////
+      field.hdr_version    = header["HDR_VERSION"];
+      field.data_type      = header["DATATYPE"];
+      field.storage_format = header["STORAGE_FORMAT"];
+  
+      field.dimension[0] = std::stol(header["DIMENSION_1"]);
+      field.dimension[1] = std::stol(header["DIMENSION_2"]);
+      field.dimension[2] = std::stol(header["DIMENSION_3"]);
+      field.dimension[3] = std::stol(header["DIMENSION_4"]);
+
+      assert(grid->_ndimension == 4);
+      for(int d=0;d<4;d++){
+      assert(grid->_fdimensions[d]==field.dimension[d]);
+    }
+
+      field.link_trace = std::stod(header["LINK_TRACE"]);
+      field.plaquette  = std::stod(header["PLAQUETTE"]);
+
+      field.boundary[0] = header["BOUNDARY_1"];
+      field.boundary[1] = header["BOUNDARY_2"];
+      field.boundary[2] = header["BOUNDARY_3"];
+      field.boundary[3] = header["BOUNDARY_4"];
+
+      field.checksum = std::stoul(header["CHECKSUM"],0,16);
+      field.ensemble_id      = header["ENSEMBLE_ID"];
+      field.ensemble_label   = header["ENSEMBLE_LABEL"];
+      field.sequence_number  = std::stol(header["SEQUENCE_NUMBER"]);
+      field.creator          = header["CREATOR"];
+      field.creator_hardware = header["CREATOR_HARDWARE"];
+      field.creation_date    = header["CREATION_DATE"];
+      field.archive_date     = header["ARCHIVE_DATE"];
+      field.floating_point   = header["FLOATING_POINT"];
+
+      return field.data_start;
+    }
+
+    /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+    // Now the meat: the object readers
+    /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+    template<class vsimd>
+    static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+					 FieldMetaData& header,
+					 std::string file)
+    {
+      typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+
+      GridBase *grid = Umu._grid;
+      uint64_t offset = readHeader(file,Umu._grid,header);
+
+      FieldMetaData clone(header);
+
+      std::string format(header.floating_point);
+
+      int ieee32big = (format == std::string("IEEE32BIG"));
+      int ieee32    = (format == std::string("IEEE32"));
+      int ieee64big = (format == std::string("IEEE64BIG"));
+      int ieee64    = (format == std::string("IEEE64"));
+
+      uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+      // depending on datatype, set up munger;
+      // munger is a function of <floating point, Real, data_type>
+      if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
+	if ( ieee32 || ieee32big ) {
+	  BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F> 
+	    (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+	}
+	if ( ieee64 || ieee64big ) {
+	  BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D> 
+	    (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+	}
+      } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
+	if ( ieee32 || ieee32big ) {
+	  BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
+	    (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+	}
+	if ( ieee64 || ieee64big ) {
+	  BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
+	    (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
+	     nersc_csum,scidac_csuma,scidac_csumb);
+	}
+      } else {
+	assert(0);
+      }
+
+      GaugeStatistics(Umu,clone);
+
+      std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
+	       <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl;
+      std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" plaquette "<<clone.plaquette
+	       <<" header    "<<header.plaquette<<std::endl;
+      std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" link_trace "<<clone.link_trace
+	       <<" header    "<<header.link_trace<<std::endl;
+
+      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;
+	std::cerr << " plaqs " << clone.plaquette << " " << header.plaquette << std::endl;
+	std::cerr << " trace " << clone.link_trace<< " " << header.link_trace<< std::endl;
+	std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
+	exit(0);
+      }
+      assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
+      assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
+      assert(nersc_csum == header.checksum );
+      
+      std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
+    }
+
+      template<class vsimd>
+      static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+					    std::string file, 
+					    int two_row,
+					    int bits32)
+      {
+	typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+
+	typedef iLorentzColourMatrix<vsimd> vobj;
+	typedef typename vobj::scalar_object sobj;
+
+	FieldMetaData header;
+	///////////////////////////////////////////
+	// Following should become arguments
+	///////////////////////////////////////////
+	header.sequence_number = 1;
+	header.ensemble_id     = "UKQCD";
+	header.ensemble_label  = "DWF";
+
+	typedef LorentzColourMatrixD fobj3D;
+	typedef LorentzColour2x3D    fobj2D;
+  
+	GridBase *grid = Umu._grid;
+
+	GridMetaData(grid,header);
+	assert(header.nd==4);
+	GaugeStatistics(Umu,header);
+	MachineCharacteristics(header);
+
+	uint64_t offset;
+
+	// Sod it -- always write 3x3 double
+	header.floating_point = std::string("IEEE64BIG");
+	header.data_type      = std::string("4D_SU3_GAUGE_3x3");
+	GaugeSimpleUnmunger<fobj3D,sobj> munge;
+	if ( grid->IsBoss() ) { 
+	  truncate(file);
+	  offset = writeHeader(header,file);
+	}
+	grid->Broadcast(0,(void *)&offset,sizeof(offset));
+
+	uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+	BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
+								  nersc_csum,scidac_csuma,scidac_csumb);
+	header.checksum = nersc_csum;
+	if ( grid->IsBoss() ) { 
+	  writeHeader(header,file);
+	}
+
+	std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
+		 <<std::hex<<header.checksum
+		 <<std::dec<<" plaq "<< header.plaquette <<std::endl;
+
+      }
+      ///////////////////////////////
+      // RNG state
+      ///////////////////////////////
+      static inline void writeRNGState(GridSerialRNG &serial,GridParallelRNG &parallel,std::string file)
+      {
+	typedef typename GridParallelRNG::RngStateType RngStateType;
+
+	// Following should become arguments
+	FieldMetaData header;
+	header.sequence_number = 1;
+	header.ensemble_id     = "UKQCD";
+	header.ensemble_label  = "DWF";
+
+	GridBase *grid = parallel._grid;
+
+	GridMetaData(grid,header);
+	assert(header.nd==4);
+	header.link_trace=0.0;
+	header.plaquette=0.0;
+	MachineCharacteristics(header);
+
+	uint64_t offset;
+  
+#ifdef RNG_RANLUX
+	header.floating_point = std::string("UINT64");
+	header.data_type      = std::string("RANLUX48");
+#endif
+#ifdef RNG_MT19937
+	header.floating_point = std::string("UINT32");
+	header.data_type      = std::string("MT19937");
+#endif
+#ifdef RNG_SITMO
+	header.floating_point = std::string("UINT64");
+	header.data_type      = std::string("SITMO");
+#endif
+
+	if ( grid->IsBoss() ) { 
+	  truncate(file);
+	  offset = writeHeader(header,file);
+	}
+	grid->Broadcast(0,(void *)&offset,sizeof(offset));
+	
+	uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+	BinaryIO::writeRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
+	header.checksum = nersc_csum;
+	if ( grid->IsBoss() ) { 
+	  offset = writeHeader(header,file);
+	}
+
+	std::cout<<GridLogMessage 
+		 <<"Written NERSC RNG STATE "<<file<< " checksum "
+		 <<std::hex<<header.checksum
+		 <<std::dec<<std::endl;
+
+      }
+    
+      static inline void readRNGState(GridSerialRNG &serial,GridParallelRNG & parallel,FieldMetaData& header,std::string file)
+      {
+	typedef typename GridParallelRNG::RngStateType RngStateType;
+
+	GridBase *grid = parallel._grid;
+
+	uint64_t offset = readHeader(file,grid,header);
+
+	FieldMetaData clone(header);
+
+	std::string format(header.floating_point);
+	std::string data_type(header.data_type);
+
+#ifdef RNG_RANLUX
+	assert(format == std::string("UINT64"));
+	assert(data_type == std::string("RANLUX48"));
+#endif
+#ifdef RNG_MT19937
+	assert(format == std::string("UINT32"));
+	assert(data_type == std::string("MT19937"));
+#endif
+#ifdef RNG_SITMO
+	assert(format == std::string("UINT64"));
+	assert(data_type == std::string("SITMO"));
+#endif
+
+	// depending on datatype, set up munger;
+	// munger is a function of <floating point, Real, data_type>
+	uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+	BinaryIO::readRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
+
+	if ( nersc_csum != header.checksum ) { 
+	  std::cerr << "checksum mismatch "<<std::hex<< nersc_csum <<" "<<header.checksum<<std::dec<<std::endl;
+	  exit(0);
+	}
+	assert(nersc_csum == header.checksum );
+
+	std::cout<<GridLogMessage <<"Read NERSC RNG file "<<file<< " format "<< data_type <<std::endl;
+      }
+
+    };
+
+  }}
+#endif

Grid/perfmon/PerfCount.cc

@@ -40,7 +40,7 @@ const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::Performan
   { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES          ,  "CPUCYCLES.........." , INSTRUCTIONS},
   { PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS        ,  "INSTRUCTIONS......." , CPUCYCLES   },
     // 4
-#ifdef AVX512
+#ifdef KNL
     { PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", CPUCYCLES    },
     { PERF_TYPE_RAW, RawConfig(0x01,0x04), "L1_MISS_LOADS......", L1D_READ_ACCESS  },
     { PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", L1D_READ_ACCESS    },