Hadrons: faster A2A matrix load

This reverts commit 10fc263675.

.gitignore

@@ -83,6 +83,7 @@ ltmain.sh
 .Trashes
 ehthumbs.db
 Thumbs.db
+.dirstamp
 
 # build directory #
 ###################
@@ -97,11 +98,8 @@ build.sh
 
 # Eigen source #
 ################
-lib/Eigen/*
+Grid/Eigen
-
+Eigen/*
-# FFTW source #
-################
-lib/fftw/*
 
 # libtool macros #
 ##################
@@ -112,20 +110,7 @@ m4/libtool.m4
 ################
 gh-pages/
 
-# Buck files #
-##############
-.buck*
-buck-out
-BUCK
-make-bin-BUCK.sh
-
 # generated sources #
 #####################
-lib/qcd/spin/gamma-gen/*.h
+Grid/qcd/spin/gamma-gen/*.h
-lib/qcd/spin/gamma-gen/*.cc
+Grid/qcd/spin/gamma-gen/*.cc
-
-# vs code editor files #
-########################
-.vscode/
-.vscode/settings.json
-settings.json

.travis.yml

@@ -9,6 +9,11 @@ matrix:
     - os:        osx
       osx_image: xcode8.3
       compiler: clang
+      env: PREC=single
+    - os:        osx
+      osx_image: xcode8.3
+      compiler: clang
+      env: PREC=double
       
 before_install:
     - export GRIDDIR=`pwd`
@@ -16,9 +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
@@ -31,17 +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
-    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
+    - make install
-    - echo make clean
+    - cd $CWD/build
-    - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
+    - ../configure --enable-precision=$PREC --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
     - make -j4 
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - make check
-

Grid/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/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

@@ -21,6 +21,32 @@ if BUILD_HDF5
   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
 #
@@ -30,8 +56,8 @@ include Eigen.inc
 lib_LIBRARIES = libGrid.a
 
 CCFILES += $(extra_sources)
-HFILES  += $(extra_headers)
+HFILES  += $(extra_headers) Config.h Version.h
 
 libGrid_a_SOURCES              = $(CCFILES)
-libGrid_adir                   = $(pkgincludedir)
+libGrid_adir                   = $(includedir)/Grid
-nobase_dist_pkginclude_HEADERS = $(HFILES) $(eigen_files) Config.h
+nobase_dist_pkginclude_HEADERS = $(HFILES) $(eigen_files) $(eigen_unsupp_files)

Grid/algorithms/Algorithms.h

@@ -39,6 +39,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #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>
@@ -47,6 +48,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
+#include <Grid/algorithms/iterative/MinimalResidual.h>
+#include <Grid/algorithms/iterative/GeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/CoarsenedMatrix.h>
 #include <Grid/algorithms/FFT.h>

Grid/algorithms/CoarsenedMatrix.h

@@ -211,6 +211,7 @@ namespace Grid {
 
       for(int b=0;b<nn;b++){
 	
+	subspace[b] = zero;
 	gaussian(RNG,noise);
 	scale = std::pow(norm2(noise),-0.5); 
 	noise=noise*scale;
@@ -296,12 +297,57 @@ namespace Grid {
     };
 
     RealD Mdag (const CoarseVector &in, CoarseVector &out){
-      return M(in,out);
+      // // corresponds to Petrov-Galerkin coarsening
+      // return M(in,out);
+
+      // corresponds to Galerkin coarsening
+      CoarseVector tmp(Grid());
+      G5C(tmp, in);
+      M(tmp, out);
+      G5C(out, out);
+      return norm2(out);
     };
 
-    // Defer support for further coarsening for now
+    void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
-    void Mdiag    (const CoarseVector &in,  CoarseVector &out){};
+
-    void Mdir     (const CoarseVector &in,  CoarseVector &out,int dir, int disp){};
+      conformable(_grid,in._grid);
+      conformable(in._grid,out._grid);
+
+      SimpleCompressor<siteVector> compressor;
+      Stencil.HaloExchange(in,compressor);
+
+      auto point = [dir, disp](){
+        if(dir == 0 and disp == 0)
+          return 8;
+        else
+          return (4 * dir + 1 - disp) / 2;
+      }();
+
+      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) 	: 
 
@@ -417,7 +463,7 @@ namespace Grid {
       std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
 #endif
       //      ForceHermitian();
-      AssertHermitian();
+      // AssertHermitian();
       // ForceDiagonal();
     }
     void ForceDiagonal(void) {

Grid/algorithms/LinearOperator.h

@@ -309,36 +309,59 @@ namespace Grid {
       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:
-      SchurStaggeredOperator (Matrix &Mat): _Mat(Mat){};
+      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){
-	GridLogIterative.TimingMode(1);
+	ncall++;
-	std::cout << GridLogIterative << " HermOpAndNorm "<<std::endl;
+	tMpc-=usecond();
 	n2 = Mpc(in,out);
-	std::cout << GridLogIterative << " HermOpAndNorm.Mpc "<<std::endl;
+	tMpc+=usecond();
+	tIP-=usecond();
 	ComplexD dot= innerProduct(in,out);
-	std::cout << GridLogIterative << " HermOpAndNorm.innerProduct "<<std::endl;
+	tIP+=usecond();
 	n1 = real(dot);
       }
       virtual void HermOp(const Field &in, Field &out){
-	std::cout << GridLogIterative << " HermOp "<<std::endl;
+	ncall++;
-	Mpc(in,out);
+	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) {
-	Field tmp(in._grid);
+	tMeo-=usecond();
-	Field tmp2(in._grid);
-
-	std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
-	_Mat.Mooee(in,out);
-	_Mat.Mooee(out,tmp);
-	std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
-
 	_Mat.Meooe(in,out);
-	_Mat.Meooe(out,tmp2);
+	_Mat.Meooe(out,tmp);
-	std::cout << GridLogIterative << " HermOp.MeooeMeooe "<<std::endl;
+	tMeo+=usecond();
-
+	taxpby_norm-=usecond();
-	RealD nn=axpy_norm(out,-1.0,tmp2,tmp);
+	RealD nn=axpby_norm(out,-1.0,mass*mass,tmp,in);
-	std::cout << GridLogIterative << " HermOp.axpy_norm "<<std::endl;
+	taxpby_norm+=usecond();
 	return nn;
       }
       virtual  RealD MpcDag   (const Field &in, Field &out){
@@ -357,6 +380,12 @@ 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 {

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 0; }; // 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/iterative/BlockConjugateGradient.h

@@ -33,7 +33,7 @@ directory
 
 namespace Grid {
 
-enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS };
+enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
 
 //////////////////////////////////////////////////////////////////////////
 // Block conjugate gradient. Dimension zero should be the block direction
@@ -42,7 +42,6 @@ template <class Field>
 class BlockConjugateGradient : public OperatorFunction<Field> {
  public:
 
-
   typedef typename Field::scalar_type scomplex;
 
   int blockDim ;
@@ -54,21 +53,15 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
   RealD Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  Integer PrintInterval; //GridLogMessages or Iterative
   
   BlockConjugateGradient(BlockCGtype cgtype,int _Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : Tolerance(tol), CGtype(cgtype),   blockDim(_Orthog),  MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv)
+    : Tolerance(tol), CGtype(cgtype),   blockDim(_Orthog),  MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv),PrintInterval(100)
   {};
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Thin QR factorisation (google it)
 ////////////////////////////////////////////////////////////////////////////////////////////////////
-void ThinQRfact (Eigen::MatrixXcd &m_rr,
-		 Eigen::MatrixXcd &C,
-		 Eigen::MatrixXcd &Cinv,
-		 Field & Q,
-		 const Field & R)
-{
-  int Orthog = blockDim; // First dimension is block dim; this is an assumption
   ////////////////////////////////////////////////////////////////////////////////////////////////////
   //Dimensions
   // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
@@ -85,22 +78,20 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   // Cdag C = Rdag R ; passes.
   // QdagQ  = 1      ; passes
   ////////////////////////////////////////////////////////////////////////////////////////////////////
+void ThinQRfact (Eigen::MatrixXcd &m_rr,
+		 Eigen::MatrixXcd &C,
+		 Eigen::MatrixXcd &Cinv,
+		 Field & Q,
+		 const Field & R)
+{
+  int Orthog = blockDim; // First dimension is block dim; this is an assumption
   sliceInnerProductMatrix(m_rr,R,R,Orthog);
 
   // Force manifest hermitian to avoid rounding related
   m_rr = 0.5*(m_rr+m_rr.adjoint());
 
-#if 0
-  std::cout << " Calling Cholesky  ldlt on m_rr "  << m_rr <<std::endl;
-  Eigen::MatrixXcd L_ldlt = m_rr.ldlt().matrixL(); 
-  std::cout << " Called Cholesky  ldlt on m_rr "  << L_ldlt <<std::endl;
-  auto  D_ldlt = m_rr.ldlt().vectorD(); 
-  std::cout << " Called Cholesky  ldlt on m_rr "  << D_ldlt <<std::endl;
-#endif
-
-  //  std::cout << " Calling Cholesky  llt on m_rr "  <<std::endl;
   Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
-  //  std::cout << " Called Cholesky  llt on m_rr "  << L <<std::endl;
+
   C    = L.adjoint();
   Cinv = C.inverse();
   ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -112,6 +103,25 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   ////////////////////////////////////////////////////////////////////////////////////////////////////
   sliceMulMatrix(Q,Cinv,R,Orthog);
 }
+// see comments above
+void ThinQRfact (Eigen::MatrixXcd &m_rr,
+		 Eigen::MatrixXcd &C,
+		 Eigen::MatrixXcd &Cinv,
+		 std::vector<Field> & Q,
+		 const std::vector<Field> & R)
+{
+  InnerProductMatrix(m_rr,R,R);
+
+  m_rr = 0.5*(m_rr+m_rr.adjoint());
+
+  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
+
+  C    = L.adjoint();
+  Cinv = C.inverse();
+
+  MulMatrix(Q,Cinv,R);
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Call one of several implementations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -119,14 +129,20 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
 {
   if ( CGtype == BlockCGrQ ) {
     BlockCGrQsolve(Linop,Src,Psi);
-  } else if (CGtype == BlockCG ) {
-    BlockCGsolve(Linop,Src,Psi);
   } else if (CGtype == CGmultiRHS ) {
     CGmultiRHSsolve(Linop,Src,Psi);
   } else {
     assert(0);
   }
 }
+virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Field> &Src, std::vector<Field> &Psi) 
+{
+  if ( CGtype == BlockCGrQVec ) {
+    BlockCGrQsolveVec(Linop,Src,Psi);
+  } else {
+    assert(0);
+  }
+}
 
 ////////////////////////////////////////////////////////////////////////////
 // BlockCGrQ implementation:
@@ -139,7 +155,8 @@ 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;
@@ -202,15 +219,10 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   std::cout << GridLogMessage<<"BlockCGrQ algorithm initialisation " <<std::endl;
 
   //1.  QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
-
   Linop.HermOp(X, AD);
   tmp = B - AD;  
-  //std::cout << GridLogMessage << " initial tmp " << norm2(tmp)<< std::endl;
+
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
-  //std::cout << GridLogMessage << " initial Q " << norm2(Q)<< std::endl;
-  //std::cout << GridLogMessage << " m_rr " << m_rr<<std::endl;
-  //std::cout << GridLogMessage << " m_C " << m_C<<std::endl;
-  //std::cout << GridLogMessage << " m_Cinv " << m_Cinv<<std::endl;
   D=Q;
 
   std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
@@ -232,14 +244,12 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     MatrixTimer.Start();
     Linop.HermOp(D, Z);      
     MatrixTimer.Stop();
-    //std::cout << GridLogMessage << " norm2 Z " <<norm2(Z)<<std::endl;
 
     //4. M  = [D^dag Z]^{-1}
     sliceInnerTimer.Start();
     sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
     sliceInnerTimer.Stop();
     m_M       = m_DZ.inverse();
-    //std::cout << GridLogMessage << " m_DZ " <<m_DZ<<std::endl;
     
     //5. X  = X + D MC
     m_tmp     = m_M * m_C;
@@ -257,6 +267,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     
     //7. D  = Q + D S^dag
     m_tmp = m_S.adjoint();
+
     sliceMaddTimer.Start();
     sliceMaddMatrix(D,m_tmp,D,Q,Orthog);
     sliceMaddTimer.Stop();
@@ -317,152 +328,6 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   IterationsToComplete = k;
 }
 //////////////////////////////////////////////////////////////////////////
-// Block conjugate gradient; Original O'Leary Dimension zero should be the block direction
-//////////////////////////////////////////////////////////////////////////
-void BlockCGsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
-{
-  int Orthog = blockDim; // First dimension is block dim; this is an assumption
-  Nblock = Src._grid->_fdimensions[Orthog];
-
-  std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
-
-  Psi.checkerboard = Src.checkerboard;
-  conformable(Psi, Src);
-
-  Field P(Src);
-  Field AP(Src);
-  Field R(Src);
-  
-  Eigen::MatrixXcd m_pAp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_pAp_inv= Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_rr_inv = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  Eigen::MatrixXcd m_alpha      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_beta   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  // Initial residual computation & set up
-  std::vector<RealD> residuals(Nblock);
-  std::vector<RealD> ssq(Nblock);
-
-  sliceNorm(ssq,Src,Orthog);
-  RealD sssum=0;
-  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
-
-  sliceNorm(residuals,Src,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  sliceNorm(residuals,Psi,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  // Initial search dir is guess
-  Linop.HermOp(Psi, AP);
-  
-
-  /************************************************************************
-   * Block conjugate gradient (Stephen Pickles, thesis 1995, pp 71, O Leary 1980)
-   ************************************************************************
-   * O'Leary : R = B - A X
-   * O'Leary : P = M R ; preconditioner M = 1
-   * O'Leary : alpha = PAP^{-1} RMR
-   * O'Leary : beta  = RMR^{-1}_old RMR_new
-   * O'Leary : X=X+Palpha
-   * O'Leary : R_new=R_old-AP alpha
-   * O'Leary : P=MR_new+P beta
-   */
-
-  R = Src - AP;  
-  P = R;
-  sliceInnerProductMatrix(m_rr,R,R,Orthog);
-
-  GridStopWatch sliceInnerTimer;
-  GridStopWatch sliceMaddTimer;
-  GridStopWatch MatrixTimer;
-  GridStopWatch SolverTimer;
-  SolverTimer.Start();
-
-  int k;
-  for (k = 1; k <= MaxIterations; k++) {
-
-    RealD rrsum=0;
-    for(int b=0;b<Nblock;b++) rrsum+=real(m_rr(b,b));
-
-    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
-	      <<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
-
-    MatrixTimer.Start();
-    Linop.HermOp(P, AP);
-    MatrixTimer.Stop();
-
-    // Alpha
-    sliceInnerTimer.Start();
-    sliceInnerProductMatrix(m_pAp,P,AP,Orthog);
-    sliceInnerTimer.Stop();
-    m_pAp_inv = m_pAp.inverse();
-    m_alpha   = m_pAp_inv * m_rr ;
-
-    // Psi, R update
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(Psi,m_alpha, P,Psi,Orthog);     // add alpha *  P to psi
-    sliceMaddMatrix(R  ,m_alpha,AP,  R,Orthog,-1.0);// sub alpha * AP to resid
-    sliceMaddTimer.Stop();
-
-    // Beta
-    m_rr_inv = m_rr.inverse();
-    sliceInnerTimer.Start();
-    sliceInnerProductMatrix(m_rr,R,R,Orthog);
-    sliceInnerTimer.Stop();
-    m_beta = m_rr_inv *m_rr;
-
-    // Search update
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(AP,m_beta,P,R,Orthog);
-    sliceMaddTimer.Stop();
-    P= AP;
-
-    /*********************
-     * convergence monitor
-     *********************
-     */
-    RealD max_resid=0;
-    RealD rr;
-    for(int b=0;b<Nblock;b++){
-      rr = real(m_rr(b,b))/ssq[b];
-      if ( rr > max_resid ) max_resid = rr;
-    }
-    
-    if ( max_resid < Tolerance*Tolerance ) { 
-
-      SolverTimer.Stop();
-
-      std::cout << GridLogMessage<<"BlockCG converged in "<<k<<" iterations"<<std::endl;
-      for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "
-		  << std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
-      }
-      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
-
-      Linop.HermOp(Psi, AP);
-      AP = AP-Src;
-      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
-
-      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
-      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
-      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
-	    
-      IterationsToComplete = k;
-      return;
-    }
-
-  }
-  std::cout << GridLogMessage << "BlockConjugateGradient did NOT converge" << std::endl;
-
-  if (ErrorOnNoConverge) assert(0);
-  IterationsToComplete = k;
-}
-//////////////////////////////////////////////////////////////////////////
 // multiRHS conjugate gradient. Dimension zero should be the block direction
 // Use this for spread out across nodes
 //////////////////////////////////////////////////////////////////////////
@@ -600,6 +465,233 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   IterationsToComplete = k;
 }
 
+void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
+  for(int b=0;b<Nblock;b++){
+  for(int bp=0;bp<Nblock;bp++) {
+    m(b,bp) = innerProduct(X[b],Y[bp]);  
+  }}
+}
+void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
+  // Should make this cache friendly with site outermost, parallel_for
+  // Deal with case AP aliases with either Y or X
+  std::vector<Field> tmp(Nblock,X[0]);
+  for(int b=0;b<Nblock;b++){
+    tmp[b]   = Y[b];
+    for(int bp=0;bp<Nblock;bp++) {
+      tmp[b] = tmp[b] + (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;
+}
+
+
+
 };
 
 }

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

@@ -54,6 +54,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
+
     psi.checkerboard = src.checkerboard;
     conformable(psi, src);
 
@@ -70,7 +71,6 @@ class ConjugateGradient : public OperatorFunction<Field> {
     
     Linop.HermOpAndNorm(psi, mmp, d, b);
 
-
     r = src - mmp;
     p = r;
 
@@ -96,38 +96,44 @@ class ConjugateGradient : public OperatorFunction<Field> {
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
     GridStopWatch LinalgTimer;
+    GridStopWatch InnerTimer;
+    GridStopWatch AxpyNormTimer;
+    GridStopWatch LinearCombTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
 
     SolverTimer.Start();
     int k;
-    for (k = 1; k <= MaxIterations; k++) {
+    for (k = 1; k <= MaxIterations*1000; k++) {
       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
+      LinearCombTimer.Start();
-      psi = a * p + psi;
+      parallel_for(int ss=0;ss<src._grid->oSites();ss++){
-      p = p * b + r;
+	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;
+                << " residual^2 " << sqrt(cp/ssq) << " target " << Tolerance << 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;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -148,6 +154,9 @@ 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);
 

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;
 
@@ -164,6 +165,15 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     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/Deflation.h

@@ -0,0 +1,104 @@
+    /*************************************************************************************
+
+    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 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

@@ -57,7 +57,8 @@ void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, i
       
   parallel_region
   {
-    std::vector < vobj > B(Nm); // Thread private
+
+    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.;
@@ -149,19 +150,6 @@ void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, boo
   basisReorderInPlace(_v,sort_vals,idx);
 }
 
-// PAB: faster to compute the inner products first then fuse loops.
-// If performance critical can improve.
-template<class Field>
-void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
-  result = zero;
-  assert(_v.size()==eval.size());
-  int N = (int)_v.size();
-  for (int i=0;i<N;i++) {
-    Field& tmp = _v[i];
-    axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
-  }
-}
-
 /////////////////////////////////////////////////////////////
 // Implicitly restarted lanczos
 /////////////////////////////////////////////////////////////
@@ -181,6 +169,7 @@ enum IRLdiagonalisation {
 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)
@@ -243,6 +232,7 @@ class ImplicitlyRestartedLanczos {
   /////////////////////////
   
 public:       
+
   //////////////////////////////////////////////////////////////////
   // PAB:
   //////////////////////////////////////////////////////////////////
@@ -490,15 +480,13 @@ until convergence
 	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) ) {
+	  if( !_Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) {
-	    if ( j > Nconv ) {
+	    allconv=0;
-	      Nconv=j+1;
-	      jj=Nstop; // Terminate the scan
-	    }
 	  }
 	}
 	// Do evec[0] for good measure
@@ -506,8 +494,10 @@ until convergence
 	  int j=0;
 	  RealD e = eval2_copy[0]; 
 	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
-	  _Tester.TestConvergence(j,eresid,B,e,evalMaxApprox);
+	  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){

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -28,7 +28,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     /*  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,
@@ -45,6 +48,7 @@ struct LanczosParams : Serializable {
 struct LocalCoherenceLanczosParams : Serializable {
  public:
   GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
+				  bool, saveEvecs,
 				  bool, doFine,
 				  bool, doFineRead,
 				  bool, doCoarse,
@@ -70,21 +74,24 @@ public:
   typedef Lattice<Fobj>          FineField;
 
   LinearOperatorBase<FineField> &_Linop;
-  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+  std::vector<FineField>        &subspace;
 
-  ProjectedHermOp(LinearOperatorBase<FineField>& linop,  Aggregation<Fobj,CComplex,nbasis> &aggregate) : 
+  ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) : 
-    _Linop(linop),
+    _Linop(linop), subspace(_subspace)
-    _Aggregate(aggregate)  {  };
+  {  
+    assert(subspace.size() >0);
+  };
 
   void operator()(const CoarseField& in, CoarseField& out) {
+    GridBase *FineGrid = subspace[0]._grid;    
+    int   checkerboard = subspace[0].checkerboard;
       
-    GridBase *FineGrid = _Aggregate.FineGrid;
+    FineField fin (FineGrid);     fin.checkerboard= checkerboard;
-    FineField fin(FineGrid);
+    FineField fout(FineGrid);   fout.checkerboard = checkerboard;
-    FineField fout(FineGrid);
 
-    _Aggregate.PromoteFromSubspace(in,fin);    std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
+    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;
-    _Aggregate.ProjectToSubspace(out,fout);    std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
+    blockProject(out,fout,subspace);     std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
   }
 };
 
@@ -99,24 +106,27 @@ public:
 
   OperatorFunction<FineField>   & _poly;
   LinearOperatorBase<FineField> &_Linop;
-  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+  std::vector<FineField>        &subspace;
 
-  ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,LinearOperatorBase<FineField>& linop, 
+  ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,
-			  Aggregation<Fobj,CComplex,nbasis> &aggregate) : 
+			  LinearOperatorBase<FineField>& linop, 
+			  std::vector<FineField> & _subspace) :
     _poly(poly),
     _Linop(linop),
-    _Aggregate(aggregate)  {  };
+    subspace(_subspace)
+  {  };
 
   void operator()(const CoarseField& in, CoarseField& out) {
     
-    GridBase *FineGrid = _Aggregate.FineGrid;
+    GridBase *FineGrid = subspace[0]._grid;    
+    int   checkerboard = subspace[0].checkerboard;
 
-    FineField fin(FineGrid) ;fin.checkerboard  =_Aggregate.checkerboard;
+    FineField fin (FineGrid); fin.checkerboard =checkerboard;
-    FineField fout(FineGrid);fout.checkerboard =_Aggregate.checkerboard;
+    FineField fout(FineGrid);fout.checkerboard =checkerboard;
     
-    _Aggregate.PromoteFromSubspace(in,fin);    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
+    blockPromote(in,fin,subspace);             std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
     _poly(_Linop,fin,fout);                    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
-    _Aggregate.ProjectToSubspace(out,fout);    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
+    blockProject(out,fout,subspace);           std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
   }
 };
 
@@ -132,19 +142,23 @@ class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanc
   LinearFunction<CoarseField> & _Poly;
   OperatorFunction<FineField>   & _smoother;
   LinearOperatorBase<FineField> &_Linop;
-  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
   RealD                          _coarse_relax_tol;
+  std::vector<FineField>        &_subspace;
+  
   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
 					   OperatorFunction<FineField>   &smoother,
 					   LinearOperatorBase<FineField> &Linop,
-					   Aggregation<Fobj,CComplex,nbasis> &Aggregate,
+					   std::vector<FineField>        &subspace,
 					   RealD coarse_relax_tol=5.0e3) 
-    : _smoother(smoother), _Linop(Linop),_Aggregate(Aggregate), _Poly(Poly), _coarse_relax_tol(coarse_relax_tol)  {    };
+    : _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);
 
@@ -168,14 +182,13 @@ class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanc
   }
   int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
-    GridBase *FineGrid = _Aggregate.FineGrid;
+    GridBase *FineGrid = _subspace[0]._grid;    
-
+    int checkerboard   = _subspace[0].checkerboard;
-    int checkerboard   = _Aggregate.checkerboard;
-
     FineField fB(FineGrid);fB.checkerboard =checkerboard;
     FineField fv(FineGrid);fv.checkerboard =checkerboard;
 
-    _Aggregate.PromoteFromSubspace(B,fv);
+    blockPromote(B,fv,_subspace);  
+    
     _smoother(_Linop,fv,fB); 
 
     RealD eval_poly = eval;
@@ -217,27 +230,67 @@ protected:
   int _checkerboard;
   LinearOperatorBase<FineField>                 & _FineOp;
   
-  // FIXME replace Aggregation with vector of fine; the code reuse is too small for
+  std::vector<RealD>                              &evals_fine;
-  // the hassle and complexity of cross coupling.
+  std::vector<RealD>                              &evals_coarse; 
-  Aggregation<Fobj,CComplex,nbasis>               _Aggregate;  
+  std::vector<FineField>                          &subspace;
-  std::vector<RealD>                              evals_fine;
+  std::vector<CoarseField>                        &evec_coarse;
-  std::vector<RealD>                              evals_coarse; 
+
-  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),
-    _Aggregate(CoarseGrid,FineGrid,checkerboard),
     _FineOp(FineOp),
-    _checkerboard(checkerboard)
+    _checkerboard(checkerboard),
+    evals_fine  (_evals_fine),
+    evals_coarse(_evals_coarse),
+    subspace    (_subspace),
+    evec_coarse(_evec_coarse)
   {
     evals_fine.resize(0);
     evals_coarse.resize(0);
   };
-  void Orthogonalise(void ) { _Aggregate.Orthogonalise(); }
+  //////////////////////////////////////////////////////////////////////////
+  // 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) 
   {
@@ -246,43 +299,44 @@ public:
     v = v * (1.0/nn);
     return nn;
   }
-
+  /*
   void fakeFine(void)
   {
     int Nk = nbasis;
-    _Aggregate.subspace.resize(Nk,_FineGrid);
+    subspace.resize(Nk,_FineGrid);
-    _Aggregate.subspace[0]=1.0;
+    subspace[0]=1.0;
-    _Aggregate.subspace[0].checkerboard=_checkerboard;
+    subspace[0].checkerboard=_checkerboard;
-    normalise(_Aggregate.subspace[0]);
+    normalise(subspace[0]);
     PlainHermOp<FineField>    Op(_FineOp);
     for(int k=1;k<Nk;k++){
-      _Aggregate.subspace[k].checkerboard=_checkerboard;
+      subspace[k].checkerboard=_checkerboard;
-      Op(_Aggregate.subspace[k-1],_Aggregate.subspace[k]);
+      Op(subspace[k-1],subspace[k]);
-      normalise(_Aggregate.subspace[k]);
+      normalise(subspace[k]);
     }
   }
+  */
 
   void testFine(RealD resid) 
   {
     assert(evals_fine.size() == nbasis);
-    assert(_Aggregate.subspace.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,_Aggregate.subspace[k],evals_fine[k],1.0)==1);
+      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(_Aggregate.subspace.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,_Aggregate);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,subspace);
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
 
     for(int k=0;k<evec_coarse.size();k++){
       if ( k < nbasis ) { 
@@ -302,34 +356,34 @@ public:
     PlainHermOp<FineField>    Op(_FineOp);
 
     evals_fine.resize(Nm);
-    _Aggregate.subspace.resize(Nm,_FineGrid);
+    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,_Aggregate.subspace,src,Nconv,false);
+    IRL.calc(evals_fine,subspace,src,Nconv,false);
     
     // Shrink down to number saved
     assert(Nstop>=nbasis);
     assert(Nconv>=nbasis);
     evals_fine.resize(nbasis);
-    _Aggregate.subspace.resize(nbasis,_FineGrid);
+    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,_Aggregate);
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace);
-    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,_Aggregate);
+    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,_Aggregate,relax);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
 
     evals_coarse.resize(Nm);
     evec_coarse.resize(Nm,_CoarseGrid);

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/PrecGeneralisedConjugateResidual.h

@@ -139,7 +139,10 @@ namespace Grid {
       MatTimer.Start();
       Linop.HermOpAndNorm(psi,Az,zAz,zAAz); 
       MatTimer.Stop();
+
+      LinalgTimer.Start();
       r=src-Az;
+      LinalgTimer.Stop();
 
       /////////////////////
       // p = Prec(r)
@@ -152,8 +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);
+        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 +1,473 @@
+    /*************************************************************************************
+
+    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;
+  public:
+
+    SchurRedBlackBase(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false)  :
+    _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise = 0;
+      subtractGuess(initSubGuess);
+    };
+    void subtractGuess(const bool initSubGuess)
+    {
+      subGuess = initSubGuess;
+    }
+    bool isSubtractGuess(void)
+    {
+      return subGuess;
+    }
+
+    /////////////////////////////////////////////////////////////
+    // 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++){
+	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
+      ////////////////////////////////
+      Field   tmp(grid);
+      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. Not virtual as template methods
+    /////////////////////////////////////////////////////////////
+    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) 
+      :    SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess) 
+    {
+    }
+
+    //////////////////////////////////////////////////////
+    // 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)  
+      : SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess) {};
+
+
+    //////////////////////////////////////////////////////
+    // 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)  
+    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess) {};
+
+    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.h

@@ -277,7 +277,9 @@ public:
     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;
       }

Grid/communicator/Communicator_mpi3.cc

@@ -44,11 +44,13 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
     MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
-    assert (provided == MPI_THREAD_MULTIPLE);
+    //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);
   }
 
-  Grid_quiesce_nodes();
+  // Never clean up as done once.
-
   MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
 
   GlobalSharedMemory::Init(communicator_world);
@@ -85,9 +87,17 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
 {
   MPI_Comm optimal_comm;
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm); // Remap using the shared memory optimising routine
+  ////////////////////////////////////////////////////
+  // 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);
 }
 
 //////////////////////////////////
@@ -112,10 +122,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   //  int Nparent = parent._processors ; 
-  //  std::cout << " splitting from communicator "<<parent.communicator <<std::endl;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
-  //  std::cout << " Parent size  "<<Nparent <<std::endl;
 
   int childsize=1;
   for(int d=0;d<processors.size();d++) {
@@ -124,8 +132,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   int Nchild = Nparent/childsize;
   assert (childsize * Nchild == Nparent);
 
-  //  std::cout << " child size  "<<childsize <<std::endl;
-
   std::vector<int> ccoor(_ndimension); // coor within subcommunicator
   std::vector<int> scoor(_ndimension); // coor of split within parent
   std::vector<int> ssize(_ndimension); // coor of split within parent
@@ -183,8 +189,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
 
   } else {
     srank = 0;
-    comm_split    = parent.communicator;
+    int ierr = MPI_Comm_dup (parent.communicator,&comm_split);
-    //    std::cout << " Inherited communicator " <<comm_split <<std::endl;
+    assert(ierr==0);
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -197,6 +203,11 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   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++){
@@ -210,6 +221,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
 
 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);
 

Grid/communicator/SharedMemory.h

@@ -133,6 +133,7 @@ class SharedMemory
 
  public:
   SharedMemory() {};
+  ~SharedMemory();
   ///////////////////////////////////////////////////////////////////////////////////////
   // set the buffers & sizes
   ///////////////////////////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryMPI.cc

@@ -27,6 +27,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
+#include <pwd.h>
 
 namespace Grid { 
 
@@ -113,19 +114,150 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   assert(WorldNode!=-1);
   _ShmSetup=1;
 }
-
+// Gray encode support 
-void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+int BinaryToGray (int  binary) {
+  int gray = (binary>>1)^binary;
+  return gray;
+}
+int Log2Size(int TwoToPower,int MAXLOG2)
 {
-  ////////////////////////////////////////////////////////////////
-  // Assert power of two shm_size.
-  ////////////////////////////////////////////////////////////////
   int log2size = -1;
-  for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){  
+  for(int i=0;i<=MAXLOG2;i++){
-    if ( (0x1<<i) == WorldShmSize ) {
+    if ( (0x1<<i) == TwoToPower ) {
       log2size = i;
       break;
     }
   }
+  return log2size;
+}
+void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+#ifdef HYPERCUBE
+  ////////////////////////////////////////////////////////////////
+  // 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);
+#else 
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
   assert(log2size != -1);
 
   ////////////////////////////////////////////////////////////////
@@ -174,7 +306,69 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
+#endif
 }
+////////////////////////////////////////////////////////////////////////////////////////////
+// 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  %lld\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
@@ -182,6 +376,7 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
 #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);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -191,7 +386,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   WorldShmCommBufs.resize(WorldShmSize);
   
   ////////////////////////////////////////////////////////////////////////////////////////////
-  // Hugetlbf and others map filesystems as mappable huge pages
+  // Hugetlbfs and others map filesystems as mappable huge pages
   ////////////////////////////////////////////////////////////////////////////////////////////
   char shm_name [NAME_MAX];
   for(int r=0;r<WorldShmSize;r++){
@@ -218,6 +413,49 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+  }
+  _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;
@@ -232,6 +470,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
+  std::cout << "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
@@ -243,7 +482,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	
       size_t size = bytes;
       
-      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldNode,r);
+      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);
@@ -259,7 +499,11 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      if ( ptr == (void * )MAP_FAILED ) {       perror("failed mmap");      assert(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;
@@ -274,7 +518,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 
       size_t size = bytes ;
       
-      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldNode,r);
+      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);      }
@@ -292,6 +537,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 }
 #endif
 
+
+
+
   ////////////////////////////////////////////////////////
   // Global shared functionality finished
   // Now move to per communicator functionality
@@ -318,11 +566,12 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   heap_size = GlobalSharedMemory::ShmAllocBytes();
   for(int r=0;r<ShmSize;r++){
 
-    uint32_t sr = (r==ShmRank) ? GlobalSharedMemory::WorldRank : 0 ;
+    uint32_t wsr = (r==ShmRank) ? GlobalSharedMemory::WorldShmRank : 0 ;
 
-    MPI_Allreduce(MPI_IN_PLACE,&sr,1,MPI_UINT32_T,MPI_SUM,comm);
+    MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
 
-    ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[sr];
+    ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
+    //    std::cout << "SetCommunicator ShmCommBufs ["<< r<< "] = "<< ShmCommBufs[r]<< "  wsr = "<<wsr<<std::endl;
   }
   ShmBufferFreeAll();
 
@@ -391,5 +640,12 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
     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

@@ -122,5 +122,7 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 {
   return NULL;
 }
+SharedMemory::~SharedMemory()
+{};
 
 }

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];
   }
 }
 
@@ -141,32 +143,36 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
   int e2=rhs._grid->_slice_block[dimension];
   int stride=rhs._grid->_slice_stride[dimension];
 
+  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+  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;
+
+  parallel_for(int i=0;i<ent;i++){
     rhs._odata[table[i].first]=buffer[table[i].second];
   }
 }
-}
 
 //////////////////////////////////////////////////////
 // Scatter for when there *is* need to SIMD split
@@ -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];
-  if(cbmask == 0x3 ){
+  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
-    parallel_for_nest2(int n=0;n<e1;n++){
+  int ent=0;
-      for(int b=0;b<e2;b++){
 
+  if(cbmask == 0x3 ){
+    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];
+	table[ent++] = std::pair<int,int>(lo+o,ro+o);
-	vstream(lhs._odata[lo+o],rhs._odata[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];
+	  table[ent++] = std::pair<int,int>(lo+o,ro+o);
-	  vstream(lhs._odata[lo+o],rhs._odata[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++){
+  static std::vector<std::pair<int,int> > table;  table.resize(e1*e2);
-  for(int b=0;b<e2;b++){
+  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 ) {
+      if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
-	permute(lhs._odata[lo+o+b],rhs._odata[ro+o+b],permute_type);
+    }}
   }
 
-  }}
+  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];
@@ -326,10 +349,6 @@ 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.
     // 
@@ -366,9 +385,7 @@ 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;
+    return axpy_norm_fast(ret,a,x,y);
-    conformable(ret,x);
-    conformable(x,y);
-    axpy(ret,a,x,y);
-    return norm2(ret);
   }
   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;
+    return axpby_norm_fast(ret,a,b,x,y);
-    conformable(ret,x);
-    conformable(x,y);
-    axpby(ret,a,b,x,y);
-    return norm2(ret); // FIXME implement parallel norm in ss loop
   }
 
 }

Grid/lattice/Lattice_base.h

@@ -257,7 +257,40 @@ public:
     }  	
   }
 
+  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;
     
@@ -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

@@ -33,7 +33,7 @@ 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); 
 }
 
@@ -43,32 +43,85 @@ inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &righ
 {
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
-  scalar_type  nrm;
-  
   GridBase *grid = left._grid;
+  const int pad = 8;
 
-  std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
+  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(inner);
-  right._grid->GlobalSum(nrm);
+  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)
   ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
@@ -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) 
 {

Grid/lattice/Lattice_rng.h

@@ -158,10 +158,19 @@ 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 = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
+      const int shift = 30;
+
       uint64_t skip = site;
-      skip = skip<<40;
+
+      skip = skip<<shift;
+
+      assert((skip >> shift)==site); // check for overflow
+
       eng.discard(skip);
       //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
     } 
@@ -308,6 +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 {
@@ -368,6 +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

Grid/lattice/Lattice_transfer.h

@@ -464,9 +464,11 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
+    if ( d!=orthog ) {
       assert(lg->_processors[d]  == hg->_processors[d]);
       assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
     }
+  }
 
   // the above should guarantee that the operations are local
   parallel_for(int idx=0;idx<lg->lSites();idx++){
@@ -485,7 +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;
 
@@ -499,9 +501,11 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slic
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
+    if ( d!=orthog ) {
       assert(lg->_processors[d]  == hg->_processors[d]);
       assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
     }
+  }
 
   // the above should guarantee that the operations are local
   parallel_for(int idx=0;idx<lg->lSites();idx++){
@@ -599,6 +603,51 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
     extract1(in_vobj, out_ptrs, 0);
   }
 }
+
+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 
@@ -648,10 +697,59 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   }
 }
 
+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;

Grid/log/Log.cc

@@ -59,6 +59,7 @@ void GridLogTimestamp(int on){
 }
 
 Colours GridLogColours(0);
+GridLogger GridLogMG     (1, "MG"    , GridLogColours, "NORMAL");
 GridLogger GridLogIRL    (1, "IRL"   , GridLogColours, "NORMAL");
 GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");

Grid/log/Log.h

@@ -86,7 +86,7 @@ protected:
   Colours &Painter;
   int active;
   int timing_mode;
-  int topWidth{-1};
+  int topWidth{-1}, chanWidth{-1};
   static int timestamp;
   std::string name, topName;
   std::string COLOUR;
@@ -126,6 +126,7 @@ public:
     }
   }
   void setTopWidth(const int w) {topWidth = w;}
+  void setChanWidth(const int w) {chanWidth = w;}
 
   friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
@@ -136,13 +137,20 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      stream << log.colour() <<  std::left << log.name << log.background() << " : ";
+      stream << log.colour() <<  std::left;
+      if (log.chanWidth > 0)
+      {
+        stream << std::setw(log.chanWidth);
+      }
+      stream << log.name << log.background() << " : ";
       if ( log.timestamp ) {
 	log.StopWatch->Stop();
 	GridTime now = log.StopWatch->Elapsed();
+	
 	if ( log.timing_mode==1 ) log.StopWatch->Reset();
 	log.StopWatch->Start();
-	stream << log.evidence()<< std::setw(6)<<now << log.background() << " : " ;
+	stream << log.evidence()
+	       << now	       << log.background() << " : " ;
       }
       stream << log.colour();
       return stream;
@@ -161,6 +169,7 @@ public:
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
 
+extern GridLogger GridLogMG;
 extern GridLogger GridLogIRL;
 extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;

Grid/parallelIO/BinaryIO.cc

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

Grid/parallelIO/BinaryIO.h

@@ -81,6 +81,7 @@ inline void removeWhitespace(std::string &key)
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 class BinaryIO {
  public:
+  static int latticeWriteMaxRetry;
 
   /////////////////////////////////////////////////////////////////////////////
   // more byte manipulation helpers
@@ -91,7 +92,7 @@ class BinaryIO {
     typedef typename vobj::scalar_object sobj;
 
     GridBase *grid = lat._grid;
-    int lsites = grid->lSites();
+    uint64_t lsites = grid->lSites();
 
     std::vector<sobj> scalardata(lsites); 
     unvectorizeToLexOrdArray(scalardata,lat);    
@@ -110,11 +111,11 @@ class BinaryIO {
       lsites = 1;
     }
 
-    #pragma omp parallel
+PARALLEL_REGION
     {
       uint32_t nersc_csum_thr = 0;
 
-      #pragma omp for
+PARALLEL_FOR_LOOP_INTERN
       for (uint64_t local_site = 0; local_site < lsites; local_site++)
       {
         uint32_t *site_buf = (uint32_t *)&fbuf[local_site];
@@ -124,7 +125,7 @@ class BinaryIO {
         }
       }
 
-      #pragma omp critical
+PARALLEL_CRITICAL
       {
         nersc_csum += nersc_csum_thr;
       }
@@ -146,21 +147,23 @@ class BinaryIO {
     std::vector<int> local_start =grid->LocalStarts();
     std::vector<int> global_vol  =grid->FullDimensions();
 
-#pragma omp parallel
+PARALLEL_REGION
     { 
       std::vector<int> coor(nd);
       uint32_t scidac_csuma_thr=0;
       uint32_t scidac_csumb_thr=0;
       uint32_t site_crc=0;
 
-#pragma omp for
+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);
@@ -181,7 +184,7 @@ class BinaryIO {
 	scidac_csumb_thr ^= site_crc<<gsite31 | site_crc>>(32-gsite31);
       }
 
-#pragma omp critical
+PARALLEL_CRITICAL
       {
 	scidac_csuma^= scidac_csuma_thr;
 	scidac_csumb^= scidac_csumb_thr;
@@ -261,7 +264,7 @@ class BinaryIO {
 			      GridBase *grid,
 			      std::vector<fobj> &iodata,
 			      std::string file,
-			      Integer offset,
+			      uint64_t& offset,
 			      const std::string &format, int control,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
@@ -368,7 +371,7 @@ class BinaryIO {
 #endif
       } else {
 	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
+                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
         std::ifstream fin;
 	fin.open(file, std::ios::binary | std::ios::in);
         if (control & BINARYIO_MASTER_APPEND)
@@ -429,14 +432,20 @@ class BinaryIO {
           MPI_Abort(MPI_COMM_WORLD, 1); //assert(ierr == 0);
         }
 
-        std::cout << GridLogDebug << "MPI read I/O set view " << file << std::endl;
+        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 read I/O write all " << file << std::endl;
+        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);
@@ -446,16 +455,20 @@ class BinaryIO {
       } else { 
 
         std::cout << GridLogMessage << "IOobject: C++ write I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
+                  << 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;
+	  //	  std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
 #ifdef USE_MPI_IO
 	  MPI_Abort(MPI_COMM_WORLD,1);
 #else
@@ -489,6 +502,7 @@ class BinaryIO {
 	  exit(1);
 #endif
 	}
+  offset  = fout.tellp();
 	fout.close();
       }
       timer.Stop();
@@ -523,7 +537,7 @@ class BinaryIO {
   static inline void readLatticeObject(Lattice<vobj> &Umu,
 				       std::string file,
 				       munger munge,
-				       Integer offset,
+				       uint64_t offset,
 				       const std::string &format,
 				       uint32_t &nersc_csum,
 				       uint32_t &scidac_csuma,
@@ -533,7 +547,7 @@ class BinaryIO {
     typedef typename vobj::Realified::scalar_type word;    word w=0;
 
     GridBase *grid = Umu._grid;
-    int lsites = grid->lSites();
+    uint64_t lsites = grid->lSites();
 
     std::vector<sobj> scalardata(lsites); 
     std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
@@ -544,7 +558,7 @@ class BinaryIO {
     GridStopWatch timer; 
     timer.Start();
 
-    parallel_for(int x=0;x<lsites;x++) munge(iodata[x], scalardata[x]);
+    parallel_for(uint64_t x=0;x<lsites;x++) munge(iodata[x], scalardata[x]);
 
     vectorizeFromLexOrdArray(scalardata,Umu);    
     grid->Barrier();
@@ -560,7 +574,7 @@ class BinaryIO {
     static inline void writeLatticeObject(Lattice<vobj> &Umu,
 					  std::string file,
 					  munger munge,
-					  Integer offset,
+					  uint64_t offset,
 					  const std::string &format,
 					  uint32_t &nersc_csum,
 					  uint32_t &scidac_csuma,
@@ -569,7 +583,9 @@ class BinaryIO {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
     GridBase *grid = Umu._grid;
-    int lsites = grid->lSites();
+    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
@@ -580,13 +596,39 @@ class BinaryIO {
     GridStopWatch timer; timer.Start();
     unvectorizeToLexOrdArray(scalardata,Umu);    
 
-    parallel_for(int x=0;x<lsites;x++) munge(scalardata[x],iodata[x]);
+    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;
+        }
+        else
+        {
+          std::cout << GridLogMessage << "writeLatticeObject: read test checksum correct" << std::endl;
+          break;
+        }
+      }
+      attemptsLeft--;
+    }
+    
 
     std::cout<<GridLogMessage<<"writeLatticeObject: unvectorize overhead "<<timer.Elapsed()  <<std::endl;
   }
@@ -597,7 +639,7 @@ class BinaryIO {
   static inline void readRNG(GridSerialRNG &serial,
 			     GridParallelRNG &parallel,
 			     std::string file,
-			     Integer offset,
+			     uint64_t offset,
 			     uint32_t &nersc_csum,
 			     uint32_t &scidac_csuma,
 			     uint32_t &scidac_csumb)
@@ -610,8 +652,8 @@ class BinaryIO {
     std::string format = "IEEE32BIG";
 
     GridBase *grid = parallel._grid;
-    int gsites = grid->gSites();
+    uint64_t gsites = grid->gSites();
-    int lsites = grid->lSites();
+    uint64_t lsites = grid->lSites();
 
     uint32_t nersc_csum_tmp   = 0;
     uint32_t scidac_csuma_tmp = 0;
@@ -626,7 +668,7 @@ class BinaryIO {
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     timer.Start();
-    parallel_for(int lidx=0;lidx<lsites;lidx++){
+    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);
@@ -659,7 +701,7 @@ class BinaryIO {
   static inline void writeRNG(GridSerialRNG &serial,
 			      GridParallelRNG &parallel,
 			      std::string file,
-			      Integer offset,
+			      uint64_t offset,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
 			      uint32_t &scidac_csumb)
@@ -670,8 +712,8 @@ class BinaryIO {
     typedef std::array<RngStateType,RngStateCount> RNGstate;
 
     GridBase *grid = parallel._grid;
-    int gsites = grid->gSites();
+    uint64_t gsites = grid->gSites();
-    int lsites = grid->lSites();
+    uint64_t lsites = grid->lSites();
 
     uint32_t nersc_csum_tmp;
     uint32_t scidac_csuma_tmp;
@@ -684,7 +726,7 @@ class BinaryIO {
 
     timer.Start();
     std::vector<RNGstate> iodata(lsites);
-    parallel_for(int lidx=0;lidx<lsites;lidx++){
+    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());
@@ -693,7 +735,6 @@ class BinaryIO {
 
     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);
@@ -713,5 +754,6 @@ class BinaryIO {
     std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
   }
 };
+
 }
 #endif

Grid/parallelIO/IldgIO.h

@@ -182,6 +182,11 @@ class GridLimeReader : public BinaryIO {
    {
      filename= _filename;
      File = fopen(filename.c_str(), "r");
+     if (File == nullptr)
+     {
+       std::cerr << "cannot open file '" << filename << "'" << std::endl;
+       abort();
+     }
      LimeR = limeCreateReader(File);
    }
    /////////////////////////////////////////////
@@ -228,7 +233,8 @@ class GridLimeReader : public BinaryIO {
 	//	std::cout << " ReadLatticeObject from offset "<<offset << std::endl;
 	BinarySimpleMunger<sobj,sobj> munge;
 	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
-
+  std::cout << GridLogMessage << "SciDAC checksum A " << std::hex << scidac_csuma << std::dec << std::endl;
+  std::cout << GridLogMessage << "SciDAC checksum B " << std::hex << scidac_csumb << std::dec << std::endl;
 	/////////////////////////////////////////////
 	// Insist checksum is next record
 	/////////////////////////////////////////////
@@ -245,10 +251,8 @@ class GridLimeReader : public BinaryIO {
   ////////////////////////////////////////////
   // Read a generic serialisable object
   ////////////////////////////////////////////
-  template<class serialisable_object>
+  void readLimeObject(std::string &xmlstring,std::string record_name)
-  void readLimeObject(serialisable_object &object,std::string object_name,std::string record_name)
   {
-    std::string xmlstring;
     // should this be a do while; can we miss a first record??
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
 
@@ -262,18 +266,29 @@ class GridLimeReader : public BinaryIO {
 	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
 	//	std::cout << GridLogMessage<< " readLimeObject matches XML " << &xmlc[0] <<std::endl;
 
-	XmlReader RD(&xmlc[0],"");
+   xmlstring = std::string(&xmlc[0]);
-	read(RD,object_name,object);
 	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 {
+class GridLimeWriter : public BinaryIO 
+{
  public:
+
    ///////////////////////////////////////////////////
    // FIXME: format for RNG? Now just binary out instead
    // FIXME: collective calls or not ?
@@ -282,17 +297,24 @@ class GridLimeWriter : public BinaryIO {
    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);
    }
   ///////////////////////////////////////////////////////
@@ -300,24 +322,22 @@ class GridLimeWriter : public BinaryIO {
   ///////////////////////////////////////////////////////
   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
   ////////////////////////////////////////////
-  template<class serialisable_object>
+  void writeLimeObject(int MB,int ME,XmlWriter &writer,std::string object_name,std::string record_name)
-  void writeLimeObject(int MB,int ME,serialisable_object &object,std::string object_name,std::string record_name)
   {
-    std::string xmlstring;
+    if ( boss_node ) {
-    {
+      std::string xmlstring = writer.docString();
-      XmlWriter WR("","");
+
-      write(WR,object_name,object);
-      xmlstring = WR.XmlString();
-    }
       //    std::cout << "WriteLimeObject" << record_name <<std::endl;
       uint64_t nbytes = xmlstring.size();
       //    std::cout << " xmlstring "<< nbytes<< " " << xmlstring <<std::endl;
@@ -329,48 +349,95 @@ class GridLimeWriter : public BinaryIO {
       err=limeWriteRecordData(&xmlstring[0], &nbytes, LimeW); assert(err>=0);
       err=limeWriterCloseRecord(LimeW);                       assert(err>=0);
       limeDestroyHeader(h);
-    //    std::cout << " File offset is now"<<ftello(File) << std::endl;
     }
-  ////////////////////////////////////////////
+  }
+
+  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)
   {
-    ////////////////////////////////////////////
-    // Create record header
-    ////////////////////////////////////////////
-    typedef typename vobj::scalar_object sobj;
-    int err;
-    uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    uint64_t PayloadSize = sizeof(sobj) * field._grid->_gsites;
-    createLimeRecordHeader(record_name, 0, 0, PayloadSize);
-
-    //    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;
-
     ////////////////////////////////////////////////////////////////////
     // 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. 
+    // 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.
     ////////////////////////////////////////////////////////////////////
-    uint64_t offset = ftello(File);
+    
-    //    std::cout << " Writing to offset "<<offset << std::endl;
+    GridBase *grid = field._grid;
+    assert(boss_node == field._grid->IsBoss() );
+
+    ////////////////////////////////////////////
+    // 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, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb);
-    //    fseek(File,0,SEEK_END);    offset = ftello(File);std::cout << " offset now "<<offset << std::endl;
-    err=limeWriterCloseRecord(LimeW);  assert(err>=0);
 
+    ///////////////////////////////////////////
+    // 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
@@ -380,26 +447,32 @@ class GridLimeWriter : public BinaryIO {
     std::stringstream streamb; streamb << std::hex << scidac_csumb;
     checksum.suma= streama.str();
     checksum.sumb= streamb.str();
-    //    std::cout << GridLogMessage<<" writing scidac checksums "<<std::hex<<scidac_csuma<<"/"<<scidac_csumb<<std::dec<<std::endl;
+    if ( boss_node ) { 
       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) 
+  void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
+                              const unsigned int recordScientificPrec = 0) 
   {
     GridBase * grid = field._grid;
 
@@ -415,9 +488,12 @@ class ScidacWriter : public GridLimeWriter {
     //////////////////////////////////////////////
     // 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));
+      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
   }
 };
@@ -485,6 +561,8 @@ class ScidacReader : public GridLimeReader {
 class IldgWriter : public ScidacWriter {
  public:
   
+  IldgWriter(bool isboss) : ScidacWriter(isboss) {};
+
   ///////////////////////////////////
   // A little helper
   ///////////////////////////////////
@@ -568,7 +646,6 @@ class IldgWriter : public ScidacWriter {
     writeLimeIldgLFN(header.ildg_lfn);                                                 // rec
     writeLimeLatticeBinaryObject(Umu,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
     //    limeDestroyWriter(LimeW);
-    fclose(File);
   }
 };
 
@@ -644,9 +721,11 @@ class IldgReader : public GridLimeReader {
 
 	//////////////////////////////////
 	// ILDG format record
+
+  std::string xmlstring(&xmlc[0]);
 	if ( !strncmp(limeReaderType(LimeR), ILDG_FORMAT,strlen(ILDG_FORMAT)) ) { 
 
-	  XmlReader RD(&xmlc[0],"");
+	  XmlReader RD(xmlstring, true, "");
 	  read(RD,"ildgFormat",ildgFormat_);
 
 	  if ( ildgFormat_.precision == 64 ) format = std::string("IEEE64BIG");
@@ -661,13 +740,13 @@ class IldgReader : public GridLimeReader {
 	}
 
 	if ( !strncmp(limeReaderType(LimeR), ILDG_DATA_LFN,strlen(ILDG_DATA_LFN)) ) {
-	  FieldMetaData_.ildg_lfn = std::string(&xmlc[0]);
+	  FieldMetaData_.ildg_lfn = xmlstring;
 	  found_ildgLFN = 1;
 	}
 
 	if ( !strncmp(limeReaderType(LimeR), GRID_FORMAT,strlen(ILDG_FORMAT)) ) { 
 
-	  XmlReader RD(&xmlc[0],"");
+	  XmlReader RD(xmlstring, true, "");
 	  read(RD,"FieldMetaData",FieldMetaData_);
 
 	  format = FieldMetaData_.floating_point;
@@ -681,18 +760,17 @@ class IldgReader : public GridLimeReader {
 	}
 
 	if ( !strncmp(limeReaderType(LimeR), SCIDAC_RECORD_XML,strlen(SCIDAC_RECORD_XML)) ) { 
-	  std::string xmls(&xmlc[0]);
 	  // is it a USQCD info field
-	  if ( xmls.find(std::string("usqcdInfo")) != std::string::npos ) { 
+	  if ( xmlstring.find(std::string("usqcdInfo")) != std::string::npos ) { 
 	    //	    std::cout << GridLogMessage<<"...found a usqcdInfo field"<<std::endl;
-	    XmlReader RD(&xmlc[0],"");
+	    XmlReader RD(xmlstring, true, "");
 	    read(RD,"usqcdInfo",usqcdInfo_);
 	    found_usqcdInfo = 1;
 	  }
 	}
 
 	if ( !strncmp(limeReaderType(LimeR), SCIDAC_CHECKSUM,strlen(SCIDAC_CHECKSUM)) ) { 
-	  XmlReader RD(&xmlc[0],"");
+	  XmlReader RD(xmlstring, true, "");
 	  read(RD,"scidacChecksum",scidacChecksum_);
 	  found_scidacChecksum = 1;
 	}

Grid/parallelIO/IldgIOtypes.h

@@ -136,8 +136,9 @@ struct scidacRecord : Serializable {
 				  int, typesize,
 				  int, datacount);
 
-  scidacRecord() { version =1.0; }
+  scidacRecord()
-
+  : version(1.0), recordtype(0), colors(0), spins(0), typesize(0), datacount(0)
+  {}
 };
 
 ////////////////////////

Grid/parallelIO/MetaData.h

@@ -81,18 +81,16 @@ namespace Grid {
 				      std::string, creation_date,
 				      std::string, archive_date,
 				      std::string, floating_point);
-      FieldMetaData(void) { 
+      // WARNING: non-initialised values might lead to twisted parallel IO
-	nd=4;
+      // issues, std::string are fine because they initliase to size 0
-	dimension.resize(4);
+      // as per C++ standard.
-	boundary.resize(4);
+      FieldMetaData(void) 
-	scidac_checksuma=0;
+      : nd(4), dimension(4,0), boundary(4, ""), data_start(0),
-	scidac_checksumb=0;
+      link_trace(0.), plaquette(0.), checksum(0),
-	checksum=0;
+      scidac_checksuma(0), scidac_checksumb(0), sequence_number(0)
-      }
+      {}
     };
 
-
-
   namespace QCD {
 
     using namespace Grid;

Grid/parallelIO/NerscIO.h

@@ -57,7 +57,7 @@ namespace Grid {
       // for the header-reader
       static inline int readHeader(std::string file,GridBase *grid,  FieldMetaData &field)
       {
-      int offset=0;
+      uint64_t offset=0;
       std::map<std::string,std::string> header;
       std::string line;
 
@@ -139,7 +139,7 @@ namespace Grid {
       typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
       GridBase *grid = Umu._grid;
-      int offset = readHeader(file,Umu._grid,header);
+      uint64_t offset = readHeader(file,Umu._grid,header);
 
       FieldMetaData clone(header);
 
@@ -236,21 +236,25 @@ namespace Grid {
 	GaugeStatistics(Umu,header);
 	MachineCharacteristics(header);
 
-	int offset;
+	uint64_t offset;
-  
-	truncate(file);
 
 	// 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
@@ -278,7 +282,7 @@ namespace Grid {
 	header.plaquette=0.0;
 	MachineCharacteristics(header);
 
-	int offset;
+	uint64_t offset;
   
 #ifdef RNG_RANLUX
 	header.floating_point = std::string("UINT64");
@@ -293,12 +297,18 @@ namespace Grid {
 	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 "
@@ -313,7 +323,7 @@ namespace Grid {
 
 	GridBase *grid = parallel._grid;
 
-	int offset = readHeader(file,grid,header);
+	uint64_t offset = readHeader(file,grid,header);
 
 	FieldMetaData clone(header);
 

Grid/perfmon/Timer.h

@@ -49,14 +49,38 @@ inline double usecond(void) {
 
 typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
-typedef  std::chrono::milliseconds          GridTime;
-typedef  std::chrono::microseconds          GridUsecs;
 
-inline std::ostream& operator<< (std::ostream & stream, const std::chrono::milliseconds & time)
+typedef  std::chrono::seconds               GridSecs;
+typedef  std::chrono::milliseconds          GridMillisecs;
+typedef  std::chrono::microseconds          GridUsecs;
+typedef  std::chrono::microseconds          GridTime;
+
+inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
-  stream << time.count()<<" ms";
+  stream << time.count()<<" s";
   return stream;
 }
+inline std::ostream& operator<< (std::ostream & stream, const GridMillisecs & now)
+{
+  GridSecs second(1);
+  auto     secs       = now/second ; 
+  auto     subseconds = now%second ;
+  auto     fill       = stream.fill();
+  stream << secs<<"."<<std::setw(3)<<std::setfill('0')<<subseconds.count()<<" s";
+  stream.fill(fill);
+  return stream;
+}
+inline std::ostream& operator<< (std::ostream & stream, const GridUsecs & now)
+{
+  GridSecs second(1);
+  auto     seconds    = now/second ; 
+  auto     subseconds = now%second ;
+  auto     fill       = stream.fill();
+  stream << seconds<<"."<<std::setw(6)<<std::setfill('0')<<subseconds.count()<<" s";
+  stream.fill(fill);
+  return stream;
+}
+
 
 class GridStopWatch {
 private:
@@ -96,6 +120,9 @@ public:
     assert(running == false);
     return (uint64_t) accumulator.count();
   }
+  bool isRunning(void){
+    return running;
+  }
 };
 
 }

Grid/pugixml/pugiconfig.hpp

@@ -1,7 +1,7 @@
 /**
- * pugixml parser - version 1.6
+ * pugixml parser - version 1.9
  * --------------------------------------------------------
- * Copyright (C) 2006-2015, by Arseny Kapoulkine (arseny.kapoulkine@gmail.com)
+ * Copyright (C) 2006-2018, by Arseny Kapoulkine (arseny.kapoulkine@gmail.com)
  * Report bugs and download new versions at http://pugixml.org/
  *
  * This library is distributed under the MIT License. See notice at the end
@@ -17,6 +17,9 @@
 // Uncomment this to enable wchar_t mode
 // #define PUGIXML_WCHAR_MODE
 
+// Uncomment this to enable compact mode
+// #define PUGIXML_COMPACT
+
 // Uncomment this to disable XPath
 // #define PUGIXML_NO_XPATH
 
@@ -46,7 +49,7 @@
 #endif
 
 /**
- * Copyright (c) 2006-2015 Arseny Kapoulkine
+ * Copyright (c) 2006-2018 Arseny Kapoulkine
  *
  * Permission is hereby granted, free of charge, to any person
  * obtaining a copy of this software and associated documentation