Reverted restart to inline in KrylovSchur

Making constructUR to use basisRotate()

Grid/GridStd.h

@@ -42,6 +42,7 @@ void GridAbort(void);
 #define ASSLOG(A) ::write(STDERR_FILENO,A,::strlen(A));
 #ifdef HAVE_EXECINFO_H
 #define GRID_ASSERT(b) if(!(b)) {					\
+    fflush(stdout); \
     ASSLOG(" GRID_ASSERT failure: ");					\
     ASSLOG(__FILE__);							\
     ASSLOG(" : ");							\

Grid/Makefile.am

@@ -54,21 +54,25 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
-extra_sources+=$(WILS_FERMION_FILES)
-extra_sources+=$(STAG_FERMION_FILES)
+if BUILD_FERMION_INSTANTIATIONS
+  extra_sources+=$(WILS_FERMION_FILES)
+  extra_sources+=$(STAG_FERMION_FILES)
 if BUILD_ZMOBIUS
-  extra_sources+=$(ZWILS_FERMION_FILES)
+    extra_sources+=$(ZWILS_FERMION_FILES)
 endif
 if BUILD_GPARITY
-  extra_sources+=$(GP_FERMION_FILES)
+    extra_sources+=$(GP_FERMION_FILES)
 endif
 if BUILD_FERMION_REPS
-  extra_sources+=$(ADJ_FERMION_FILES)
-  extra_sources+=$(TWOIND_FERMION_FILES)
+    extra_sources+=$(ADJ_FERMION_FILES)
+    extra_sources+=$(TWOIND_FERMION_FILES)
 endif
 if BUILD_SP
     extra_sources+=$(SP_FERMION_FILES)
-    extra_sources+=$(SP_TWOIND_FERMION_FILES)
+if BUILD_FERMION_REPS
+      extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
+endif
 endif
 
 lib_LIBRARIES = libGrid.a

Grid/algorithms/Algorithms.h

@@ -55,6 +55,7 @@ NAMESPACE_CHECK(approx);
 #include <Grid/algorithms/blas/MomentumProject.h>
 NAMESPACE_CHECK(deflation);
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
+//#include <Grid/algorithms/iterative/EigCG.h>
 NAMESPACE_CHECK(ConjGrad);
 #include <Grid/algorithms/iterative/BiCGSTAB.h>
 NAMESPACE_CHECK(BiCGSTAB);
@@ -75,6 +76,7 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
+#include <Grid/algorithms/iterative/SimpleLanczos.h>
 #include <Grid/algorithms/iterative/PowerMethod.h>
 #include <Grid/algorithms/iterative/AdefGeneric.h>
 #include <Grid/algorithms/iterative/AdefMrhs.h>
@@ -83,4 +85,17 @@ NAMESPACE_CHECK(PowerMethod);
 NAMESPACE_CHECK(multigrid);
 #include <Grid/algorithms/FFT.h>
 
+#include <Grid/algorithms/iterative/KrylovSchur.h>
+#include <Grid/algorithms/iterative/BlockKrylovSchur.h>
+#include <Grid/algorithms/iterative/SplitGridBlockKrylovSchur.h>
+#include <Grid/algorithms/iterative/HarmonicBlockKrylovSchur.h>
+#include <Grid/algorithms/iterative/TrueHarmonicBlockKrylovSchur.h>
+#include <Grid/algorithms/iterative/Gamma5BlockLanczos.h>
+//#include <Grid/algorithms/iterative/Gamma5ScalarLanczos.h>
+#include <Grid/algorithms/iterative/Arnoldi.h>
+#include <Grid/algorithms/iterative/LanczosBidiagonalization.h>
+#include <Grid/algorithms/iterative/RestartedLanczosBidiagonalization.h>
+//#include <Grid/algorithms/iterative/GCR.h>
+#include <Grid/algorithms/iterative/MultiSplittingPreconditionedCG.h>
+
 #endif

Grid/algorithms/FFT.h

@@ -28,6 +28,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_FFT_H_
 #define _GRID_FFT_H_
 
+#ifdef GRID_CUDA
+#include <cufft.h>
+#endif
+
+#ifdef GRID_HIP
+#include <hipfft/hipfft.h>
+#endif
+
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
 #ifdef HAVE_FFTW
 #if defined(USE_MKL) || defined(GRID_SYCL)
 #include <fftw/fftw3.h>
@@ -35,88 +44,190 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <fftw3.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
-template<class scalar> struct FFTW { };
+#ifndef FFTW_FORWARD
+#define FFTW_FORWARD (-1)
+#define FFTW_BACKWARD (+1)
+#define FFTW_ESTIMATE (0)
+#endif
 
-#ifdef HAVE_FFTW	
+template<class scalar> struct FFTW {
+};
+
+#ifdef GRID_HIP
 template<> struct FFTW<ComplexD> {
 public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef hipfftDoubleComplex FFTW_scalar;
+  typedef hipfftHandle        FFTW_plan;
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    auto rv = hipfftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,HIPFFT_Z2Z,howmany);
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    hipfftResult rv;
+    if ( sign == forward ) rv =hipfftExecZ2Z(p,in,out,HIPFFT_FORWARD);
+    else                   rv =hipfftExecZ2Z(p,in,out,HIPFFT_BACKWARD);
+    accelerator_barrier();
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    hipfftDestroy(p);
+  }
+};
+template<> struct FFTW<ComplexF> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef hipfftComplex      FFTW_scalar;
+  typedef hipfftHandle        FFTW_plan;
 
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    auto rv = hipfftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,HIPFFT_C2C,howmany);
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    hipfftResult rv;
+    if ( sign == forward ) rv =hipfftExecC2C(p,in,out,HIPFFT_FORWARD);
+    else                   rv =hipfftExecC2C(p,in,out,HIPFFT_BACKWARD);
+    accelerator_barrier();
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    hipfftDestroy(p);
+  }
+};
+#endif
+
+#ifdef GRID_CUDA
+template<> struct FFTW<ComplexD> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef cufftDoubleComplex FFTW_scalar;
+  typedef cufftHandle        FFTW_plan;
+
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    cufftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,CUFFT_Z2Z,howmany);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    if ( sign == forward ) cufftExecZ2Z(p,in,out,CUFFT_FORWARD);
+    else                   cufftExecZ2Z(p,in,out,CUFFT_INVERSE);
+    accelerator_barrier();
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    cufftDestroy(p);
+  }
+};
+template<> struct FFTW<ComplexF> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef cufftComplex FFTW_scalar;
+  typedef cufftHandle        FFTW_plan;
+
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    cufftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,CUFFT_C2C,howmany);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    if ( sign == forward ) cufftExecC2C(p,in,out,CUFFT_FORWARD);
+    else                   cufftExecC2C(p,in,out,CUFFT_INVERSE);
+    accelerator_barrier();
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    cufftDestroy(p);
+  }
+};
+#endif
+
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
+#ifdef HAVE_FFTW
+template<> struct FFTW<ComplexD> {
+public:
   typedef fftw_complex FFTW_scalar;
   typedef fftw_plan    FFTW_plan;
-
-  static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
-				      FFTW_scalar *in, const int *inembed,		
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
 				      int istride, int idist,		
-				      FFTW_scalar *out, const int *onembed,		
+				      FFTW_scalar *out, int *onembed,		
 				      int ostride, int odist,		
 				      int sign, unsigned flags) {
     return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
   }	  
     
-  static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
-    ::fftw_flops(p,add,mul,fmas);
-  }
-
-  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     ::fftw_execute_dft(p,in,out);
   }
   inline static void fftw_destroy_plan(const FFTW_plan p) {
     ::fftw_destroy_plan(p);
   }
 };
-
 template<> struct FFTW<ComplexF> {
 public:
-
   typedef fftwf_complex FFTW_scalar;
   typedef fftwf_plan    FFTW_plan;
-
-  static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
-				      FFTW_scalar *in, const int *inembed,		
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
 				      int istride, int idist,		
-				      FFTW_scalar *out, const int *onembed,		
+				      FFTW_scalar *out, int *onembed,		
 				      int ostride, int odist,		
 				      int sign, unsigned flags) {
     return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
   }	  
     
-  static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
-    ::fftwf_flops(p,add,mul,fmas);
-  }
-
-  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     ::fftwf_execute_dft(p,in,out);
   }
   inline static void fftw_destroy_plan(const FFTW_plan p) {
     ::fftwf_destroy_plan(p);
   }
 };
-
 #endif
-
-#ifndef FFTW_FORWARD
-#define FFTW_FORWARD (-1)
-#define FFTW_BACKWARD (+1)
 #endif
 
 class FFT {
 private:
     
-  GridCartesian *vgrid;
-  GridCartesian *sgrid;
-    
-  int Nd;
   double flops;
   double flops_call;
   uint64_t usec;
     
-  Coordinate dimensions;
-  Coordinate processors;
-  Coordinate processor_coor;
-    
 public:
     
   static const int forward=FFTW_FORWARD;
@@ -126,31 +237,25 @@ public:
   double MFlops(void) {return flops/usec;}
   double USec(void)   {return (double)usec;}    
 
-  FFT ( GridCartesian * grid ) :
-    vgrid(grid),
-    Nd(grid->_ndimension),
-    dimensions(grid->_fdimensions),
-    processors(grid->_processors),
-    processor_coor(grid->_processor_coor)
+  FFT ( GridCartesian * grid ) 
   {
     flops=0;
     usec =0;
-    Coordinate layout(Nd,1);
-    sgrid = new GridCartesian(dimensions,layout,processors,*grid);
   };
     
   ~FFT ( void)  {
-    delete sgrid;
+    //    delete sgrid;
   }
     
   template<class vobj>
   void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
 
-    conformable(result.Grid(),vgrid);
-    conformable(source.Grid(),vgrid);
-    Lattice<vobj> tmp(vgrid);
-    tmp = source;
-    for(int d=0;d<Nd;d++){
+    //    vgrid=result.Grid();
+    //    conformable(result.Grid(),vgrid);
+    //    conformable(source.Grid(),vgrid);
+    const int Ndim = source.Grid()->Nd();
+    Lattice<vobj> tmp = source;
+    for(int d=0;d<Ndim;d++){
       if( mask[d] ) {
 	FFT_dim(result,tmp,d,sign);
 	tmp=result;
@@ -160,62 +265,70 @@ public:
 
   template<class vobj>
   void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
-    Coordinate mask(Nd,1);
+    const int Ndim = source.Grid()->Nd();
+    Coordinate mask(Ndim,1);
     FFT_dim_mask(result,source,mask,sign);
   }
 
 
   template<class vobj>
   void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
-#ifndef HAVE_FFTW
-    std::cerr << "FFTW is not compiled but is called"<<std::endl;
-    GRID_ASSERT(0);
-#else
-    conformable(result.Grid(),vgrid);
-    conformable(source.Grid(),vgrid);
+    const int Ndim = source.Grid()->Nd();
+    GridBase *grid = source.Grid();
+    conformable(result.Grid(),source.Grid());
 
-    int L = vgrid->_ldimensions[dim];
-    int G = vgrid->_fdimensions[dim];
-      
-    Coordinate layout(Nd,1);
-    Coordinate pencil_gd(vgrid->_fdimensions);
-      
-    pencil_gd[dim] = G*processors[dim];
-      
-    // Pencil global vol LxLxGxLxL per node
-    GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid);
+    int L = grid->_ldimensions[dim];
+    int G = grid->_fdimensions[dim];
       
+    Coordinate layout(Ndim,1);
+    
     // Construct pencils
     typedef typename vobj::scalar_object sobj;
-    typedef typename sobj::scalar_type   scalar;
+    typedef typename vobj::scalar_type   scalar;
+    typedef typename vobj::scalar_type   scalar_type;
+    typedef typename vobj::vector_type   vector_type;
       
-    Lattice<sobj> pgbuf(&pencil_g);
-    autoView(pgbuf_v , pgbuf, CpuWrite);
     //std::cout << "CPU view" << std::endl;
     
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
       
     int Ncomp = sizeof(sobj)/sizeof(scalar);
-    int Nlow  = 1;
+    int64_t Nlow  = 1;
+    int64_t Nhigh = 1;
+
     for(int d=0;d<dim;d++){
-      Nlow*=vgrid->_ldimensions[d];
+      Nlow*=grid->_ldimensions[d];
     }
+    for(int d=dim+1;d<Ndim;d++){
+      Nhigh*=grid->_ldimensions[d];
+    }
+    int64_t Nperp=Nlow*Nhigh;
+    
+    deviceVector<scalar> pgbuf; // Layout is [perp][component][dim]
+    pgbuf.resize(Nperp*Ncomp*G);
+    scalar *pgbuf_v = &pgbuf[0];
       
     int rank = 1;  /* 1d transforms */
     int n[] = {G}; /* 1d transforms of length G */
-    int howmany = Ncomp;
+    int howmany = Ncomp * Nperp;
     int odist,idist,istride,ostride;
-    idist   = odist   = 1;          /* Distance between consecutive FT's */
-    istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
+    idist   = odist   = G;            /* Distance between consecutive FT's */
+    istride = ostride = 1;            /* Distance between two elements in the same FT */
     int *inembed = n, *onembed = n;
       
     scalar div;
     if ( sign == backward ) div = 1.0/G;
     else if ( sign == forward ) div = 1.0;
     else GRID_ASSERT(0);
-      
-    //std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
+
+    double t_pencil=0;
+    double t_fft   =0;
+    double t_total =-usecond();
+    //    std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
+    /*
+     *
+     */
     FFTW_plan p;
     {
       FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
@@ -229,72 +342,154 @@ public:
     }
       
     // Barrel shift and collect global pencil
-    //std::cout << GridLogPerformance<<"Making pencil" << std::endl;
-    Coordinate lcoor(Nd), gcoor(Nd);
+    //    std::cout << GridLogPerformance<<"Making pencil" << std::endl;
+    Coordinate lcoor(Ndim), gcoor(Ndim);
+    double t_copy=0;
+    double t_shift=0;
+    t_pencil = -usecond();
     result = source;
-    int pc = processor_coor[dim];
+    int pc = grid->_processor_coor[dim];
+
+    const Coordinate ldims = grid->_ldimensions;
+    const Coordinate rdims = grid->_rdimensions;
+    const Coordinate sdims = grid->_simd_layout;
+
+    Coordinate processors = grid->_processors;
+    Coordinate pgdims(Ndim);
+    pgdims[0] = G;
+    for(int d=0, dd=1;d<Ndim;d++){
+      if ( d!=dim ) pgdims[dd++] = ldims[d];
+    }
+    int64_t pgvol=1;
+    for(int d=0;d<Ndim;d++) pgvol*=pgdims[d];
+    
+    const int Nsimd = vobj::Nsimd();
     for(int p=0;p<processors[dim];p++) {
+      t_copy-=usecond();
+      autoView(r_v,result,AcceleratorRead);
+      accelerator_for(idx, grid->oSites(), vobj::Nsimd(), {
+#ifdef GRID_SIMT
       {
-	autoView(r_v,result,CpuRead);
-	autoView(p_v,pgbuf,CpuWrite);
-	thread_for(idx, sgrid->lSites(),{
-          Coordinate cbuf(Nd);
-          sobj s;
-	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,r_v,cbuf);
-	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  pokeLocalSite(s,p_v,cbuf);
-        });
+	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	Coordinate icoor;
+	Coordinate ocoor;
+	Coordinate pgcoor;
+
+	Lexicographic::CoorFromIndex(icoor,lane,sdims);
+	Lexicographic::CoorFromIndex(ocoor,idx,rdims);
+
+	pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + ((pc+p)%processors[dim])*L;
+	for(int d=0,dd=1;d<Ndim;d++){
+	  if ( d!=dim ) {
+	    pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d];
+	    dd++;
+	  }
+	}
+
+	// Map coordinates in lattice layout to FFTW index
+	int64_t pgidx;
+	Lexicographic::IndexFromCoor(pgcoor,pgidx,pgdims);
+
+	vector_type *from = (vector_type *)&r_v[idx];
+	scalar_type stmp;
+	for(int w=0;w<Ncomp;w++){
+	  int64_t pg_idx = pgidx + w*pgvol;
+	  stmp = getlane(from[w], lane);
+	  pgbuf_v[pg_idx] = stmp;
+	}
+#ifdef GRID_SIMT
       }
+#else
+      }
+#endif
+      });
+
+      t_copy+=usecond();
       if (p != processors[dim] - 1) {
-	result = Cshift(result,dim,L);
+	Lattice<vobj> temp(grid);
+	t_shift-=usecond();
+	temp = Cshift(result,dim,L); result = temp;
+	t_shift+=usecond();
       }
     }
+    t_pencil += usecond();
       
-    //std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
-    // Loop over orthog coords
-    int NN=pencil_g.lSites();
-    GridStopWatch timer;
-    timer.Start();
-    thread_for( idx,NN,{
-        Coordinate cbuf(Nd);
-	pencil_g.LocalIndexToLocalCoor(idx, cbuf);
-	if ( cbuf[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
-	  FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[idx];
-	  FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[idx];
-	  FFTW<scalar>::fftw_execute_dft(p,in,out);
-	}
-    });
-    timer.Stop();
-      
+    FFTW_scalar *in = (FFTW_scalar *)pgbuf_v;
+    FFTW_scalar *out= (FFTW_scalar *)pgbuf_v;
+    t_fft = -usecond();
+    FFTW<scalar>::fftw_execute_dft(p,in,out,sign);
+    t_fft += usecond();
+    
     // performance counting
-    double add,mul,fma;
-    FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
-    flops_call = add+mul+2.0*fma;
-    usec += timer.useconds();
-    flops+= flops_call*NN;
-      
-    //std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
-    // writing out result
+    flops_call = 5.0*howmany*G*log2(G);
+    usec = t_fft;
+    flops= flops_call;
+
+    result = Zero();
+    
+    double t_insert = -usecond();
     {
-      autoView(pgbuf_v,pgbuf,CpuRead);
-      autoView(result_v,result,CpuWrite);
-      thread_for(idx,sgrid->lSites(),{
-	Coordinate clbuf(Nd), cgbuf(Nd);
-	sobj s;
-	sgrid->LocalIndexToLocalCoor(idx,clbuf);
-	cgbuf = clbuf;
-	cgbuf[dim] = clbuf[dim]+L*pc;
-	peekLocalSite(s,pgbuf_v,cgbuf);
-	pokeLocalSite(s,result_v,clbuf);
+      autoView(r_v,result,AcceleratorWrite);
+      accelerator_for(idx,grid->oSites(),Nsimd,{
+#ifdef GRID_SIMT
+      {
+	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	Coordinate icoor(Ndim);
+	Coordinate ocoor(Ndim);
+	Coordinate pgcoor(Ndim);
+
+	Lexicographic::CoorFromIndex(icoor,lane,sdims);
+	Lexicographic::CoorFromIndex(ocoor,idx,rdims);
+
+	pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + pc*L;
+	for(int d=0,dd=1;d<Ndim;d++){
+	  if ( d!=dim ) {
+	    pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d];
+	    dd++;
+	  }
+	}
+	// Map coordinates in lattice layout to FFTW index
+	int64_t pgidx;
+	Lexicographic::IndexFromCoor(pgcoor,pgidx,pgdims);
+
+	vector_type *to = (vector_type *)&r_v[idx];
+	scalar_type stmp;
+	for(int w=0;w<Ncomp;w++){
+	  int64_t pg_idx = pgidx + w*pgvol;
+	  stmp = pgbuf_v[pg_idx];
+	  putlane(to[w], stmp, lane);
+	}
+	
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
       });
     }
+
     result = result*div;
-      
-    //std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
+
+    t_insert +=usecond();
+    
     // destroying plan
     FFTW<scalar>::fftw_destroy_plan(p);
-#endif
+
+    t_total +=usecond();
+
+    std::cout <<GridLogPerformance<< " FFT took   "<<t_total/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< " FFT pencil "<<t_pencil/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< "  of which copy "<<t_copy/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< "  of which shift"<<t_shift/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< " FFT kernels "<<t_fft/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< " FFT insert  "<<t_insert/1.0e6 <<" s" << std::endl;
+    
   }
 };
 

Grid/algorithms/blas/BatchedBlas.h

@@ -28,6 +28,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #pragma once
 
 #ifdef GRID_HIP
+#include <hip/hip_version.h>
 #include <hipblas/hipblas.h>
 #endif
 #ifdef GRID_CUDA
@@ -255,16 +256,29 @@ public:
     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasZgemmBatched(gridblasHandle,
 				   hOpA,
 				   hOpB,
 				   m,n,k,
-				   (hipblasDoubleComplex *) &alpha_p[0],
-				   (hipblasDoubleComplex **)&Amk[0], lda,
-				   (hipblasDoubleComplex **)&Bkn[0], ldb,
-				   (hipblasDoubleComplex *) &beta_p[0],
-				   (hipblasDoubleComplex **)&Cmn[0], ldc,
+				   (hipDoubleComplex *) &alpha_p[0],
+				   (hipDoubleComplex **)&Amk[0], lda,
+				   (hipDoubleComplex **)&Bkn[0], ldb,
+				   (hipDoubleComplex *) &beta_p[0],
+				   (hipDoubleComplex **)&Cmn[0], ldc,
 				   batchCount);
+#else
+    auto err = hipblasZgemmBatched(gridblasHandle,
+                                   hOpA,
+                                   hOpB,
+                                   m,n,k,
+                                   (hipblasDoubleComplex *) &alpha_p[0],
+                                   (hipblasDoubleComplex **)&Amk[0], lda,
+                                   (hipblasDoubleComplex **)&Bkn[0], ldb,
+                                   (hipblasDoubleComplex *) &beta_p[0],
+                                   (hipblasDoubleComplex **)&Cmn[0], ldc,
+                                   batchCount);
+#endif
     //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
@@ -503,17 +517,30 @@ public:
     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasCgemmBatched(gridblasHandle,
 				   hOpA,
 				   hOpB,
 				   m,n,k,
-				   (hipblasComplex *) &alpha_p[0],
-				   (hipblasComplex **)&Amk[0], lda,
-				   (hipblasComplex **)&Bkn[0], ldb,
-				   (hipblasComplex *) &beta_p[0],
-				   (hipblasComplex **)&Cmn[0], ldc,
+				   (hipComplex *) &alpha_p[0],
+				   (hipComplex **)&Amk[0], lda,
+				   (hipComplex **)&Bkn[0], ldb,
+				   (hipComplex *) &beta_p[0],
+				   (hipComplex **)&Cmn[0], ldc,
 				   batchCount);
+#else
+    auto err = hipblasCgemmBatched(gridblasHandle,
+                                   hOpA,
+                                   hOpB,
+                                   m,n,k,
+                                   (hipblasComplex *) &alpha_p[0],
+                                   (hipblasComplex **)&Amk[0], lda,
+                                   (hipblasComplex **)&Bkn[0], ldb,
+                                   (hipblasComplex *) &beta_p[0],
+                                   (hipblasComplex **)&Cmn[0], ldc,
+                                   batchCount);
 
+#endif
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1094,11 +1121,19 @@ public:
     GRID_ASSERT(info.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasZgetrfBatched(gridblasHandle,(int)n,
-				    (hipblasDoubleComplex **)&Ann[0], (int)n,
+				    (hipDoubleComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasZgetrfBatched(gridblasHandle,(int)n,
+                                    (hipblasDoubleComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1124,11 +1159,20 @@ public:
     GRID_ASSERT(info.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasCgetrfBatched(gridblasHandle,(int)n,
-				    (hipblasComplex **)&Ann[0], (int)n,
+				    (hipComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasCgetrfBatched(gridblasHandle,(int)n,
+                                    (hipblasComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1201,12 +1245,22 @@ public:
     GRID_ASSERT(Cnn.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasZgetriBatched(gridblasHandle,(int)n,
-				    (hipblasDoubleComplex **)&Ann[0], (int)n,
+				    (hipDoubleComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
-				    (hipblasDoubleComplex **)&Cnn[0], (int)n,
+				    (hipDoubleComplex **)&Cnn[0], (int)n,
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasZgetriBatched(gridblasHandle,(int)n,
+                                    (hipblasDoubleComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (hipblasDoubleComplex **)&Cnn[0], (int)n,
+                                    (int*) &info[0],
+                                    (int)batchCount);
+
+#endif
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1235,12 +1289,21 @@ public:
     GRID_ASSERT(Cnn.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasCgetriBatched(gridblasHandle,(int)n,
-				    (hipblasComplex **)&Ann[0], (int)n,
+				    (hipComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
-				    (hipblasComplex **)&Cnn[0], (int)n,
+				    (hipComplex **)&Cnn[0], (int)n,
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasCgetriBatched(gridblasHandle,(int)n,
+                                    (hipblasComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (hipblasComplex **)&Cnn[0], (int)n,
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA

Grid/algorithms/iterative/AdefMrhs.h

@@ -92,8 +92,8 @@ class TwoLevelCGmrhs
   // Vector case
   virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
   {
-    //    SolveSingleSystem(src,x);
-    SolvePrecBlockCG(src,x);
+    SolveSingleSystem(src,x);
+	// SolvePrecBlockCG(src,x);
   }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////

Grid/algorithms/iterative/Arnoldi.h

@@ -0,0 +1,433 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: ./lib/algorithms/iterative/Arnoldi.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Patrick Oare <poare@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_ARNOLDI_H
+#define GRID_ARNOLDI_H
+
+NAMESPACE_BEGIN(Grid); 
+
+//Moved to KrylovSchur
+#if 0
+/**
+<<<<<<< HEAD
+ * Options for which Ritz values to keep in implicit restart.
+ */
+enum RitzFilter {
+  EvalNormSmall,           // Keep evals with smallest norm
+  EvalNormLarge,           // Keep evals with largest norm
+  EvalReSmall,             // Keep evals with smallest real part
+  EvalReLarge              // Keep evals with largest real part
+};
+
+// Select comparison function from RitzFilter
+struct ComplexComparator
+{
+  RitzFilter f;
+  ComplexComparator (RitzFilter _f) : f(_f) {}
+  bool operator()(std::complex<double> z1, std::complex<double> z2) { 
+    switch (f) {
+      RealD tmp1, tmp2;
+      tmp1=std::abs(std::imag(z1));
+      tmp2=std::abs(std::imag(z2));
+      case EvalNormSmall:
+        return std::abs(z1) < std::abs(z2);
+      case EvalNormLarge:
+        return std::abs(z1) > std::abs(z2);
+// Terrible hack
+//        return std::abs(std::real(z1)) < std::abs(std::real(z2));
+//	if ( std::abs(std::real(z1))  >4.) tmp1 +=1.;
+//	if ( std::abs(std::real(z2))  >4.) tmp2 +=1.;
+      case EvalReSmall:
+	  return tmp1 < tmp2;
+//        return std::abs(std::imag(z1)) < std::abs(std::imag(z2));
+      case EvalReLarge:
+	  return tmp1 > tmp2;
+//        return std::abs(std::real(z1)) > std::abs(std::real(z2));
+      default:
+        assert(0);
+    }
+  }
+};
+
+=======
+>>>>>>> 68af1bba67dd62881ead5ab1e54962a5486a0791
+#endif
+
+/**
+ * Implementation of the Arnoldi algorithm.
+ */
+template<class Field> 
+class Arnoldi {
+
+  private:
+  
+    std::string cname = std::string("Arnoldi");
+    int MaxIter;   // Max iterations
+    RealD Tolerance;
+    RealD ssq;
+    RealD rtol;
+    int Nm;           // Number of basis vectors to track (equals MaxIter if no restart)
+    int Nk;           // Number of basis vectors to keep every restart (equals -1 if no restart)
+    int Nstop;       // Stop after converging Nstop eigenvectors.
+
+    LinearOperatorBase<Field> &Linop;
+    GridBase *Grid;
+
+    RealD approxLambdaMax;
+    RealD beta_k;
+    Field f;
+    std::vector<Field> basis;               // orthonormal Arnoldi basis
+    Eigen::MatrixXcd Hess;                  // Hessenberg matrix of size Nbasis (after construction)
+    Eigen::MatrixXcd Qt;                    // Transpose of basis rotation which projects out high modes.
+
+    Eigen::VectorXcd evals;                 // evals of Hess
+    Eigen::MatrixXcd littleEvecs;           // Nm x Nm evecs matrix
+    std::vector<Field> evecs;               // Vector of evec fields
+
+    RitzFilter ritzFilter;                        // how to sort evals
+
+  public:       
+
+    Arnoldi(LinearOperatorBase<Field> &_Linop, GridBase *_Grid, RealD _Tolerance, RitzFilter filter = EvalReSmall)
+      : Linop(_Linop), Grid(_Grid), Tolerance(_Tolerance), ritzFilter(filter), f(_Grid), MaxIter(-1), Nm(-1), Nk(-1), 
+          Nstop (-1), evals (0), evecs (), ssq (0.0), rtol (0.0), beta_k (0.0), approxLambdaMax (0.0)
+    {
+      f = Zero();
+    };
+
+    /**
+     * Runs the Arnoldi loop with(out) implicit restarting. For each iteration:
+     *   - Runs an Arnoldi step.
+     *   - Computes the eigensystem of the Hessenberg matrix.
+     *   - Performs implicit restarting.
+     */
+    void operator()(const Field& v0, int _maxIter, int _Nm, int _Nk, int _Nstop, bool doubleOrthog = false) {
+      MaxIter = _maxIter;
+      Nm = _Nm; Nk = _Nk;
+      Nstop = _Nstop;
+      
+      ssq = norm2(v0);
+      RealD approxLambdaMax = approxMaxEval(v0);
+      rtol = Tolerance * approxLambdaMax;
+
+      ComplexComparator compareComplex (ritzFilter);
+      std::cout << GridLogMessage << "Comparing Ritz values with: " << ritzFilter << std::endl;
+
+      int start = 1;
+      Field startVec = v0;
+      littleEvecs = Eigen::MatrixXcd::Zero(Nm, Nm);
+      for (int i = 0; i < MaxIter; i++) {
+        std::cout << GridLogMessage << "Restart Iteration " << i << std::endl;
+
+        // Perform Arnoldi steps to compute Krylov basis and Rayleigh quotient (Hess)
+        arnoldiIteration(startVec, Nm, start, doubleOrthog);
+        startVec = f;
+
+        // compute eigensystem and sort evals
+        // compute_eigensystem();
+        compute_eigensystem(Hess);
+        std::cout << GridLogMessage << "Eigenvalues after Arnoldi step: " << std::endl << evals << std::endl;
+
+        std::sort(evals.begin(), evals.end(), compareComplex);
+        std::cout << GridLogMessage << "Ritz values after sorting (first Nk preserved): " << std::endl << evals << std::endl;
+        // SU(N)::tepidConfiguration
+
+        // Implicit restart to de-weight unwanted eigenvalues
+        implicitRestart(_Nm, _Nk);      // probably can delete _Nm and _Nk from function args
+        start = Nk;
+
+        // check convergence and return if needed.
+        int Nconv = converged();
+        std::cout << GridLogMessage << "Number of evecs converged: " << Nconv << std::endl;
+        if (Nconv >= Nstop || i == MaxIter - 1) {
+          std::cout << GridLogMessage << "Converged with " << Nconv << " / " << Nstop << " eigenvectors on iteration " 
+                        << i << "." << std::endl;
+          basisRotate(evecs, Qt, 0, Nk, 0, Nk, Nm);
+          std::cout << GridLogMessage << "Eigenvalues [first " << Nconv << " converged]: " << std::endl << evals << std::endl;
+          return;
+        }
+      }      
+    }
+
+    /**
+     * Approximates the maximum eigenvalue of Linop.Op to normalize the residual and test for convergence. 
+     * 
+     * Parameters
+     * ----------
+     * Field& v0
+     *  Source field to start with. Must have non-zero norm.
+     * int MAX_ITER (default = 50)
+     *  Maximum number of iterations for power approximation. 
+     * 
+     * Returns
+     * -------
+     * RealD lamApprox
+     *  Approximation of largest eigenvalue. 
+     */
+    RealD approxMaxEval(const Field& v0, int MAX_ITER = 50) {
+      assert (norm2(v0) > 1e-8);                        // must have relatively large source norm to start
+      RealD lamApprox = 0.0;
+      RealD denom = 1.0; RealD num = 1.0;
+      Field v0cp (Grid); Field tmp (Grid);
+      v0cp = v0;
+      denom = std::sqrt(norm2(v0cp));
+      for (int i = 0; i < MAX_ITER; i++) {
+        Linop.Op(v0cp, tmp);                               // CAREFUL: do not do Op(tmp, tmp)
+        v0cp = tmp;
+        num = std::sqrt(norm2(v0cp));                      // num = |A^{n+1} v0|
+        lamApprox = num / denom;                           // lam = |A^{n+1} v0| / |A^n v0|
+        std::cout << GridLogDebug << "Approx for max eval: " << lamApprox << std::endl;
+        denom = num;                                       // denom = |A^{n} v0|
+      }
+      return lamApprox;
+    }
+
+    /**
+     * Constructs the Arnoldi basis for the Krylov space K_n(D, src). (TODO make private)
+     * 
+     * Parameters
+     * ----------
+     * v0 : Field&
+     *  Source to generate Krylov basis. 
+     * Nm : int
+     *  Final size of the basis desired. If the basis becomes complete before a basis of size Nm is constructed 
+     *  (determined by relative tolerance Tolerance), stops iteration there. 
+     * doubleOrthog : bool (default = false)
+     *  Whether to double orthogonalize the basis (for numerical cancellations) or not. 
+     * start        : int (default = 0)
+     *  If non-zero, assumes part of the Arnoldi basis has already been constructed. 
+     */
+    void arnoldiIteration(const Field& v0, int Nm, int start = 1, bool doubleOrthog = false)
+    {
+
+      ComplexD coeff;
+      Field w (Grid);           // A acting on last Krylov vector. 
+
+      if (start == 1) {       // initialize everything that we need.
+        RealD v0Norm = 1 / std::sqrt(ssq);
+        basis.push_back(v0Norm * v0);                // normalized source
+
+        Hess = Eigen::MatrixXcd::Zero(Nm, Nm);
+        f = Zero();
+      } else {
+        assert( start == basis.size() );      // should be starting at the end of basis (start = Nk)
+        Eigen::MatrixXcd HessCp = Hess;
+        Hess = Eigen::MatrixXcd::Zero(Nm, Nm);
+        Hess(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk)) = HessCp;
+      }
+
+      // Construct next Arnoldi vector by normalizing w_i = Dv_i - \sum_j v_j h_{ji}
+      for (int i = start - 1; i < Nm; i++) {
+
+        Linop.Op(basis.back(), w);
+        for (int j = 0; j < basis.size(); j++) {
+          coeff = innerProduct(basis[j], w);       // coeff = h_{ij}. Note that since {vi} is ONB it's OK to subtract it off after. 
+          Hess(j, i) = coeff;
+          w -= coeff * basis[j];
+        }
+
+        if (doubleOrthog) {
+          // TODO implement
+        }
+
+        // add w_i to the pile
+        if (i < Nm - 1) {
+          coeff = std::sqrt(norm2(w));
+          Hess(i+1, i) = coeff;
+          basis.push_back(
+            (1.0/coeff) * w
+          );
+        }
+
+        // after iterations, update f and beta_k = ||f||
+        f = w;                                // make sure f is not normalized
+        beta_k = std::sqrt(norm2(f));         // beta_k = ||f_k|| determines convergence.
+      }
+
+      std::cout << GridLogMessage << "|f|^2 after Arnoldi step = " << norm2(f) << std::endl;
+      std::cout << GridLogDebug << "Computed Hessenberg matrix = " << std::endl << Hess << std::endl;
+
+      return;
+    }
+
+    /**
+     * Approximates the eigensystem of the linear operator by computing the eigensystem of 
+     * the Hessenberg matrix. Assumes that the Hessenberg matrix has already been constructed (by 
+     * calling the operator() function).
+     * 
+     * TODO implement in parent class eventually.
+     * 
+     * Parameters
+     * ----------
+     * Eigen::MatrixXcd& S
+     *  Schur matrix (upper triangular) similar to original Rayleigh quotient.
+     */
+    void compute_eigensystem(Eigen::MatrixXcd& S)
+    {
+
+      std::cout << GridLogMessage << "Computing eigenvalues." << std::endl;
+
+      evecs.clear();
+
+      Eigen::ComplexEigenSolver<Eigen::MatrixXcd> es;
+      es.compute(S);
+      evals = es.eigenvalues();
+      littleEvecs = es.eigenvectors();
+
+      // Convert evecs to lattice fields
+      for (int k = 0; k < evals.size(); k++) {
+        Eigen::VectorXcd vec = littleEvecs.col(k);
+        Field tmp (basis[0].Grid());
+        tmp = Zero();
+        for (int j = 0; j < basis.size(); j++) {
+          tmp = tmp + vec[j] * basis[j];
+        }
+        evecs.push_back(tmp);
+      }
+
+      std::cout << GridLogMessage << "Eigenvalues: " << std::endl << evals << std::endl;
+
+    }
+
+    /**
+     * Verifies the factorization DV = V^\dag H + f e^\dag with the last-computed 
+     * V, H, f. 
+     */
+    // RealD verifyFactorization() {
+    //   int k = basis.size();         // number of basis vectors, also the size of H.
+    //   std::vector<Field> factorized (k, Zero());
+    //   Field tmp (FGrid); tmp = Zero();
+    //   for (int i = 0; i < basis.size(); i++) {
+    //     Linop.Op(basis[i], tmp);
+    //   }
+    //   // basisRotate(basis, Q, 0, Nk, 0, Nk, Nm);
+    //   // Linop.Op(, )
+    // }
+
+    /* Getters */
+    Eigen::MatrixXcd    getHessenbergMat()  { return Hess; }
+    Field               getF()              { return f; }
+    std::vector<Field>  getBasis()          { return basis; }
+    Eigen::VectorXcd    getEvals()          { return evals; }
+    std::vector<Field>  getEvecs()          { return evecs; }
+
+    /**
+     * Implements implicit restarting for Arnoldi. Assumes eigenvalues are sorted. 
+     * 
+     * Parameters
+     * ----------
+     * int _Nm
+     *  Size of basis to keep (Hessenberg is MxM).
+     * int Nk
+     *  Number of basis vectors to keep at each restart.
+     */
+    void implicitRestart(int _Nm, int _Nk) {
+      assert ( _Nk <= _Nm );
+      Nm = _Nm; Nk = _Nk;
+      int Np = Nm - Nk;       // keep Nk smallest (or largest, depends on sort function) evecs
+      
+      std::cout << GridLogMessage << "Computing QR Factorizations." << std::endl;
+
+      Eigen::MatrixXcd Q = Eigen::MatrixXcd::Identity(Nm, Nm);
+      Eigen::MatrixXcd Qi (Nm, Nm);
+      Eigen::MatrixXcd R (Nm, Nm);
+
+      for (int i = Nk; i < Nm; i++) {        // keep the first Nk eigenvalues and iterate through the last Np. Should loop Np times
+
+        // Useful debugging output
+        std::cout << GridLogDebug << "Computing QR factorization for i = " << i << std::endl;
+        std::cout << GridLogDebug << "Eval shift = " << evals[i] << std::endl;
+        std::cout << GridLogDebug << "Hess before rotation: " << Hess << std::endl;
+
+        // QR factorize 
+        Eigen::HouseholderQR<Eigen::MatrixXcd> QR (Hess - evals[i] * Eigen::MatrixXcd::Identity(Nm, Nm));
+        Qi = QR.householderQ();
+        Q = Q * Qi;
+        Hess = Qi.adjoint() * Hess * Qi;
+
+        std::cout << GridLogDebug << "Qt up to i = " << Q.transpose() << std::endl;
+
+      }
+
+      std::cout << GridLogDebug << "Hess after all rotations: " << std::endl << Hess << std::endl; 
+
+      // form Arnoldi vector f: f is normal to the basis vectors and its norm \beta is used to determine the Ritz estimate. 
+      std::complex<double> beta = Hess(Nk, Nk-1);
+      std::complex<double> sigma = Q(Nm-1, Nk-1);
+      f = basis[Nk] * beta + f * sigma;
+      RealD betak = std::sqrt(norm2(f));
+      std::cout << GridLogMessage << "|f|^2 after implicit restart = " << norm2(f) << std::endl;
+
+      // Rotate basis by Qt
+      Qt = Q.transpose();
+      basisRotate(basis, Qt, 0, Nk + 1, 0, Nm, Nm);
+
+      // rotate
+      basisRotate(evecs, Qt, 0, Nk + 1, 0, Nm, Nm);
+
+      // Truncate the basis and restart
+      basis = std::vector<Field> (basis.begin(), basis.begin() + Nk);
+      // evecs = std::vector<Field> (evecs.begin(), evecs.begin() + Nk);
+      Hess = Hess(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk));
+
+      std::cout << "evecs size: " << evecs.size() << std::endl;
+
+    }
+  
+    /**
+     * Computes the number of Arnoldi eigenvectors that have converged. An eigenvector s is considered converged 
+     * for a tolerance epsilon if 
+     *    r(s) := |\beta e_m^T s| < epsilon
+     * where beta is the norm of f_{m+1}.
+     * 
+     * Parameters
+     * ----------
+     * 
+     * Returns
+     * -------
+     * int : Number of converged eigenvectors.
+     */
+    int converged() {
+      int Nconv = 0;
+      for (int k = 0; k < evecs.size(); k++) {
+        RealD emTs = std::abs(littleEvecs(Nm - 1, k));           // e_m^T s
+        RealD ritzEstimate = beta_k * emTs;
+        // TODO should be ritzEstimate < Tolerance * lambda_max
+        std::cout << GridLogMessage << "Ritz estimate for evec " << k << " = " << ritzEstimate << std::endl;
+        if (ritzEstimate < rtol) {
+          Nconv++;
+        }
+      }
+      return Nconv;
+    }
+
+};
+    
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/BlockKrylovSchur.h

@@ -0,0 +1,878 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/BlockKrylovSchur.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_BLOCKED_KRYLOV_SCHUR_H
+#define GRID_BLOCKED_KRYLOV_SCHUR_H
+
+#include <iomanip>
+#include <numeric>
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Block (block-Arnoldi) restarted Krylov-Schur eigensolver for
+ * general non-Hermitian operators.
+ *
+ * Algorithm
+ * ---------
+ * Uses a block Arnoldi factorisation of block size Nblock:
+ *
+ *   A V_k = V_k H_k + F_k B_k^dag
+ *
+ * where
+ *   V_k   = Nm orthonormal basis vectors (stored flat in basis[])
+ *   H_k   = Nm x Nm upper block-Hessenberg Rayleigh quotient
+ *   F_k   = Nblock residual vectors (the next block beyond V_k)
+ *   B_k   = Nm x Nblock coupling matrix (non-zero only in last Nblock rows)
+ *
+ * Each block Arnoldi step applies A to each of the Nblock vectors in the
+ * current block, orthogonalises against all previous basis vectors, and
+ * reduces the residual block to upper-triangular form via Householder QR
+ * (implemented here as modified Gram-Schmidt within the block).
+ *
+ * The restart is a thick restart via the Schur decomposition of H_k:
+ *   H_k = Q^dag S Q
+ * The leading Nk Schur vectors (chosen by RitzFilter) are retained,
+ * the basis and Rayleigh quotient are truncated, and block Arnoldi continues
+ * from the (Nk/Nblock)-th block.
+ *
+ * Parameters
+ * ----------
+ * Nblock  : block size p
+ * Nm      : total Krylov dimension (must be divisible by Nblock)
+ * Nk      : total vectors to keep after each restart (must be divisible by Nblock, Nk < Nm)
+ * Nstop   : declare convergence when this many eigenpairs have converged
+ * MaxIter : maximum number of outer (restart) iterations
+ * Tolerance : relative convergence tolerance (||r|| < Tolerance * |lambda_max|)
+ */
+template<class Field>
+class BlockKrylovSchur {
+
+protected:
+  //--------------------------------------------------------------------
+  // Types
+  //--------------------------------------------------------------------
+  typedef Eigen::MatrixXcd CMat;
+  typedef Eigen::VectorXcd CVec;
+
+  //--------------------------------------------------------------------
+  // Parameters (set by operator())
+  //--------------------------------------------------------------------
+  int   Nblock;    // block size
+  int   Nm;        // total Krylov dimension (multiple of Nblock)
+  int   Nk;        // total vectors retained after restart (multiple of Nblock)
+  int   Nstop;
+  int   MaxIter;
+  RealD Tolerance;
+
+  //--------------------------------------------------------------------
+  // Internal state
+  //--------------------------------------------------------------------
+  LinearOperatorBase<Field>& Linop;
+  GridBase*                  Grid_;
+  RitzFilter                 ritzFilter;
+
+  // Prefix used in log messages; derived classes override in their constructor
+  // (e.g. "HarmonicBlockKrylovSchur") so the shared code reports the right name.
+  std::string className = "BlockKrylovSchur";
+
+  // Flat storage: basis[s*Nblock + t] is the t-th vector of block s
+  // After construction: basis has Nm entries
+  std::vector<Field> basis;
+
+  // Rayleigh quotient  Nm x Nm
+  CMat H;
+
+  // Residual block: Nblock vectors (the (Nm/Nblock+1)-th block, unnormalised before
+  // QR; normalised and orthogonalised as part of block Arnoldi)
+  std::vector<Field> F;
+
+  // Coupling matrix B: Nm x Nblock.
+  // In exact arithmetic only the last Nblock rows are non-zero:
+  //   B(Nm - Nblock + t, s) = H_{Nm/Nblock+1, Nm/Nblock}(t, s)   (the subdiagonal block)
+  // We keep it as a full matrix for generality after restarts.
+  CMat B;
+
+  RealD beta_k;          // Frobenius norm of the last subdiagonal block
+  RealD rtol;            // absolute tolerance = Tolerance * approxLambdaMax
+
+  // Output
+  CVec               evals;
+  CMat               littleEvecs;   // Nm columns
+  std::vector<RealD> ritzEstimates;
+
+public:
+  std::vector<Field> evecs;
+  bool doEvalCheck  = false;
+  // When true (and Nblock even), only Nblock/2 starting vectors are required.
+  // The remaining Nblock/2 slots are filled by pairing each supplied vector
+  // with its parity-flipped partner (sign negated on odd checkerboard sites).
+  bool useParityFlip = false;
+  // Same pairing behaviour but the partner is produced by gamma5Func(v).
+  // gamma5Func must be set by the caller before operator() is called.
+  // It is a std::function so that BlockKrylovSchur.h does not need to know
+  // about the Grid QCD Gamma class (which is defined at a later include layer).
+  bool useGamma5     = false;
+  std::function<void(const Field&, Field&)> gamma5Func;
+
+  //--------------------------------------------------------------------
+  // Constructor
+  //--------------------------------------------------------------------
+  BlockKrylovSchur(LinearOperatorBase<Field>& _Linop, GridBase* _Grid,
+                     RealD _Tolerance, RitzFilter _rf = EvalReSmall)
+    : Linop(_Linop), Grid_(_Grid), Tolerance(_Tolerance), ritzFilter(_rf),
+      Nblock(-1), Nm(-1), Nk(-1), Nstop(-1), MaxIter(-1),
+      beta_k(0.0), rtol(0.0)
+  {}
+
+  virtual ~BlockKrylovSchur() = default;
+
+  //--------------------------------------------------------------------
+  // Main entry point
+  //--------------------------------------------------------------------
+  /**
+   * Run the blocked Krylov-Schur algorithm.
+   *
+   * Parameters
+   * ----------
+   * v0       : block of Nblock starting vectors (size >= Nblock)
+   * _maxIter : maximum outer (restart) iterations
+   * _Nm      : total Krylov dimension (must be divisible by _Nblock)
+   * _Nk      : total vectors to keep after restart (must be divisible by _Nblock, _Nk < _Nm)
+   * _Nstop   : stop after _Nstop eigenvalues converged
+   * _Nblock  : block size
+   */
+  virtual void operator()(const std::vector<Field>& v0, int _maxIter, int _Nm, int _Nk,
+                          int _Nstop, int _Nblock = 1, bool doubleOrthog = true,
+                          bool doVerify = false)
+  {
+    MaxIter = _maxIter;
+    Nm      = _Nm;
+    Nk      = _Nk;
+    Nstop   = _Nstop;
+    Nblock  = _Nblock;
+
+    {
+      int divisor = 1;
+      if (useParityFlip) divisor *= 2;
+      if (useGamma5)     divisor *= 2;
+      assert(Nblock % divisor == 0 && (int)v0.size() >= Nblock / divisor);
+      std::cout << GridLogMessage << className << ": divisor = " << divisor << std::endl;
+    }
+    assert(Nm % Nblock == 0);
+    assert(Nk % Nblock == 0);
+    assert(Nk < Nm);
+    if (useGamma5) assert(gamma5Func && "useGamma5: gamma5Func must be set");
+    preRun();  // hook: derived classes add parameter assertions here
+
+    int N = Nm;   // total Krylov dimension
+
+    // Approximate largest eigenvalue for tolerance normalisation
+    RealD approxLambdaMax = approxMaxEval(v0[0]);
+    rtol = Tolerance * approxLambdaMax;
+    std::cout << GridLogMessage << className << ": approx max eval = "
+              << approxLambdaMax << ", rtol = " << rtol << std::endl;
+
+    // Initialise
+    H = CMat::Zero(N, N);
+    B = CMat::Zero(N, Nblock);
+
+    int start = 0;
+    std::vector<Field> startBlock = expandStartBlock(v0);
+
+    for (int iter = 0; iter < MaxIter; iter++) {
+      std::cout << GridLogMessage << className << ": restart iteration " << iter << std::endl;
+
+      // ---- Block Arnoldi: extend from block start to block Nm/Nblock ----
+      blockArnoldiIteration(startBlock, Nm/Nblock, start, doubleOrthog);
+
+      // After first full cycle start from block Nk/Nblock
+      start = Nk/Nblock;
+
+      if (doVerify) {
+        std::string lbl = "iter " + std::to_string(iter) + " after Arnoldi";
+        verify(lbl);
+      }
+
+      // ---- Restart rotation: compute Q and the rotated+reordered Rayleigh
+      //      quotient Hnew.  Base class uses the plain Schur decomposition of
+      //      H; derived classes (e.g. HarmonicBlockKrylovSchur) override this
+      //      to target a shift.
+      CMat Q, Hnew;
+      restartRotation(Q, Hnew);
+      CMat Qt = Q.adjoint();
+
+      // Rotate Krylov basis: basis_new[i] = sum_j basis[j] * Qt(j,i)
+      std::vector<Field> basis2;
+      constructUR(basis2, basis, Qt, N);
+      basis = basis2;
+
+      // Update B and H
+      B = Q * B;
+      H = Hnew;
+
+      // ---- Truncate to Nk ----
+      int Nkeep = Nk;
+
+      CMat Htmp = H(Eigen::seqN(0, Nkeep), Eigen::seqN(0, Nkeep));
+      H = CMat::Zero(N, N);
+      H(Eigen::seqN(0, Nkeep), Eigen::seqN(0, Nkeep)) = Htmp;
+
+      std::vector<Field> basisTmp(basis.begin(), basis.begin() + Nkeep);
+      basis = basisTmp;
+
+      CMat Btmp = B(Eigen::seqN(0, Nkeep), Eigen::all);
+      B = CMat::Zero(N, Nblock);
+      B(Eigen::seqN(0, Nkeep), Eigen::all) = Btmp;
+
+      // beta_k = Frobenius norm of the effective coupling
+      beta_k = Btmp.norm();
+      std::cout << GridLogMessage << className << ": beta_k = " << beta_k << std::endl;
+
+      // Restart: the new starting block is F (the residual block from Arnoldi)
+      startBlock = F;
+
+      if (doVerify) {
+        std::string lbl = "iter " + std::to_string(iter) + " after restart+truncation";
+        verify(lbl);
+      }
+
+      // ---- Compute eigensystem of truncated H for convergence check ----
+      CMat Hk = H(Eigen::seqN(0, Nkeep), Eigen::seqN(0, Nkeep));
+      computeEigensystem(Hk, Nkeep);
+
+      int Nconv = converged(Nkeep);
+      std::cout << GridLogMessage << className << ": converged " << Nconv
+                << " / " << Nstop << std::endl;
+
+      if (Nconv >= Nstop || iter == MaxIter - 1) {
+        std::cout << GridLogMessage << className << ": done after " << iter
+                  << " restarts, " << Nconv << " converged." << std::endl;
+        std::cout << GridLogMessage << "Eigenvalues: " << evals << std::endl;
+
+        if (doEvalCheck) {
+          Field w(Grid_);
+          for (int k = 0; k < (int)evecs.size(); k++) {
+            Linop.Op(evecs[k], w);
+            ComplexD eval_est = toStdCmplx(innerProduct(evecs[k], w));
+            w -= eval_est * evecs[k];
+            RealD res = std::sqrt(norm2(w));
+            std::cout << GridLogMessage << className << ": evec[" << k << "]"
+                      << "  eval_reported = " << evals[k]
+                      << "  eval_est = " << eval_est
+                      << "  || A v - eval_est * v || = " << res << std::endl;
+          }
+        }
+
+        return;
+      }
+    }
+  }
+
+  // Accessors
+  std::vector<Field>  getEvecs()         { return evecs;         }
+  CVec                getEvals()         { return evals;         }
+  std::vector<RealD>  getRitzEstimates() { return ritzEstimates; }
+
+  // Negate the field on all odd-parity (checkerboard) sites.
+  // Uses pickCheckerboard/setCheckerboard so the SIMD layout is handled
+  // correctly and the negation runs in parallel.
+  static void parityFlippedField(const Field& v, Field& out) {
+    GridCartesian* cgrid = dynamic_cast<GridCartesian*>(v.Grid());
+    assert(cgrid != nullptr);
+    GridRedBlackCartesian rbgrid(cgrid);
+    Field veven(&rbgrid), vodd(&rbgrid);
+    pickCheckerboard(Even, veven, v);
+    pickCheckerboard(Odd,  vodd,  v);
+    vodd = -vodd;
+    setCheckerboard(out, veven);
+    setCheckerboard(out, vodd);
+  }
+
+  //--------------------------------------------------------------------
+  // Verification: print H and B, check A V = V H + F B^dag explicitly
+  //--------------------------------------------------------------------
+  /**
+   * Checks the block Arnoldi / Krylov-Schur decomposition
+   *
+   *   A V = V H + F B^dag                                      (KS)
+   *
+   * by explicit operator applications.  For each basis vector j:
+   *
+   *   w_j = A basis[j]
+   *
+   * The nBasis × nBasis matrix M of inner products is computed:
+   *
+   *   M[i, j] = <basis[i] | A basis[j]>
+   *
+   * and compared column-by-column against H.  Separately, the nBasis × Nblock
+   * residual coupling matrix R is computed:
+   *
+   *   R[j, t] = <basis[j] | F[t]> * ||F[t]||   (scaled by F-block norms)
+   *
+   * but since F is already normalised, R[j,t] = <basis[j] | F[t]>.
+   *
+   * The KS relation for column j reads:
+   *   w_j = sum_i basis[i] H[i,j]  +  sum_t F[t] B[j,t]*
+   * so the deviation in column j is
+   *   dev_j = w_j - sum_i basis[i] M[i,j]   (should be zero for exact arithmetic)
+   * augmented by the F B^dag term in the last block.
+   *
+   * Prints:
+   *   - H (current Rayleigh quotient, nBasis × nBasis)
+   *   - B (coupling matrix, nBasis × Nblock)
+   *   - M (explicit inner product matrix <V | A V>)
+   *   - max |H[i,j] - M[i,j]|  (should be O(machine epsilon))
+   *   - for each basis column j: || A v_j - V H[:,j] - F B[j,:]^* ||
+   *
+   * Parameters
+   * ----------
+   * label : string printed at the start (e.g. "after restart 2")
+   */
+  void verify(const std::string& label = "")
+  {
+    int nBasis = (int)basis.size();
+    int nF     = (int)F.size();
+
+    if (nBasis == 0) {
+      std::cout << GridLogMessage
+                << className << "::verify [" << label
+                << "]: basis is empty." << std::endl;
+      return;
+    }
+
+    // Full allocated dimension of H (= Nm); may be larger than nBasis after restart+truncation
+    int Nfull = (int)H.rows();
+
+    std::cout << GridLogMessage
+              << "======== " << className << "::verify [" << label << "] ========" << std::endl;
+    std::cout << GridLogMessage
+              << "  nBasis = " << nBasis << "  Nfull = " << Nfull
+              << "  Nblock = " << Nblock << "  nF = " << nF << std::endl;
+
+    // ---- Print H (full Nm x Nm matrix) ----
+    // After restart+truncation nBasis < Nfull; entries outside [0:nBasis,0:nBasis]
+    // should be ~0 — printing the full matrix makes this visible.
+    std::cout << GridLogMessage << "H (" << Nfull << " x " << Nfull << "):" << std::endl;
+    for (int i = 0; i < Nfull; i++) {
+      for (int j = 0; j < Nfull; j++)
+        std::cout << "  " << std::setprecision(4) << std::setw(14) << H(i, j);
+      std::cout << std::endl;
+    }
+
+    // ---- Print B (full Nm x Nblock matrix) ----
+    std::cout << GridLogMessage << "B (" << Nfull << " x " << nF << "):" << std::endl;
+    for (int i = 0; i < Nfull; i++) {
+      for (int t = 0; t < nF; t++)
+        std::cout << "  " << std::setprecision(4) << std::setw(14) << B(i, t);
+      std::cout << std::endl;
+    }
+
+    // ---- Compute M[i,j] = <basis[i] | A basis[j]> ----
+    CMat M = CMat::Zero(nBasis, nBasis);
+    Field w(Grid_);
+    for (int j = 0; j < nBasis; j++) {
+      Linop.Op(basis[j], w);
+      for (int i = 0; i < nBasis; i++)
+        M(i, j) = toStdCmplx(innerProduct(basis[i], w));
+    }
+
+    std::cout << GridLogMessage << "M = <V|AV> (" << nBasis << " x " << nBasis << "):" << std::endl;
+    for (int i = 0; i < nBasis; i++) {
+      for (int j = 0; j < nBasis; j++)
+        std::cout << "  " << std::setprecision(4) << std::setw(14) << M(i, j);
+      std::cout << std::endl;
+    }
+
+    // ---- max |H - M| (for the nBasis x nBasis Rayleigh-quotient block) ----
+    RealD maxHM = 0.0;
+    for (int i = 0; i < nBasis; i++)
+      for (int j = 0; j < nBasis; j++)
+        maxHM = std::max(maxHM, std::abs(H(i,j) - M(i,j)));
+    std::cout << GridLogMessage
+              << "  max |H[i,j] - M[i,j]| (i,j < nBasis) = " << maxHM << std::endl;
+
+    // ---- Check entries of H OUTSIDE the nBasis x nBasis block ----
+    // These live in rows i >= nBasis or columns j >= nBasis.
+    // They must be ~0 after restart+truncation (before the coupling rows are
+    // re-filled by the next blockArnoldiIteration call).
+    // During a full Arnoldi run (nBasis == Nfull) this loop is a no-op.
+    RealD maxOutside = 0.0;
+    for (int i = 0; i < Nfull; i++)
+      for (int j = 0; j < Nfull; j++)
+        if (i >= nBasis || j >= nBasis)
+          maxOutside = std::max(maxOutside, std::abs(H(i,j)));
+    if (Nfull > nBasis)
+      std::cout << GridLogMessage
+                << "  max |H[i,j]| outside [0:" << nBasis << ",0:" << nBasis
+                << "] block = " << maxOutside
+                << " (should be ~0 after restart+truncation)" << std::endl;
+
+    // ---- Check orthonormality of basis ----
+    CMat G = CMat::Zero(nBasis, nBasis);
+    for (int i = 0; i < nBasis; i++)
+      for (int j = 0; j < nBasis; j++)
+        G(i, j) = toStdCmplx(innerProduct(basis[i], basis[j]));
+    CMat Gerr = G - CMat::Identity(nBasis, nBasis);
+    std::cout << GridLogMessage
+              << "  max |<V_i|V_j> - delta_ij| = " << Gerr.cwiseAbs().maxCoeff() << std::endl;
+
+    // ---- Per-column residual: || A v_j - V H[:,j] - F B[j,:]^* || ----
+    // For each basis vector j, compute A v_j then subtract V H[:,j] and F B[j,:]^*
+    RealD maxColDev = 0.0;
+    for (int j = 0; j < nBasis; j++) {
+      Linop.Op(basis[j], w);
+
+      // subtract V H[:,j]
+      for (int i = 0; i < nBasis; i++)
+        w -= basis[i] * H(i, j);
+
+      // subtract F B[j,:]^*  (F[t] * conj(B[j,t]))
+      for (int t = 0; t < nF; t++)
+        w -= F[t] * std::conj(B(j, t));
+
+      RealD dev = std::sqrt(norm2(w));
+      std::cout << GridLogMessage
+                << "  || A v[" << j << "] - V H[:,j] - F B[j,:]* || = " << dev << std::endl;
+      maxColDev = std::max(maxColDev, dev);
+    }
+    std::cout << GridLogMessage
+              << "  max column deviation = " << maxColDev << std::endl;
+
+    // ---- Check F block orthogonality against basis ----
+    if (nF > 0) {
+      RealD maxFV = 0.0;
+      for (int t = 0; t < nF; t++)
+        for (int i = 0; i < nBasis; i++) {
+          RealD ip = std::abs(toStdCmplx(innerProduct(basis[i], F[t])));
+          maxFV = std::max(maxFV, ip);
+        }
+      std::cout << GridLogMessage
+                << "  max |<V_i | F_t>| (should be ~0) = " << maxFV << std::endl;
+    }
+
+    std::cout << GridLogMessage
+              << "======== end verify ========" << std::endl;
+  }
+
+  //--------------------------------------------------------------------
+  // Extension hooks for derived classes
+  //--------------------------------------------------------------------
+
+  /**
+   * Apply the operator (or poly-filtered operator) to the current Arnoldi
+   * block and return the results in W.
+   *
+   * Base implementation: W[t] = Linop.Op(basis[kBase + t]).
+   * Derived classes (e.g. SplitGridBlockKrylovSchur) override this to
+   * apply poly(A) via Grid_split / Grid_unsplit instead.
+   */
+  virtual void applyBlock(std::vector<Field>& W, int kBase)
+  {
+    for (int t = 0; t < Nblock; t++)
+      Linop.Op(basis[kBase + t], W[t]);
+  }
+
+  /**
+   * Called once at the start of operator(), after parameters are set.
+   * Base implementation is a no-op; derived classes override to add
+   * parameter assertions (e.g. Nblock % mrhs == 0).
+   */
+  virtual void preRun() {}
+
+  /**
+   * Compute the restart rotation for the current Rayleigh quotient H.
+   *
+   * Returns Q (unitary) and Hnew = Q H Q^dag reordered so that the wanted
+   * Nk eigenvalues occupy the leading block, ready for truncation.
+   *
+   * Base implementation: the plain Schur decomposition of H, reordered by
+   * ritzFilter.  HarmonicBlockKrylovSchur overrides this to Schur-decompose
+   * the shifted quotient (H - sigma I) so that eigenvalues nearest sigma are
+   * selected.
+   */
+  virtual void restartRotation(CMat& Q, CMat& Hnew)
+  {
+    ComplexSchurDecomposition schur(H, false, ritzFilter);
+    std::cout << GridLogMessage << className << ": Schur decomposed." << std::endl;
+    schur.schurReorder(Nk);
+    std::cout << GridLogMessage << className << ": Schur reordered." << std::endl;
+    Q    = schur.getMatrixQ();
+    Hnew = schur.getMatrixS();
+  }
+
+protected:
+
+  //--------------------------------------------------------------------
+  // Starting-block expansion (parity flip / gamma5 partners)
+  //--------------------------------------------------------------------
+  /**
+   * Expand Nblock/divisor seed vectors into the full Nblock starting block by
+   * pairing each seed with its parity-flipped and/or gamma5 partners.  Order:
+   * {v, flip(v)} then doubled by {., g5(.)} giving {v, flip(v), g5(v),
+   * g5(flip(v))} when both flags are set.  Requires Nblock, Grid_ and the
+   * flags to be set.
+   */
+  std::vector<Field> expandStartBlock(const std::vector<Field>& v0)
+  {
+    int divisor = (useParityFlip ? 2 : 1) * (useGamma5 ? 2 : 1);
+    std::vector<Field> startBlock;
+    startBlock.reserve(Nblock);
+    for (int i = 0; i < Nblock / divisor; i++) {
+      std::vector<Field> group;
+      group.push_back(v0[i]);
+      if (useParityFlip) {
+        int n = (int)group.size();
+        for (int j = 0; j < n; j++) {
+          Field fp(Grid_);
+          parityFlippedField(group[j], fp);
+          group.push_back(std::move(fp));
+        }
+      }
+      if (useGamma5) {
+        int n = (int)group.size();
+        for (int j = 0; j < n; j++) {
+          Field g5v(Grid_);
+          gamma5Func(group[j], g5v);
+          group.push_back(std::move(g5v));
+        }
+      }
+      for (auto& f : group) startBlock.push_back(std::move(f));
+    }
+    return startBlock;
+  }
+
+protected:
+
+  //--------------------------------------------------------------------
+  // Block Arnoldi iteration
+  //--------------------------------------------------------------------
+  /**
+   * Extends the block Arnoldi factorisation from block index 'start' to
+   * block index 'Nm'.
+   *
+   * On entry (start > 0): basis[0..start*Nblock-1] already set,
+   *   H[0..start*Nblock-1, 0..start*Nblock-1] already set,
+   *   B[start*Nblock-1, :] set (coupling from prior residual block).
+   *   startBlock = the normalised residual block F from the previous cycle.
+   *
+   * On entry (start == 0): initialises everything from startBlock.
+   */
+  void blockArnoldiIteration(std::vector<Field>& startBlock, int endBlock,
+                             int startIdx, bool doubleOrthog)
+  {
+    int N = Nm;
+
+    if (startIdx == 0) {
+      basis.clear();
+      F.clear();
+      H = CMat::Zero(N, N);
+      B = CMat::Zero(N, Nblock);
+
+      // Orthonormalise starting block via modified Gram-Schmidt
+      std::vector<Field> V0 = startBlock;
+      blockOrthonormalise(V0);
+      for (auto& v : V0) basis.push_back(v);
+    } else {
+      // Append residual block (startBlock = F_old) to basis.
+      // The truncated KS relation after restart is:
+      //
+      //   A V_k = V_k S_k + F_old B_old^dag          (*)
+      //
+      // where V_k = basis[0:Nkeep], S_k is stored in H[0:Nkeep,0:Nkeep],
+      // B_old = B[0:Nkeep,:], F_old = startBlock.
+      //
+      // Once F_old is appended as basis[Nkeep:Nkeep+Nblock], (*) becomes
+      // a statement about the extended H matrix:
+      //
+      //   H[Nkeep+t, j] = (B_old^dag)[t,j] = conj(B_old[j,t])
+      //                   for t=0..Nblock-1, j=0..Nkeep-1
+      //
+      // These entries are the "restart coupling rows" that connect the new
+      // block to all retained Schur vectors and must be set before Arnoldi
+      // continues, otherwise A V_k = V_k H[:,0:Nkeep] would be missing the
+      // F_old B_old^dag term for those columns.
+
+      int Nkeep = startIdx * Nblock;
+      for (auto& v : startBlock) basis.push_back(v);
+
+      // Fill restart coupling rows into H
+      for (int t = 0; t < Nblock; t++)
+        for (int j = 0; j < Nkeep; j++)
+          H(Nkeep + t, j) = std::conj(B(j, t));
+
+      // Zero out B for the retained columns now that the coupling is in H
+      for (int j = 0; j < Nkeep; j++)
+        for (int t = 0; t < Nblock; t++)
+          B(j, t) = 0.0;
+    }
+
+    // Main block Arnoldi loop
+    for (int k = startIdx; k < endBlock; k++) {
+      blockArnoldiStep(k, doubleOrthog);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  /**
+   * One block Arnoldi step: extends by one block (Nblock vectors).
+   *
+   * Computes block column k of H and the next basis block V_{k+1}.
+   *
+   * Layout of basis (flat):
+   *   basis[j*Nblock + t]  = t-th vector of j-th block,  j = 0..k
+   *
+   * After this call:
+   *   H[i, k*Nblock : (k+1)*Nblock] filled for i = 0..(k+1)*Nblock - 1
+   *   basis[k*Nblock .. (k+1)*Nblock - 1] normalised (already set on entry)
+   *   F = residual block (to become V_{k+1} after this step if k < Nm-1)
+   *
+   * If k < Nm-1, also:
+   *   H[(k+1)*Nblock : (k+2)*Nblock, k*Nblock : (k+1)*Nblock] = subdiag block (from QR of residual)
+   *   basis extended by Nblock (the normalised residual vectors)
+   */
+  void blockArnoldiStep(int k, bool doubleOrthog)
+  {
+    int kBase  = k * Nblock;          // first flat index of current block
+    int prevN  = kBase + Nblock;      // number of basis vectors so far after this step
+    int N      = Nm;
+
+    // W[t] = op(basis[kBase + t])  — dispatches to applyBlock() virtual
+    std::vector<Field> W(Nblock, Field(Grid_));
+    applyBlock(W, kBase);
+
+    // Orthogonalise W against all current basis vectors (full reorthogonalisation)
+    // H[i, kBase + t] = <basis[i] | W[t]>
+    for (int pass = 0; pass < (doubleOrthog ? 2 : 1); pass++) {
+      for (int i = 0; i < prevN; i++) {
+        for (int t = 0; t < Nblock; t++) {
+          ComplexD coeff = innerProduct(basis[i], W[t]);
+          if (pass == 0)
+            H(i, kBase + t) = toStdCmplx(coeff);
+          else
+            H(i, kBase + t) += toStdCmplx(coeff);
+          W[t] -= coeff * basis[i];
+        }
+      }
+    }
+
+    // Store residual block F
+    F = W;
+
+    if (k == Nm/Nblock - 1) {
+      // Last block: compute coupling matrix B for KS decomp.
+      //
+      // blockQR modifies F in-place (F → Q orthonormal) and returns R
+      // such that  W_orig[t] = sum_s F[s] * R[s,t]  (W_orig = F_after * R).
+      //
+      // The KS relation for column j = kBase+t requires the coefficient of F[s]
+      // to be (B†)[s,j] = conj(B[j,s]).  Matching with R[s,t]:
+      //   conj(B[kBase+t, s]) = R[s,t]  →  B[kBase+t, s] = conj(R[s,t])
+      //
+      // Equivalently the last Nblock rows of B are R^H (Hermitian conjugate of R).
+      // Note: for Nblock=1, R is scalar real positive, so this reduces to B = R. ✓
+      CMat R = blockQR(F);   // F is modified in-place to become Q; returns R
+      for (int t = 0; t < Nblock; t++)
+        for (int s = 0; s < Nblock; s++)
+          B(kBase + t, s) = std::conj(R(s, t));    // B_block = R^H
+
+      beta_k = R.norm();
+      return;
+    }
+
+    // Not last block: QR-decompose residual to get V_{k+1}
+    CMat R = blockQR(F);   // F orthonormalised in-place, R is upper triangular
+
+    // Subdiagonal block of H: H[(k+1)*Nblock : (k+2)*Nblock, kBase : kBase+Nblock] = R
+    int nextBase = (k + 1) * Nblock;
+    for (int i = 0; i < Nblock; i++)
+      for (int j = 0; j < Nblock; j++)
+        H(nextBase + i, kBase + j) = R(i, j);
+
+    // Append normalised residual block to basis
+    for (int t = 0; t < Nblock; t++)
+      basis.push_back(F[t]);
+  }
+
+  //--------------------------------------------------------------------
+  // Block QR via modified Gram-Schmidt within the block
+  //--------------------------------------------------------------------
+  /**
+   * Given a block of Nblock vectors W (not necessarily orthonormal),
+   * orthonormalises them in-place and returns the upper-triangular R
+   * such that W_in = W_out * R.
+   *
+   * Handles (near-)linear dependence by zeroing vectors below threshold.
+   */
+  CMat blockQR(std::vector<Field>& W)
+  {
+    CMat R = CMat::Zero(Nblock, Nblock);
+    const RealD deflThresh = 1e-14;
+
+    for (int j = 0; j < Nblock; j++) {
+      // Orthogonalise W[j] against W[0..j-1]
+      for (int i = 0; i < j; i++) {
+        ComplexD coeff = innerProduct(W[i], W[j]);
+        R(i, j) = toStdCmplx(coeff);
+        W[j] -= coeff * W[i];
+      }
+      RealD nrm = std::sqrt(norm2(W[j]));
+      R(j, j)   = nrm;
+      if (nrm > deflThresh) {
+        W[j] *= (1.0 / nrm);
+      } else {
+        // deflation: zero this vector
+        W[j] = Zero();
+        std::cout << GridLogMessage
+                  << className << ": deflation at block column " << j
+                  << " (norm = " << nrm << ")" << std::endl;
+      }
+    }
+    return R;
+  }
+
+  //--------------------------------------------------------------------
+  // Orthonormalise a block against itself (no prior basis)
+  //--------------------------------------------------------------------
+  void blockOrthonormalise(std::vector<Field>& V)
+  {
+    for (int j = 0; j < (int)V.size(); j++) {
+      for (int i = 0; i < j; i++) {
+        ComplexD c = innerProduct(V[i], V[j]);
+        V[j] -= c * V[i];
+      }
+      RealD nrm = std::sqrt(norm2(V[j]));
+      assert(nrm > 1e-14);
+      V[j] *= (1.0 / nrm);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  // Basis rotation: UR[i] = sum_j U[j] * R(j, i)
+  //--------------------------------------------------------------------
+  void constructUR(std::vector<Field>& UR, std::vector<Field>& U,
+                   CMat& R, int N)
+  {
+    UR.clear();
+    Field tmp(Grid_);
+    for (int i = 0; i < N; i++) {
+      tmp = Zero();
+      for (int j = 0; j < N; j++)
+        tmp += U[j] * R(j, i);
+      UR.push_back(tmp);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  // Eigensystem of the truncated Rayleigh quotient
+  //--------------------------------------------------------------------
+  void computeEigensystem(CMat& Hk, int Nkeep)
+  {
+    Eigen::ComplexEigenSolver<CMat> es;
+    es.compute(Hk);
+
+    // Sort to match schurReorder ordering.
+    int n = es.eigenvalues().size();
+    ComplexComparator cComp(ritzFilter);
+    std::vector<int> idx(n);
+    std::iota(idx.begin(), idx.end(), 0);
+    std::sort(idx.begin(), idx.end(), [&](int a, int b){
+      return cComp(toStdCmplx(es.eigenvalues()(a)), toStdCmplx(es.eigenvalues()(b)));
+    });
+
+    evals.resize(n);
+    littleEvecs.resize(n, n);
+    for (int k = 0; k < n; k++) {
+      evals(k)           = es.eigenvalues()(idx[k]);
+      littleEvecs.col(k) = es.eigenvectors().col(idx[k]);
+    }
+
+    evecs.clear();
+    for (int k = 0; k < Nkeep; k++) {
+      CVec vec = littleEvecs.col(k);
+      Field tmp(Grid_);
+      tmp = Zero();
+      for (int j = 0; j < (int)basis.size() && j < Nkeep; j++)
+        tmp += vec[j] * basis[j];
+      evecs.push_back(tmp);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  // Convergence check
+  //--------------------------------------------------------------------
+  /**
+   * An eigenpair (lambda, y) is converged if the Ritz estimate
+   *   r = || B^dag y ||
+   * satisfies r < rtol.  Here B is the (Nkeep x Nblock) coupling matrix
+   * and y is the little eigenvector (Nkeep-vector) of H.
+   */
+  int converged(int Nkeep)
+  {
+    ritzEstimates.clear();
+    int Nconv = 0;
+
+    CMat Bk = B(Eigen::seqN(0, Nkeep), Eigen::all);  // Nkeep x Nblock
+
+    for (int k = 0; k < Nkeep; k++) {
+      CVec yk   = littleEvecs.col(k);                // Nkeep-vector
+      CVec Bty  = Bk.adjoint() * yk;                // Nblock-vector
+      RealD res = Bty.norm();
+      ritzEstimates.push_back(res);
+      std::cout << GridLogMessage << className << ": Ritz estimate[" << k
+                << "] = " << res << "  eval = " << evals[k] << std::endl;
+      if (res < rtol) Nconv++;
+      else break;
+    }
+    return Nconv;
+  }
+
+  //--------------------------------------------------------------------
+  // Approximate maximum eigenvalue via power iteration
+  //--------------------------------------------------------------------
+  RealD approxMaxEval(const Field& v0, int MAX_ITER = 50)
+  {
+    assert(norm2(v0) > 1e-8);
+    RealD lam = 0.0, denom = std::sqrt(norm2(v0));
+    Field vcur(Grid_), vtmp(Grid_);
+    vcur = v0;
+    for (int i = 0; i < MAX_ITER; i++) {
+      Linop.Op(vcur, vtmp);
+      vcur  = vtmp;
+      RealD num = std::sqrt(norm2(vcur));
+      lam   = num / denom;
+      denom = num;
+    }
+    return lam;
+  }
+
+};
+
+NAMESPACE_END(Grid);
+
+#endif  // GRID_BLOCKED_KRYLOV_SCHUR_H

Grid/algorithms/iterative/ConjugateGradientTimeslice.h

@@ -0,0 +1,277 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/ConjugateGradientTimeslice.h
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_TIMESLICE_H
+#define GRID_CONJUGATE_GRADIENT_TIMESLICE_H
+
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////////////////////////////////
+// Base classes for iterative processes based on operators
+// single input vec, single output vec.
+/////////////////////////////////////////////////////////////
+
+/**
+ * Simple modification of conjugate gradient that outputs the residual as a function 
+ * of time, in order to study the large wavelength behavior of the solver. 
+ */
+
+
+template <class Field>
+class ConjugateGradientTimeslice : public OperatorFunction<Field> {
+public:
+
+  using OperatorFunction<Field>::operator();
+  
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
+  RealD Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  RealD TrueResidual;
+  
+  ConjugateGradientTimeslice(RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : Tolerance(tol),
+      MaxIterations(maxit),
+      ErrorOnNoConverge(err_on_no_conv)
+  {};
+
+  virtual void LogIteration(int k,RealD a,RealD b){
+    //    std::cout << "ConjugageGradient::LogIteration() "<<std::endl;
+  };
+  virtual void LogBegin(void){
+    std::cout << "ConjugageGradient::LogBegin() "<<std::endl;
+  };
+
+    void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
+
+      this->LogBegin();
+
+      GRID_TRACE("ConjugateGradientTimeslice");
+    GridStopWatch PreambleTimer;
+    GridStopWatch ConstructTimer;
+    GridStopWatch NormTimer;
+    GridStopWatch AssignTimer;
+    PreambleTimer.Start();
+    psi.Checkerboard() = src.Checkerboard();
+
+    conformable(psi, src);
+
+    RealD cp, c, a, d, b, ssq, qq;
+    //RealD b_pred;
+
+    // Was doing copies
+    ConstructTimer.Start();
+    Field p  (src.Grid());
+    Field mmp(src.Grid());
+    Field r  (src.Grid());
+    ConstructTimer.Stop();
+
+    // Initial residual computation & set up
+    NormTimer.Start();
+    ssq = norm2(src);                 // Norm of source vector ||b||^2
+
+    ssqtx = localNorm2(src);          // Norm |b(x, t)|^2 as a field
+    std::vector<RealD> ssqt;          // Norm of source not summed over time slices, ssq(t) = \sum_x |b(x, t)|^2
+    sliceSum(ssqtx, ssqt, Tdir);      // TODO make sure Tdir is globally defined
+
+    RealD guess = norm2(psi);         // Norm of initial guess ||psi||^2
+    NormTimer.Stop();
+    assert(std::isnan(guess) == 0);
+    AssignTimer.Start();
+    if ( guess == 0.0 ) {
+      r = src;
+      p = r;
+      a = ssq;
+    } else { 
+      Linop.HermOpAndNorm(psi, mmp, d, b);        // 
+      r = src - mmp;      // Initial residual r0 = b - A guess
+      p = r;              // initial conj vector p0 = r0
+      a = norm2(p);
+    }
+    cp = a;
+    AssignTimer.Stop();
+
+    // Handle trivial case of zero src
+    if (ssq == 0.){
+      psi = Zero();
+      IterationsToComplete = 1;
+      TrueResidual = 0.;
+      return;
+    }
+
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:     p " << a << std::endl;
+
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    // Check if guess is really REALLY good :)
+    if (cp <= rsq) {
+      TrueResidual = std::sqrt(a/ssq);
+      std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
+      IterationsToComplete = 0;	
+      return;
+    }
+
+    std::cout << GridLogIterative << std::setprecision(8)
+              << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
+
+    PreambleTimer.Stop();
+    GridStopWatch LinalgTimer;
+    GridStopWatch InnerTimer;
+    GridStopWatch AxpyNormTimer;
+    GridStopWatch LinearCombTimer;
+    GridStopWatch MatrixTimer;
+    GridStopWatch SolverTimer;
+
+    RealD usecs = -usecond();
+    SolverTimer.Start();
+    int k;
+    for (k = 1; k <= MaxIterations; k++) {
+
+      GridStopWatch IterationTimer;
+      IterationTimer.Start();
+      c = cp;
+
+      MatrixTimer.Start();
+      Linop.HermOp(p, mmp);         // Computes mmp = Ap
+      MatrixTimer.Stop();
+
+      LinalgTimer.Start();
+
+      InnerTimer.Start();
+      ComplexD dc  = innerProduct(p,mmp);         // p^\dagger A p
+      InnerTimer.Stop();
+      d = dc.real();
+      a = c / d;
+
+      // What is axpy? Some accelerator or something? Check Lattice_arith.h
+      AxpyNormTimer.Start();
+
+      // axpy_norm computes ax+by for vectors x and y compatible with a GPU. Here b is set to 1 (see the function in Lattice_reduction.h). 
+      // The first argument passes r by reference, so it stores r --> -a * Ap + 1 * r, i.e. it performs an update on 
+      // r_k --> r_{k+1} = r_k - \alpha_k A p_k. The function returns the norm squared of the first variable, i.e. ||r_{k+1}||^2.
+      cp = axpy_norm(r, -a, mmp, r);
+      AxpyNormTimer.Stop();
+      b = cp / c;
+
+      LinearCombTimer.Start();
+      {
+        autoView( psi_v , psi, AcceleratorWrite);
+        autoView( p_v   , p,   AcceleratorWrite);
+        autoView( r_v   , r,   AcceleratorWrite);
+        accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+            coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+            coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+        });
+      }
+      LinearCombTimer.Stop();
+      LinalgTimer.Stop();
+      LogIteration(k,a,b);
+
+      IterationTimer.Stop();
+      if ( (k % 500) == 0 ) {
+        std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
+                << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
+      } else { 
+        std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+                << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
+      }
+
+      // Stopping condition
+      if (cp <= rsq) {
+        usecs +=usecond();
+        SolverTimer.Stop();
+        Linop.HermOpAndNorm(psi, mmp, d, qq);
+        p = mmp - src;
+        GridBase *grid = src.Grid();
+        RealD DwfFlops = (1452. )*grid->gSites()*4*k
+   	               + (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra
+        RealD srcnorm = std::sqrt(norm2(src));
+        RealD resnorm = std::sqrt(norm2(p));
+        RealD true_residual = resnorm / srcnorm;
+        std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k 
+          << "\tComputed residual " << std::sqrt(cp / ssq)
+          << "\tTrue residual " << true_residual
+          << "\tTarget " << Tolerance << std::endl;
+
+        // GridLogMessage logs the message to the terminal output; GridLogPerformance probably writes to a log file?
+        //	std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
+        std::cout << GridLogMessage << "\tSolver Elapsed    " << SolverTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "Time breakdown "<<std::endl;
+        std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+
+        std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
+
+        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+
+        IterationsToComplete = k;	
+        TrueResidual = true_residual;
+
+        return;
+      }
+    }
+    // Failed. Calculate true residual before giving up                                                         
+    // Linop.HermOpAndNorm(psi, mmp, d, qq);
+    //    p = mmp - src;
+    //TrueResidual = sqrt(norm2(p)/ssq);
+    //    TrueResidual = 1;
+
+    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations
+    	      <<" residual "<< std::sqrt(cp / ssq)<< std::endl;
+    SolverTimer.Stop();
+    std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tConstruct  " << ConstructTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tNorm       " << NormTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tAssign     " << AssignTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tSolver     " << SolverTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "Solver breakdown "<<std::endl;
+    std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage<< "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+
+    if (ErrorOnNoConverge) assert(0);
+    IterationsToComplete = k;
+
+  }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/Gamma5ScalarLanczos.h

@@ -0,0 +1,230 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    γ5-Scalar Lanczos algorithm for γ5-Hermitian operators (Wilson Dirac).
+
+    Single-vector Lanczos using the γ5-inner product (u,v)_{γ5} = u†γ5v.
+    Builds the Krylov space {q, D_W q, D_W² q, ...} and projects D_W onto it,
+    producing a real tridiagonal matrix T_m whose eigenvalues approximate those
+    of D_W.  Complex conjugate pairs of D_W eigenvalues appear as complex
+    eigenvalues of the real non-symmetric T_m.
+
+    Recurrence (G_k = sign(q_k†γ5 q_k) = ±1):
+        r_k      = D_W q_k − α_k q_k − β_sup[k-1] q_{k-1}
+        β_sub[k] = √|r_k†γ5 r_k|           (lower subdiag T_{k+1,k})
+        β_sup[k] = β_sub[k] G_{k+1}/G_k    (upper superdiag T_{k,k+1})
+        q_{k+1}  = r_k / β_sub[k]
+
+    T_m is real but not symmetric in general; eigenvalues via EigenSolver.
+
+*************************************************************************************/
+#ifndef GRID_GAMMA5_SCALAR_LANCZOS_H
+#define GRID_GAMMA5_SCALAR_LANCZOS_H
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Field>
+class Gamma5ScalarLanczos {
+public:
+  using Gamma5Func = std::function<void(const Field&, Field&)>;
+
+private:
+  typedef Eigen::MatrixXcd CMat;
+  typedef Eigen::VectorXcd CVec;
+  typedef Eigen::MatrixXd  RMat;
+
+  LinearOperatorBase<Field>& Linop;
+  GridBase*                  Grid_;
+  Gamma5Func                 applyGamma5;
+  RealD                      Tolerance;
+  int                        nSteps;
+
+  std::vector<Field>  basis;
+  std::vector<RealD>  alpha_;    // diagonal of T_m
+  std::vector<RealD>  beta_sub_; // lower subdiagonal: T_{k+1,k}
+  std::vector<RealD>  beta_sup_; // upper superdiagonal: T_{k,k+1}
+  std::vector<int>    gsign_;    // G_k = ±1
+
+  CVec               evals_;
+  std::vector<Field> evecs_;
+  std::vector<RealD> residuals_;
+
+public:
+  bool doEvalCheck = false;
+
+  Gamma5ScalarLanczos(LinearOperatorBase<Field>& op, GridBase* grid,
+                      Gamma5Func g5, RealD tol = 1e-8)
+    : Linop(op), Grid_(grid), applyGamma5(g5), Tolerance(tol), nSteps(0)
+  {}
+
+  CVec               getEvals()     { return evals_;     }
+  std::vector<Field> getEvecs()     { return evecs_;     }
+  std::vector<RealD> getResiduals() { return residuals_; }
+
+  // v0       : starting vector
+  // maxSteps : maximum Lanczos steps
+  // Nstop    : target Ritz pairs to return (0 = all)
+  // reorthog : full γ5-reorthogonalisation at each step
+  // filter   : eigenvalue selection criterion
+  void operator()(const Field& v0, int maxSteps, int Nstop,
+                  bool reorthog = false, RitzFilter filter = EvalImNormSmall)
+  {
+    basis.clear(); alpha_.clear(); beta_sub_.clear(); beta_sup_.clear(); gsign_.clear();
+    nSteps = 0;
+
+    // Normalise starting vector in γ5-norm: q_0 = v0 / √|v0†γ5 v0|
+    Field q(Grid_); q = v0;
+    Field g5q(Grid_);
+    applyGamma5(q, g5q);
+    RealD omega = std::real(toStdCmplx(innerProduct(g5q, q)));
+    RealD beta0 = std::sqrt(std::abs(omega));
+    assert(beta0 > 1e-14 && "starting vector has zero γ5-norm");
+    q *= (1.0 / beta0);
+    gsign_.push_back(omega >= 0 ? +1 : -1);
+    basis.push_back(q);
+
+    std::cout << GridLogMessage
+              << "Gamma5ScalarLanczos: start  (v0†γ5 v0)=" << omega
+              << "  G[0]=" << gsign_[0] << std::endl;
+
+    for (int k = 0; k < maxSteps; k++) {
+      // ── Apply operator ─────────────────────────────────────────────────
+      Field p(Grid_);
+      Linop.Op(basis[k], p);
+
+      // ── α_k = (q_k†γ5 D_W q_k) / G_k  (real) ─────────────────────────
+      applyGamma5(basis[k], g5q);
+      RealD a = std::real(toStdCmplx(innerProduct(g5q, p))) / RealD(gsign_[k]);
+      alpha_.push_back(a);
+
+      // ── r = D_W q_k − α_k q_k − β_sup[k-1] q_{k-1} ───────────────────
+      Field r(Grid_); r = p;
+      r -= basis[k] * a;
+      if (k > 0)
+        r -= basis[k-1] * beta_sup_[k-1];
+
+      // ── Optional γ5-reorthogonalisation ────────────────────────────────
+      if (reorthog) {
+        for (int i = 0; i <= k; i++) {
+          applyGamma5(basis[i], g5q);
+          ComplexD c = toStdCmplx(innerProduct(g5q, r)) / RealD(gsign_[i]);
+          r -= basis[i] * c;
+        }
+      }
+
+      // ── β_sub[k] = √|r†γ5 r| ───────────────────────────────────────────
+      applyGamma5(r, g5q);
+      omega = std::real(toStdCmplx(innerProduct(g5q, r)));
+      RealD beta = std::sqrt(std::abs(omega));
+
+      std::cout << GridLogMessage
+                << "Gamma5ScalarLanczos: k=" << k
+                << "  alpha=" << a
+                << "  beta=" << beta
+                << "  G[k]=" << gsign_[k]
+                << "  (r†γ5r)=" << omega << std::endl;
+
+      beta_sub_.push_back(beta);
+
+      if (beta < Tolerance) {
+        std::cout << GridLogMessage
+                  << "Gamma5ScalarLanczos: invariant subspace at step " << k << std::endl;
+        nSteps = k + 1;
+        break;
+      }
+
+      int G_next = (omega >= 0) ? +1 : -1;
+      gsign_.push_back(G_next);
+      // β_sup[k] = T_{k,k+1} = β_sub[k] * G_{k+1} / G_k
+      beta_sup_.push_back(beta * RealD(G_next) / RealD(gsign_[k]));
+
+      Field qnew(Grid_); qnew = r; qnew *= (1.0 / beta);
+      basis.push_back(qnew);
+      nSteps = k + 1;
+    }
+
+    if (nSteps == 0) return;
+    computeRitzPairs(nSteps, Nstop, filter);
+  }
+
+private:
+  void computeRitzPairs(int m, int Nstop, RitzFilter filter)
+  {
+    // Assemble real tridiagonal T_m
+    RMat Tm = RMat::Zero(m, m);
+    for (int k = 0; k < m; k++) {
+      Tm(k, k) = alpha_[k];
+      if (k + 1 < m) {
+        Tm(k+1, k) = beta_sub_[k];
+        Tm(k,   k+1) = beta_sup_[k];
+      }
+    }
+
+    std::cout << GridLogMessage
+              << "Gamma5ScalarLanczos: T_m (" << m << "×" << m << "):" << std::endl;
+    for (int i = 0; i < m; i++) {
+      for (int j = 0; j < m; j++)
+        std::cout << "  " << std::setprecision(8) << std::setw(16) << Tm(i,j);
+      std::cout << std::endl;
+    }
+
+    Eigen::EigenSolver<RMat> es(Tm);
+    CVec lambdas = es.eigenvalues();
+    CMat Y       = es.eigenvectors();
+
+    // Sort by filter criterion
+    ComplexComparator cComp(filter);
+    std::vector<int> idx(m);
+    std::iota(idx.begin(), idx.end(), 0);
+    std::sort(idx.begin(), idx.end(), [&](int a, int b){
+      return cComp(toStdCmplx(lambdas(a)), toStdCmplx(lambdas(b)));
+    });
+
+    int Nout = (Nstop > 0 && Nstop < m) ? Nstop : m;
+    evals_.resize(Nout);
+    evecs_.clear();   evecs_.reserve(Nout);
+    residuals_.clear(); residuals_.reserve(Nout);
+
+    // Trailing β for residual estimate: β_sub[m-1] * |y_j[m-1]|
+    RealD beta_trail = (m - 1 < (int)beta_sub_.size()) ? beta_sub_[m-1] : 0.0;
+
+    for (int ji = 0; ji < Nout; ji++) {
+      int  j  = idx[ji];
+      evals_(ji) = lambdas(j);
+      CVec yj = Y.col(j);
+
+      // Ritz vector: u_j = Σ_k q_k y_j[k]
+      Field uj(Grid_); uj = Zero();
+      for (int k = 0; k < m && k < (int)basis.size(); k++)
+        uj += basis[k] * toStdCmplx(yj(k));
+      evecs_.push_back(uj);
+
+      RealD res = beta_trail * std::abs(toStdCmplx(yj(m-1)));
+      if (!std::isfinite(res)) res = std::numeric_limits<RealD>::infinity();
+      residuals_.push_back(res);
+
+      std::cout << GridLogMessage
+                << "  Ritz[" << std::setw(3) << ji << "]"
+                << "  lambda=" << evals_(ji)
+                << "  |res|=" << res;
+
+      if (doEvalCheck) {
+        Field w(Grid_);
+        Linop.Op(uj, w);
+        w -= uj * evals_(ji);
+        RealD actual = std::sqrt(norm2(w));
+        std::cout << "  ||D_W u - λu||=" << actual;
+      }
+      std::cout << std::endl;
+    }
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif // GRID_GAMMA5_SCALAR_LANCZOS_H

Grid/algorithms/iterative/HarmonicBlockKrylovSchur.h

@@ -0,0 +1,719 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/HarmonicBlockKrylovSchur.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_HARMONIC_BLOCKED_KRYLOV_SCHUR_H
+#define GRID_HARMONIC_BLOCKED_KRYLOV_SCHUR_H
+
+#include <iomanip>
+#include <numeric>
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Block shift-targeted Krylov-Schur eigensolver.
+ *
+ * Algorithm
+ * ---------
+ * Uses a block Arnoldi factorisation:
+ *
+ *   A V = V H + F B^dag                                         (1)
+ *
+ * with V orthonormal (Nm columns), H the Nm×Nm block upper-Hessenberg
+ * Rayleigh quotient, F the Nblock residual vectors and B the Nm×Nblock
+ * coupling matrix.
+ *
+ * Thick restart
+ * -------------
+ * To target eigenvalues near shift σ, the Schur decomposition is computed
+ * for the shifted Rayleigh quotient:
+ *
+ *   (H - σI) = Q^dag S Q                                        (2)
+ *
+ * Sorting the Schur values of (H - σI) by smallest |S(i,i)| = |λ - σ| and
+ * retaining the leading Nk is equivalent to selecting the Ritz values of H
+ * closest to σ.  Since Q diagonalises H - σI (and hence H itself), the
+ * rotated Rayleigh quotient is exactly upper triangular:
+ *
+ *   H_new = Q H Q^dag = S + σI  (upper triangular)              (3)
+ *
+ * Truncation to Nk is therefore exact: the off-diagonal coupling block
+ * H_new[Nk:, :Nk] = 0 by triangularity.  The Krylov-Schur relation after
+ * restart is exact and block Arnoldi resumes cleanly from F.
+ *
+ * Convergence
+ * -----------
+ * Convergence is declared when || B^H y_k || < Tolerance * approxLambdaMax
+ * for each Ritz pair (λ_k, y_k) of the truncated H.
+ *
+ * Parameters
+ * ----------
+ * shift    : target shift σ (default 0.0); Schur values sorted by |λ - σ|
+ * Nblock   : block size p
+ * Nm       : total Krylov dimension (must be divisible by Nblock)
+ * Nk       : total vectors to retain after each restart (must be divisible by Nblock, Nk < Nm)
+ * Nstop    : stop when this many eigenpairs converge
+ * MaxIter  : maximum outer (restart) iterations
+ * Tolerance: relative convergence tolerance
+ *
+ * Usage
+ * -----
+ *   HarmonicBlockKrylovSchur<Field> hbks(LinOp, Grid, tol, shift, EvalNormSmall);
+ *   std::vector<Field> v0(Nblock, Field(Grid));
+ *   // fill v0 with random starting vectors
+ *   hbks(v0, maxIter, Nm, Nk, Nstop, Nblock);
+ *   auto evals = hbks.getEvals();
+ *   auto evecs = hbks.getEvecs();
+ */
+template<class Field>
+class HarmonicBlockKrylovSchur {
+
+protected:
+  typedef Eigen::MatrixXcd CMat;
+  typedef Eigen::VectorXcd CVec;
+
+  //--------------------------------------------------------------------
+  // Parameters
+  //--------------------------------------------------------------------
+  int    Nblock;
+  int    Nm;        // total Krylov dimension (multiple of Nblock)
+  int    Nk;        // total vectors retained after restart (multiple of Nblock)
+  int    Nstop;
+  int    MaxIter;
+  RealD  Tolerance;
+  ComplexD shift;       // target shift σ
+
+  //--------------------------------------------------------------------
+  // Internal state
+  //--------------------------------------------------------------------
+  LinearOperatorBase<Field>& Linop;
+  GridBase*                  Grid_;
+  RitzFilter                 ritzFilter;
+
+  std::vector<Field> basis;   // Nm*Nblock flat basis
+  CMat               H;       // (Nm*Nblock)² block-Hessenberg Rayleigh quotient
+  std::vector<Field> F;       // Nblock residual vectors
+  CMat               B;       // (Nm*Nblock) × Nblock coupling matrix
+  RealD              beta_k;
+  RealD              rtol;
+
+  CVec               evals;
+  CMat               littleEvecs;
+  std::vector<RealD> ritzEstimates;
+
+public:
+  std::vector<Field>  evecs;
+  bool doEvalCheck  = false;
+  bool useParityFlip = false;
+  bool useGamma5     = false;
+  std::function<void(const Field&, Field&)> gamma5Func;
+
+  //--------------------------------------------------------------------
+  // Constructor
+  //--------------------------------------------------------------------
+  HarmonicBlockKrylovSchur(LinearOperatorBase<Field>& _Linop, GridBase* _Grid,
+                             RealD _Tolerance, ComplexD _shift = 0.0,
+                             RitzFilter _rf = EvalNormSmall)
+    : Linop(_Linop), Grid_(_Grid), Tolerance(_Tolerance), shift(_shift),
+      ritzFilter(_rf),
+      Nblock(-1), Nm(-1), Nk(-1), Nstop(-1), MaxIter(-1),
+      beta_k(0.0), rtol(0.0)
+  {}
+
+  virtual ~HarmonicBlockKrylovSchur() = default;
+
+  //--------------------------------------------------------------------
+  // Main entry point
+  //--------------------------------------------------------------------
+  virtual void operator()(const std::vector<Field>& v0, int _maxIter, int _Nm, int _Nk,
+                          int _Nstop, int _Nblock = 1, bool doubleOrthog = true,
+                          bool doVerify = false)
+  {
+    MaxIter = _maxIter;
+    Nm      = _Nm;
+    Nk      = _Nk;
+    Nstop   = _Nstop;
+    Nblock  = _Nblock;
+
+    {
+      int divisor = 1;
+      if (useParityFlip) divisor *= 2;
+      if (useGamma5)     divisor *= 2;
+      assert(Nblock % divisor == 0 && (int)v0.size() >= Nblock / divisor);
+    std::cout << GridLogMessage << "divisor= " << divisor << std::endl;
+
+    }
+    assert(Nm % Nblock == 0);
+    assert(Nk % Nblock == 0);
+    assert(Nk < Nm);
+    if (useGamma5) assert(gamma5Func && "useGamma5: gamma5Func must be set");
+
+    int N = Nm;
+
+    RealD approxLambdaMax = approxMaxEval(v0[0]);
+    rtol = Tolerance * approxLambdaMax;
+    std::cout << GridLogMessage
+              << "HarmonicBlockKrylovSchur: approx max eval = " << approxLambdaMax
+              << ", rtol = " << rtol
+              << ", shift = " << shift << std::endl;
+
+    H = CMat::Zero(N, N);
+    B = CMat::Zero(N, Nblock);
+
+    int start = 0;
+    std::vector<Field> startBlock = expandStartBlock(v0);
+
+    for (int iter = 0; iter < MaxIter; iter++) {
+      std::cout << GridLogMessage
+                << "HarmonicBlockKrylovSchur: restart iteration " << iter << std::endl;
+
+      // ---- Block Arnoldi: extend from block 'start' to block Nm/Nblock ----
+      blockArnoldiIteration(startBlock, Nm/Nblock, start, doubleOrthog);
+      start = Nk/Nblock;
+
+      if (doVerify) {
+        std::string lbl = "iter " + std::to_string(iter) + " after Arnoldi";
+        verify(lbl);
+      }
+
+      // ---- Schur decompose (H - σI) to select Schur vectors closest to σ ----
+      // Sorting the Schur values of (H - σI) by |S(i,i)| = |λ - σ| gives the
+      // Ritz values of H nearest the target shift without any matrix inversion.
+      // Because Q diagonalises (H - σI), it also diagonalises H:
+      //   H_new = Q H Q^dag = S + σI  (upper triangular)
+      // Truncation is therefore exact.
+      CMat Hshift = H - shift * CMat::Identity(N, N);
+      ComplexSchurDecomposition schur(Hshift, false, ritzFilter);
+      schur.schurReorder(Nk);
+
+      std::cout << GridLogMessage
+                << "HarmonicBlockKrylovSchur: Ritz values nearest shift (first Nk):" << std::endl;
+      CMat S = schur.getMatrixS();
+      for (int i = 0; i < Nk; i++)
+        std::cout << GridLogMessage << "  [" << i << "] " << S(i, i) + shift << std::endl;
+
+      CMat Q  = schur.getMatrixQ();
+      CMat Qt = Q.adjoint();
+
+      // ---- Rotate Krylov basis ----
+      std::vector<Field> basis2;
+      constructUR(basis2, basis, Qt, N);
+      basis = basis2;
+
+      // ---- Update H and B ----
+      // H_new = S + σI is upper triangular; off-diagonal block H_new[Nk:,:Nk] = 0
+      H = S + shift * CMat::Identity(N, N);
+      B = Q * B;
+
+      // ---- Truncate to Nk (exact: H upper triangular) ----
+      int Nkeep = Nk;
+
+      CMat Htmp = H(Eigen::seqN(0, Nkeep), Eigen::seqN(0, Nkeep));
+      H = CMat::Zero(N, N);
+      H(Eigen::seqN(0, Nkeep), Eigen::seqN(0, Nkeep)) = Htmp;
+
+      std::vector<Field> basisTmp(basis.begin(), basis.begin() + Nkeep);
+      basis = basisTmp;
+
+      CMat Btmp = B(Eigen::seqN(0, Nkeep), Eigen::all);
+      B = CMat::Zero(N, Nblock);
+      B(Eigen::seqN(0, Nkeep), Eigen::all) = Btmp;
+
+      beta_k = Btmp.norm();
+      std::cout << GridLogMessage
+                << "HarmonicBlockKrylovSchur: beta_k = " << beta_k << std::endl;
+
+      // Restart from F (exact: no discarded-basis correction needed)
+      startBlock = F;
+
+      if (doVerify) {
+        std::string lbl = "iter " + std::to_string(iter) + " after restart+truncation";
+        verify(lbl);
+      }
+
+      // ---- Eigensystem of truncated H for convergence ----
+      CMat Hk = H(Eigen::seqN(0, Nkeep), Eigen::seqN(0, Nkeep));
+      computeEigensystem(Hk, Nkeep);
+
+      int Nconv = converged(Nkeep);
+      std::cout << GridLogMessage
+                << "HarmonicBlockKrylovSchur: converged " << Nconv
+                << " / " << Nstop << std::endl;
+
+      if (Nconv >= Nstop || iter == MaxIter - 1) {
+        std::cout << GridLogMessage
+                  << "HarmonicBlockKrylovSchur: done after " << iter
+                  << " restarts, " << Nconv << " converged." << std::endl;
+        std::cout << GridLogMessage << "Eigenvalues: " << evals.transpose() << std::endl;
+
+        if (doEvalCheck) {
+          Field w(Grid_);
+          for (int k = 0; k < (int)evecs.size(); k++) {
+            Linop.Op(evecs[k], w);
+            ComplexD eval_est = toStdCmplx(innerProduct(evecs[k], w));
+            w -= eval_est * evecs[k];
+            RealD res = std::sqrt(norm2(w));
+            std::cout << GridLogMessage << "HarmonicBlockKrylovSchur: evec[" << k << "]"
+                      << "  eval_reported = " << evals[k]
+                      << "  eval_est = " << eval_est
+                      << "  || A v - eval_est * v || = " << res << std::endl;
+          }
+        }
+
+        return;
+      }
+    }
+  }
+
+  // Accessors
+  std::vector<Field>  getEvecs()         { return evecs;         }
+  CVec                getEvals()         { return evals;         }
+  std::vector<RealD>  getRitzEstimates() { return ritzEstimates; }
+
+  static void parityFlippedField(const Field& v, Field& out) {
+    GridCartesian* cgrid = dynamic_cast<GridCartesian*>(v.Grid());
+    assert(cgrid != nullptr);
+    GridRedBlackCartesian rbgrid(cgrid);
+    Field veven(&rbgrid), vodd(&rbgrid);
+    pickCheckerboard(Even, veven, v);
+    pickCheckerboard(Odd,  vodd,  v);
+    vodd = -vodd;
+    setCheckerboard(out, veven);
+    setCheckerboard(out, vodd);
+  }
+
+  //--------------------------------------------------------------------
+  // Verification: check A V = V H + F B^dag explicitly
+  //--------------------------------------------------------------------
+  /**
+   * Checks the block Arnoldi / Krylov-Schur decomposition
+   *
+   *   A V = V H + F B^dag                                      (KS)
+   *
+   * by explicit operator applications.  H here is the standard Rayleigh
+   * quotient, so the KS relation is the same as for BlockKrylovSchur.
+   *
+   * Prints:
+   *   - H (current Rayleigh quotient, nBasis × nBasis)
+   *   - B (coupling matrix, nBasis × Nblock)
+   *   - M (explicit inner product matrix <V | A V>)
+   *   - max |H[i,j] - M[i,j]|  (should be O(machine epsilon))
+   *   - max |<V_i|V_j> - delta_ij|  (orthonormality check)
+   *   - for each basis column j: || A v_j - V H[:,j] - F B[j,:]^* ||
+   *   - max |<V_i | F_t>|  (F orthogonal to basis)
+   */
+  void verify(const std::string& label = "")
+  {
+    int nBasis = (int)basis.size();
+    int nF     = (int)F.size();
+
+    if (nBasis == 0) {
+      std::cout << GridLogMessage
+                << "HarmonicBlockKrylovSchur::verify [" << label
+                << "]: basis is empty." << std::endl;
+      return;
+    }
+
+    std::cout << GridLogMessage
+              << "======== HarmonicBlockKrylovSchur::verify [" << label << "] ========" << std::endl;
+    std::cout << GridLogMessage
+              << "  nBasis = " << nBasis << "  Nblock = " << Nblock
+              << "  nF = " << nF << std::endl;
+
+    // ---- Print H ----
+    std::cout << GridLogMessage << "H (" << nBasis << " x " << nBasis << "):" << std::endl;
+    for (int i = 0; i < nBasis; i++) {
+      for (int j = 0; j < nBasis; j++)
+        std::cout << "  " << std::setprecision(4) << std::setw(14) << H(i, j);
+      std::cout << std::endl;
+    }
+
+    // ---- Print B ----
+    std::cout << GridLogMessage << "B (" << nBasis << " x " << nF << "):" << std::endl;
+    for (int i = 0; i < nBasis; i++) {
+      for (int t = 0; t < nF; t++)
+        std::cout << "  " << std::setprecision(4) << std::setw(14) << B(i, t);
+      std::cout << std::endl;
+    }
+
+    // ---- Compute M[i,j] = <basis[i] | A basis[j]> ----
+    CMat M = CMat::Zero(nBasis, nBasis);
+    Field w(Grid_);
+    for (int j = 0; j < nBasis; j++) {
+      Linop.Op(basis[j], w);
+      for (int i = 0; i < nBasis; i++)
+        M(i, j) = toStdCmplx(innerProduct(basis[i], w));
+    }
+
+    std::cout << GridLogMessage << "M = <V|AV> (" << nBasis << " x " << nBasis << "):" << std::endl;
+    for (int i = 0; i < nBasis; i++) {
+      for (int j = 0; j < nBasis; j++)
+        std::cout << "  " << std::setprecision(4) << std::setw(14) << M(i, j);
+      std::cout << std::endl;
+    }
+
+    // ---- max |H - M| ----
+    RealD maxHM = 0.0;
+    for (int i = 0; i < nBasis; i++)
+      for (int j = 0; j < nBasis; j++)
+        maxHM = std::max(maxHM, std::abs(H(i,j) - M(i,j)));
+    std::cout << GridLogMessage
+              << "  max |H[i,j] - M[i,j]| = " << maxHM << std::endl;
+
+    // ---- Check orthonormality of basis ----
+    CMat G = CMat::Zero(nBasis, nBasis);
+    for (int i = 0; i < nBasis; i++)
+      for (int j = 0; j < nBasis; j++)
+        G(i, j) = toStdCmplx(innerProduct(basis[i], basis[j]));
+    CMat Gerr = G - CMat::Identity(nBasis, nBasis);
+    std::cout << GridLogMessage
+              << "  max |<V_i|V_j> - delta_ij| = " << Gerr.cwiseAbs().maxCoeff() << std::endl;
+
+    // ---- Per-column residual: || A v_j - V H[:,j] - F B[j,:]^* || ----
+    RealD maxColDev = 0.0;
+    for (int j = 0; j < nBasis; j++) {
+      Linop.Op(basis[j], w);
+
+      // subtract V H[:,j]
+      for (int i = 0; i < nBasis; i++)
+        w -= basis[i] * H(i, j);
+
+      // subtract F B[j,:]^*  (F[t] * conj(B[j,t]))
+      for (int t = 0; t < nF; t++)
+        w -= F[t] * std::conj(B(j, t));
+
+      RealD dev = std::sqrt(norm2(w));
+      std::cout << GridLogMessage
+                << "  || A v[" << j << "] - V H[:,j] - F B[j,:]* || = " << dev << std::endl;
+      maxColDev = std::max(maxColDev, dev);
+    }
+    std::cout << GridLogMessage
+              << "  max column deviation = " << maxColDev << std::endl;
+
+    // ---- Check F block orthogonality against basis ----
+    if (nF > 0) {
+      RealD maxFV = 0.0;
+      for (int t = 0; t < nF; t++)
+        for (int i = 0; i < nBasis; i++) {
+          RealD ip = std::abs(toStdCmplx(innerProduct(basis[i], F[t])));
+          maxFV = std::max(maxFV, ip);
+        }
+      std::cout << GridLogMessage
+                << "  max |<V_i | F_t>| (should be ~0) = " << maxFV << std::endl;
+    }
+
+    std::cout << GridLogMessage
+              << "======== end verify ========" << std::endl;
+  }
+
+protected:
+
+  //--------------------------------------------------------------------
+  // Starting-block expansion (parity flip / gamma5 partners)
+  //--------------------------------------------------------------------
+  // Expands Nblock/divisor seed vectors into the full Nblock starting
+  // block, pairing each seed with its parity-flipped and/or gamma5
+  // partners.  Requires Nblock, Grid_ and the flags to be set.
+  std::vector<Field> expandStartBlock(const std::vector<Field>& v0)
+  {
+    int divisor = (useParityFlip ? 2 : 1) * (useGamma5 ? 2 : 1);
+    std::vector<Field> startBlock;
+    startBlock.reserve(Nblock);
+    for (int i = 0; i < Nblock / divisor; i++) {
+      std::vector<Field> group;
+      group.push_back(v0[i]);
+      if (useParityFlip) {
+        int n = (int)group.size();
+        for (int j = 0; j < n; j++) {
+          Field fp(Grid_);
+          parityFlippedField(group[j], fp);
+          group.push_back(std::move(fp));
+        }
+      }
+      if (useGamma5) {
+        int n = (int)group.size();
+        for (int j = 0; j < n; j++) {
+          Field g5v(Grid_);
+          gamma5Func(group[j], g5v);
+          group.push_back(std::move(g5v));
+        }
+      }
+      for (auto& f : group) startBlock.push_back(std::move(f));
+    }
+    return startBlock;
+  }
+
+  //--------------------------------------------------------------------
+  // Block Arnoldi iteration
+  //--------------------------------------------------------------------
+  void blockArnoldiIteration(std::vector<Field>& startBlock, int endBlock,
+                             int startIdx, bool doubleOrthog)
+  {
+    int N = Nm;
+
+    if (startIdx == 0) {
+      basis.clear();
+      F.clear();
+      H = CMat::Zero(N, N);
+      B = CMat::Zero(N, Nblock);
+
+      std::vector<Field> V0 = startBlock;
+      blockOrthonormalise(V0);
+      for (auto& v : V0) basis.push_back(v);
+    } else {
+      // Append the new starting block to the retained basis.
+      //
+      // The exact truncated KS relation is  A V_k = V_k H_k + F B_k^dag,
+      // so the coupling rows are  H[Nkeep+t, j] = conj(B_k[j,t]).
+      // Since H_new = S + σI is upper triangular, the off-diagonal block
+      // H_new[Nkeep:, :Nkeep] = 0 and the restart from F is exact.
+      int Nkeep = startIdx * Nblock;
+      for (auto& v : startBlock) basis.push_back(v);
+
+      // Fill restart coupling rows into H (exact: H_new is upper triangular,
+      // so B encodes the only non-zero coupling to the new block).
+      for (int t = 0; t < Nblock; t++)
+        for (int j = 0; j < Nkeep; j++)
+          H(Nkeep + t, j) = std::conj(B(j, t));
+
+      // Zero out B for the retained columns now that the coupling is in H
+      for (int j = 0; j < Nkeep; j++)
+        for (int t = 0; t < Nblock; t++)
+          B(j, t) = 0.0;
+    }
+
+    for (int k = startIdx; k < endBlock; k++)
+      blockArnoldiStep(k, doubleOrthog);
+  }
+
+  //--------------------------------------------------------------------
+  // One block Arnoldi step
+  //--------------------------------------------------------------------
+  void blockArnoldiStep(int k, bool doubleOrthog)
+  {
+    int kBase  = k * Nblock;
+    int prevN  = kBase + Nblock;
+    int N      = Nm;
+
+    std::vector<Field> W(Nblock, Field(Grid_));
+    for (int t = 0; t < Nblock; t++)
+      Linop.Op(basis[kBase + t], W[t]);
+
+    // Full reorthogonalisation against all current basis vectors
+    for (int pass = 0; pass < (doubleOrthog ? 2 : 1); pass++) {
+      for (int i = 0; i < prevN; i++) {
+        for (int t = 0; t < Nblock; t++) {
+          ComplexD coeff = innerProduct(basis[i], W[t]);
+          if (pass == 0)
+            H(i, kBase + t) = toStdCmplx(coeff);
+          else
+            H(i, kBase + t) += toStdCmplx(coeff);
+          W[t] -= coeff * basis[i];
+        }
+      }
+    }
+
+    F = W;
+
+    if (k == Nm/Nblock - 1) {
+      // Last block: record coupling in B as R^H (Hermitian conjugate of QR factor)
+      // KS relation requires B[kBase+t, s] = conj(R[s,t])
+      CMat R = blockQR(F);
+      for (int t = 0; t < Nblock; t++)
+        for (int s = 0; s < Nblock; s++)
+          B(kBase + t, s) = std::conj(R(s, t));    // B_block = R^H
+      beta_k = R.norm();
+      return;
+    }
+
+    // Not last: QR the residual, extend basis
+    CMat R = blockQR(F);
+
+    int nextBase = (k + 1) * Nblock;
+    for (int i = 0; i < Nblock; i++)
+      for (int j = 0; j < Nblock; j++)
+        H(nextBase + i, kBase + j) = R(i, j);
+
+    for (int t = 0; t < Nblock; t++)
+      basis.push_back(F[t]);
+
+  }
+
+  //--------------------------------------------------------------------
+  // Block QR (modified Gram-Schmidt within the block)
+  //--------------------------------------------------------------------
+  CMat blockQR(std::vector<Field>& W)
+  {
+    CMat R = CMat::Zero(Nblock, Nblock);
+    const RealD deflThresh = 1e-14;
+
+    for (int j = 0; j < Nblock; j++) {
+      for (int i = 0; i < j; i++) {
+        ComplexD coeff = innerProduct(W[i], W[j]);
+        R(i, j) = toStdCmplx(coeff);
+        W[j] -= coeff * W[i];
+      }
+      RealD nrm = std::sqrt(norm2(W[j]));
+      R(j, j)   = nrm;
+      if (nrm > deflThresh) {
+        W[j] *= (1.0 / nrm);
+      } else {
+        W[j] = Zero();
+        std::cout << GridLogMessage
+                  << "HarmonicBlockKrylovSchur: deflation at block column " << j
+                  << " (norm = " << nrm << ")" << std::endl;
+      }
+    }
+    return R;
+  }
+
+  //--------------------------------------------------------------------
+  // Orthonormalise starting block
+  //--------------------------------------------------------------------
+  void blockOrthonormalise(std::vector<Field>& V)
+  {
+    for (int j = 0; j < (int)V.size(); j++) {
+      for (int i = 0; i < j; i++) {
+        ComplexD c = innerProduct(V[i], V[j]);
+        V[j] -= c * V[i];
+      }
+      RealD nrm = std::sqrt(norm2(V[j]));
+      assert(nrm > 1e-14);
+      V[j] *= (1.0 / nrm);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  // Basis rotation: UR[i] = sum_j U[j] * R(j,i)
+  //--------------------------------------------------------------------
+  void constructUR(std::vector<Field>& UR, std::vector<Field>& U,
+                   CMat& R, int N)
+  {
+    UR.clear();
+    Field tmp(Grid_);
+    for (int i = 0; i < N; i++) {
+      tmp = Zero();
+      for (int j = 0; j < N; j++)
+        tmp += U[j] * R(j, i);
+      UR.push_back(tmp);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  // Eigensystem of the truncated H (not Hhat)
+  //--------------------------------------------------------------------
+  /**
+   * Eigenvalues of H_k are the standard Ritz values in the retained
+   * subspace.  After convergence has been declared via harmonic estimates,
+   * the final reported eigenvalues and vectors come from H_k (not Hhat_k),
+   * since H_k contains the true projected operator.
+   */
+  void computeEigensystem(CMat& Hk, int Nkeep)
+  {
+    Eigen::ComplexEigenSolver<CMat> es;
+    es.compute(Hk);
+
+    // Sort to match schurReorder ordering.
+    int n = es.eigenvalues().size();
+    ComplexComparator cComp(ritzFilter);
+    std::vector<int> idx(n);
+    std::iota(idx.begin(), idx.end(), 0);
+    std::sort(idx.begin(), idx.end(), [&](int a, int b){
+      return cComp(toStdCmplx(es.eigenvalues()(a)), toStdCmplx(es.eigenvalues()(b)));
+    });
+
+    evals.resize(n);
+    littleEvecs.resize(n, n);
+    for (int k = 0; k < n; k++) {
+      evals(k)           = es.eigenvalues()(idx[k]);
+      littleEvecs.col(k) = es.eigenvectors().col(idx[k]);
+    }
+
+    evecs.clear();
+    for (int k = 0; k < Nkeep; k++) {
+      CVec vec = littleEvecs.col(k);
+      Field tmp(Grid_);
+      tmp = Zero();
+      for (int j = 0; j < (int)basis.size() && j < Nkeep; j++)
+        tmp += vec[j] * basis[j];
+      evecs.push_back(tmp);
+    }
+  }
+
+  //--------------------------------------------------------------------
+  // Convergence check
+  //--------------------------------------------------------------------
+  /**
+   * Ritz estimate for eigenpair k: || B^H y_k ||
+   * where y_k is the k-th eigenvector of the truncated H.
+   * The same bound applies whether using Ritz or harmonic Ritz restart.
+   */
+  int converged(int Nkeep)
+  {
+    ritzEstimates.clear();
+    int Nconv = 0;
+
+    CMat Bk = B(Eigen::seqN(0, Nkeep), Eigen::all);
+
+    for (int k = 0; k < Nkeep; k++) {
+      CVec yk   = littleEvecs.col(k);
+      CVec Bty  = Bk.adjoint() * yk;
+      RealD res = Bty.norm();
+      ritzEstimates.push_back(res);
+      std::cout << GridLogMessage
+                << "HarmonicBlockKrylovSchur: Ritz estimate[" << k
+                << "] = " << res << "  eval = " << evals[k] << std::endl;
+      if (res < rtol) Nconv++;
+      else break;
+    }
+    return Nconv;
+  }
+
+  //--------------------------------------------------------------------
+  // Approximate maximum eigenvalue (power iteration)
+  //--------------------------------------------------------------------
+  RealD approxMaxEval(const Field& v0, int MAX_ITER = 50)
+  {
+    assert(norm2(v0) > 1e-8);
+    RealD lam = 0.0, denom = std::sqrt(norm2(v0));
+    Field vcur(Grid_), vtmp(Grid_);
+    vcur = v0;
+    for (int i = 0; i < MAX_ITER; i++) {
+      Linop.Op(vcur, vtmp);
+      vcur  = vtmp;
+      RealD num = std::sqrt(norm2(vcur));
+      lam   = num / denom;
+      denom = num;
+    }
+    return lam;
+  }
+
+};
+
+NAMESPACE_END(Grid);
+
+#endif  // GRID_HARMONIC_BLOCKED_KRYLOV_SCHUR_H

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -53,6 +53,18 @@ enum IRLdiagonalisation {
   IRLdiagonaliseWithEigen
 };
 
+enum IRLeigsort { 
+  IRLeigsortMax,
+  IRLeigsortSqMin
+};
+
+#if 0
+bool square_comp(RealD a, RealD b){
+	if (a*a<b*b) return true;
+	return false;
+}
+#endif
+
 template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public ImplicitlyRestartedLanczosTester<Field>
 {
  public:
@@ -119,9 +131,10 @@ class ImplicitlyRestartedLanczos {
   /////////////////////////
   // Constructor
   /////////////////////////
-  
-public:       
+ public:
+  IRLeigsort EigSort;
 
+  
   //////////////////////////////////////////////////////////////////
   // PAB:
   //////////////////////////////////////////////////////////////////
@@ -154,6 +167,7 @@ public:
     Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
     eresid(_eresid),      betastp(_betastp),
     MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    EigSort(IRLeigsortMax), 
     orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
 
     ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
@@ -170,6 +184,7 @@ public:
     Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
     eresid(_eresid),      betastp(_betastp),
     MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    EigSort(IRLeigsortMax), 
     orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
 
   ////////////////////////////////
@@ -316,8 +331,12 @@ until convergence
       // sorting
       //////////////////////////////////
       eval2_copy = eval2;
+//      if (EigSort==IRLeigsortMax)
+//      std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),square_comp);
+//      else
       std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
       std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
+//      eval2_copy = eval2;
       const int chunk=8;
       for(int io=0; io<k2;io+=chunk){
 	std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
@@ -333,6 +352,7 @@ until convergence
       //////////////////////////////////
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       for(int ip=k2; ip<Nm; ++ip){ 
+//        std::cout<<GridLogIRL <<"QR decompose "<<eval2[ip]<<std::endl;
 	QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
       }
       std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
@@ -375,7 +395,8 @@ until convergence
 
 	//  power of two search pattern;  not every evalue in eval2 is assessed.
 	int allconv =1;
-	for(int jj = 1; jj<=Nstop; jj*=2){
+//	for(int jj = 1; jj<=Nstop; jj*=2){
+	for(int jj = 1; jj<=Nstop; jj++){
 	  int j = Nstop-jj;
 	  RealD e = eval2_copy[j]; // Discard the evalue
 	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    

Grid/algorithms/iterative/KrylovSchur.h

@@ -0,0 +1,994 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: ./lib/algorithms/iterative/KrylovSchur.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Patrick Oare <poare@bnl.gov>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_KRYLOVSCHUR_H
+#define GRID_KRYLOVSCHUR_H
+
+NAMESPACE_BEGIN(Grid);
+
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+using thrust::abs;
+using thrust::conj;
+inline std::complex<double> toStdCmplx(const ComplexD& c) { return {c.real(), c.imag()}; }
+#else
+using std::abs;
+using std::conj;
+inline const ComplexD& toStdCmplx(const ComplexD& c) { return c; }
+#endif
+
+/**
+ * Small dense coefficient matrix of ComplexD (thrust::complex on GPU builds)
+ * for passing to basisRotate, whose device kernel cannot multiply Grid tensors
+ * by std::complex. Must be namespace-scope: nvcc rejects function-local types
+ * as template arguments to functions containing extended device lambdas.
+ */
+struct KSCoeffMat {
+  std::vector<ComplexD> a; int n;
+  KSCoeffMat(int _n) : a(_n*_n), n(_n) {}
+  ComplexD&       operator()(int i, int j)       { return a[i*n+j]; }
+  const ComplexD& operator()(int i, int j) const { return a[i*n+j]; }
+};
+
+/**
+ * Options for which Ritz values to keep in implicit restart. TODO move this and utilities into a new file
+ */
+enum RitzFilter {
+  EvalNormSmall,           // Keep evals with smallest norm
+  EvalNormLarge,           // Keep evals with largest norm
+  EvalReSmall,             // Keep evals with smallest real part
+  EvalReLarge,             // Keep evals with largest real part
+  EvalReNormLarge,         // Keep evals with largest real part
+  EvalImSmall,             // Keep evals with smallest imaginary part
+  EvalImLarge,             // Keep evals with largest imaginary part
+  EvalImNormSmall,         // Keep evals with smallest |imaginary| part
+  EvalImNormLarge,         // Keep evals with largest |imaginary| part
+};
+
+/** Selects the RitzFilter corresponding to the input string. */
+inline RitzFilter selectRitzFilter(std::string s) {
+  if (s == "EvalNormSmall")   { return EvalNormSmall; }   else
+  if (s == "EvalNormLarge")   { return EvalNormLarge; }   else
+  if (s == "EvalReSmall")     { return EvalReSmall; }     else
+  if (s == "EvalReLarge")     { return EvalReLarge; }     else
+  if (s == "EvalReNormLarge")     { return EvalReNormLarge; }     else
+  if (s == "EvalImSmall")     { return EvalImSmall; }     else
+  if (s == "EvalImLarge")     { return EvalImLarge; }     else 
+  if (s == "EvalImNormSmall") { return EvalImNormSmall; } else
+  if (s == "EvalImNormLarge") { return EvalImNormLarge; } else 
+  { assert(0); }
+}
+
+/** Returns a string saying which RitzFilter it is. */
+inline std::string rfToString(RitzFilter RF) {
+  switch (RF) {
+    case EvalNormSmall:
+      return "EvalNormSmall";
+    case EvalNormLarge:
+      return "EvalNormLarge";
+    case EvalReSmall:
+      return "EvalReSmall";
+    case EvalReLarge:
+      return "EvalReLarge";
+    case EvalReNormLarge:
+      return "EvalReNormLarge";
+    case EvalImSmall:
+      return "EvalImSmall";
+    case EvalImLarge:
+      return "EvalImLarge";
+    case EvalImNormSmall:
+      return "EvalImNormSmall";
+    case EvalImNormLarge:
+      return "EvalImNormLarge";
+    default:
+      assert(0);
+  }
+}
+
+// Select comparison function from RitzFilter
+struct ComplexComparator
+{
+  RitzFilter RF;
+  ComplexComparator (RitzFilter _rf) : RF(_rf) {}
+  bool operator()(std::complex<double> z1, std::complex<double> z2) { 
+	  RealD tmp1=std::abs(std::imag(z1));
+	  RealD tmp2=std::abs(std::imag(z2));
+	  if ( std::abs(std::real(z1)) >2.) tmp1 += 100.;
+	  if ( std::abs(std::real(z2)) >2.) tmp2 += 100.;
+    switch (RF) {
+      case EvalNormSmall:
+        return std::abs(z1) < std::abs(z2);
+      case EvalNormLarge:
+        return std::abs(z1) > std::abs(z2);
+      case EvalReSmall:
+        return std::real(z1) < std::real(z2);     // DELETE THE ABS HERE!!!
+      case EvalReLarge:
+        return std::real(z1) > std::real(z2);
+      case EvalReNormLarge:
+        return std::abs(std::real(z1)) < std::abs(std::real(z2));
+      case EvalImSmall:
+        return std::imag(z1) < std::imag(z2);
+      case EvalImLarge:
+        return std::imag(z1) > std::imag(z2);
+      case EvalImNormSmall:
+//        return std::abs(std::imag(z1)) < std::abs(std::imag(z2));
+          return tmp1 < tmp2 ;
+      case EvalImNormLarge:
+        return std::abs(std::imag(z1)) > std::abs(std::imag(z2));
+      default:
+        assert(0);
+    }
+  }
+};
+
+/**
+ * Computes a complex Schur decomposition of a complex matrix A using Eigen's matrix library. The Schur decomposition,
+ *    A = Q^\dag S Q
+ * factorizes A into a unitary matrix Q and an upper triangular matrix S. The eigenvalues of A lie on the diagonal of the upper triangular matrix S. 
+ * The Schur decomposition is not unique: in particular, any ordering of the eigenvalues of A can be used as the diagonal of the matrix S. 
+ * This class supports eigenvalue reordering by swapping two adjacent eigenvalues with a unitary transformation. 
+ */
+class ComplexSchurDecomposition {
+
+  private:
+
+    typedef Eigen::MatrixXcd CMat;
+
+    CMat A;           // Matrix to decompose, A = Q^\dag S Q
+    CMat Q;           // Unitary matrix Q
+    CMat S;           // Upper triangular matrix S
+
+    // Placeholders for Givens rotation
+    CMat Givens;      // Givens rotation
+    ComplexD s;       // off-diagonal element
+    ComplexD lam1;    // First eval for swap
+    ComplexD lam2;    // Second eval for swap
+    ComplexD phi;     // phase of s
+    RealD r;          // norm of s and lam2 - lam1
+    
+    int Nm;                               // size of matrix problem
+
+    ComplexComparator cCompare;           // function to sort the Schur matrix. 
+  
+  public:
+
+    /**
+     * If the input matrix _A is in Hessenberg form (upper triangular + first subdiagonal non-zero), then the Schur 
+     * decomposition is easier to compute. 
+     */
+    ComplexSchurDecomposition(CMat _A, bool isHess, RitzFilter ritzFilter = EvalReSmall) : A(_A), Nm (_A.rows()), cCompare (ritzFilter)
+    {
+      Eigen::ComplexSchur<CMat> schur (Nm);
+      if (isHess) {
+        schur.computeFromHessenberg(_A, CMat::Identity(Nm, Nm), true);
+      } else {
+        schur.compute(_A, true);
+      }
+      S = schur.matrixT();
+      Q = schur.matrixU().adjoint();      // Eigen computes A = Q S Q^\dag, we want A = Q^\dag S Q
+    }
+
+    // Getters
+    int  getNm()      { return Nm; }         // size of matrix problem
+    CMat getMatrixA() { return A;  }         // matrix for decomposition
+    CMat getMatrixQ() { return Q;  }         // unitary matrix Q
+    CMat getMatrixS() { return S;  }         // Schur matrix (upper triangular) S
+    CMat getRitz()    { return S.diagonal(); }
+
+    /**
+     * Checks the Schur decomposition A = Q^\dag S Q holds for the computed matrices. Returns if the relative 
+     * Frobenius norm || A - Q^\dag S Q || / || A || is less than rtol. 
+     */
+    bool checkDecomposition(RealD rtol = 1e-8) {
+      RealD Anorm = A.norm();
+      if (Anorm < rtol) {
+        std::cout << GridLogMessage << "Zero matrix" << std::endl;
+        return true;
+      }
+
+      std::cout << GridLogDebug << "S = " << std::endl << S << std::endl;
+      std::cout << GridLogDebug << "Q = " << std::endl << Q << std::endl;
+
+      CMat A2 = Q.adjoint() * S * Q;
+      std::cout << GridLogDebug << "Q^dag S Q = " << std::endl << A2 << std::endl;
+      RealD dA = (A - A2).norm() / Anorm;
+      return (dA < rtol);
+    }
+
+    /**
+     * Swaps the components on the diagonal of the Schur matrix at index i with index i + 1. 
+     * Updates the orthogonal matrix Q accordingly. 
+     */
+    void swapEvals(int i) {
+
+      assert(0 <= i && i <= Nm - 1);        // can only swap blocks with upper left index between 0 and Nm - 1
+
+      // get parameters for rotation
+      s    = S(i, i+1);
+      lam1 = S(i, i);
+      lam2 = S(i+1, i+1);
+      phi  = s / abs(s);
+      r    = std::sqrt(std::pow(abs(s), 2) + std::pow(abs(lam2 - lam1), 2));
+
+      // // Original code which performs Givens rotations by manual matrix multiplication
+      // // compute Givens rotation corresponding to these parameters
+      // Givens = CMat::Identity(Nm, Nm);
+      // Givens(i, i)     = std::abs(s) / r;
+      // Givens(i+1, i+1) = Givens(i, i);
+      // Givens(i, i+1)   = (phi / r) * std::conj(lam2 - lam1);
+      // Givens(i+1, i)   = -std::conj(Givens(i, i+1));
+      // // rotate Schur matrix and unitary change of basis matrix Q
+      // S = Givens * S * Givens.adjoint();
+      // Q = Givens * Q;
+
+      // Modified code
+      Givens = CMat::Identity(2, 2);
+      Givens(0, 0)     = abs(s) / r;
+      Givens(1, 1) = Givens(0, 0);
+      Givens(0, 1)   = (phi / r) * conj(lam2 - lam1);
+      Givens(1, 0)   = -conj(Givens(0, 1));
+
+      // TODO: make sure these are correct
+      Eigen::MatrixXcd tmp;
+      tmp = Givens * S.block(i, 0, 2, Nm);
+      S.block(i, 0, 2, Nm) = tmp;
+
+      // S_{:, i:i+2} = S.block<Nm, 2>(0, i);
+      tmp = S.block(0, i, Nm, 2) * Givens.adjoint();
+      S.block(0, i, Nm, 2) = tmp;
+
+      // Q_{i:i+2, :} = Q.block<2, Nm>(i, 0);
+      tmp = Givens * Q.block(i, 0, 2, Nm);
+      Q.block(i, 0, 2, Nm) = tmp;
+
+      return;
+    }
+
+    /**
+     * Reorders a Schur matrix &Schur to have the Ritz values that we would like to keep for 
+     * restart as the first Nk elements on the diagonal. 
+     * 
+     * This algorithm is implemented as Nk iterations of a a reverse bubble sort with comparator compare.
+     * TODO: pass in compare function as an argument, default to compare with <. 
+     */
+    // void schurReorder(int Nk, std::function compare) {
+    void schurReorder(int Nk) {
+     std::cout << GridLogMessage << "schurReorder start " << std::endl;
+     int nswap=0;
+      for (int i = 0; i < Nk; i++) {
+        for (int k = 0; k <= Nm - 2; k++) {
+          int idx = Nm - 2 - k;
+          // TODO use RitzFilter enum here
+          // if (std::abs(S(idx, idx)) < std::abs(S(idx+1, idx+1))) {            // sort by largest modulus of eigenvalue
+          // if (std::real(S(idx, idx)) > std::real(S(idx+1, idx+1))) {        // sort by smallest real eigenvalue
+          if ( cCompare(S(idx+1, idx+1), S(idx, idx)) ) {            // sort by largest modulus of eigenvalue
+            swapEvals(idx);
+	    nswap++;
+          }
+        }
+      }
+     std::cout << GridLogMessage << "schurReorder end "<< nswap << std::endl;
+      for (int i = 0; i < Nk; i++) {
+      	std::cout << GridLogMessage << "schurR "<<i<<" "<<S(i,i) << std::endl;
+      }
+      return;
+    }
+
+    void schurReorderBlock() {
+      // TODO method stub
+      return;
+    }
+
+};
+
+// template<class Field>
+// inline void writeFile(const Field &field, const std::string &fname) {
+//   emptyUserRecord record;
+//   ScidacWriter WR(field.Grid()->IsBoss());
+//   WR.open(fname);
+//   WR.writeScidacFieldRecord(field, record, 0); // 0 = Lexico
+//   WR.close();
+// }
+
+/**
+ * Implementation of the Krylov-Schur algorithm.
+ */
+template<class Field> 
+class KrylovSchur {
+
+  private:
+  
+    std::string cname = std::string("KrylovSchur");
+    int MaxIter;   // Max iterations
+    RealD Tolerance;
+    RealD ssq;
+    RealD rtol;
+    int Nm;           // Number of basis vectors to track (equals MaxIter if no restart)
+    int Nk;           // Number of basis vectors to keep every restart (equals -1 if no restart)
+    int Nstop;       // Stop after converging Nstop eigenvectors.
+
+    LinearOperatorBase<Field> &Linop;
+    GridBase *Grid;
+
+    RealD approxLambdaMax;
+    RealD beta_k;
+    Field u;                                // Residual vector perpendicular to Krylov space (u_{k+1} in notes)
+    Eigen::VectorXcd   b;                   // b vector in Schur decomposition (e_{k+1} in Arnoldi).
+    std::vector<Field> basis;               // orthonormal Krylov basis
+    Eigen::MatrixXcd   Rayleigh;            // Rayleigh quotient of size Nbasis (after construction)
+    Eigen::MatrixXcd   Qt;                  // Transpose of basis rotation which projects out high modes.
+
+    Eigen::VectorXcd   evals;               // evals of Rayleigh quotient
+    std::vector<RealD> ritzEstimates;       // corresponding ritz estimates for evals
+    Eigen::MatrixXcd   littleEvecs;         // Nm x Nm evecs matrix
+
+    RitzFilter ritzFilter;                  // how to sort evals
+
+  public:
+    bool doEvalCheck = false;
+
+    KrylovSchur(LinearOperatorBase<Field> &_Linop, GridBase *_Grid, RealD _Tolerance, RitzFilter filter = EvalReSmall)
+      : Linop(_Linop), Grid(_Grid), Tolerance(_Tolerance), ritzFilter(filter), u(_Grid), MaxIter(-1), Nm(-1), Nk(-1), Nstop (-1),
+        evals (0), ritzEstimates (), evecs (), ssq (0.0), rtol (0.0), beta_k (0.0), approxLambdaMax (0.0)
+    {
+      u = Zero();
+    };
+    std::vector<Field> evecs;               // Vector of evec fields
+
+
+    /* Getters */
+    int                 getNk()                { return Nk;            }
+    Eigen::MatrixXcd    getRayleighQuotient()  { return Rayleigh;      }
+    Field               getU()                 { return u;             }
+    std::vector<Field>  getBasis()             { return basis;         }
+    Eigen::VectorXcd    getEvals()             { return evals;         }
+    std::vector<RealD>  getRitzEstimates()     { return ritzEstimates; }
+    std::vector<Field>  getEvecs()             { return evecs;         }
+
+    /**
+     * Runs the non-harmonic Krylov-Schur loop.
+     *   - Runs an Arnoldi step to generate the Rayleigh quotient and Krylov basis.
+     *   - Schur decompose the Rayleigh quotient.
+     *   - Permutes the Rayleigh quotient according to the eigenvalues.
+     *   - Truncate the Krylov-Schur expansion.
+     */
+    void operator()(const Field& v0, int _maxIter, int _Nm, int _Nk, int _Nstop, bool doubleOrthog = true) {
+
+      MaxIter = _maxIter;
+      Nm = _Nm; Nk = _Nk;
+      Nstop = _Nstop;
+
+      ssq = norm2(v0);
+      RealD approxLambdaMax = approxMaxEval(v0);
+      rtol = Tolerance * approxLambdaMax;
+      std::cout << GridLogMessage << "Approximate max eigenvalue: " << approxLambdaMax << std::endl;
+
+      b = Eigen::VectorXcd::Zero(Nm);       // start as e_{k+1}
+      b(Nm-1) = 1.0;
+
+      int start = 0;
+      Field startVec = v0;
+      littleEvecs = Eigen::MatrixXcd::Zero(Nm, Nm);
+      for (int i = 0; i < MaxIter; i++) {
+        std::cout << GridLogMessage << "Restart Iteration " << i << std::endl;
+
+        // Perform Arnoldi steps to compute Krylov basis and Rayleigh quotient (Hess)
+        arnoldiIteration(startVec, Nm, start, doubleOrthog);
+        startVec = u;
+        start = Nk;
+
+        std::cout << GridLogDebug << "b after Arnoldi " << b << std::endl;
+
+        // --- Restart: Schur-decompose Rayleigh, reorder by ritzFilter, rotate
+        //     the basis into the Schur vectors, and truncate to the leading Nk.
+        ComplexSchurDecomposition schur (Rayleigh, false, ritzFilter);
+        std::cout << GridLogDebug << "Schur decomp holds? " << schur.checkDecomposition() << std::endl;
+
+        // Rearrange Schur matrix so wanted evals are on top left (like MATLAB's ordschur)
+        std::cout << GridLogMessage << "Reordering Schur eigenvalues" << std::endl;
+        schur.schurReorder(Nk);
+        std::cout << GridLogMessage << "Schur eigenvalues reordered." << std::endl;
+
+        Eigen::MatrixXcd Q = schur.getMatrixQ();
+        Qt = Q.adjoint();                           // TODO should Q be real?
+
+        std::cout << GridLogMessage << "*** ROTATING TO SCHUR BASIS *** " << std::endl;
+
+        // Rotate Krylov basis, b vector, redefine Rayleigh quotient and evecs, and truncate.
+        Rayleigh = schur.getMatrixS();
+        b = Q * b;            // b^\dag = b^\dag * Q^\dag <==> b = Q*b
+
+        std::vector<Field> basis2;
+        constructUR(basis2, basis, Qt, Nm,Nk);
+
+        std::cout << GridLogMessage << "*** TRUNCATING FOR RESTART *** " << std::endl;
+        std::cout << GridLogDebug << "Rayleigh before truncation: " << std::endl << Rayleigh << std::endl;
+
+        Eigen::MatrixXcd RayTmp = Rayleigh(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk));
+        Rayleigh = RayTmp;
+
+        basis = std::vector<Field> (basis2.begin(), basis2.begin() + Nk);
+
+        b = b.head(Nk).eval();
+
+        std::cout << GridLogDebug << "Rayleigh after truncation: " << std::endl << Rayleigh << std::endl;
+
+        checkKSDecomposition();
+
+        // Compute eigensystem of Rayleigh. Note the eigenvectors correspond to the sorted eigenvalues.
+        computeEigensystem(Rayleigh);
+        std::cout << GridLogMessage << "Eigenvalues (first Nk sorted): " << std::endl << evals << std::endl;
+
+        if (checkConvergedAndReport(i)) return;
+      }
+    }
+
+    /**
+     * Runs the harmonic (shifted) Krylov-Schur loop: extracts Ritz values of a
+     * shift-augmented Rayleigh quotient so that eigenvalues near `*_shift` are
+     * reordered to the top instead of the extremal ones.
+     */
+    void operator()(const Field& v0, int _maxIter, int _Nm, int _Nk, int _Nstop, RealD *_shift, bool doubleOrthog = true) {
+
+      assert(_shift && "harmonic KrylovSchur: shift must be non-null");
+      RealD shiftVal = *_shift;
+
+      MaxIter = _maxIter;
+      Nm = _Nm; Nk = _Nk;
+      Nstop = _Nstop;
+
+      ssq = norm2(v0);
+      RealD approxLambdaMax = approxMaxEval(v0);
+      rtol = Tolerance * approxLambdaMax;
+      std::cout << GridLogMessage << "Approximate max eigenvalue: " << approxLambdaMax << std::endl;
+
+      b = Eigen::VectorXcd::Zero(Nm);       // start as e_{k+1}
+      b(Nm-1) = 1.0;
+
+      int start = 0;
+      Field startVec = v0;
+      littleEvecs = Eigen::MatrixXcd::Zero(Nm, Nm);
+      for (int i = 0; i < MaxIter; i++) {
+        std::cout << GridLogMessage << "Restart Iteration " << i << std::endl;
+
+        // Perform Arnoldi steps to compute Krylov basis and Rayleigh quotient (Hess)
+        arnoldiIteration(startVec, Nm, start, doubleOrthog);
+        startVec = u;
+        start = Nk;
+
+        std::cout << GridLogDebug << "b after Arnoldi " << b << std::endl;
+
+        // --- Harmonic restart: Schur-decompose the shift-augmented Rayleigh
+        //     quotient so Ritz values close to shiftVal are reordered to the
+        //     top, then rotate and truncate as in the non-harmonic case.
+        Eigen::MatrixXcd temp = Rayleigh;
+        for (int m=0;m<Nm;m++) temp(m,m) -= shiftVal;
+        Eigen::MatrixXcd RayleighS = temp.inverse(); // (B-tI)^-1
+        Eigen::MatrixXcd temp2 = RayleighS.adjoint(); //(B-tI)^-1*
+        Eigen::VectorXcd g  = temp2*b; //g = (B-tI)^-1* * b
+        Eigen::MatrixXcd Btilde = Rayleigh + g*(b.adjoint());
+
+        Field utilde(Grid);
+        utilde  = u;
+        for (int j = 0; j<Nm; j++){
+          utilde -= basis[j]*g(j);
+        }
+
+        ComplexSchurDecomposition schurS (Btilde, false, ritzFilter);
+        std::cout << GridLogMessage << "Shifted Schur eigenvalues shift = "<<shiftVal  << std::endl;
+        schurS.schurReorder(Nk);
+
+        Eigen::MatrixXcd Q_s = schurS.getMatrixQ();
+        Eigen::MatrixXcd Qt_s = Q_s.adjoint();                           // TODO should Q be real?
+
+        Btilde=schurS.getMatrixS();
+        Eigen::VectorXcd b_s = Q_s*b; // Q is Qt in SlepC, b_s=bhat
+
+        std::vector<Field> basis2_s;
+        constructUR(basis2_s, basis, Qt_s, Nm,Nk);
+
+        Eigen::MatrixXcd RayTmp_s = Btilde(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk));
+        Btilde = RayTmp_s;
+
+        basis2_s = std::vector<Field> (basis2_s.begin(), basis2_s.begin() + Nk);
+
+        b_s = b_s.head(Nk).eval();
+
+        Eigen::VectorXcd ghat = -Q_s * g;
+        ghat = ghat.head(Nk).eval();
+
+        Field uhat(Grid);
+        uhat  = utilde;
+        for (int j = 0; j<Nk; j++){
+          uhat -= basis2_s[j]*ghat(j);
+        }
+
+        RealD gamma = std::sqrt(norm2(uhat));
+        uhat *= 1.0/gamma;
+        std::cout << GridLogMessage << " gamma "<<gamma << std::endl;
+
+        Btilde  += ghat*(b_s.adjoint());
+        b_s *=gamma;
+
+        Rayleigh = Btilde;
+        basis= basis2_s;
+        b  = b_s;
+        beta_k = gamma;
+        u= uhat;
+
+        checkKSDecomposition();
+        computeEigensystem(Rayleigh);
+        std::cout << GridLogMessage << "Eigenvalues (first Nk sorted): " << std::endl << evals << std::endl;
+
+        if (checkConvergedAndReport(i)) return;
+      }
+    }
+
+  private:
+
+    /**
+     * Shared post-restart convergence check. On convergence (or the final
+     * iteration) reports eigenvalues and optionally the explicit residual
+     * check. Returns true if operator() should return.
+     */
+    bool checkConvergedAndReport(int i) {
+      int Nconv = converged();
+      std::cout << GridLogMessage << "Number of evecs converged: " << Nconv << std::endl;
+      if (Nconv >= Nstop || i == MaxIter - 1) {
+        std::cout << GridLogMessage << "Converged with " << Nconv << " / " << Nstop << " eigenvectors on iteration "
+                      << i << "." << std::endl;
+        std::cout << GridLogMessage << "Eigenvalues: " << std::endl << evals << std::endl;
+
+        if (doEvalCheck) {
+          Field w(Grid);
+          for (int k = 0; k < (int)evecs.size(); k++) {
+            Linop.Op(evecs[k], w);
+            ComplexD eval_est = toStdCmplx(innerProduct(evecs[k], w));
+            w -= eval_est * evecs[k];
+            RealD res = std::sqrt(norm2(w));
+            std::cout << GridLogMessage << "KrylovSchur: evec[" << k << "]"
+                      << "  eval_reported = " << evals[k]
+                      << "  eval_est = " << eval_est
+                      << "  || A v - eval_est * v || = " << res << std::endl;
+          }
+        }
+        return true;
+      }
+      return false;
+    }
+
+  public:
+
+    /**
+     * Constructs the Arnoldi basis for the Krylov space K_n(D, src). (TODO make private)
+     * 
+     * Parameters
+     * ----------
+     * v0 : Field&
+     *  Source to generate Krylov basis. 
+     * Nm : int
+     *  Final size of the basis desired. If the basis becomes complete before a basis of size Nm is constructed 
+     *  (determined by relative tolerance Tolerance), stops iteration there. 
+     * doubleOrthog : bool (default = false)
+     *  Whether to double orthogonalize the basis (for numerical cancellations) or not. 
+     * start        : int (default = 0)
+     *  If non-zero, assumes part of the Arnoldi basis has already been constructed. 
+     */
+    void arnoldiIteration(const Field& v0, int Nm, int start = 0, bool doubleOrthog = true)
+    {
+
+      ComplexD coeff;
+      Field w (Grid);           // A acting on last Krylov vector. 
+
+      // basis.reserve(Nm);
+      // for (int i = start; i < Nm; i++) {
+      //   basis.emplace_back(Grid);
+      // }
+      // basis.assign(Nm, Field(Grid));
+      // basis.resize(Nm);
+      // for (int i = start; i < Nm; i++) {
+      //   basis[i] = Field(Grid);
+      // }
+
+      if (start == 0) {       // initialize everything that we need.
+        RealD v0Norm = 1 / std::sqrt(ssq);
+        basis.push_back(v0Norm * v0);                // normalized source
+        // basis[0] = v0Norm * v0;                // normalized source
+
+        Rayleigh = Eigen::MatrixXcd::Zero(Nm, Nm);
+        u = Zero();
+      } else {
+        // assert( start == basis.size() );      // should be starting at the end of basis (start = Nk)
+        std::cout << GridLogMessage << "Resetting Rayleigh and b" << std::endl;
+        Eigen::MatrixXcd RayleighCp = Rayleigh;
+        Rayleigh = Eigen::MatrixXcd::Zero(Nm, Nm);
+        Rayleigh(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk)) = RayleighCp;
+
+        // append b^\dag to Rayleigh, add u to basis
+        Rayleigh(Nk, Eigen::seqN(0, Nk)) = b.adjoint();
+        basis.push_back(u);
+        // basis[start] = u;       // TODO make sure this is correct
+        b = Eigen::VectorXcd::Zero(Nm);
+      }
+
+      // Construct next Arnoldi vector by normalizing w_i = Dv_i - \sum_j v_j h_{ji}
+      for (int i = start; i < Nm; i++) {
+        Linop.Op(basis.back(), w);
+        // Linop.Op(basis[i], w);
+        for (int j = 0; j < basis.size(); j++) {
+          coeff = innerProduct(basis[j], w);       // coeff = h_{ij}. Note that since {vi} is ONB it's OK to subtract it off after. 
+          Rayleigh(j, i) = toStdCmplx(coeff);
+          w -= coeff * basis[j];
+        }
+
+        if (doubleOrthog) {
+          std::cout << GridLogDebug << "Double orthogonalizing." << std::endl;
+          for (int j = 0; j < basis.size(); j++) {
+            coeff = innerProduct(basis[j], w);      // see if there is any residual component in basis[j] direction
+            Rayleigh(j, i) += toStdCmplx(coeff);   // if coeff is non-zero, adjust Rayleigh
+            w -= coeff * basis[j];
+          }
+        }
+
+        // add w_i to the pile
+        if (i < Nm - 1) {
+          coeff = std::sqrt(norm2(w));
+          Rayleigh(i+1, i) = toStdCmplx(coeff);
+          basis.push_back(
+            (1.0/coeff) * w
+          );
+          // basis[i+1] = (1.0/coeff) * w;
+        }
+
+        // after iterations, update u and beta_k = ||u|| before norm
+        u = w;                                // make sure u is normalized
+        beta_k = std::sqrt(norm2(u));         // beta_k = ||f_k|| determines convergence.
+        u = (1/beta_k) * u;
+      }
+
+      b(Nm - 1) = beta_k;
+
+      // std::cout << GridLogMessage << "|f|^2 after Arnoldi step = " << norm2(f) << std::endl;
+      std::cout << GridLogMessage << "beta_k = |u| (before norm) after Arnoldi step = " << beta_k << std::endl;
+      std::cout << GridLogDebug << "Computed Rayleigh quotient = " << std::endl << Rayleigh << std::endl;
+
+      return;
+    }
+
+    /**
+     * Approximates the eigensystem of the linear operator by computing the eigensystem of 
+     * the Rayleigh quotient. Assumes that the Rayleigh quotient has already been constructed (by 
+     * calling the operator() function).
+     * 
+     * Parameters
+     * ----------
+     * Eigen::MatrixXcd& S
+     *  Schur matrix (upper triangular) similar to original Rayleigh quotient.
+     */
+    void computeEigensystem(Eigen::MatrixXcd& S)
+    {
+      std::cout << GridLogMessage << "Computing eigenvalues." << std::endl;
+
+      evecs.clear();
+
+      Eigen::ComplexEigenSolver<Eigen::MatrixXcd> es;
+      es.compute(S);
+
+      // Sort eigenvalues/evecs to match the schurReorder ordering.
+      int n = es.eigenvalues().size();
+      ComplexComparator cComp(ritzFilter);
+      std::vector<int> idx(n);
+      std::iota(idx.begin(), idx.end(), 0);
+      std::sort(idx.begin(), idx.end(), [&](int a, int b){
+        return cComp(toStdCmplx(es.eigenvalues()(a)), toStdCmplx(es.eigenvalues()(b)));
+      });
+
+      evals.resize(n);
+      littleEvecs.resize(n, n);
+      for (int k = 0; k < n; k++) {
+        evals(k)          = es.eigenvalues()(idx[k]);
+        littleEvecs.col(k) = es.eigenvectors().col(idx[k]);
+      }
+
+      // Convert evecs to lattice fields
+      for (int k = 0; k < n; k++) {
+        Eigen::VectorXcd vec = littleEvecs.col(k);
+        Field tmp (basis[0].Grid());
+        tmp = Zero();
+        for (int j = 0; j < (int)basis.size(); j++) {
+          tmp = tmp + vec[j] * basis[j];
+        }
+        evecs.push_back(tmp);
+      }
+    }
+
+    /**
+     * Approximates the maximum eigenvalue of Linop.Op to normalize the residual and test for convergence. 
+     * 
+     * TODO implement in parent class eventually
+     * 
+     * Parameters
+     * ----------
+     * Field& v0
+     *  Source field to start with. Must have non-zero norm.
+     * int MAX_ITER (default = 50)
+     *  Maximum number of iterations for power approximation. 
+     * 
+     * Returns
+     * -------
+     * RealD lamApprox
+     *  Approximation of largest eigenvalue. 
+     */
+    RealD approxMaxEval(const Field& v0, int MAX_ITER = 50) {
+      assert (norm2(v0) > 1e-8);                        // must have relatively large source norm to start
+      RealD lamApprox = 0.0;
+      RealD denom = 1.0; RealD num = 1.0;
+      Field v0cp (Grid); Field tmp (Grid);
+      v0cp = v0;
+      denom = std::sqrt(norm2(v0cp));
+      for (int i = 0; i < MAX_ITER; i++) {
+        Linop.Op(v0cp, tmp);                               // CAREFUL: do not do Op(tmp, tmp)
+        v0cp = tmp;
+        num = std::sqrt(norm2(v0cp));                      // num = |A^{n+1} v0|
+        lamApprox = num / denom;                           // lam = |A^{n+1} v0| / |A^n v0|
+        std::cout << GridLogDebug << "Approx for max eval: " << lamApprox << std::endl;
+        denom = num;                                       // denom = |A^{n} v0|
+      }
+      return lamApprox;
+    }
+
+    /**
+     * Computes the number of Krylov-Schur eigenvectors that have converged. An eigenvector s is considered converged 
+     * for a tolerance epsilon if 
+     *    r(s) := |\beta e_m^T s| < epsilon
+     * where beta is the norm of f_{m+1}.
+     * 
+     * TODO implement in parent class eventually
+     * 
+     * Parameters
+     * ----------
+     * 
+     * Returns
+     * -------
+     * int : Number of converged eigenvectors.
+     */
+    int converged() {
+      int Nconv = 0;
+      int _Nm = evecs.size();
+      std::cout << GridLogDebug << "b: " << b << std::endl;
+      Field tmp (Grid); Field fullEvec (Grid);
+      ritzEstimates.clear();
+      // ritzEstimates.resize(_Nm);
+      for (int k = 0; k < _Nm; k++) {
+        Eigen::VectorXcd evec_k = littleEvecs.col(k);
+        RealD ritzEstimate = std::abs(b.dot(evec_k));           // b^\dagger s
+        ritzEstimates.push_back(ritzEstimate);
+        std::cout << GridLogMessage << "Ritz estimate for evec " << k << " = " << ritzEstimate << std::endl;
+        if (ritzEstimate < rtol) {
+          Nconv++;
+        } else {
+          break;
+        }
+      }
+      // Check that Ritz estimate is explicitly || D (Uy) - lambda (Uy) ||
+//       checkRitzEstimate();
+      return Nconv;
+    }
+
+    /**
+     * Checks the Krylov-Schur decomposition DU = UR + f b^\dag with the last-computed 
+     * U, R, f, and b. 
+     */
+    bool checkKSDecomposition(RealD tol = 1e-8) {
+
+      std::cout << GridLogMessage << "*** CHECKING KRYLOV-SCHUR DECOMPOSITION *** " << std::endl;
+
+      int k = basis.size();         // number of basis vectors, also the size of Rayleigh.
+
+      // rotate basis by Rayleigh to construct UR
+      // std::vector<Field> rotated;
+
+      // std::cout << GridLogDebug << "Rayleigh in KSDecomposition: " << std::endl << Rayleigh << std::endl;
+
+      std::vector<Field> rotated = basis;
+      constructUR(rotated, basis, Rayleigh, k,k);             // manually rotate
+      // Eigen::MatrixXcd Rt = Rayleigh.adjoint();
+      // basisRotate(rotated, Rt, 0, k, 0, k, k);           // UR
+
+      // TODO: make a new function that I'm positive does what this is doing
+      // just take the basis U = (u1 u2 ... uNm) and form the linear combination UR from R
+      
+      // For each i, form D u(i) and subtract off (US - u b^\dag)(i)
+      RealD delta = 0.0; RealD deltaSum = 0;
+      Field tmp (Grid); tmp = Zero();
+      for (int i = 0; i < k; i++) {
+        Linop.Op(basis[i], tmp);          // tmp = D u(i)
+        delta = norm2(tmp - rotated[i] - u * std::conj(b(i)));
+        delta = delta / norm2(tmp);               // relative tolerance
+        deltaSum += delta;
+
+        // std::cout << GridLogDebug << "Iteration " << i << std::endl;
+        // std::cout << GridLogDebug << "Du = " << norm2(tmp) << std::endl;
+        // std::cout << GridLogDebug << "rotated = " << norm2(rotated[i]) << std::endl;
+        // std::cout << GridLogDebug << "b[i] = " << b(i) << std::endl;
+        std::cout << GridLogMessage << "Deviation in decomp, column " << i << ": " << delta << std::endl;
+      }
+      std::cout << GridLogMessage << "Squared sum of relative deviations in decomposition: " << deltaSum << std::endl;
+
+      // std::cout << "[DEBUG] testing basis rotate" << std::endl;
+      // std::vector<Field> rotated2;
+      // constructUR(rotated2, basis, Rayleigh, k);
+      // for (int i = 0; i < k; i++) {
+      //   std::cout << "rotated[i] - UR[i] = " << norm2(rotated[i] - rotated2[i]) << std::endl;
+      // }
+
+      return deltaSum < tol;
+    }
+
+    /**
+     * Checks the Ritz vector s was rotated correctly by explicitly recomputing the 
+     * eigenvectors of the rotated Rayleigh quotient. 
+     */
+    bool checkRitzRotation(RealD tol = 1e-8) {
+      std::cout << GridLogMessage << "*** CHECKING RITZ VECTOR ROTATION *** " << std::endl;
+
+      Eigen::ComplexEigenSolver<Eigen::MatrixXcd> es;
+      es.compute(Rayleigh);
+      Eigen::MatrixXcd littleEvecs2 = es.eigenvectors();
+      RealD dLittle = (littleEvecs2 - littleEvecs).norm() / littleEvecs.norm();
+      std::cout << GridLogMessage << "|littleEvecs2 - littleEvecs| / |littleEvecs| = " << dLittle << std::endl;
+
+      std::cout << GridLogMessage << "Forming full eigenvectors" << std::endl;
+      RealD delta = 0.0; RealD deltaSum = 0;
+      for (int k = 0; k < evals.size(); k++) {
+        Eigen::VectorXcd vec = littleEvecs.col(k);
+        Field tmpEvec (Grid);
+        tmpEvec = Zero();
+        for (int j = 0; j < basis.size(); j++) {
+          tmpEvec = tmpEvec + vec[j] * basis[j];
+        }
+        delta = norm2(tmpEvec - evecs[k]) / norm2(evecs[k]);
+        std::cout << GridLogMessage << "Deviation in evec " << k << ": " << delta << std::endl;
+        deltaSum += delta;
+      }
+      return deltaSum < tol;
+    }
+
+    /**
+     * Checks the Ritz estimate R(s) is indeed the deviation of a Ritz eigenvector from being a true eigenvector. 
+     */
+    void checkRitzEstimate(RealD tol = 1e-8) {
+      std::cout << GridLogMessage << "*** CHECKING RITZ ESTIMATE *** " << std::endl;
+
+      // The issue was that the Eigen::eigensolver occasionally returned the complex conjugate pairs in the wrong 
+      // order compared to the diagonal, which is how I was reading them out. When this happened, the Ritz estimate would 
+      // be wrong. So, just need to be more careful and actually read out the eigenvalues. 
+
+      Field tmp (Grid);
+      // std::cout << "n evecs: " << evecs.size() << std::endl;
+      for (int k = 0; k < evecs.size(); k++) {
+        tmp = Zero();
+        Linop.Op(evecs[k], tmp);        // D evec
+        RealD ritz = std::sqrt(norm2(tmp - evals[k] * evecs[k]));
+        std::cout << "Ritz estimate " << k << " = " << ritz << std::endl;
+
+        // Checking little Ritz estimate
+        // Eigen::VectorXcd littleEvec = littleEvecs.col(k);
+        // Eigen::VectorXcd dev = Rayleigh * littleEvec - evals[k] * littleEvec;
+        // std::cout << GridLogMessage << "Little Ritz estimate = " << dev.norm() << std::endl;
+      }
+      return;
+    }
+
+    /**
+     * Given a vector of fields U (equivalently, a LxN matrix, where L is the number of degrees of
+     * freedom on the lattice field) and an NxN matrix R, forms the product UR.
+     * Only the first N2 output columns are computed; columns [N2, N) are Zero.
+     *
+     * Implemented with the accelerator basisRotate kernel, which computes
+     * basis[j] <- sum_k Qt(j,k) basis[k] in place, so we copy U into UR and
+     * pass R^T. The coefficients are staged as ComplexD (thrust::complex on
+     * GPU builds) since std::complex has no device arithmetic with Grid tensors.
+     */
+    void constructUR(std::vector<Field>& UR, std::vector<Field> &U, Eigen::MatrixXcd& R, int N, int N2) {
+      KSCoeffMat Rt(N);
+      for (int i = 0; i < N; i++)
+        for (int j = 0; j < N; j++)
+          Rt(i, j) = ComplexD(R(j, i).real(), R(j, i).imag());
+
+      UR = U;
+      basisRotate(UR, Rt, 0, N2, 0, N, N);
+      for (int i = N2; i < N; i++) UR[i] = Zero();
+      return;
+    }
+
+    /**
+     * Same as constructUR but for the product order RU. 
+     */
+    void constructRU(std::vector<Field>& RU, std::vector<Field> &U, Eigen::MatrixXcd& R, int N) {
+      Field tmp (Grid);
+      RU.clear();
+      // RU.resize(N);
+      for (int i = 0; i < N; i++) {
+        tmp = Zero();
+        for (int j = 0; j < N; j++) {
+          tmp = tmp + R(i, j) * U[j];
+        }
+        RU.push_back(tmp);
+        // RU[i] = tmp;
+      }
+      return;
+    }
+
+    // void writeEvec(Field& in, std::string const fname){  
+    //   #ifdef HAVE_LIME
+    //     // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+    //     std::cout << GridLogMessage << "Writing evec to: " << fname << std::endl;
+    //     Grid::emptyUserRecord record;
+    //     Grid::ScidacWriter WR(in.Grid()->IsBoss());
+    //     WR.open(fname);
+    //     WR.writeScidacFieldRecord(in,record,0); // Lexico
+    //     WR.close();
+    //   #endif
+    //     // What is the appropriate way to throw error?
+    // }
+
+    // /**
+    //  * Writes the eigensystem of a Krylov Schur object to a directory. 
+    //  * 
+    //  * Parameters
+    //  * ----------
+    //  * std::string path
+    //  *    Directory to write to. 
+    //  */
+    // void writeEigensystem(std::string outDir) {
+    //   std::cout << GridLogMessage << "Writing output to directory: " << outDir << std::endl;
+    //   // TODO write a scidac density file so that we can easily integrate with visualization toolkit
+    //   std::string evalPath = outDir + "/evals.txt";
+    //   std::ofstream fEval;
+    //   fEval.open(evalPath);
+    //   for (int i = 0; i < Nk; i++) {
+    //     // write Eigenvalues
+    //     fEval << i << " " << evals(i);
+    //     if (i < Nk - 1) { fEval << "\n"; }
+    //   }
+    //   fEval.close();
+      
+    //   for (int i = 0; i < Nk; i++) {
+    //     std::string fName = outDir + "/evec" + std::to_string(i);
+    //     // writeFile(evecs[i], fName);     // using method from Grid/HMC/ComputeWilsonFlow.cc
+    //     // writeEvec(evecs[i], fName);
+    //   }
+      
+    // }
+
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/LanczosBidiagonalization.h

@@ -0,0 +1,276 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./Grid/algorithms/iterative/LanczosBidiagonalization.h
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_LANCZOS_BIDIAGONALIZATION_H
+#define GRID_LANCZOS_BIDIAGONALIZATION_H
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Lanczos Bidiagonalization (Golub-Kahan)
+ *
+ * For a linear operator A with adjoint A^dag, constructs the bidiagonal
+ * decomposition:
+ *
+ *   A  V_m = U_m B_m
+ *   A^dag U_m = V_m B_m^T + beta_{m+1} v_{m+1} e_m^T
+ *
+ * where:
+ *   V_m = [v_1, ..., v_m]  right Lanczos vectors (orthonormal)
+ *   U_m = [u_1, ..., u_m]  left  Lanczos vectors (orthonormal)
+ *   B_m is upper bidiagonal with diag(alpha_1,...,alpha_m) and
+ *       superdiag(beta_2,...,beta_m)
+ *
+ * The singular values of A are approximated by those of B_m.
+ * The singular values of B_m are the square roots of the eigenvalues of
+ * the symmetric tridiagonal matrix B_m^T B_m.
+ *
+ * Usage:
+ *   LanczosBidiagonalization<Field> lb(Linop, grid);
+ *   lb.run(src, Nm, tol);
+ *   // Access results via getters.
+ */
+template <class Field>
+class LanczosBidiagonalization {
+
+  public: 
+  LinearOperatorBase<Field> &Linop;
+  GridBase *Grid;
+
+  int Nm;           // number of Lanczos steps taken
+  RealD Tolerance;  // convergence threshold on beta_{k+1} / alpha_k
+
+  std::vector<Field>  V;       // right Lanczos vectors v_1 ... v_m
+  std::vector<Field>  U;       // left  Lanczos vectors u_1 ... u_m
+  std::vector<RealD>  alpha;   // diagonal of bidiagonal matrix
+  std::vector<RealD>  beta;    // super-diagonal (beta[k] couples u_k and v_{k+1})
+
+  // SVD of the bidiagonal matrix (filled after computeSVD())
+  Eigen::VectorXd  singularValues;
+  Eigen::MatrixXd  leftSVecs;   // columns are left  singular vectors of B
+  Eigen::MatrixXd  rightSVecs;  // columns are right singular vectors of B
+
+public:
+
+  LanczosBidiagonalization(LinearOperatorBase<Field> &_Linop, GridBase *_Grid,
+                           RealD _tol = 1.0e-8)
+    : Linop(_Linop), Grid(_Grid), Tolerance(_tol), Nm(0)
+  {}
+
+  /**
+   * Run the Golub-Kahan Lanczos bidiagonalization.
+   *
+   * Parameters
+   * ----------
+   * src  : starting vector (need not be normalised)
+   * Nmax : maximum number of Lanczos steps
+   * reorth : if true, full reorthogonalisation of both V and U bases
+   */
+  void run(const Field &src, int Nmax, bool reorth = true)
+  {
+    assert(norm2(src) > 0.0);
+
+    V.clear(); U.clear();
+    alpha.clear(); beta.clear();
+    Nm = 0;
+
+    Field p(Grid), r(Grid);
+
+    // --- initialise: v_1 = src / ||src|| ---
+    Field v(Grid);
+    v = src;
+    RealD nrm = std::sqrt(norm2(v));
+    v = (1.0 / nrm) * v;
+    V.push_back(v);
+
+    for (int k = 0; k < Nmax; ++k) {
+
+      // p = A v_k
+      Linop.Op(V[k], p);
+
+      // p = p - beta_k * u_{k-1}   (remove previous left vector)
+      if (k > 0) {
+        p = p - beta[k-1] * U[k-1];
+      }
+
+      // alpha_k = ||p||
+      RealD ak = std::sqrt(norm2(p));
+      if (ak < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "LanczosBidiagonalization: lucky breakdown at step "
+                  << k << " (alpha = " << ak << ")" << std::endl;
+        break;
+      }
+      alpha.push_back(ak);
+
+      // u_k = p / alpha_k
+      Field u(Grid);
+      u = (1.0 / ak) * p;
+
+      // full reorthogonalisation of u against previous U
+      if (reorth) {
+        for (int j = 0; j < (int)U.size(); ++j) {
+          ComplexD ip = innerProduct(U[j], u);
+          u = u - ip * U[j];
+        }
+        RealD unrm = std::sqrt(norm2(u));
+        if (unrm > 1.0e-14) u = (1.0 / unrm) * u;
+      }
+      U.push_back(u);
+
+      // r = A^dag u_k - alpha_k * v_k
+      Linop.AdjOp(U[k], r);
+      r = r - ak * V[k];
+
+      // full reorthogonalisation of r against previous V
+      if (reorth) {
+        for (int j = 0; j < (int)V.size(); ++j) {
+          ComplexD ip = innerProduct(V[j], r);
+          r = r - ip * V[j];
+        }
+      }
+
+      // beta_{k+1} = ||r||
+      RealD bk = std::sqrt(norm2(r));
+      beta.push_back(bk);
+
+      Nm = k + 1;
+
+      std::cout << GridLogMessage
+                << "LanczosBidiagonalization step " << k
+                << "  alpha = " << ak
+                << "  beta  = " << bk << std::endl;
+
+      // convergence: residual beta / alpha small enough
+      if (bk / ak < Tolerance) {
+        std::cout << GridLogMessage
+                  << "LanczosBidiagonalization converged at step " << k
+                  << "  (beta/alpha = " << bk / ak << ")" << std::endl;
+        break;
+      }
+
+      if (k == Nmax - 1) break;   // no v_{k+2} needed after last step
+
+      // v_{k+1} = r / beta_{k+1}
+      Field vnext(Grid);
+      vnext = (1.0 / bk) * r;
+      V.push_back(vnext);
+    }
+  }
+
+  /**
+   * Compute the SVD of the bidiagonal matrix B using Eigen.
+   * Singular values are stored in descending order.
+   */
+  void computeSVD()
+  {
+    int m = Nm;
+    Eigen::MatrixXd B = Eigen::MatrixXd::Zero(m, m);
+
+    for (int k = 0; k < m; ++k) {
+      B(k, k) = alpha[k];
+      if (k + 1 < m && k < (int)beta.size())
+        B(k, k+1) = beta[k];
+    }
+
+    Eigen::JacobiSVD<Eigen::MatrixXd> svd(B,
+        Eigen::ComputeThinU | Eigen::ComputeThinV);
+
+    singularValues = svd.singularValues();   // already sorted descending
+    leftSVecs      = svd.matrixU();
+    rightSVecs     = svd.matrixV();
+
+    std::cout << GridLogMessage
+              << "LanczosBidiagonalization: singular values of B_" << m
+              << std::endl;
+    for (int k = 0; k < m; ++k)
+      std::cout << GridLogMessage << "  sigma[" << k << "] = "
+                << singularValues(k) << std::endl;
+  }
+
+  /**
+   * Return the k-th approximate left singular vector of A in the full
+   * lattice space.  computeSVD() must have been called first.
+   */
+  Field leftSingularVector(int k)
+  {
+    assert(k < (int)leftSVecs.cols());
+    Field svec(Grid);
+    svec = Zero();
+    for (int j = 0; j < Nm; ++j)
+      svec = svec + leftSVecs(j, k) * U[j];
+    return svec;
+  }
+
+  /**
+   * Return the k-th approximate right singular vector of A in the full
+   * lattice space.  computeSVD() must have been called first.
+   */
+  Field rightSingularVector(int k)
+  {
+    assert(k < (int)rightSVecs.cols());
+    Field svec(Grid);
+    svec = Zero();
+    for (int j = 0; j < Nm; ++j)
+      svec = svec + rightSVecs(j, k) * V[j];
+    return svec;
+  }
+
+  /**
+   * Verify the bidiagonalization: returns max residual
+   *   max_k || A v_k - alpha_k u_k - beta_k u_{k-1} ||
+   */
+  RealD verify()
+  {
+    Field tmp(Grid);
+    RealD maxres = 0.0;
+    for (int k = 0; k < Nm; ++k) {
+      Linop.Op(V[k], tmp);
+      tmp = tmp - alpha[k] * U[k];
+      if (k > 0 && k-1 < (int)beta.size())
+        tmp = tmp - beta[k-1] * U[k-1];
+      RealD res = std::sqrt(norm2(tmp));
+      if (res > maxres) maxres = res;
+      std::cout << GridLogMessage
+                << "LanczosBidiagonalization verify step " << k
+                << "  ||A v_k - alpha_k u_k - beta_{k-1} u_{k-1}|| = "
+                << res << std::endl;
+    }
+    return maxres;
+  }
+
+  /* Getters */
+  int                       getNm()           const { return Nm; }
+  const std::vector<Field>& getV()            const { return V; }
+  const std::vector<Field>& getU()            const { return U; }
+  const std::vector<RealD>& getAlpha()        const { return alpha; }
+  const std::vector<RealD>& getBeta()         const { return beta; }
+  Eigen::VectorXd           getSingularValues() const { return singularValues; }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -60,18 +60,29 @@ public:
 
   void Level(int lv) { level=lv; };
 
-  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
+  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax, int _nstep) : 
     Tolerance(tol), 
     MaxIterations(maxit),
     Linop(_Linop),
     Preconditioner(Prec),
     mmax(_mmax),
-    nstep(_nstep)
+    nstep(_nstep)         // what is nstep vs mmax? one is the number of inner iterations
   { 
     level=1;
     verbose=1;
   };
 
+  // virtual method stubs for updating GCR polynomial
+  virtual void LogBegin(void){
+    std::cout << "GCR::LogBegin() "<<std::endl;
+  };
+  virtual void LogIteration(int k, ComplexD a, std::vector<ComplexD> betas){
+    std::cout << "GCR::LogIteration() "<<std::endl;
+  };
+  virtual void LogComplete(std::vector<ComplexD>& alphas, std::vector<std::vector<ComplexD>>& betas) {
+    std::cout << "GCR::LogComplete() "<<std::endl;
+  };
+
   void operator() (const Field &src, Field &psi){
 
     //    psi=Zero();
@@ -96,19 +107,18 @@ public:
       GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
 
       if(cp<rsq) {
+        SolverTimer.Stop();
 
-	SolverTimer.Stop();
+        Linop.Op(psi,r);
+        axpy(r,-1.0,src,r);
+        RealD tr = norm2(r);
+        GCRLogLevel<<"PGCR: Converged on iteration " <<steps
+          << " computed residual "<<sqrt(cp/ssq)
+          << " true residual "    <<sqrt(tr/ssq)
+          << " target "           <<Tolerance <<std::endl;
 
-	Linop.Op(psi,r);
-	axpy(r,-1.0,src,r);
-	RealD tr = norm2(r);
-	GCRLogLevel<<"PGCR: Converged on iteration " <<steps
-		 << " computed residual "<<sqrt(cp/ssq)
-		 << " true residual "    <<sqrt(tr/ssq)
-		 << " target "           <<Tolerance <<std::endl;
-
-	GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
-	return;
+        GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
+        return;
       }
 
     }
@@ -135,9 +145,9 @@ public:
     ////////////////////////////////
     // history for flexible orthog
     ////////////////////////////////
-    std::vector<Field> q(mmax,grid);
-    std::vector<Field> p(mmax,grid);
-    std::vector<RealD> qq(mmax);
+    std::vector<Field> q(mmax,grid);        // q = Ap
+    std::vector<Field> p(mmax,grid);        // store mmax conjugate momenta
+    std::vector<RealD> qq(mmax);            // qq = (Ap)^2 = <p|A^\dagger A |p> (denom of \alpha)
       
     GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
 
@@ -155,7 +165,9 @@ public:
     LinalgTimer.Start();
     r=src-Az;
     LinalgTimer.Stop();
-    GCRLogLevel<< "PGCR true residual r = src - A psi   "<<norm2(r) <<std::endl;
+    GCRLogLevel<< "PGCR true residual r = src - A psi   "<< norm2(r) <<std::endl;
+
+    this->LogBegin();       // initialize polynomial GCR if needed (TODO think about placement of this)
     
     /////////////////////
     // p = Prec(r)
@@ -178,32 +190,45 @@ public:
     p[0]= z;
     q[0]= Az;
     qq[0]= zAAz;
+
+    std::cout << "||init p - src||: " << norm2(p[0] - src) << std::endl;   // for debugging
     
     cp =norm2(r);
     LinalgTimer.Stop();
 
+    std::vector<ComplexD> all_alphas;
+    std::vector<std::vector<ComplexD>> all_betas;
+
     for(int k=0;k<nstep;k++){
 
       steps++;
 
       int kp     = k+1;
-      int peri_k = k %mmax;
+      int peri_k = k %mmax;     // only store mmax vectors; just roll around if needed
       int peri_kp= kp%mmax;
 
+      // std::cout << "peri_kp = " << peri_kp << std::endl;
+
       LinalgTimer.Start();
       rq= innerProduct(q[peri_k],r); // what if rAr not real?
-      a = rq/qq[peri_k];
+      a = rq/qq[peri_k];              // compute alpha_j
 
-      axpy(psi,a,p[peri_k],psi);         
+      all_alphas.push_back(a);
 
-      cp = axpy_norm(r,-a,q[peri_k],r);
+      axpy(psi,a,p[peri_k],psi);      // update psi --> psi + \alpha p
+
+      cp = axpy_norm(r,-a,q[peri_k],r);       // update r --> r - \alpha D p. Note q = Dp
       LinalgTimer.Stop();
 
-      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
+      // LogIterationA(k + 1, a);
 
-      if((k==nstep-1)||(cp<rsq)){
-	return cp;
-      }
+      GCRLogLevel<< "GCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
+
+      // moving this to end of loop so that it doesn't exit beforehand
+      // TODO if I want to uncomment this, I have to split the LogIteration again and put LogIterationA() beforehand
+      // if((k==nstep-1)||(cp<rsq)){
+      //   return cp;
+      // }
 
 
       PrecTimer.Start();
@@ -221,22 +246,205 @@ public:
       q[peri_kp]=Az;
       p[peri_kp]=z;
 
+      // Field Dsrc (grid);
+      // Linop.Op(src, Dsrc);
+      // std::cout << "||q[peri_kp] - D(src)||: " << norm2(q[peri_kp] - Dsrc) << std::endl;   // for debugging
+
+          // // delete after testing
+          // std::cout << "Testing Dsq on one for GCR: " << std::endl;
+          // Field myField (grid);
+          // myField = 1.0;
+          // Field out1 (grid); Field out2 (grid);
+          // Linop.HermOp(myField, out1);
+          // Linop.Op(myField, out2);
+          // std::cout << "Dsq.Hermop(ones) has norm " << norm2(out1) << std::endl;
+          // std::cout << "Dsq.Op(ones) has norm " << norm2(out2) << std::endl;
+
+      // basically northog = k+1 if mmax is large
       int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
+      // std::cout << "northog: " << northog << std::endl;
+      std::vector<ComplexD> betas (northog);
+      // std::cout << "peri_kp: " << peri_kp << std::endl;
+      // we iterate backwards counting down from the current k+1 index (peri_kp) because we 
       for(int back=0;back<northog;back++){
 
 	int peri_back=(k-back)%mmax;   	  GRID_ASSERT((k-back)>=0);
 
-	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
-	p[peri_kp]=p[peri_kp]+b*p[peri_back];
-	q[peri_kp]=q[peri_kp]+b*q[peri_back];
+        // b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
+        b=-(innerProduct(q[peri_back],Az))/qq[peri_back];     // TODO try complex beta
+        p[peri_kp]=p[peri_kp]+b*p[peri_back];
+        q[peri_kp]=q[peri_kp]+b*q[peri_back];
+
+        // LogIterationB(peri_back, b);
+        // betas[back] = b;    // may need to change the indexing if I ever do it with restarts
+        // std::cout << "[DEBUG] pushing beta for back = " << back << ", peri_back = " << peri_back << std::endl;
+
+        betas[peri_back] = b;    // may need to change the indexing if I ever do it with restarts
 
       }
       qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
+
+      // log iteration and update GCR polynomial if necessary.
+      all_betas.push_back(betas);
+      LogIteration(k + 1, a, betas);
+
+      // finish if necessary
+      if((k==nstep-1)||(cp<rsq)){
+        std::cout << "All alphas: " << std::endl << all_alphas << std::endl;
+        std::cout << "All betas: " << std::endl << all_betas << std::endl;
+        LogComplete(all_alphas, all_betas);
+        std::cout << "Exiting GCR." << std::endl;
+        return cp;
+      }
+
     }
     GRID_ASSERT(0); // never reached
     return cp;
   }
 };
+
+class PolynomialFile: Serializable {
+  public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(PolynomialFile, 
+      std::vector<std::vector<std::complex<double>>>, data,
+      std::vector<std::vector<std::complex<double>>>, betas,
+      std::vector<std::complex<double>>,              alphas
+    );
+};
+
+// Optionally record the GCR polynomial. [PO]: TODO
+template <class Field>
+class PGCRPolynomial : public PrecGeneralisedConjugateResidualNonHermitian<Field> {
+public:
+  std::vector<ComplexD> ak;
+  std::vector<std::vector<ComplexD>> bk;
+  // std::vector<ComplexD> poly_p;
+  std::vector<std::vector<ComplexD>> poly_p;
+  std::vector<ComplexD> poly_Ap;        // polynomial in Ap_j (only store it for last p)
+  std::vector<ComplexD> poly_r;
+  std::vector<ComplexD> polynomial;
+
+  PolynomialFile& PF;
+
+public:
+  PGCRPolynomial(RealD tol, Integer maxit,LinearOperatorBase<Field> &_Linop, LinearFunction<Field> &Prec, int _mmax, int _nstep, PolynomialFile& _PF)
+    : PrecGeneralisedConjugateResidualNonHermitian<Field>(tol, maxit, _Linop, Prec, _mmax, _nstep), PF(_PF)
+  {};
+
+  // think this applies the polynomial in A = Linop to a field src. The coeffs are 
+  // stored in the vector `polynomial`.
+  void PolyOp(const Field &src, Field &psi)
+  {
+    Field tmp(src.Grid());
+    Field AtoN(src.Grid());
+    AtoN = src;
+    psi=AtoN*polynomial[0];
+    for(int n=1;n<polynomial.size();n++){
+      tmp = AtoN;
+      this->Linop.Op(tmp,AtoN);               // iterate A^n
+      psi = psi + polynomial[n]*AtoN;       // psi += poly_n A^n src
+    }
+  }
+
+  // [PO TODO] debug this
+  void PGCRsequence(const Field &src, Field &x)
+  {
+    Field Ap(src.Grid());
+    Field r(src.Grid());
+    // Field p(src.Grid());
+    // p=src;
+    std::vector<Field> p;
+    p.push_back(src);
+    r=src;
+    x=Zero();
+    x.Checkerboard()=src.Checkerboard();
+    for(int k=0;k<ak.size();k++){
+      x = x + ak[k]*p[k];
+      this->Linop.Op(p[k], Ap);
+      r = r - ak[k] * Ap;
+      // p[k] = r;
+      p.push_back(r);
+      for (int i = 0; i < k; i++) {     // [PO TODO] check indices
+        p[k+1] += bk[i, k+1] * p[i];
+      }
+      // p = r + bk[k] * p;
+    }
+  }
+
+  void Solve(const Field &src, Field &psi)
+  {
+    psi=Zero();
+    this->operator()(src, psi);
+  }
+
+  virtual void LogBegin(void)
+  {
+    std::cout << "PGCR::LogBegin() "<<std::endl;
+    ak.resize(0);
+    bk.resize(0);
+    polynomial.resize(0);
+    poly_Ap.push_back(0.0);     // start with (0.0); during first iteration should change to (0.0, 1.0)
+    std::vector<ComplexD> p0_tmp;
+    p0_tmp.push_back(1.0);
+    poly_p.push_back(p0_tmp);
+    poly_r.push_back(1.0);
+  };
+
+  // Updates vector psi and r and initializes vector p[k+1]
+  virtual void LogIteration(int k, ComplexD a, std::vector<ComplexD> betas){
+    std::cout << "PGCR::LogIteration(k = " << k << ")" << std::endl;
+    ak.push_back(a);
+    bk.push_back(betas);
+
+    // update Ap by pushing p[k] to the right
+    poly_Ap.push_back(0.0);   // need to pad the end with an element
+    poly_Ap[0] = 0.0;         // technically this should be unnecessary, as the first component is never set
+    for(int i = 0; i < k; i++){
+      poly_Ap[i+1]=poly_p[k-1][i];        // A\vec{p} = (0, \vec{p}) bc A shifts components of p to the right
+    }
+
+    // update psi_{k+1} --> psi_k + a_k p_k
+    polynomial.push_back(0.0);
+    for(int i = 0; i < k; i++) {
+      polynomial[i] += a * poly_p[k-1][i];
+    }
+    {
+      std::vector<std::complex<double>> poly_stdcmplx(polynomial.begin(), polynomial.end());
+      PF.data.push_back(poly_stdcmplx);
+    }
+
+    //  r_{k+1} --> r_k - a_k A p_k
+    //  p_{k+1} --> r_k + \sum_{i=0}^k \beta_{ik} p_i, input betas = (\beta_{ik})_i
+    poly_r.push_back(0.0);        // should be of size k+1 if we start with k = 1
+    std::vector<ComplexD> p_next (k + 1, ComplexD(0.0));     // p_{k+1} = same size as r_{k+1}
+    for(int i = 0; i < k + 1; i++){
+      poly_r[i] = poly_r[i] - a * poly_Ap[i];     // update r_{k+1} --> r_k - \alpha_k A p_k
+      p_next[i] = poly_r[i];                 // init new vector as r_{k+1}
+    }
+
+    // p_{k+1} --> p_{k+1} + \sum_i \beta_{ij} p_i
+    int nbeta = betas.size();
+    std::cout << "Betas: " << betas << std::endl;
+    for (int j = 0; j < nbeta; j++) {
+      for (int i = 0; i < j+1; i++) {
+        p_next[i] += betas[j] * poly_p[j][i];
+      }
+    }
+    poly_p.push_back(p_next);                 // add p_{k+1} to the list of p's
+  }
+
+  virtual void LogComplete(std::vector<ComplexD>& alphas, std::vector<std::vector<ComplexD>>& betas) {
+    /** Logs all alphas and betas to complete the iterations. */
+    std::cout << "PGCR::LogComplete() "<<std::endl;
+    for (int i = 0; i < alphas.size(); i++) {
+      PF.alphas.push_back(std::complex<double>(alphas[i].real(), alphas[i].imag()));
+      std::vector<std::complex<double>> beta_stdcmplx(betas[i].begin(), betas[i].end());
+      PF.betas.push_back(beta_stdcmplx);
+    }
+  };
+
+};
+
 NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/RestartedLanczosBidiagonalization.h

@@ -0,0 +1,753 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./Grid/algorithms/iterative/RestartedLanczosBidiagonalization.h
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_RESTARTED_LANCZOS_BIDIAGONALIZATION_H
+#define GRID_RESTARTED_LANCZOS_BIDIAGONALIZATION_H
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Implicitly Restarted Lanczos Bidiagonalization (IRLBA)
+ *
+ * Computes the p largest (or p smallest) singular triplets of a linear
+ * operator A using the Golub-Kahan-Lanczos bidiagonalization with implicit
+ * restart via thick-restart / QR shifts.
+ *
+ * Algorithm (Baglama & Reichel, SIAM J. Sci. Comput. 27(1):19-42, 2005):
+ *
+ *   Outer loop:
+ *     1. Extend the p-step (or seed) bidiagonalization to k steps:
+ *           A  V_k = U_k B_k
+ *           A^dag U_k = V_k B_k^T + beta_{k+1} v_{k+1} e_k^T
+ *     2. Compute SVD:  B_k = X Sigma Y^T
+ *     3. Check convergence of the p desired singular values via
+ *           |beta_{k+1} * y_{k,i}|  <  tol * sigma_i
+ *        where y_{k,i} is the last component of the i-th right singular vector.
+ *     4. Apply k-p implicit QR shifts to implicitly compress the basis
+ *        to p steps (Sorensen-Lehoucq thick restart):
+ *           B_p^+ = X_p^T B_k Y_p   (upper bidiagonal, p x p)
+ *        and update the lattice vectors:
+ *           V_p^+ = V_k Y_p
+ *           U_p^+ = U_k X_p
+ *        The new residual coupling is
+ *           beta_p^+ v_{p+1}^+ = beta_{k+1} v_{k+1} * (e_k^T Y_p)_p
+ *                               + B_k(p,p+1) * (orthogonal tail from QR)
+ *     5. Go to step 1.
+ *
+ * Template parameter
+ * ------------------
+ *   Field : lattice field type (must support Grid algebra operations)
+ *
+ * Usage
+ * -----
+ *   RestartedLanczosBidiagonalization<Field> irlba(Linop, grid, p, k, tol, maxIter);
+ *   irlba.run(src);
+ *   // Results available via getters.
+ */
+template <class Field>
+class RestartedLanczosBidiagonalization {
+
+public:
+  LinearOperatorBase<Field> &Linop;
+  GridBase *Grid;
+
+  int    Nk;       // number of desired singular triplets
+  int    Nm;       // Lanczos basis size (Nm > Nk)
+  RealD  Tolerance;
+  int    MaxIter;
+  bool   largest; // if true, target largest singular values; otherwise smallest
+
+  // Converged singular triplets (filled after run())
+  std::vector<RealD>  singularValues;   // sigma_0 >= sigma_1 >= ...
+  std::vector<Field>  leftVectors;      // approximate left singular vectors
+  std::vector<Field>  rightVectors;     // approximate right singular vectors
+
+private:
+  // Working bases (size up to Nm+1)
+  std::vector<Field>  V;    // right Lanczos vectors
+  std::vector<Field>  U;    // left  Lanczos vectors
+  std::vector<RealD>  alpha;
+  std::vector<RealD>  beta;
+
+  // After a thick restart, the column at index restart_col of U^dag A V
+  // has extra non-zero entries (rows 0..restart_col-2) beyond what the
+  // upper bidiagonal captures.  fvec[j] = <U[j] | A V[restart_col]> for
+  // j = 0..restart_col-1.  (fvec[restart_col-1] == beta[restart_col-1].)
+  // reset_col == -1 means no restart has occurred yet (pure bidiagonal).
+  std::vector<RealD>  fvec;
+  int                 restart_col;
+
+public:
+
+  RestartedLanczosBidiagonalization(LinearOperatorBase<Field> &_Linop,
+                                    GridBase *_Grid,
+                                    int _Nk, int _Nm,
+                                    RealD _tol   = 1.0e-8,
+                                    int   _maxIt = 300,
+                                    bool  _largest = true)
+    : Linop(_Linop), Grid(_Grid),
+      Nk(_Nk), Nm(_Nm),
+      Tolerance(_tol), MaxIter(_maxIt),
+      largest(_largest)
+  {
+    assert(Nm > Nk);
+  }
+
+  /**
+   * Run IRLBA starting from src.
+   * On exit, singularValues, leftVectors, rightVectors are filled with
+   * the Nk converged singular triplets.
+   */
+  void run(const Field &src)
+  {
+    assert(norm2(src) > 0.0);
+
+    singularValues.clear();
+    leftVectors.clear();
+    rightVectors.clear();
+
+    // Allocate working bases
+    V.clear(); U.clear();
+    alpha.clear(); beta.clear();
+    fvec.clear(); restart_col = -1;
+    V.reserve(Nm + 1);
+    U.reserve(Nm);
+
+    // Seed: v_0 = src / ||src||
+    Field vtmp(Grid);
+    vtmp = src;
+    RealD nrm = std::sqrt(norm2(vtmp));
+    vtmp = (1.0 / nrm) * vtmp;
+    V.push_back(vtmp);
+
+    int pStart = 0;  // current basis size at start of extension
+    RealD betaRestart = 0.0; // coupling from previous restart
+
+    for (int iter = 0; iter < MaxIter; ++iter) {
+
+      // ----------------------------------------------------------------
+      // Step 1: extend from pStart steps to Nm steps
+      // ----------------------------------------------------------------
+      extendBasis(pStart, Nm, betaRestart);
+//      verify();
+
+      // ----------------------------------------------------------------
+      // Step 2: SVD of the Nm x Nm B matrix.
+      // iter=0 (pStart==0): B is exactly bidiagonal — use buildBidiagonal.
+      // iter>0 (pStart==Nk): after a thick restart, column restart_col of
+      // U^dag A V has extra off-diagonal entries captured by fvec; use
+      // buildFullB so the Ritz values and restart vectors are computed from
+      // the exact projected matrix A V = U B_full.
+      // ----------------------------------------------------------------
+      Eigen::MatrixXd B = (pStart == 0) ? buildBidiagonal(Nm) : buildFullB(Nm);
+      Eigen::JacobiSVD<Eigen::MatrixXd> svd(B,
+          Eigen::ComputeThinU | Eigen::ComputeThinV);
+
+      Eigen::VectorXd sigma = svd.singularValues();  // descending
+      Eigen::MatrixXd X     = svd.matrixU();          // Nm x Nm left SVecs of B
+      Eigen::MatrixXd Y     = svd.matrixV();          // Nm x Nm right SVecs of B
+
+      // If targeting smallest, reorder so desired ones come first
+      Eigen::VectorXi order = sortOrder(sigma);
+
+      // ----------------------------------------------------------------
+      // Step 3: check convergence of the Nk desired singular values
+      // ----------------------------------------------------------------
+      RealD betaK = beta.back();  // beta_{k+1}
+      // In our convention A V = U B (exact), the residual is in the A^dag
+      // direction: A^dag u_j - sigma_j v_j = betaK * X[Nm-1,j] * V[Nm].
+      // Convergence criterion: |betaK * X[Nm-1, idx]| < tol * sigma_idx.
+      int nconv = 0;
+      for (int i = 0; i < Nk; ++i) {
+        int idx = order(i);
+        RealD res = std::abs(betaK * X(Nm - 1, idx));
+        RealD thr = Tolerance * std::max(sigma(idx), 1.0e-14);
+        std::cout << GridLogMessage
+                  << "IRLBA iter " << iter
+                  << "  sigma[" << i << "] = " << sigma(idx)
+                  << "  res = " << res
+                  << "  thr = " << thr << std::endl;
+        if (res < thr) ++nconv;
+        else break;  // residuals not strictly ordered but break is conservative
+      }
+
+      if (nconv >= Nk) {
+        std::cout << GridLogMessage
+                  << "IRLBA converged: " << nconv << " singular values after "
+                  << iter + 1 << " iterations." << std::endl;
+        // Collect converged triplets
+        extractTriplets(Nm, sigma, X, Y, order, Nk);
+        return;
+      }
+
+      // ----------------------------------------------------------------
+      // Step 4: implicit restart — compress to Nk steps
+      // ----------------------------------------------------------------
+      implicitRestart(Nm, Nk, sigma, X, Y, order, betaK, betaRestart);
+//      verify();
+
+      // Lucky breakdown: exact invariant subspace found; convergence is exact.
+      // B_p^+ = diag(alpha[0..Nk-1]); extract directly from restart basis.
+      if (betaRestart < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "IRLBA: lucky breakdown after restart (betaRestart = 0)."
+                  << " Extracting " << Nk << " exact Ritz triplets." << std::endl;
+        // Re-run SVD on the p-step diagonal B^+ to get sorted Ritz triplets.
+        Eigen::MatrixXd Bp = buildBidiagonal(Nk);
+        Eigen::JacobiSVD<Eigen::MatrixXd> svdp(Bp,
+            Eigen::ComputeThinU | Eigen::ComputeThinV);
+        Eigen::VectorXi ordp = sortOrder(svdp.singularValues());
+        extractTriplets(Nk, svdp.singularValues(), svdp.matrixU(),
+                        svdp.matrixV(), ordp, Nk);
+        return;
+      }
+
+      pStart = Nk;
+    }
+
+    std::cout << GridLogMessage
+              << "IRLBA: did not converge in " << MaxIter
+              << " iterations. Returning best approximations." << std::endl;
+
+    // Return best available approximations
+    Eigen::MatrixXd B = buildFullB((int)alpha.size());
+    Eigen::JacobiSVD<Eigen::MatrixXd> svd(B,
+        Eigen::ComputeThinU | Eigen::ComputeThinV);
+    Eigen::VectorXd sigma = svd.singularValues();
+    Eigen::MatrixXd X     = svd.matrixU();
+    Eigen::MatrixXd Y     = svd.matrixV();
+    Eigen::VectorXi order = sortOrder(sigma);
+    int nout = std::min(Nk, (int)alpha.size());
+    extractTriplets((int)alpha.size(), sigma, X, Y, order, nout);
+  }
+
+  /* Getters */
+  int getNk() const { return (int)singularValues.size(); }
+  const std::vector<RealD>&  getSingularValues() const { return singularValues; }
+  const std::vector<Field>&  getLeftVectors()    const { return leftVectors; }
+  const std::vector<Field>&  getRightVectors()   const { return rightVectors; }
+
+  /**
+   * Print B_k and U^dag A V to verify the bidiagonalization relation
+   *   A V_m = U_m B_m   (exact in our GK convention)
+   * On the first call (pStart=0), max|B - U^dag A V| should be ~machine epsilon.
+   * After a restart and extension, the column p of U^dag A V deviates from B
+   * by O(betaK): this is expected because the thick restart breaks the Krylov
+   * structure at column p, introducing off-diagonal terms proportional to betaK.
+   * These terms vanish as betaK -> 0 (convergence), so the algorithm is correct.
+   */
+  void verify()
+  {
+    int m  = (int)alpha.size();
+    int nU = (int)U.size();
+    int nV = (int)V.size();
+    if (m == 0) { std::cout << GridLogMessage << "IRLBA verify: empty basis" << std::endl; return; }
+
+    // Build reference matrix Bref (nU x nV):
+    //   Columns 0..m-1 : buildFullB(m)  (bidiagonal + fvec column at restart_col)
+    //   Column  m      : residual column, two cases:
+    //     (a) restart_col == m (right after implicitRestart, before extendBasis):
+    //         V[m] = sgn*V_old[Nm], so <U[i]|A|V[m]> = fvec[i] for all i
+    //     (b) otherwise (pure GK or after extendBasis):
+    //         only entry (m-1, m) = beta[m-1]  (GK recurrence residual)
+    Eigen::MatrixXd Bref = Eigen::MatrixXd::Zero(nU, nV);
+    {
+      Eigen::MatrixXd Bfull = buildFullB(m);
+      int cols = std::min(m, nV);
+      Bref.block(0, 0, m, cols) = Bfull.block(0, 0, m, cols);
+    }
+    if (nV > m && m > 0) {
+      if (restart_col == m && (int)fvec.size() == m) {
+        // Case (a): right after implicitRestart
+        for (int i = 0; i < m; ++i) Bref(i, m) = fvec[i];
+      } else if ((int)beta.size() >= m) {
+        // Case (b): standard GK residual column
+        Bref(m - 1, m) = beta[m - 1];
+      }
+    }
+
+    // Compute M[i,j] = <U[i] | A | V[j]>
+    Eigen::MatrixXd M = Eigen::MatrixXd::Zero(nU, nV);
+    Field Avj(Grid);
+    for (int j = 0; j < nV; ++j) {
+      Linop.Op(V[j], Avj);
+      for (int i = 0; i < nU; ++i) {
+        ComplexD ip = innerProduct(U[i], Avj);
+        M(i, j) = ip.real();
+      }
+    }
+
+    // Print Bref
+    std::cout << GridLogMessage
+              << "IRLBA verify: Bref (" << nU << "x" << nV << "):" << std::endl;
+    for (int i = 0; i < nU; ++i) {
+      std::cout << GridLogMessage << "  row " << i << ": ";
+      for (int j = 0; j < nV; ++j) std::cout << Bref(i,j) << " ";
+      std::cout << std::endl;
+    }
+
+    // Print U^dag A V
+    std::cout << GridLogMessage
+              << "IRLBA verify: U^dag A V (" << nU << "x" << nV << "):" << std::endl;
+    for (int i = 0; i < nU; ++i) {
+      std::cout << GridLogMessage << "  row " << i << ": ";
+      for (int j = 0; j < nV; ++j) std::cout << M(i,j) << " ";
+      std::cout << std::endl;
+    }
+
+    // Max deviation over the full nU x nV matrix
+    RealD maxdev = (Bref - M).cwiseAbs().maxCoeff();
+    std::cout << GridLogMessage
+              << "IRLBA verify: max|Bref - U^dag A V| = " << maxdev << std::endl;
+
+    // Beta
+    std::cout << GridLogMessage << "IRLBA verify: beta[0.." << (int)beta.size()-1 << "] = ";
+    for (auto b : beta) std::cout << b << " ";
+    std::cout << std::endl;
+  }
+
+private:
+
+  // ------------------------------------------------------------------
+  // Build the m x m upper-bidiagonal matrix from alpha[0..m-1], beta[0..m-2]
+  // ------------------------------------------------------------------
+  Eigen::MatrixXd buildBidiagonal(int m) const
+  {
+    Eigen::MatrixXd B = Eigen::MatrixXd::Zero(m, m);
+    for (int k = 0; k < m; ++k) {
+      B(k, k) = alpha[k];
+      if (k + 1 < m && k < (int)beta.size())
+        B(k, k + 1) = beta[k];
+    }
+    return B;
+  }
+
+  // ------------------------------------------------------------------
+  // Build the full m x m B matrix, including the non-bidiagonal column
+  // at restart_col that arises after a thick restart.
+  //
+  // After restart, A V[restart_col] has projections onto all U[0..restart_col-1]
+  // (not just U[restart_col-1]).  These are stored in fvec[0..restart_col-1]
+  // and make column restart_col of U^dag A V non-bidiagonal.
+  // ------------------------------------------------------------------
+  Eigen::MatrixXd buildFullB(int m) const
+  {
+    Eigen::MatrixXd B = buildBidiagonal(m);
+    if (restart_col >= 0 && restart_col < m && (int)fvec.size() > 0) {
+      for (int j = 0; j < restart_col && j < (int)fvec.size(); ++j){
+        B(j, restart_col) = fvec[j];
+        std::cout << GridLogDebug << "buildFullB: B  " <<j<<" "<<restart_col<<B(j, restart_col)<<std::endl;
+      }
+    }
+    return B;
+  }
+
+  // ------------------------------------------------------------------
+  // Return a permutation vector that puts the desired Nk singular values
+  // first (largest first if largest==true, smallest first otherwise).
+  // Eigen's JacobiSVD already returns sigma in descending order, so for
+  // largest we just return 0,1,...,m-1; for smallest we reverse.
+  // ------------------------------------------------------------------
+  Eigen::VectorXi sortOrder(const Eigen::VectorXd &sigma) const
+  {
+    int m = (int)sigma.size();
+    Eigen::VectorXi ord(m);
+    if (largest) {
+      for (int i = 0; i < m; ++i) ord(i) = i;
+    } else {
+      for (int i = 0; i < m; ++i) ord(i) = m - 1 - i;
+    }
+    return ord;
+  }
+
+  // ------------------------------------------------------------------
+  // Extend the Lanczos bidiagonalization from pStart to kEnd steps.
+  // On first call pStart==0 (V[0] already set).
+  // On restart calls V[0..pStart], U[0..pStart-1], alpha[0..pStart-1],
+  // beta[0..pStart-1] are already set; betaRestart is the coupling
+  // beta_{pStart} that drives the first new U step.
+  // ------------------------------------------------------------------
+  void extendBasis(int pStart, int kEnd, RealD betaRestart)
+  {
+    // Truncate containers to pStart (Lattice has no default constructor)
+    if ((int)V.size() > pStart + 1) V.erase(V.begin() + pStart + 1, V.end());
+    if ((int)U.size() > pStart)     U.erase(U.begin() + pStart,     U.end());
+    alpha.resize(pStart);
+    beta.resize(pStart);
+
+    Field p(Grid), r(Grid);
+
+    for (int k = pStart; k < kEnd; ++k) {
+
+      // p = A v_k
+      Linop.Op(V[k], p);
+
+      // Remove previous left vector coupling
+      if (k > 0) {
+        p = p - beta[k - 1] * U[k - 1];
+      }
+      // On the first step after a restart, beta[pStart-1] was already set;
+      // but V[pStart] was already constructed including the beta correction,
+      // so no extra subtraction needed here beyond the standard recurrence.
+
+      // Reorthogonalize p against U, then alpha_k = ||p||, u_k = p/alpha_k
+      reorthogonalize(p, U);
+      RealD ak = std::sqrt(norm2(p));
+      if (ak < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "IRLBA extendBasis: lucky breakdown at step " << k
+                  << " (alpha = " << ak << ")" << std::endl;
+        alpha.push_back(ak);
+        Field zero(Grid); zero = Zero();
+        U.push_back(zero);
+        beta.push_back(0.0);
+        V.push_back(zero);
+        break;
+      }
+      alpha.push_back(ak);
+
+      Field u(Grid);
+      u = (1.0 / ak) * p;
+      U.push_back(u);
+
+      // r = A^dag u_k - alpha_k v_k, reorthogonalize, then beta_{k+1} = ||r||
+      Linop.AdjOp(U[k], r);
+      r = r - ak * V[k];
+      reorthogonalize(r, V);
+
+      RealD bk = std::sqrt(norm2(r));
+      beta.push_back(bk);
+
+      std::cout << GridLogMessage
+                << "IRLBA extend step " << k
+                << "  alpha = " << ak
+                << "  beta  = " << bk << std::endl;
+
+      // Always push v_{k+1} (needed as residual direction for restart)
+      if (bk < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "IRLBA extendBasis: lucky breakdown (beta = 0) at step "
+                  << k << std::endl;
+        Field zero(Grid); zero = Zero();
+        V.push_back(zero);
+        break;
+      }
+      Field vnext(Grid);
+      vnext = (1.0 / bk) * r;
+      V.push_back(vnext);
+
+      if (k == kEnd - 1) break;  // v_{k+1} pushed above; stop here
+    }
+  }
+
+public:
+  // ------------------------------------------------------------------
+  // Block reorthogonalization helpers.
+  // Declared public because CUDA extended lambdas cannot live inside
+  // private/protected member functions.
+  //
+  // batchInnerProducts: computes c[j] = <basis[j], vec> for all j
+  //   in a single GPU pass (one accelerator_barrier instead of n).
+  //   Queues n pairs of (per-site kernel, reduceKernel) to computeStream
+  //   without intermediate CPU syncs, then syncs once at the end.
+  //
+  // batchUpdate: computes vec -= sum_j c[j]*basis[j] in one GPU kernel.
+  //
+  // reorthogonalize: two-pass Classical Gram-Schmidt (CGS2) using the
+  //   two helpers above.  Each pass costs 2 GPU syncs (1 IP + 1 update)
+  //   instead of 2n syncs per pass in the old sequential MGS.
+  // ------------------------------------------------------------------
+
+  void batchInnerProducts(const Field &vec,
+                          const std::vector<Field> &basis,
+                          std::vector<ComplexD> &c)
+  {
+    int n = (int)basis.size();
+    c.resize(n);
+    if (n == 0) return;
+
+    typedef typename Field::vector_object         vobj;
+    typedef decltype(innerProduct(vobj(), vobj())) inner_t;
+    typedef decltype(basis[0].View(AcceleratorRead)) View;
+
+    GridBase *grid = vec.Grid();
+    uint64_t oSites = grid->oSites();
+    uint64_t nsimd  = grid->Nsimd();
+
+    // all_ip[j * oSites + ss] = per-site inner product of basis[j] and vec at site ss.
+    // Layout: n contiguous blocks of oSites each.
+    deviceVector<inner_t> all_ip((uint64_t)n * oSites);
+    inner_t *all_ip_p = &all_ip[0];
+
+    hostVector<View>   h_basis_v(n);
+    deviceVector<View> d_basis_v(n);
+    for (int j = 0; j < n; ++j) {
+      h_basis_v[j] = basis[j].View(AcceleratorRead);
+      acceleratorPut(d_basis_v[j], h_basis_v[j]);
+    }
+    View *basis_vp = &d_basis_v[0];
+
+    // Queue n per-site kernels to the accelerator stream — no intermediate barriers.
+    autoView(vec_v, vec, AcceleratorRead);
+    for (int j = 0; j < n; ++j) {
+      int      jj      = j;
+      uint64_t oSites_ = oSites;
+      accelerator_for(ss, oSites, nsimd, {
+        auto x = coalescedRead(basis_vp[jj][ss]);
+        auto y = coalescedRead(vec_v[ss]);
+        coalescedWrite(all_ip_p[jj * oSites_ + ss], innerProduct(x, y));
+      });
+    }
+
+    // ONE sync after all n kernels
+    accelerator_barrier();
+
+    // Copy all per-site results to host
+    hostVector<inner_t> all_ip_h((uint64_t)n * oSites);
+    acceleratorCopyFromDevice(all_ip_p, &all_ip_h[0], (uint64_t)n * oSites * sizeof(inner_t));
+
+    // Reduce on host: sum over oSites, then collapse SIMD lanes via Reduce(TensorRemove(...))
+    // TensorRemove strips the iSinglet tensor wrapper to expose the SIMD scalar type.
+    // Reduce sums all nsimd lanes and returns a plain scalar (RealD or ComplexD).
+    std::vector<ComplexD> raw(n);
+    for (int j = 0; j < n; ++j) {
+      inner_t sum = Zero();
+      for (uint64_t ss = 0; ss < oSites; ++ss)
+        sum += all_ip_h[(uint64_t)j * oSites + ss];
+      raw[j] = ComplexD(Reduce(TensorRemove(sum)));
+    }
+    grid->GlobalSumVector(&raw[0], n);
+    for (int j = 0; j < n; ++j) c[j] = raw[j];
+
+    for (int j = 0; j < n; ++j) h_basis_v[j].ViewClose();
+  }
+
+  void batchUpdate(Field &vec,
+                   const std::vector<Field> &basis,
+                   const std::vector<ComplexD> &c)
+  {
+    int n = (int)basis.size();
+    if (n == 0) return;
+
+    typedef typename Field::vector_object vobj;
+    typedef decltype(basis[0].View(AcceleratorRead)) View;
+
+    GridBase *grid = vec.Grid();
+    uint64_t oSites = grid->oSites();
+    uint64_t nsimd  = grid->Nsimd();
+
+    // Split complex coefficients into real/imag double arrays on device.
+    // Using doubles avoids potential ComplexD-device-code compatibility issues.
+    hostVector<double>   h_re(n), h_im(n);
+    deviceVector<double> d_re(n), d_im(n);
+    for (int k = 0; k < n; ++k) {
+      h_re[k] = c[k].real();
+      h_im[k] = c[k].imag();
+    }
+    acceleratorCopyToDevice(&h_re[0], &d_re[0], n * sizeof(double));
+    acceleratorCopyToDevice(&h_im[0], &d_im[0], n * sizeof(double));
+    double *re_p = &d_re[0];
+    double *im_p = &d_im[0];
+
+    // Basis views
+    hostVector<View>   h_basis_v(n);
+    deviceVector<View> d_basis_v(n);
+    for (int k = 0; k < n; ++k) {
+      h_basis_v[k] = basis[k].View(AcceleratorRead);
+      acceleratorPut(d_basis_v[k], h_basis_v[k]);
+    }
+    View *basis_vp = &d_basis_v[0];
+
+    // Single kernel: vec[ss] -= sum_k (re[k] + i*im[k]) * basis[k][ss]
+    autoView(vec_v, vec, AcceleratorWrite);
+    accelerator_for(ss, oSites, nsimd, {
+      auto v = coalescedRead(vec_v[ss]);
+      for (int k = 0; k < n; ++k) {
+        auto b = coalescedRead(basis_vp[k][ss]);
+        v = v - re_p[k] * b - timesI(im_p[k] * b);
+      }
+      coalescedWrite(vec_v[ss], v);
+    });
+
+    for (int k = 0; k < n; ++k) h_basis_v[k].ViewClose();
+  }
+
+  // ------------------------------------------------------------------
+  // Full reorthogonalization using two-pass Classical Gram-Schmidt (CGS2).
+  // Each pass calls batchInnerProducts (1 GPU sync) + batchUpdate (1 sync),
+  // replacing the old 2n GPU syncs per pass from sequential MGS.
+  // ------------------------------------------------------------------
+  void reorthogonalize(Field &vec, const std::vector<Field> &basis)
+  {
+    if (basis.empty()) return;
+    std::vector<ComplexD> c;
+    for (int pass = 0; pass < 2; ++pass) {
+      batchInnerProducts(vec, basis, c);
+      batchUpdate(vec, basis, c);
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Implicit restart: given the Nm-step bidiagonalization and its SVD,
+  // compress to Nk steps via implicit QR shifts applied to B_k.
+  //
+  // The "shifts" are the Nm - Nk singular values we want to deflate
+  // (those NOT in the desired set).  We apply them as implicit QR steps
+  // to the bidiagonal matrix, then update the lattice bases accordingly.
+  //
+  // After this call:
+  //   V[0..Nk],  U[0..Nk-1],  alpha[0..Nk-1],  beta[0..Nk-1]  are updated.
+  //   betaRestart  ← new beta_Nk coupling for the next extension.
+  // ------------------------------------------------------------------
+  void implicitRestart(int k, int p,
+                       const Eigen::VectorXd &sigma,
+                       const Eigen::MatrixXd &X,
+                       const Eigen::MatrixXd &Y,
+                       const Eigen::VectorXi &order,
+                       RealD betaK,
+                       RealD &betaRestart)
+  {
+    // Thick restart (Baglama & Reichel, Sec. 2.2):
+    //
+    // Given B_k = X Sigma Y^T, define the new p-step basis by:
+    //   V^+_i = V_k * y_{order(i)}      (right sing. vec. of B_k)
+    //   U^+_i = U_k * x_{order(i)}      (left  sing. vec. of B_k)
+    //
+    // Then A V^+_i = A V_k y_{order(i)} = U_k B_k y_{order(i)}
+    //             = sigma_{order(i)} U_k x_{order(i)} = sigma_{order(i)} U^+_i
+    //
+    // So B_p^+ = diag(sigma_{order(0)}, ..., sigma_{order(p-1)}) — DIAGONAL,
+    // all internal betas are zero.
+    //
+    // The residual coupling comes from A^dag U_k = V_k B_k^T + betaK V[k] e_{k-1}^T:
+    //   A^dag U^+_{p-1} - sigma_{order(p-1)} V^+_{p-1}
+    //     = V_k (B_k^T x_{order(p-1)} - sigma_{order(p-1)} y_{order(p-1)})
+    //       + betaK * X(k-1, order(p-1)) * V[k]
+    //     = betaK * X(k-1, order(p-1)) * V[k]   (since B_k^T x_j = sigma_j y_j)
+    //
+    // Therefore: betaRestart = |betaK * X(k-1, order(p-1))|
+    //            V[p] = sign(X(k-1, order(p-1))) * V[k]
+
+    // ---- Build new lattice vectors ----
+    std::vector<Field> Vnew, Unew;
+    Vnew.reserve(p + 1);
+    Unew.reserve(p);
+
+    for (int i = 0; i < p; ++i) {
+      int idx = order(i);
+      Field vi(Grid); vi = Zero();
+      for (int j = 0; j < k; ++j)
+        vi = vi + Y(j, idx) * V[j];
+      Vnew.push_back(vi);
+    }
+
+    for (int i = 0; i < p; ++i) {
+      int idx = order(i);
+      Field ui(Grid); ui = Zero();
+      for (int j = 0; j < k; ++j)
+        ui = ui + X(j, idx) * U[j];
+      Unew.push_back(ui);
+    }
+
+    // New v_{p} (0-indexed: V[p]) = sign * V[k]
+    // From A^dag U_k = V_k B_k^T + betaK V[k] e_{k-1}^T:
+    //   A^dag U^+_j - sigma_j V^+_j = betaK * X(k-1, order(j)) * V[k]
+    // The last Ritz pair (j=p-1) defines betaRestart and the sign of V[p].
+    // All p couplings (j=0..p-1) are stored in fvec so that buildFullB can
+    // reconstruct the exact column p of U^dag A V after the next extension.
+    RealD coeff = betaK * X(k - 1, order(p - 1));
+    betaRestart  = std::abs(coeff);
+    RealD sgn = (coeff >= 0.0) ? 1.0 : -1.0;
+
+    fvec.resize(p);
+    for (int j = 0; j < p; ++j)
+      fvec[j] = betaK * X(k - 1, order(j)) * sgn;
+    // fvec[p-1] == betaRestart by construction
+    restart_col = p;
+
+    Field vp(Grid);
+    if (betaRestart > 1.0e-14) {
+      vp = sgn * V[k];
+    } else {
+      betaRestart = 0.0;
+      vp = Zero();
+    }
+    Vnew.push_back(vp);  // V[p]
+
+    // ---- New alpha, beta ----
+    // B_p^+ is diagonal: alpha^+_i = sigma_{order(i)}, all internal beta = 0
+    std::vector<RealD> alpha_new(p), beta_new(p);
+    for (int i = 0; i < p; ++i) alpha_new[i] = sigma(order(i));
+    for (int i = 0; i < p - 1; ++i) beta_new[i] = 0.0;
+    beta_new[p - 1] = betaRestart;
+
+    // ---- Commit new state ----
+    V = Vnew;
+    U = Unew;
+    alpha = alpha_new;
+    beta  = beta_new;
+
+    std::cout << GridLogMessage
+              << "IRLBA restart: compressed to " << p << " steps,"
+              << "  new beta_p = " << betaRestart << std::endl;
+  }
+
+  // ------------------------------------------------------------------
+  // Extract the desired singular triplets into the public output vectors.
+  // ------------------------------------------------------------------
+  void extractTriplets(int m,
+                       const Eigen::VectorXd &sigma,
+                       const Eigen::MatrixXd &X,
+                       const Eigen::MatrixXd &Y,
+                       const Eigen::VectorXi &order,
+                       int nout)
+  {
+    singularValues.resize(nout);
+    leftVectors.clear();   leftVectors.reserve(nout);
+    rightVectors.clear();  rightVectors.reserve(nout);
+
+    for (int i = 0; i < nout; ++i) {
+      int idx = order(i);
+      singularValues[i] = sigma(idx);
+
+      // Left singular vector of A:  svec_L = U_m * x_i
+      Field svL(Grid); svL = Zero();
+      for (int j = 0; j < m && j < (int)U.size(); ++j)
+        svL = svL + X(j, idx) * U[j];
+      leftVectors.push_back(svL);
+
+      // Right singular vector of A:  svec_R = V_m * y_i
+      Field svR(Grid); svR = Zero();
+      for (int j = 0; j < m && j < (int)V.size(); ++j)
+        svR = svR + Y(j, idx) * V[j];
+      rightVectors.push_back(svR);
+    }
+  }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/SimpleLanczos.h

@@ -0,0 +1,931 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
+
+    Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#ifndef GRID_LANC_H
+#define GRID_LANC_H
+
+#include <string.h>		//memset
+
+#ifdef USE_LAPACK
+#ifdef USE_MKL
+#include<mkl_lapack.h>
+#else
+void LAPACK_dstegr (char *jobz, char *range, int *n, double *d, double *e,
+		    double *vl, double *vu, int *il, int *iu, double *abstol,
+		    int *m, double *w, double *z, int *ldz, int *isuppz,
+		    double *work, int *lwork, int *iwork, int *liwork,
+		    int *info);
+//#include <lapacke/lapacke.h>
+#endif
+#endif
+
+//#include <Grid/algorithms/densematrix/DenseMatrix.h>
+
+// eliminate temorary vector in calc()
+#define MEM_SAVE
+
+namespace Grid
+{
+
+  struct Bisection
+  {
+
+#if 0
+    static void get_eig2 (int row_num, std::vector < RealD > &ALPHA,
+			  std::vector < RealD > &BETA,
+			  std::vector < RealD > &eig)
+    {
+      int i, j;
+        std::vector < RealD > evec1 (row_num + 3);
+        std::vector < RealD > evec2 (row_num + 3);
+      RealD eps2;
+        ALPHA[1] = 0.;
+        BETHA[1] = 0.;
+      for (i = 0; i < row_num - 1; i++)
+	{
+	  ALPHA[i + 1] = A[i * (row_num + 1)].real ();
+	  BETHA[i + 2] = A[i * (row_num + 1) + 1].real ();
+	}
+      ALPHA[row_num] = A[(row_num - 1) * (row_num + 1)].real ();
+        bisec (ALPHA, BETHA, row_num, 1, row_num, 1e-10, 1e-10, evec1, eps2);
+        bisec (ALPHA, BETHA, row_num, 1, row_num, 1e-16, 1e-16, evec2, eps2);
+
+      // Do we really need to sort here?
+      int begin = 1;
+      int end = row_num;
+      int swapped = 1;
+      while (swapped)
+	{
+	  swapped = 0;
+	  for (i = begin; i < end; i++)
+	    {
+	      if (mag (evec2[i]) > mag (evec2[i + 1]))
+		{
+		  swap (evec2 + i, evec2 + i + 1);
+		  swapped = 1;
+		}
+	    }
+	  end--;
+	  for (i = end - 1; i >= begin; i--)
+	    {
+	      if (mag (evec2[i]) > mag (evec2[i + 1]))
+		{
+		  swap (evec2 + i, evec2 + i + 1);
+		  swapped = 1;
+		}
+	    }
+	  begin++;
+	}
+
+      for (i = 0; i < row_num; i++)
+	{
+	  for (j = 0; j < row_num; j++)
+	    {
+	      if (i == j)
+		H[i * row_num + j] = evec2[i + 1];
+	      else
+		H[i * row_num + j] = 0.;
+	    }
+	}
+    }
+#endif
+
+    static void bisec (std::vector < RealD > &c,
+		       std::vector < RealD > &b,
+		       int n,
+		       int m1,
+		       int m2,
+		       RealD eps1,
+		       RealD relfeh, std::vector < RealD > &x, RealD & eps2)
+    {
+      std::vector < RealD > wu (n + 2);
+
+      RealD h, q, x1, xu, x0, xmin, xmax;
+      int i, a, k;
+
+      b[1] = 0.0;
+      xmin = c[n] - fabs (b[n]);
+      xmax = c[n] + fabs (b[n]);
+      for (i = 1; i < n; i++)
+	{
+	  h = fabs (b[i]) + fabs (b[i + 1]);
+	  if (c[i] + h > xmax)
+	    xmax = c[i] + h;
+	  if (c[i] - h < xmin)
+	    xmin = c[i] - h;
+	}
+      xmax *= 2.;
+
+      eps2 = relfeh * ((xmin + xmax) > 0.0 ? xmax : -xmin);
+      if (eps1 <= 0.0)
+	eps1 = eps2;
+      eps2 = 0.5 * eps1 + 7.0 * (eps2);
+      x0 = xmax;
+      for (i = m1; i <= m2; i++)
+	{
+	  x[i] = xmax;
+	  wu[i] = xmin;
+	}
+
+      for (k = m2; k >= m1; k--)
+	{
+	  xu = xmin;
+	  i = k;
+	  do
+	    {
+	      if (xu < wu[i])
+		{
+		  xu = wu[i];
+		  i = m1 - 1;
+		}
+	      i--;
+	    }
+	  while (i >= m1);
+	  if (x0 > x[k])
+	    x0 = x[k];
+	  while ((x0 - xu) > 2 * relfeh * (fabs (xu) + fabs (x0)) + eps1)
+	    {
+	      x1 = (xu + x0) / 2;
+
+	      a = 0;
+	      q = 1.0;
+	      for (i = 1; i <= n; i++)
+		{
+		  q =
+		    c[i] - x1 -
+		    ((q != 0.0) ? b[i] * b[i] / q : fabs (b[i]) / relfeh);
+		  if (q < 0)
+		    a++;
+		}
+//      printf("x1=%0.14e a=%d\n",x1,a);
+	      if (a < k)
+		{
+		  if (a < m1)
+		    {
+		      xu = x1;
+		      wu[m1] = x1;
+		    }
+		  else
+		    {
+		      xu = x1;
+		      wu[a + 1] = x1;
+		      if (x[a] > x1)
+			x[a] = x1;
+		    }
+		}
+	      else
+		x0 = x1;
+	    }
+	  printf ("x0=%0.14e xu=%0.14e k=%d\n", x0, xu, k);
+	  x[k] = (x0 + xu) / 2;
+	}
+    }
+  };
+
+/////////////////////////////////////////////////////////////
+// Implicitly restarted lanczos
+/////////////////////////////////////////////////////////////
+
+
+  template < class Field > class SimpleLanczos
+  {
+
+    const RealD small = 1.0e-16;
+  public:
+    int lock;
+    int get;
+    int Niter;
+    int converged;
+
+    int Nstop;			// Number of evecs checked for convergence
+    int Nk;			// Number of converged sought
+    int Np;			// Np -- Number of spare vecs in kryloc space
+    int Nm;			// Nm -- total number of vectors
+
+
+    RealD OrthoTime;
+
+    RealD eresid;
+
+//    SortEigen < Field > _sort;
+
+    LinearFunction < Field > &_Linop;
+
+//    OperatorFunction < Field > &_poly;
+
+    /////////////////////////
+    // Constructor
+    /////////////////////////
+    void init (void)
+    {
+    };
+//    void Abort (int ff, std::vector < RealD > &evals, DenseVector < Denstd::vector  < RealD > >&evecs);
+
+    SimpleLanczos (LinearFunction < Field > &Linop,	// op
+//		   OperatorFunction < Field > &poly,	// polynmial
+		   int _Nstop,	// sought vecs
+		   int _Nk,	// sought vecs
+		   int _Nm,	// spare vecs
+		   RealD _eresid,	// resid in lmdue deficit 
+		   int _Niter):	// Max iterations
+     
+      _Linop (Linop),
+ //     _poly (poly),
+      Nstop (_Nstop), Nk (_Nk), Nm (_Nm), eresid (_eresid), Niter (_Niter)
+    {
+      Np = Nm - Nk;
+      assert (Np > 0);
+    };
+
+    /////////////////////////
+    // Sanity checked this routine (step) against Saad.
+    /////////////////////////
+    void RitzMatrix (std::vector < Field > &evec, int k)
+    {
+
+      if (1)
+	return;
+
+      GridBase *grid = evec[0].Grid();
+      Field w (grid);
+      std::cout << GridLogMessage << "RitzMatrix " << std::endl;
+      for (int i = 0; i < k; i++)
+	{
+	  _Linop(evec[i], w);
+//      _poly(_Linop,evec[i],w);
+	  std::cout << GridLogMessage << "[" << i << "] ";
+	  for (int j = 0; j < k; j++)
+	    {
+	      ComplexD in = innerProduct (evec[j], w);
+	      if (fabs ((double) i - j) > 1)
+		{
+		  if (abs (in) > 1.0e-9)
+		    {
+		      std::cout << GridLogMessage << "oops" << std::endl;
+		      abort ();
+		    }
+		  else
+		    std::cout << GridLogMessage << " 0 ";
+		}
+	      else
+		{
+		  std::cout << GridLogMessage << " " << in << " ";
+		}
+	    }
+	  std::cout << GridLogMessage << std::endl;
+	}
+    }
+
+    void step (std::vector < RealD > &lmd,
+	       std::vector < RealD > &lme,
+	       Field & last, Field & current, Field & next, uint64_t k)
+    {
+      if (lmd.size () <= k)
+	lmd.resize (k + Nm);
+      if (lme.size () <= k)
+	lme.resize (k + Nm);
+
+
+//      _poly(_Linop,current,next );   // 3. wk:=Avk−βkv_{k−1}
+      _Linop(current, next);	// 3. wk:=Avk−βkv_{k−1}
+      if (k > 0)
+	{
+	  next -= lme[k - 1] * last;
+	}
+//      std::cout<<GridLogMessage << "<last|next>" << innerProduct(last,next) <<std::endl;
+
+      ComplexD zalph = innerProduct (current, next);	// 4. αk:=(wk,vk)
+      RealD alph = real (zalph);
+
+      next = next - alph * current;	// 5. wk:=wk−αkvk
+//      std::cout<<GridLogMessage << "<current|next>" << innerProduct(current,next) <<std::endl;
+
+      RealD beta = normalise (next);	// 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+      // 7. vk+1 := wk/βk+1
+//       norm=beta;
+
+      int interval = Nm / 100 + 1;
+      if ((k % interval) == 0)
+	std::
+	  cout << GridLogMessage << k << " : alpha = " << zalph << " beta " <<
+	  beta << std::endl;
+      const RealD tiny = 1.0e-20;
+      if (beta < tiny)
+	{
+	  std::cout << GridLogMessage << " beta is tiny " << beta << std::
+	    endl;
+	}
+      lmd[k] = alph;
+      lme[k] = beta;
+
+    }
+
+    void qr_decomp (std::vector < RealD > &lmd,
+		    std::vector  < RealD > &lme,
+		    int Nk,
+		    int Nm,
+		    std::vector  < RealD > &Qt, RealD Dsh, int kmin, int kmax)
+    {
+      int k = kmin - 1;
+      RealD x;
+
+      RealD Fden = 1.0 / hypot (lmd[k] - Dsh, lme[k]);
+      RealD c = (lmd[k] - Dsh) * Fden;
+      RealD s = -lme[k] * Fden;
+
+      RealD tmpa1 = lmd[k];
+      RealD tmpa2 = lmd[k + 1];
+      RealD tmpb = lme[k];
+
+      lmd[k] = c * c * tmpa1 + s * s * tmpa2 - 2.0 * c * s * tmpb;
+      lmd[k + 1] = s * s * tmpa1 + c * c * tmpa2 + 2.0 * c * s * tmpb;
+      lme[k] = c * s * (tmpa1 - tmpa2) + (c * c - s * s) * tmpb;
+      x = -s * lme[k + 1];
+      lme[k + 1] = c * lme[k + 1];
+
+      for (int i = 0; i < Nk; ++i)
+	{
+	  RealD Qtmp1 = Qt[i + Nm * k];
+	  RealD Qtmp2 = Qt[i + Nm * (k + 1)];
+	  Qt[i + Nm * k] = c * Qtmp1 - s * Qtmp2;
+	  Qt[i + Nm * (k + 1)] = s * Qtmp1 + c * Qtmp2;
+	}
+
+      // Givens transformations
+      for (int k = kmin; k < kmax - 1; ++k)
+	{
+
+	  RealD Fden = 1.0 / hypot (x, lme[k - 1]);
+	  RealD c = lme[k - 1] * Fden;
+	  RealD s = -x * Fden;
+
+	  RealD tmpa1 = lmd[k];
+	  RealD tmpa2 = lmd[k + 1];
+	  RealD tmpb = lme[k];
+
+	  lmd[k] = c * c * tmpa1 + s * s * tmpa2 - 2.0 * c * s * tmpb;
+	  lmd[k + 1] = s * s * tmpa1 + c * c * tmpa2 + 2.0 * c * s * tmpb;
+	  lme[k] = c * s * (tmpa1 - tmpa2) + (c * c - s * s) * tmpb;
+	  lme[k - 1] = c * lme[k - 1] - s * x;
+
+	  if (k != kmax - 2)
+	    {
+	      x = -s * lme[k + 1];
+	      lme[k + 1] = c * lme[k + 1];
+	    }
+
+	  for (int i = 0; i < Nk; ++i)
+	    {
+	      RealD Qtmp1 = Qt[i + Nm * k];
+	      RealD Qtmp2 = Qt[i + Nm * (k + 1)];
+	      Qt[i + Nm * k] = c * Qtmp1 - s * Qtmp2;
+	      Qt[i + Nm * (k + 1)] = s * Qtmp1 + c * Qtmp2;
+	    }
+	}
+    }
+
+#if 0
+#ifdef USE_LAPACK
+#ifdef USE_MKL
+#define LAPACK_INT MKL_INT
+#else
+#define LAPACK_INT long long
+#endif
+    void diagonalize_lapack (std::vector  < RealD > &lmd, std::vector  < RealD > &lme, int N1,	// all
+			     int N2,	// get
+			     GridBase * grid)
+    {
+      const int size = Nm;
+      LAPACK_INT NN = N1;
+      double evals_tmp[NN];
+      double DD[NN];
+      double EE[NN];
+      for (int i = 0; i < NN; i++)
+	for (int j = i - 1; j <= i + 1; j++)
+	  if (j < NN && j >= 0)
+	    {
+	      if (i == j)
+		DD[i] = lmd[i];
+	      if (i == j)
+		evals_tmp[i] = lmd[i];
+	      if (j == (i - 1))
+		EE[j] = lme[j];
+	    }
+      LAPACK_INT evals_found;
+      LAPACK_INT lwork =
+	((18 * NN) >
+	 (1 + 4 * NN + NN * NN) ? (18 * NN) : (1 + 4 * NN + NN * NN));
+      LAPACK_INT liwork = 3 + NN * 10;
+      LAPACK_INT iwork[liwork];
+      double work[lwork];
+      LAPACK_INT isuppz[2 * NN];
+      char jobz = 'N';		// calculate evals only
+      char range = 'I';		// calculate il-th to iu-th evals
+      //    char range = 'A'; // calculate all evals
+      char uplo = 'U';		// refer to upper half of original matrix
+      char compz = 'I';		// Compute eigenvectors of tridiagonal matrix
+      int ifail[NN];
+      LAPACK_INT info;
+//  int total = QMP_get_number_of_nodes();
+//  int node = QMP_get_node_number();
+//  GridBase *grid = evec[0]._grid;
+      int total = grid->_Nprocessors;
+      int node = grid->_processor;
+      int interval = (NN / total) + 1;
+      double vl = 0.0, vu = 0.0;
+      LAPACK_INT il = interval * node + 1, iu = interval * (node + 1);
+      if (iu > NN)
+	iu = NN;
+      double tol = 0.0;
+      if (1)
+	{
+	  memset (evals_tmp, 0, sizeof (double) * NN);
+	  if (il <= NN)
+	    {
+	      printf ("total=%d node=%d il=%d iu=%d\n", total, node, il, iu);
+#ifdef USE_MKL
+	      dstegr (&jobz, &range, &NN,
+#else
+	      LAPACK_dstegr (&jobz, &range, &NN,
+#endif
+			     (double *) DD, (double *) EE, &vl, &vu, &il, &iu,	// these four are ignored if second parameteris 'A'
+			     &tol,	// tolerance
+			     &evals_found, evals_tmp, (double *) NULL, &NN,
+			     isuppz, work, &lwork, iwork, &liwork, &info);
+	      for (int i = iu - 1; i >= il - 1; i--)
+		{
+		  printf ("node=%d evals_found=%d evals_tmp[%d] = %g\n", node,
+			  evals_found, i - (il - 1), evals_tmp[i - (il - 1)]);
+		  evals_tmp[i] = evals_tmp[i - (il - 1)];
+		  if (il > 1)
+		    evals_tmp[i - (il - 1)] = 0.;
+		}
+	    }
+	  {
+	    grid->GlobalSumVector (evals_tmp, NN);
+	  }
+	}
+// cheating a bit. It is better to sort instead of just reversing it, but the document of the routine says evals are sorted in increasing order. qr gives evals in decreasing order.
+    }
+#undef LAPACK_INT
+#endif
+
+
+    void diagonalize (std::vector  < RealD > &lmd,
+		      std::vector  < RealD > &lme,
+		      int N2, int N1, GridBase * grid)
+    {
+
+#ifdef USE_LAPACK
+      const int check_lapack = 0;	// just use lapack if 0, check against lapack if 1
+
+      if (!check_lapack)
+	return diagonalize_lapack (lmd, lme, N2, N1, grid);
+
+//      diagonalize_lapack(lmd2,lme2,Nm2,Nm,Qt,grid);
+#endif
+    }
+#endif
+
+    static RealD normalise (Field & v)
+    {
+      RealD nn = norm2 (v);
+      nn = sqrt (nn);
+      v = v * (1.0 / nn);
+      return nn;
+    }
+
+    void orthogonalize (Field & w, std::vector < Field > &evec, int k)
+    {
+      double t0 = -usecond () / 1e6;
+      typedef typename Field::scalar_type MyComplex;
+      MyComplex ip;
+
+      if (0)
+	{
+	  for (int j = 0; j < k; ++j)
+	    {
+	      normalise (evec[j]);
+	      for (int i = 0; i < j; i++)
+		{
+		  ip = innerProduct (evec[i], evec[j]);	// are the evecs normalised? ; this assumes so.
+		  evec[j] = evec[j] - ip * evec[i];
+		}
+	    }
+	}
+
+      for (int j = 0; j < k; ++j)
+	{
+	  ip = innerProduct (evec[j], w);	// are the evecs normalised? ; this assumes so.
+	  w = w - ip * evec[j];
+	}
+      normalise (w);
+      t0 += usecond () / 1e6;
+      OrthoTime += t0;
+    }
+
+    void setUnit_Qt (int Nm, std::vector < RealD > &Qt)
+    {
+      for (int i = 0; i < Qt.size (); ++i)
+	Qt[i] = 0.0;
+      for (int k = 0; k < Nm; ++k)
+	Qt[k + k * Nm] = 1.0;
+    }
+
+
+    void calc (std::vector < RealD > &eval, const Field & src, int &Nconv)
+    {
+
+      GridBase *grid = src.Grid();
+//      assert(grid == src._grid);
+
+      std::
+	cout << GridLogMessage << " -- Nk = " << Nk << " Np = " << Np << std::
+	endl;
+      std::cout << GridLogMessage << " -- Nm = " << Nm << std::endl;
+      std::cout << GridLogMessage << " -- size of eval   = " << eval.
+	size () << std::endl;
+
+//      assert(c.size() && Nm == eval.size());
+
+      std::vector < RealD > lme (Nm);
+      std::vector < RealD > lmd (Nm);
+
+
+      Field current (grid);
+      Field last (grid);
+      Field next (grid);
+
+      Nconv = 0;
+
+      RealD beta_k;
+
+      // Set initial vector
+      // (uniform vector) Why not src??
+      //      evec[0] = 1.0;
+      current = src;
+      std::cout << GridLogMessage << "norm2(src)= " << norm2 (src) << std::
+	endl;
+      normalise (current);
+      std::
+	cout << GridLogMessage << "norm2(evec[0])= " << norm2 (current) <<
+	std::endl;
+
+      // Initial Nk steps
+      OrthoTime = 0.;
+      double t0 = usecond () / 1e6;
+      RealD norm;		// sqrt norm of last vector
+
+      uint64_t iter = 0;
+
+      bool initted = false;
+      std::vector < RealD > low (Nstop * 10);
+      std::vector < RealD > high (Nstop * 10);
+      RealD cont = 0.;
+      while (1) {
+	  cont = 0.;
+	  std::vector < RealD > lme2 (Nm);
+	  std::vector < RealD > lmd2 (Nm);
+	  for (uint64_t k = 0; k < Nm; ++k, iter++) {
+	      step (lmd, lme, last, current, next, iter);
+	      last = current;
+	      current = next;
+	    }
+	  double t1 = usecond () / 1e6;
+	  std::cout << GridLogMessage << "IRL::Initial steps: " << t1 -
+	    t0 << "seconds" << std::endl;
+	  t0 = t1;
+	  std::
+	    cout << GridLogMessage << "IRL::Initial steps:OrthoTime " <<
+	    OrthoTime << "seconds" << std::endl;
+
+	  // getting eigenvalues
+	  lmd2.resize (iter + 2);
+	  lme2.resize (iter + 2);
+	  for (uint64_t k = 0; k < iter; ++k) {
+	      lmd2[k + 1] = lmd[k];
+	      lme2[k + 2] = lme[k];
+	    }
+	  t1 = usecond () / 1e6;
+	  std::cout << GridLogMessage << "IRL:: copy: " << t1 -
+	    t0 << "seconds" << std::endl;
+	  t0 = t1;
+	  {
+	    int total = grid->_Nprocessors;
+	    int node = grid->_processor;
+	    int interval = (Nstop / total) + 1;
+	    int iu = (iter + 1) - (interval * node + 1);
+	    int il = (iter + 1) - (interval * (node + 1));
+	    std::vector < RealD > eval2 (iter + 3);
+	    RealD eps2;
+	    Bisection::bisec (lmd2, lme2, iter, il, iu, 1e-16, 1e-10, eval2,
+			      eps2);
+//        diagonalize(eval2,lme2,iter,Nk,grid);
+	    RealD diff = 0.;
+	    for (int i = il; i <= iu; i++) {
+		if (initted)
+		  diff =
+		    fabs (eval2[i] - high[iu-i]) / (fabs (eval2[i]) +
+						      fabs (high[iu-i]));
+		if (initted && (diff > eresid))
+		  cont = 1.;
+		if (initted)
+		  printf ("eval[%d]=%0.14e %0.14e, %0.14e\n", i, eval2[i],
+			  high[iu-i], diff);
+		high[iu-i] = eval2[i];
+	      }
+	    il = (interval * node + 1);
+	    iu = (interval * (node + 1));
+	    Bisection::bisec (lmd2, lme2, iter, il, iu, 1e-16, 1e-10, eval2,
+			      eps2);
+	    for (int i = il; i <= iu; i++) {
+		if (initted)
+		  diff =
+		    fabs (eval2[i] - low[i]) / (fabs (eval2[i]) +
+						fabs (low[i]));
+		if (initted && (diff > eresid))
+		  cont = 1.;
+		if (initted)
+		  printf ("eval[%d]=%0.14e %0.14e, %0.14e\n", i, eval2[i],
+			  low[i], diff);
+		low[i] = eval2[i];
+	      }
+	    t1 = usecond () / 1e6;
+	    std::cout << GridLogMessage << "IRL:: diagonalize: " << t1 -
+	      t0 << "seconds" << std::endl;
+	    t0 = t1;
+	  }
+
+	  for (uint64_t k = 0; k < Nk; ++k) {
+//          eval[k] = eval2[k];
+	    }
+	  if (initted)
+	    {
+	      grid->GlobalSumVector (&cont, 1);
+	      if (cont < 1.) return;
+	    }
+	  initted = true;
+	}
+
+    }
+
+
+
+
+
+#if 0
+
+/**
+   There is some matrix Q such that for any vector y
+   Q.e_1 = y and Q is unitary.
+**/
+    template < class T >
+      static T orthQ (DenseMatrix < T > &Q, std::vector < T > y)
+    {
+      int N = y.size ();	//Matrix Size
+      Fill (Q, 0.0);
+      T tau;
+      for (int i = 0; i < N; i++)
+	{
+	  Q[i][0] = y[i];
+	}
+      T sig = conj (y[0]) * y[0];
+      T tau0 = fabs (sqrt (sig));
+
+      for (int j = 1; j < N; j++)
+	{
+	  sig += conj (y[j]) * y[j];
+	  tau = abs (sqrt (sig));
+
+	  if (abs (tau0) > 0.0)
+	    {
+
+	      T gam = conj ((y[j] / tau) / tau0);
+	      for (int k = 0; k <= j - 1; k++)
+		{
+		  Q[k][j] = -gam * y[k];
+		}
+	      Q[j][j] = tau0 / tau;
+	    }
+	  else
+	    {
+	      Q[j - 1][j] = 1.0;
+	    }
+	  tau0 = tau;
+	}
+      return tau;
+    }
+
+/**
+	There is some matrix Q such that for any vector y
+	Q.e_k = y and Q is unitary.
+**/
+    template < class T >
+      static T orthU (DenseMatrix < T > &Q, std::vector < T > y)
+    {
+      T tau = orthQ (Q, y);
+      SL (Q);
+      return tau;
+    }
+
+
+/**
+	Wind up with a matrix with the first con rows untouched
+
+say con = 2
+	Q is such that Qdag H Q has {x, x, val, 0, 0, 0, 0, ...} as 1st colum
+	and the matrix is upper hessenberg
+	and with f and Q appropriately modidied with Q is the arnoldi factorization
+
+**/
+
+    template < class T > static void Lock (DenseMatrix < T > &H,	///Hess mtx     
+					   DenseMatrix < T > &Q,	///Lock Transform
+					   T val,	///value to be locked
+					   int con,	///number already locked
+					   RealD small, int dfg, bool herm)
+    {
+      //ForceTridiagonal(H);
+
+      int M = H.dim;
+      DenseVector < T > vec;
+      Resize (vec, M - con);
+
+      DenseMatrix < T > AH;
+      Resize (AH, M - con, M - con);
+      AH = GetSubMtx (H, con, M, con, M);
+
+      DenseMatrix < T > QQ;
+      Resize (QQ, M - con, M - con);
+
+      Unity (Q);
+      Unity (QQ);
+
+      DenseVector < T > evals;
+      Resize (evals, M - con);
+      DenseMatrix < T > evecs;
+      Resize (evecs, M - con, M - con);
+
+      Wilkinson < T > (AH, evals, evecs, small);
+
+      int k = 0;
+      RealD cold = abs (val - evals[k]);
+      for (int i = 1; i < M - con; i++)
+	{
+	  RealD cnew = abs (val - evals[i]);
+	  if (cnew < cold)
+	    {
+	      k = i;
+	      cold = cnew;
+	    }
+	}
+      vec = evecs[k];
+
+      ComplexD tau;
+      orthQ (QQ, vec);
+      //orthQM(QQ,AH,vec);
+
+      AH = Hermitian (QQ) * AH;
+      AH = AH * QQ;
+
+      for (int i = con; i < M; i++)
+	{
+	  for (int j = con; j < M; j++)
+	    {
+	      Q[i][j] = QQ[i - con][j - con];
+	      H[i][j] = AH[i - con][j - con];
+	    }
+	}
+
+      for (int j = M - 1; j > con + 2; j--)
+	{
+
+	  DenseMatrix < T > U;
+	  Resize (U, j - 1 - con, j - 1 - con);
+	  DenseVector < T > z;
+	  Resize (z, j - 1 - con);
+	  T nm = norm (z);
+	  for (int k = con + 0; k < j - 1; k++)
+	    {
+	      z[k - con] = conj (H (j, k + 1));
+	    }
+	  normalise (z);
+
+	  RealD tmp = 0;
+	  for (int i = 0; i < z.size () - 1; i++)
+	    {
+	      tmp = tmp + abs (z[i]);
+	    }
+
+	  if (tmp < small / ((RealD) z.size () - 1.0))
+	    {
+	      continue;
+	    }
+
+	  tau = orthU (U, z);
+
+	  DenseMatrix < T > Hb;
+	  Resize (Hb, j - 1 - con, M);
+
+	  for (int a = 0; a < M; a++)
+	    {
+	      for (int b = 0; b < j - 1 - con; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += H[a][con + 1 + c] * U[c][b];
+		    }		//sum += H(a,con+1+c)*U(c,b);}
+		  Hb[b][a] = sum;
+		}
+	    }
+
+	  for (int k = con + 1; k < j; k++)
+	    {
+	      for (int l = 0; l < M; l++)
+		{
+		  H[l][k] = Hb[k - 1 - con][l];
+		}
+	    }			//H(Hb[k-1-con][l] , l,k);}}
+
+	  DenseMatrix < T > Qb;
+	  Resize (Qb, M, M);
+
+	  for (int a = 0; a < M; a++)
+	    {
+	      for (int b = 0; b < j - 1 - con; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += Q[a][con + 1 + c] * U[c][b];
+		    }		//sum += Q(a,con+1+c)*U(c,b);}
+		  Qb[b][a] = sum;
+		}
+	    }
+
+	  for (int k = con + 1; k < j; k++)
+	    {
+	      for (int l = 0; l < M; l++)
+		{
+		  Q[l][k] = Qb[k - 1 - con][l];
+		}
+	    }			//Q(Qb[k-1-con][l] , l,k);}}
+
+	  DenseMatrix < T > Hc;
+	  Resize (Hc, M, M);
+
+	  for (int a = 0; a < j - 1 - con; a++)
+	    {
+	      for (int b = 0; b < M; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += conj (U[c][a]) * H[con + 1 + c][b];
+		    }		//sum += conj( U(c,a) )*H(con+1+c,b);}
+		  Hc[b][a] = sum;
+		}
+	    }
+
+	  for (int k = 0; k < M; k++)
+	    {
+	      for (int l = con + 1; l < j; l++)
+		{
+		  H[l][k] = Hc[k][l - 1 - con];
+		}
+	    }			//H(Hc[k][l-1-con] , l,k);}}
+
+	}
+    }
+#endif
+
+
+  };
+
+}
+#endif

Grid/algorithms/iterative/SplitGridBlockKrylovSchur.h

@@ -0,0 +1,124 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/SplitGridBlockKrylovSchur.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_SPLIT_BLOCK_KRYLOV_SCHUR_H
+#define GRID_SPLIT_BLOCK_KRYLOV_SCHUR_H
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Block Krylov-Schur eigensolver with Split-Grid batched operator application.
+ *
+ * Derives from BlockKrylovSchur<Field> and overrides only the operator
+ * application: instead of calling Linop.Op() once per vector, mrhs vectors
+ * are packed onto a smaller split grid with Grid_split, the polynomial-
+ * filtered operator poly(SLinop) is applied once to the combined field, and
+ * the results are unpacked with Grid_unsplit.  All other algorithmic logic
+ * (Arnoldi orthogonalisation, Schur restart, convergence test, etc.) is
+ * inherited unchanged from the base class.
+ *
+ * Constructor extras (beyond BlockKrylovSchur)
+ * -------------------------------------------
+ * SLinop  : split-grid linear operator used inside poly
+ * poly    : polynomial filter OperatorFunction applied on the split grid
+ * SGrid   : split-grid GridBase (mrhs fields packed here)
+ * mrhs    : RHS batched per split-grid call; Nblock must be divisible by mrhs
+ *
+ * Notes
+ * -----
+ * - The Rayleigh quotient H is built from poly(A), so its Ritz values are
+ *   eigenvalues of poly(A).  Convergence is assessed against the full-grid
+ *   Linop (inherited from base), giving true A-eigenvalues.
+ * - Grid_ (inherited) is the full grid; SGrid is the split grid.
+ */
+template<class Field>
+class SplitGridBlockKrylovSchur : public BlockKrylovSchur<Field> {
+
+  using Base = BlockKrylovSchur<Field>;
+
+  // Bring protected base members into scope
+  using Base::Nblock;
+  using Base::Grid_;
+  using Base::basis;
+
+  // Split-grid extras
+  LinearOperatorBase<Field>& SLinop;
+  OperatorFunction<Field>&   poly;
+  GridBase*                  SGrid;
+  int                        mrhs;
+
+public:
+
+  SplitGridBlockKrylovSchur(LinearOperatorBase<Field>& _Linop,
+                             LinearOperatorBase<Field>& _SLinop,
+                             OperatorFunction<Field>&   _poly,
+                             GridBase*                  _FGrid,
+                             GridBase*                  _SGrid,
+                             int                        _mrhs,
+                             RealD                      _Tolerance,
+                             RitzFilter                 _rf = EvalReSmall)
+    : Base(_Linop, _FGrid, _Tolerance, _rf),
+      SLinop(_SLinop), poly(_poly), SGrid(_SGrid), mrhs(_mrhs)
+  {}
+
+protected:
+
+  // Validate mrhs divisibility before the Arnoldi loop starts
+  void preRun() override
+  {
+    assert(Nblock % mrhs == 0 && "Nblock must be divisible by mrhs");
+  }
+
+  // Apply poly(A) to basis[kBase .. kBase+Nblock-1] via Grid_split / Grid_unsplit,
+  // batching mrhs vectors per split-grid call.
+  void applyBlock(std::vector<Field>& W, int kBase) override
+  {
+    std::vector<Field> in(mrhs, Field(Grid_));
+    Field s_in(SGrid);
+    Field s_out(SGrid);
+
+    int k_start = 0;
+    while (k_start < Nblock) {
+      for (int u = 0; u < mrhs; u++)
+        in[u] = basis[kBase + k_start + u];
+
+      Grid_split(in, s_in);
+      poly(SLinop, s_in, s_out);
+      Grid_unsplit(in, s_out);
+
+      for (int u = 0; u < mrhs; u++)
+        W[k_start + u] = in[u];
+
+      k_start += mrhs;
+    }
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif  // GRID_SPLIT_BLOCK_KRYLOV_SCHUR_H

Grid/algorithms/iterative/TrueHarmonicBlockKrylovSchur.h

@@ -0,0 +1,366 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/TrueHarmonicBlockKrylovSchur.h
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_TRUE_HARMONIC_BLOCKED_KRYLOV_SCHUR_H
+#define GRID_TRUE_HARMONIC_BLOCKED_KRYLOV_SCHUR_H
+
+#include <Grid/algorithms/iterative/HarmonicBlockKrylovSchur.h>
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * True harmonic-Ritz block Krylov-Schur eigensolver.
+ *
+ * Difference from the base class
+ * ------------------------------
+ * HarmonicBlockKrylovSchur is a *shift-targeted* Krylov-Schur: it Schur-
+ * decomposes (H - sigma I) and sorts the standard Ritz values by |lambda -
+ * sigma|.  For interior eigenvalues of Hermitian indefinite operators the
+ * standard Ritz values can be spurious ("ghosts"), so the sort can select
+ * junk.  This variant performs genuine harmonic Ritz extraction instead.
+ *
+ * Harmonic extraction
+ * -------------------
+ * From the exact block Arnoldi relation
+ *
+ *   A V = V H + F B^dag                                            (1)
+ *
+ * harmonic Ritz pairs (theta, g) w.r.t. shift sigma satisfy the Petrov-
+ * Galerkin condition (A - sigma)Vg - (theta - sigma)Vg  perp  (A - sigma)V,
+ * which reduces (with Hs = H - sigma I) to the small eigenproblem
+ *
+ *   Hhat g = (theta - sigma) g,   Hhat = Hs + Hs^{-H} B B^dag      (2)
+ *
+ * Exact thick restart
+ * -------------------
+ * From (2), H g = theta g - Hs^{-H} B (B^dag g), so the residual of every
+ * harmonic pair lies in the Nblock-dimensional column space of the single
+ * block
+ *
+ *   W = F - V Hs^{-H} B :
+ *   A (V g) - theta (V g) = W (B^dag g)                            (3)
+ *
+ * Hence {V g_1 .. g_Nk} + cols(W) admits an EXACT truncated Krylov-Schur
+ * relation  A Y = Y H_k + What Bc^dag  with dense H_k, and block Arnoldi
+ * resumes cleanly from What.  Krylov-Schur (Stewart, SIMAX 23(3), 2001)
+ * does not require H to be triangular after restart -- only that the
+ * decomposition holds exactly, which (3) guarantees.
+ *
+ * This is the block analogue of the GMRES-DR restart (Morgan, SISC 24(1),
+ * 2002; block version: Morgan, Appl. Numer. Math. 54, 2005).  The exact
+ * composition "block harmonic Krylov-Schur" is a derivation from these
+ * ingredients rather than a published algorithm; the relation (1) is
+ * checked explicitly at every restart when doVerify=true, and identity (3)
+ * was validated to machine precision in a standalone Eigen prototype.
+ *
+ * Caveats
+ * -------
+ * - If sigma is (numerically) an eigenvalue of H, Hs^{-H} is singular;
+ *   an assert fires -- perturb the shift slightly.
+ * - Reported eigenvalues come from the eigensystem of the retained H_k
+ *   (the true projected Rayleigh quotient Y^dag A Y), as in the base
+ *   class, so convergence tests and reporting are inherited unchanged.
+ *
+ * Usage: identical to HarmonicBlockKrylovSchur.
+ */
+template<class Field>
+class TrueHarmonicBlockKrylovSchur : public HarmonicBlockKrylovSchur<Field> {
+
+  typedef HarmonicBlockKrylovSchur<Field> Base;
+  typedef typename Base::CMat CMat;
+  typedef typename Base::CVec CVec;
+
+  // Members of the dependent base class need explicit scope.
+  using Base::Nblock;
+  using Base::Nm;
+  using Base::Nk;
+  using Base::Nstop;
+  using Base::MaxIter;
+  using Base::Tolerance;
+  using Base::shift;
+  using Base::Linop;
+  using Base::Grid_;
+  using Base::ritzFilter;
+  using Base::basis;
+  using Base::H;
+  using Base::F;
+  using Base::B;
+  using Base::beta_k;
+  using Base::rtol;
+  using Base::evals;
+  using Base::littleEvecs;
+  using Base::ritzEstimates;
+
+  using Base::expandStartBlock;
+  using Base::blockArnoldiIteration;
+  using Base::blockQR;
+  using Base::computeEigensystem;
+  using Base::converged;
+  using Base::approxMaxEval;
+
+public:
+  using Base::evecs;
+  using Base::doEvalCheck;
+  using Base::useParityFlip;
+  using Base::useGamma5;
+  using Base::gamma5Func;
+  using Base::verify;
+
+  TrueHarmonicBlockKrylovSchur(LinearOperatorBase<Field>& _Linop, GridBase* _Grid,
+                               RealD _Tolerance, ComplexD _shift = 0.0,
+                               RitzFilter _rf = EvalNormSmall)
+    : Base(_Linop, _Grid, _Tolerance, _shift, _rf)
+  {}
+
+  //--------------------------------------------------------------------
+  // Main entry point (same signature and outer structure as the base;
+  // the Schur-sort restart is replaced by harmonicRestart()).
+  //--------------------------------------------------------------------
+  virtual void operator()(const std::vector<Field>& v0, int _maxIter, int _Nm, int _Nk,
+                          int _Nstop, int _Nblock = 1, bool doubleOrthog = true,
+                          bool doVerify = false) override
+  {
+    MaxIter = _maxIter;
+    Nm      = _Nm;
+    Nk      = _Nk;
+    Nstop   = _Nstop;
+    Nblock  = _Nblock;
+
+    {
+      int divisor = 1;
+      if (useParityFlip) divisor *= 2;
+      if (useGamma5)     divisor *= 2;
+      assert(Nblock % divisor == 0 && (int)v0.size() >= Nblock / divisor);
+    }
+    assert(Nm % Nblock == 0);
+    assert(Nk % Nblock == 0);
+    assert(Nk < Nm);
+    if (useGamma5) assert(gamma5Func && "useGamma5: gamma5Func must be set");
+
+    int N = Nm;
+
+    RealD approxLambdaMax = approxMaxEval(v0[0]);
+    rtol = Tolerance * approxLambdaMax;
+    std::cout << GridLogMessage
+              << "TrueHarmonicBlockKrylovSchur: approx max eval = " << approxLambdaMax
+              << ", rtol = " << rtol
+              << ", shift = " << shift << std::endl;
+
+    H = CMat::Zero(N, N);
+    B = CMat::Zero(N, Nblock);
+
+    int start = 0;
+    std::vector<Field> startBlock = expandStartBlock(v0);
+
+    for (int iter = 0; iter < MaxIter; iter++) {
+      std::cout << GridLogMessage
+                << "TrueHarmonicBlockKrylovSchur: restart iteration " << iter << std::endl;
+
+      // ---- Block Arnoldi: extend from block 'start' to block Nm/Nblock ----
+      blockArnoldiIteration(startBlock, Nm/Nblock, start, doubleOrthog);
+      std::cout << GridLogMessage << "blockArnoldiIteration done " << std::endl;
+      start = Nk/Nblock;
+
+      if (doVerify) {
+        std::string lbl = "iter " + std::to_string(iter) + " after Arnoldi";
+        verify(lbl);
+      }
+
+      // ---- Harmonic extraction + exact thick restart ----
+      harmonicRestart();
+      std::cout << GridLogMessage << "harmonicRestart done " << std::endl;
+
+      // Restart from the residual block W (exact by identity (3))
+      startBlock = F;
+
+      if (doVerify) {
+        std::string lbl = "iter " + std::to_string(iter) + " after harmonic restart";
+        verify(lbl);
+      }
+
+      // ---- Eigensystem of retained H_k = Y^dag A Y for convergence ----
+      CMat Hk = H(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk));
+      computeEigensystem(Hk, Nk);
+
+      int Nconv = converged(Nk);
+      std::cout << GridLogMessage
+                << "TrueHarmonicBlockKrylovSchur: converged " << Nconv
+                << " / " << Nstop << std::endl;
+
+      if (Nconv >= Nstop || iter == MaxIter - 1) {
+        std::cout << GridLogMessage
+                  << "TrueHarmonicBlockKrylovSchur: done after " << iter
+                  << " restarts, " << Nconv << " converged." << std::endl;
+        std::cout << GridLogMessage << "Eigenvalues: " << evals.transpose() << std::endl;
+
+        if (doEvalCheck) {
+          Field w(Grid_);
+          for (int k = 0; k < (int)evecs.size(); k++) {
+            Linop.Op(evecs[k], w);
+            ComplexD eval_est = toStdCmplx(innerProduct(evecs[k], w));
+            w -= eval_est * evecs[k];
+            RealD res = std::sqrt(norm2(w));
+            std::cout << GridLogMessage << "TrueHarmonicBlockKrylovSchur: evec[" << k << "]"
+                      << "  eval_reported = " << evals[k]
+                      << "  eval_est = " << eval_est
+                      << "  || A v - eval_est * v || = " << res << std::endl;
+          }
+        }
+
+        return;
+      }
+    }
+  }
+
+protected:
+
+  //--------------------------------------------------------------------
+  // Harmonic thick restart
+  //--------------------------------------------------------------------
+  // On entry:  A V = V H + F B^dag exact, V = basis (Nm fields), F Nblock
+  //            residual fields, H/B the Nm-sized small matrices.
+  // On exit:   basis = Y (Nk orthonormal fields spanning the harmonic
+  //            Ritz space), F = What (Nblock orthonormal fields, perp Y),
+  //            H top-left Nk x Nk = Y^dag A Y (dense), B top Nk rows the
+  //            new coupling; relation A Y = Y H_k + F B^dag again exact.
+  void harmonicRestart()
+  {
+    int N = Nm;
+
+    // ---- Small harmonic eigenproblem:  Hhat g = (theta - sigma) g ----
+    CMat Hs = H - shift * CMat::Identity(N, N);
+    Eigen::FullPivLU<CMat> lu(Hs.adjoint());
+    assert(lu.isInvertible() &&
+           "TrueHarmonicBlockKrylovSchur: H - shift*I singular; perturb the shift");
+    CMat M    = lu.solve(B);              // N x Nblock :  Hs^{-H} B
+    CMat Hhat = Hs + M * B.adjoint();     // N x N
+
+    Eigen::ComplexEigenSolver<CMat> es(Hhat);
+
+    // Sort harmonic values (theta - sigma) by the Ritz filter
+    // (EvalNormSmall -> closest to the shift).
+    ComplexComparator cComp(ritzFilter);
+    std::vector<int> idx(N);
+    std::iota(idx.begin(), idx.end(), 0);
+    std::sort(idx.begin(), idx.end(), [&](int a, int b){
+      return cComp(toStdCmplx(es.eigenvalues()(a)), toStdCmplx(es.eigenvalues()(b)));
+    });
+
+    CMat G(N, Nk);
+    CVec theta(Nk);
+    for (int k = 0; k < Nk; k++) {
+      G.col(k) = es.eigenvectors().col(idx[k]);
+      theta(k) = es.eigenvalues()(idx[k]) + toStdCmplx(shift);
+    }
+
+    std::cout << GridLogMessage
+              << "TrueHarmonicBlockKrylovSchur: harmonic Ritz values nearest shift:" << std::endl;
+    for (int k = 0; k < Nk; k++)
+      std::cout << GridLogMessage << "  [" << k << "] " << theta(k) << std::endl;
+
+    // ---- Orthonormalise the little vectors:  G = Gt Rg ----
+    Eigen::HouseholderQR<CMat> qr(G);
+    CMat Gt = qr.householderQ() * CMat::Identity(N, Nk);
+    CMat Rg = qr.matrixQR().topLeftCorner(Nk, Nk).template triangularView<Eigen::Upper>();
+    for (int k = 0; k < Nk; k++)
+      assert(std::abs(Rg(k, k)) > 1e-14 &&
+             "TrueHarmonicBlockKrylovSchur: harmonic vectors linearly dependent");
+    CMat RgInv = Rg.inverse();
+
+    // A (V G) = (V G) diag(theta) + W (B^dag G),  W = F - V M
+    //   =>  A (V Gt) = (V Gt) T + W Ct
+    CMat T  = Rg * theta.asDiagonal() * RgInv;   // Nk x Nk
+    CMat Ct = (B.adjoint() * G) * RgInv;         // Nblock x Nk
+
+    // ---- Large-field work ----
+    // Y = V Gt : new orthonormal basis (Nk fields)
+    std::vector<Field> Y;
+    constructThin(Y, basis, Gt, Nk);
+
+    // W = F - V M : Nblock residual directions containing every harmonic
+    // residual (identity (3) in the class comment)
+    std::vector<Field> W = F;
+    for (int t = 0; t < Nblock; t++)
+      for (int j = 0; j < N; j++)
+        W[t] -= basis[j] * M(j, t);
+
+    // Orthogonalise W against Y (two passes), recording P = Y^dag W
+    CMat P = CMat::Zero(Nk, Nblock);
+    for (int pass = 0; pass < 2; pass++) {
+      for (int j = 0; j < Nk; j++) {
+        for (int t = 0; t < Nblock; t++) {
+          ComplexD c = innerProduct(Y[j], W[t]);
+          P(j, t) += toStdCmplx(c);
+          W[t] -= c * Y[j];
+        }
+      }
+    }
+
+    // Block QR of the projected residual: W <- What, returns Rw
+    CMat Rw = blockQR(W);
+
+    // ---- Reassemble the exact Krylov-Schur relation ----
+    //   A Y = Y (T + P Ct) + What (Rw Ct)
+    CMat Hk = T + P * Ct;                        // Nk x Nk, dense
+    CMat Bc = Rw * Ct;                           // Nblock x Nk
+
+    H = CMat::Zero(N, N);
+    H(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk)) = Hk;
+    B = CMat::Zero(N, Nblock);
+    for (int j = 0; j < Nk; j++)
+      for (int t = 0; t < Nblock; t++)
+        B(j, t) = std::conj(Bc(t, j));
+
+    basis = Y;
+    F     = W;
+    beta_k = Bc.norm();
+
+    std::cout << GridLogMessage
+              << "TrueHarmonicBlockKrylovSchur: beta_k = " << beta_k << std::endl;
+  }
+
+  //--------------------------------------------------------------------
+  // Thin basis combination: out[i] = sum_j U[j] * R(j,i), i < ncol
+  //--------------------------------------------------------------------
+  void constructThin(std::vector<Field>& out, const std::vector<Field>& U,
+                     const CMat& R, int ncol)
+  {
+    out.clear();
+    Field tmp(Grid_);
+    for (int i = 0; i < ncol; i++) {
+      tmp = Zero();
+      for (int j = 0; j < (int)U.size(); j++)
+        tmp += U[j] * R(j, i);
+      out.push_back(tmp);
+    }
+  }
+
+};
+
+NAMESPACE_END(Grid);
+
+#endif  // GRID_TRUE_HARMONIC_BLOCKED_KRYLOV_SCHUR_H

Grid/algorithms/multigrid/Aggregates.h

@@ -97,7 +97,7 @@ public:
 
     RealD scale;
 
-    ConjugateGradient<FineField> CG(1.0e-3,400,false);
+    ConjugateGradient<FineField> CG(1.0e-4,2000,false);
     FineField noise(FineGrid);
     FineField Mn(FineGrid);
 
@@ -131,7 +131,10 @@ public:
     RealD scale;
 
     TrivialPrecon<FineField> simple_fine;
-    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,10,DiracOp,simple_fine,30,30);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,10,DiracOp,simple_fine,12,12);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,10,10);
     FineField noise(FineGrid);
     FineField src(FineGrid);
     FineField guess(FineGrid);
@@ -146,16 +149,16 @@ public:
       
       DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
 
-      for(int i=0;i<2;i++){
+      for(int i=0;i<3;i++){
 	//  void operator() (const Field &src, Field &psi){
 #if 1
-	std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	if (i==0)std::cout << GridLogMessage << " inverting on noise "<<std::endl;
 	src = noise;
 	guess=Zero();
 	GCR(src,guess);
 	subspace[b] = guess;
 #else
-	std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	if (i==0)std::cout << GridLogMessage << " inverting on zero "<<std::endl;
 	src=Zero();
 	guess = noise;
 	GCR(src,guess);
@@ -167,7 +170,7 @@ public:
 
       }
 
-      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<std::endl;
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<" <f|OpDagOp|f>"<<norm2(Mn)<<std::endl;
       subspace[b]   = noise;
 
     }

Grid/communicator/Communicator_base.h

@@ -108,7 +108,7 @@ public:
   // very VERY rarely (Log, serial RNG) we need world without a grid
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
-  static void BroadcastWorld(int root,void* data, int bytes);
+  static void BroadcastWorld(int root,void* data, uint64_t bytes);
   static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
@@ -175,27 +175,27 @@ public:
 			   int dest,
 			   void *recv,
 			   int from,
-			   int bytes,int dir);
+			   uint64_t bytes,int dir);
   
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,
 		      int recv_from_rank,
-		      int bytes);
+		      uint64_t bytes);
   
   int IsOffNode(int rank);
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,int do_xmit,
 			       void *recv,
 			       int recv_from_rank,int do_recv,
-			       int bytes,int dir);
+			       uint64_t bytes,int dir);
 
   double StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
 				      void *xmit,
 				      int xmit_to_rank,int do_xmit,
 				      void *recv,
 				      int recv_from_rank,int do_recv,
-				      int xbytes,int rbytes,int dir);
+				      uint64_t xbytes,uint64_t rbytes,int dir);
 
   // Could do a PollHtoD and have a CommsMerge dependence
   void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
@@ -206,7 +206,7 @@ public:
 				    int xmit_to_rank,int do_xmit,
 				    void *recv,void *recv_comp,
 				    int recv_from_rank,int do_recv,
-				    int xbytes,int rbytes,int dir);
+				    uint64_t xbytes,uint64_t rbytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
@@ -220,7 +220,7 @@ public:
   ////////////////////////////////////////////////////////////
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
-  void Broadcast(int root,void* data, int bytes);
+  void Broadcast(int root,void* data, uint64_t bytes);
 
   ////////////////////////////////////////////////////////////
   // All2All down one dimension

Grid/communicator/Communicator_mpi3.cc

@@ -342,23 +342,22 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &
 						int dest,
 						void *recv,
 						int from,
-						int bytes,int dir)
+						uint64_t bytes,int dir)
 {
   MPI_Request xrq;
   MPI_Request rrq;
 
   GRID_ASSERT(dest != _processor);
   GRID_ASSERT(from != _processor);
-
   int tag;
 
   tag= dir+from*32;
-  int ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator,&rrq);
+  int ierr=MPI_Irecv(recv,(int)( bytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator,&rrq);
   GRID_ASSERT(ierr==0);
   list.push_back(rrq);
   
   tag= dir+_processor*32;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator,&xrq);
+  ierr =MPI_Isend(xmit,(int)(bytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator,&xrq);
   GRID_ASSERT(ierr==0);
   list.push_back(xrq);
 }
@@ -379,7 +378,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
 					   void *recv,
 					   int from,
-					   int bytes)
+					   uint64_t bytes)
 {
   std::vector<MpiCommsRequest_t> reqs(0);
 
@@ -392,8 +391,8 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 
   // Give the CPU to MPI immediately; can use threads to overlap optionally
   //  printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
-  ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		    recv,bytes,MPI_CHAR,from, from,
+  ierr=MPI_Sendrecv(xmit,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,dest,myrank,
+		    recv,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,from, from,
 		    communicator,MPI_STATUS_IGNORE);
   GRID_ASSERT(ierr==0);
 
@@ -403,7 +402,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int dest, int dox,
 						     void *recv,
 						     int from, int dor,
-						     int bytes,int dir)
+						     uint64_t bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
   double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
@@ -426,7 +425,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
 							   int dest,int dox,
 							   void *recv,
 							   int from,int dor,
-							   int xbytes,int rbytes,int dir)
+							   uint64_t xbytes,uint64_t rbytes,int dir)
 {
   return 0.0; // Do nothing -- no preparation required
 }
@@ -435,7 +434,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int dest,int dox,
 							 void *recv,void *recv_comp,
 							 int from,int dor,
-							 int xbytes,int rbytes,int dir)
+							 uint64_t xbytes,uint64_t rbytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -458,7 +457,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
       //      std::cout << " StencilSendToRecvFrom "<<dir<<" MPI_Irecv "<<std::hex<<recv<<std::dec<<std::endl;
-      ierr=MPI_Irecv(recv_comp, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      ierr=MPI_Irecv(recv_comp,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
       GRID_ASSERT(ierr==0);
       list.push_back(rrq);
       off_node_bytes+=rbytes;
@@ -476,7 +475,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit_comp, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      ierr =MPI_Isend(xmit_comp,(int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
       GRID_ASSERT(ierr==0);
       list.push_back(xrq);
       off_node_bytes+=xbytes;
@@ -543,7 +542,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
 							   int dest,int dox,
 							   void *recv,
 							   int from,int dor,
-							   int xbytes,int rbytes,int dir)
+							   uint64_t xbytes,uint64_t rbytes,int dir)
 {
 /*
  * Bring sequence from Stencil.h down to lower level.
@@ -584,7 +583,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
       host_recv = this->HostBufferMalloc(rbytes);
-      ierr=MPI_Irecv(host_recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      ierr=MPI_Irecv(host_recv,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
       GRID_ASSERT(ierr==0);
       CommsRequest_t srq;
       srq.PacketType = InterNodeRecv;
@@ -686,7 +685,7 @@ void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsReque
 	if ( acceleratorEventIsComplete(list[idx].ev) ) {
 
 	  void *host_xmit = list[idx].host_buf;
-	  uint32_t xbytes = list[idx].bytes;
+	  uint64_t xbytes = list[idx].bytes;
 	  int dest        = list[idx].dest;
 	  int tag         = list[idx].tag;
 	  int commdir     = list[idx].commdir;
@@ -697,7 +696,7 @@ void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsReque
 	  //	  std::cout << " DtoH is complete for index "<<idx<<" calling MPI_Isend "<<std::endl;
 	  
 	  MPI_Request xrq;
-	  int ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+	  int ierr =MPI_Isend(host_xmit, (int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
 	  GRID_ASSERT(ierr==0);
 
 	  list[idx].req        = xrq; // Update the MPI request in the list
@@ -718,7 +717,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int dest,int dox,
 							 void *recv,void *recv_comp,
 							 int from,int dor,
-							 int xbytes,int rbytes,int dir)
+							 uint64_t xbytes,uint64_t rbytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -884,11 +883,11 @@ void CartesianCommunicator::Barrier(void)
   int ierr = MPI_Barrier(communicator);
   GRID_ASSERT(ierr==0);
 }
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+void CartesianCommunicator::Broadcast(int root,void* data,uint64_t bytes)
 {
   FlightRecorder::StepLog("Broadcast");
   int ierr=MPI_Bcast(data,
-		     bytes,
+		     (int)bytes,
 		     MPI_BYTE,
 		     root,
 		     communicator);
@@ -904,11 +903,11 @@ void CartesianCommunicator::BarrierWorld(void){
   int ierr = MPI_Barrier(communicator_world);
   GRID_ASSERT(ierr==0);
 }
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes)
 {
   FlightRecorder::StepLog("BroadcastWorld");
   int ierr= MPI_Bcast(data,
-		      bytes,
+		      (int)bytes,
 		      MPI_BYTE,
 		      root,
 		      communicator_world);

Grid/communicator/Communicator_none.cc

@@ -27,6 +27,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #include <Grid/GridCore.h>
 
+void GridAbort(void) { abort(); }
+
 NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -34,7 +36,6 @@ NAMESPACE_BEGIN(Grid);
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
-void GridAbort(void) { abort(); }
 
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
@@ -89,7 +90,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
 					   void *recv,
 					   int from,
-					   int bytes)
+					   uint64_t bytes)
 {
   GRID_ASSERT(0);
 }
@@ -99,7 +100,7 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
 						int dest,
 						void *recv,
 						int from,
-						int bytes,int dir)
+						uint64_t bytes,int dir)
 {
   GRID_ASSERT(0);
 }
@@ -115,8 +116,8 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 
 int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
+void CartesianCommunicator::Broadcast(int root,void* data, uint64_t bytes) {}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes) { }
 void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
@@ -132,7 +133,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int xmit_to_rank,int dox,
 						     void *recv,
 						     int recv_from_rank,int dor,
-						     int bytes, int dir)
+						     uint64_t bytes, int dir)
 {
   return 2.0*bytes;
 }
@@ -143,16 +144,16 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
 							   int xmit_to_rank,int dox,
 							   void *recv,
 							   int recv_from_rank,int dor,
-							   int xbytes,int rbytes, int dir)
+							   uint64_t xbytes,uint64_t rbytes, int dir)
 {
   return 0.0;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,
+							 void *xmit, void *xmit_comp,
 							 int xmit_to_rank,int dox,
-							 void *recv,
+							 void *recv, void *recv_comp,
 							 int recv_from_rank,int dor,
-							 int xbytes,int rbytes, int dir)
+							 uint64_t xbytes,uint64_t rbytes, int dir)
 {
   return xbytes+rbytes;
 }

Grid/cshift/Cshift_mpi.h

@@ -49,6 +49,20 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   // Map to always positive shift modulo global full dimension.
   shift = (shift+fd)%fd;
 
+  if( shift ==0 ) {
+    ret = rhs;
+    return ret;
+  }
+  //
+  // Potential easy fast cases:
+  // Shift is a multiple of the local lattice extent.
+  // Then need only to shift whole subvolumes
+  int L = rhs.Grid()->_ldimensions[dimension];
+  if ( (shift%L )==0 && !rhs.Grid()->CheckerBoarded(dimension) ) {
+    Cshift_simple(ret,rhs,dimension,shift);
+    return ret;
+  }
+  
   ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
         
   // the permute type
@@ -73,6 +87,55 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   return ret;
 }
 
+template<class vobj> void Cshift_simple(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
+{
+  GridBase *grid=rhs.Grid();
+  int comm_proc, xmit_to_rank, recv_from_rank;
+  
+  int fd              = rhs.Grid()->_fdimensions[dimension];
+  int rd              = rhs.Grid()->_rdimensions[dimension];
+  int ld              = rhs.Grid()->_ldimensions[dimension];
+  int pd              = rhs.Grid()->_processors[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
+
+  comm_proc = ((shift)/ld)%pd;
+
+  grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+  if(comm_dim) {
+
+    int64_t bytes = sizeof(vobj) * grid->oSites();
+
+    autoView(rhs_v , rhs, AcceleratorRead);
+    autoView(ret_v , ret, AcceleratorWrite);
+    void *send_buf  = (void *)&rhs_v[0];
+    void *recv_buf  = (void *)&ret_v[0];
+
+#ifdef ACCELERATOR_AWARE_MPI
+    grid->SendToRecvFrom(send_buf,
+			 xmit_to_rank,
+			 recv_buf,
+			 recv_from_rank,
+			 bytes);
+#else
+    static hostVector<vobj> hrhs; hrhs.resize(grid->oSites());
+    static hostVector<vobj> hret; hret.resize(grid->oSites());
+
+    void *hsend_buf = (void *)&hrhs[0];
+    void *hrecv_buf = (void *)&hret[0];
+
+    acceleratorCopyFromDevice(send_buf,hsend_buf,bytes);
+
+    grid->SendToRecvFrom(hsend_buf,
+			 xmit_to_rank,
+			 hrecv_buf,
+			 recv_from_rank,
+			 bytes);
+
+    acceleratorCopyToDevice(hrecv_buf,recv_buf,bytes);
+#endif
+  }
+}
 template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
 {
   int sshift[2];

Grid/parallelIO/IldgIO.h

@@ -260,7 +260,8 @@ class GridLimeReader : public BinaryIO {
               << " / field= " << n2ck << " / rdiff= " << GRID_FIELD_NORM_CALC(FieldNormMetaData_,n2ck) << std::endl;
 	  GRID_FIELD_NORM_CHECK(FieldNormMetaData_,n2ck);
 	}
-	assert(scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb)==1);
+//	assert(scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb)==1);
+	scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb);
 
 	// find out if next field is a GridFieldNorm
 	return;

Grid/parallelIO/NerscIO.h

@@ -122,7 +122,7 @@ public:
     field.checksum = std::stoul(header["CHECKSUM"],0,16);
     field.ensemble_id      = header["ENSEMBLE_ID"];
     field.ensemble_label   = header["ENSEMBLE_LABEL"];
-    field.sequence_number  = std::stol(header["SEQUENCE_NUMBER"]);
+//    field.sequence_number  = std::stol(header["SEQUENCE_NUMBER"]);
     field.creator          = header["CREATOR"];
     field.creator_hardware = header["CREATOR_HARDWARE"];
     field.creation_date    = header["CREATION_DATE"];

Grid/perfmon/Timer.h

@@ -60,12 +60,16 @@ inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 }
 inline std::ostream& operator<< (std::ostream & stream, const GridMillisecs & now)
 {
+  double secs = 1.0*now.count()*1.0e-3;
+  stream << secs<<" s";
+  /*
   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)

Grid/qcd/QCD.h

@@ -596,16 +596,32 @@ template<int Index,class vobj> inline vobj transposeColour(const vobj &lhs){
 //////////////////////////////////////////
 // Trace lattice and non-lattice
 //////////////////////////////////////////
+#define GRID_UNOP(name)   name
+#define GRID_DEF_UNOP(op, name)						\
+  template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
+  inline auto op(const T1 &arg) ->decltype(LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg)) \
+  {									\
+    return     LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg); \
+  }
+
 template<int Index,class vobj>
 inline auto traceSpin(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(vobj()))>
 {
   return traceIndex<SpinIndex>(lhs);
 }
+
+GridUnopClass(UnaryTraceSpin, traceIndex<SpinIndex>(a));
+GRID_DEF_UNOP(traceSpin, UnaryTraceSpin);
+
 template<int Index,class vobj>
 inline auto traceColour(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(vobj()))>
 {
   return traceIndex<ColourIndex>(lhs);
 }
+
+GridUnopClass(UnaryTraceColour, traceIndex<ColourIndex>(a));
+GRID_DEF_UNOP(traceColour, UnaryTraceColour);
+
 template<int Index,class vobj>
 inline auto traceSpin(const vobj &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(lhs))>
 {
@@ -617,6 +633,8 @@ inline auto traceColour(const vobj &lhs) -> Lattice<decltype(traceIndex<ColourIn
   return traceIndex<ColourIndex>(lhs);
 }
 
+#undef GRID_UNOP
+#undef GRID_DEF_UNOP
 //////////////////////////////////////////
 // Current types
 //////////////////////////////////////////

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

@@ -411,7 +411,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
-#ifdef GRID_SYCL
+#if 0
 extern "C" {
     ulong SYCL_EXTERNAL __attribute__((overloadable)) intel_get_cycle_counter( void );
     uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_active_channel_mask( void );

Grid/qcd/action/gauge/GaugeImplTypes.h

@@ -138,10 +138,13 @@ public:
     //auto start = std::chrono::high_resolution_clock::now();
     autoView(U_v,U,AcceleratorWrite);
     autoView(P_v,P,AcceleratorRead);
-    accelerator_for(ss, P.Grid()->oSites(),1,{
+    typedef typename Field::vector_object vobj;
+    const int Nsimd = vobj::Nsimd();
+    accelerator_for(ss, P.Grid()->oSites(),Nsimd,{
       for (int mu = 0; mu < Nd; mu++) {
-          U_v[ss](mu) = Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu);
-          U_v[ss](mu) = Group::ProjectOnGeneralGroup(U_v[ss](mu));
+	auto tmp = Exponentiate(P_v(ss)(mu), ep, Nexp) * U_v(ss)(mu);
+	tmp      = Group::ProjectOnGeneralGroup(tmp);
+	coalescedWrite(U_v[ss](mu),tmp);
       }
     });
    //auto end = std::chrono::high_resolution_clock::now();
@@ -176,6 +179,8 @@ public:
     Group::ColdConfiguration(pRNG, U);
   }
 
+  static const int num_colours = Group::Dimension;
+
 };
 
 

Grid/qcd/modules/ObservableModules.h

@@ -103,6 +103,18 @@ class PolyakovMod: public ObservableModule<PolyakovLogger<Impl>, NoParameters>{
   PolyakovMod(): ObsBase(NoParameters()){}
 };
 
+template < class Impl >
+class SpatialPolyakovMod: public ObservableModule<SpatialPolyakovLogger<Impl>, NoParameters>{
+  typedef ObservableModule<SpatialPolyakovLogger<Impl>, NoParameters> ObsBase;
+  using ObsBase::ObsBase; // for constructors
+
+  // acquire resource
+  virtual void initialize(){
+    this->ObservablePtr.reset(new SpatialPolyakovLogger<Impl>());
+  }
+  public:
+  SpatialPolyakovMod(): ObsBase(NoParameters()){}
+};
 
 template < class Impl >
 class TopologicalChargeMod: public ObservableModule<TopologicalCharge<Impl>, TopologyObsParameters>{

Grid/qcd/observables/polyakov_loop.h

@@ -2,11 +2,12 @@
 
 Grid physics library, www.github.com/paboyle/Grid
 
-Source file: ./lib/qcd/modules/polyakov_line.h
+Source file: ./Grid/qcd/observables/polyakov_loop.h
 
-Copyright (C) 2017
+Copyright (C) 2025
 
 Author: David Preti <david.preti@csic.es>
+Author: Alexis Verney-Provatas <2414441@swansea.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
@@ -60,4 +61,43 @@ class PolyakovLogger : public HmcObservable<typename Impl::Field> {
   }
 };
 
+template <class Impl>
+class SpatialPolyakovLogger : public HmcObservable<typename Impl::Field> {
+ public:
+    // here forces the Impl to be of gauge fields
+    // if not the compiler will complain
+    INHERIT_GIMPL_TYPES(Impl);
+
+     // necessary for HmcObservable compatibility
+    typedef typename Impl::Field Field;
+
+    void TrajectoryComplete(int traj,
+                            Field &U,
+                            GridSerialRNG &sRNG,
+                            GridParallelRNG &pRNG) {
+
+    // Save current numerical output precision
+    int def_prec = std::cout.precision();
+
+    // Assume that the dimensions are D=3+1
+    int Ndim = 3;
+    ComplexD polyakov;
+   
+    // Iterate over the spatial directions and print the average spatial polyakov loop
+    // over them 
+    for (int idx=0; idx<Ndim; idx++) {
+        polyakov = WilsonLoops<Impl>::avgPolyakovLoop(U, idx);
+    
+        std::cout << GridLogMessage
+            << std::setprecision(std::numeric_limits<Real>::digits10 + 1)
+            << "Polyakov Loop in the " << idx << " spatial direction : [ " << traj << " ] "<< polyakov << std::endl;
+
+    }
+
+    // Return to original output precision
+    std::cout.precision(def_prec);
+
+  }
+};
+
 NAMESPACE_END(Grid);

Grid/qcd/smearing/GaugeConfigurationMasked.h

@@ -291,8 +291,8 @@ public:
     int idx=0;
     for(int mu=0;mu<4;mu++){
     for(int nu=0;nu<4;nu++){
-      if ( mu!=nu) GRID_ASSERT(this->StoutSmearing->SmearRho[idx]==rho);
-      else         GRID_ASSERT(this->StoutSmearing->SmearRho[idx]==0.0);
+      if ( mu!=nu) assert(this->StoutSmearing->SmearRho[idx]==rho);
+      else         assert(this->StoutSmearing->SmearRho[idx]==0.0);
       idx++;
     }}
     //////////////////////////////////////////////////////////////////
@@ -825,6 +825,7 @@ public:
   virtual void fill_smearedSet(GaugeField &U)
   {
     this->ThinLinks = &U;  // attach the smearing routine to the field U
+    std::cout << GridLogMessage << " fill_smearedSet " << WilsonLoops<PeriodicGimplR>::avgPlaquette(U) << std::endl;
 
     // check the pointer is not null
     if (this->ThinLinks == NULL)
@@ -846,6 +847,8 @@ public:
 	ApplyMask(smeared_A,smearLvl);
 	smeared_B = previous_u;
 	ApplyMask(smeared_B,smearLvl);
+	std::cout << GridLogMessage << " smeared_A " << norm2(smeared_A) << std::endl;
+	std::cout << GridLogMessage << " smeared_B " << norm2(smeared_B) << std::endl;
 	// Replace only the masked portion
 	this->SmearedSet[smearLvl] = previous_u-smeared_B + smeared_A;
         previous_u = this->SmearedSet[smearLvl];
@@ -934,10 +937,10 @@ public:
   SmearedConfigurationMasked(GridCartesian* _UGrid, unsigned int Nsmear, Smear_Stout<Gimpl>& Stout)
     : SmearedConfiguration<Gimpl>(_UGrid, Nsmear,Stout)
   {
-    GRID_ASSERT(Nsmear%(2*Nd)==0); // Or multiply by 8??
+    assert(Nsmear%(2*Nd)==0); // Or multiply by 8??
 
     // was resized in base class
-    GRID_ASSERT(this->SmearedSet.size()==Nsmear);
+    assert(this->SmearedSet.size()==Nsmear);
     
     GridRedBlackCartesian * UrbGrid;
     UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(_UGrid);

Grid/qcd/smearing/JacobianAction.h

@@ -54,7 +54,7 @@ public:
   // Usual cases are not used
   //////////////////////////////////
   virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG){ GRID_ASSERT(0);};
-  virtual RealD S(const GaugeField &U) { GRID_ASSERT(0); }
+  virtual RealD S(const GaugeField &U) { GRID_ASSERT(0); return 0; }
   virtual void deriv(const GaugeField &U, GaugeField &dSdU) { GRID_ASSERT(0);  }
 
   //////////////////////////////////

Grid/qcd/smearing/WilsonFlow.h

@@ -51,10 +51,16 @@ protected:
 
 public:
 
+//Define the action used to evolve the plaquettes
+//(Lüscher: https://arxiv.org/pdf/1006.4518 eq. 1.4)
+//V'(t) = -g^2 * ( d/dVt S[Vt](g) ) * Vt
+//      = -g^2 * ( d/dVt (1/g^2 * sum_p Re tr{ 1 - Vt(p) } ) ) * Vt
+//      = - d/dVt ( sum_p ( Nc - Re tr Vt(p) ) * Vt
+//      = - d/dVt ( Nc * sum_p ( 1 - Re tr Vt(p)/Nc ) ) * Vt
+//      = - d/dVt SG[Vt](Nc) * Vt
   explicit WilsonFlowBase(unsigned int meas_interval =1) {
     
-    SG = (ActionBase *) new WilsonGaugeAction<Gimpl>(3.0);
-    // WilsonGaugeAction with beta 3.0
+    SG = (ActionBase *) new WilsonGaugeAction<Gimpl>(Gimpl::num_colours);
     setDefaultMeasurements(meas_interval);
   }
 
@@ -149,9 +155,17 @@ public:
 ////////////////////////////////////////////////////////////////////////////////
 // Implementations
 ////////////////////////////////////////////////////////////////////////////////
+
+//Compute t^2 <E(t)> for time from the plaquette form
+//(Lüscher: https://arxiv.org/pdf/1006.4518 eq. 3.1)
+//E(t) = 2 * sum_p Retr{ 1 - Vt(p) } =
+//     = 2 * sum_p ( Nc - Retr Vt(p) ) =
+//     = 2 * Nc * sum_p ( 1 - Retr Vt(p)/Nc )
+//     = 2 * SG[Vt](Nc)
+//We divide by the volume to get an energy density per site, as is convention
 template <class Gimpl>
 RealD WilsonFlowBase<Gimpl>::energyDensityPlaquette(const RealD t, const GaugeField& U){
-  static WilsonGaugeAction<Gimpl> SG(3.0);
+  static WilsonGaugeAction<Gimpl> SG(Gimpl::num_colours);
   return 2.0 * t * t * SG.S(U)/U.Grid()->gSites();
 }
 

Grid/qcd/utils/Sp2n.impl.h

@@ -254,9 +254,9 @@ static void testGenerators(GroupName::Sp) {
   }
 }
 
-template <int N>
-static Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > >
-ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu, GroupName::Sp) {
+template <class vtype, int N>
+static Lattice<iScalar<iScalar<iMatrix<vtype, N> > > >
+ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vtype, N> > > > &Umu, GroupName::Sp) {
   return ProjectOnSpGroup(Umu);
 }
 

Grid/qcd/utils/WilsonLoops.h

@@ -177,25 +177,43 @@ public:
   }
 
   //////////////////////////////////////////////////
-  // average over all x,y,z the temporal loop
+  // average Polyakov loop in mu direction over all directions != mu
   //////////////////////////////////////////////////
-  static ComplexD avgPolyakovLoop(const GaugeField &Umu) {  //assume Nd=4
-    GaugeMat Ut(Umu.Grid()), P(Umu.Grid());
+  static ComplexD avgPolyakovLoop(const GaugeField &Umu, const int mu) {  //assume Nd=4
+    
+    // Protect against bad value of mu [0, 3]
+    if ((mu < 0 ) || (mu > 3)) {
+      std::cout << GridLogError << "Index is not an integer inclusively between 0 and 3." << std::endl;
+      exit(1);
+    }
+
+    // U_loop is U_{mu}
+    GaugeMat U_loop(Umu.Grid()), P(Umu.Grid());
     ComplexD out;
     int T = Umu.Grid()->GlobalDimensions()[3];
     int X = Umu.Grid()->GlobalDimensions()[0];
     int Y = Umu.Grid()->GlobalDimensions()[1];
     int Z = Umu.Grid()->GlobalDimensions()[2];
 
-    Ut = peekLorentz(Umu,3); //Select temporal direction
-    P = Ut;
-    for (int t=1;t<T;t++){ 
-      P = Gimpl::CovShiftForward(Ut,3,P);
+    // Number of sites in mu direction
+    int N_mu = Umu.Grid()->GlobalDimensions()[mu];
+
+    U_loop = peekLorentz(Umu, mu); //Select direction
+    P = U_loop;
+    for (int t=1;t<N_mu;t++){ 
+      P = Gimpl::CovShiftForward(U_loop,mu,P);
     }
    RealD norm = 1.0/(Nc*X*Y*Z*T);
    out = sum(trace(P))*norm;
    return out;   
-}
+  }  
+
+  /////////////////////////////////////////////////
+  // overload for temporal Polyakov loop
+  /////////////////////////////////////////////////
+  static ComplexD avgPolyakovLoop(const GaugeField &Umu) { 
+    return avgPolyakovLoop(Umu, 3);
+  }
 
   //////////////////////////////////////////////////
   // average over traced single links

Grid/stencil/Stencil.h

@@ -751,7 +751,7 @@ public:
     obj.xbytes      = xbytes;
     obj.rbytes      = rbytes;
     obj.cb          = cb;
-
+    
     for(int i=0;i<CachedTransfers.size();i++){
       if (   (CachedTransfers[i].direction  ==direction)
 	   &&(CachedTransfers[i].OrthogPlane==OrthogPlane)
@@ -763,11 +763,13 @@ public:
 	     ){
 	// FIXME worry about duplicate with partial compression
 	// Wont happen as DWF has no duplicates, but...
-	AddCopy(CachedTransfers[i].recv_buf,recv_buf,rbytes);
-	return 1;
+	//	AddCopy(CachedTransfers[i].recv_buf,recv_buf,rbytes);
+	//	std::cout << "Duplicate dir " <<direction<<" "<<" OrthogPlane "<<OrthogPlane<<" Dest"<<DestProc <<" xbytes " <<xbytes<<" lane "<< lane<<" cb "<<cb<<std::endl;
+	return 0;
+	
+	//	return 1;
       }
     }
-
     CachedTransfers.push_back(obj);
     return 0;
   }

Grid/threads/ThreadReduction.h

@@ -28,6 +28,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once 
 
+#ifndef MIN
+#define MIN(x,y) ((x)>(y)?(y):(x))
+#endif
+
+
 // Introduce a class to gain deterministic bit reproducible reduction.
 // make static; perhaps just a namespace is required.
 NAMESPACE_BEGIN(Grid);

HMC/FTHMC2p1f.cc

@@ -24,7 +24,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
@@ -230,3 +234,4 @@ int main(int argc, char **argv)
 #endif
 } // main
 
+#endif

HMC/FTHMC2p1f_3GeV.cc

@@ -25,7 +25,11 @@ directory
 *************************************************************************************/
 /*  END LEGAL */
 
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
@@ -231,5 +235,4 @@ int main(int argc, char **argv)
 #endif
 } // main
 
-
-
+#endif

HMC/HMC2p1f_3GeV.cc

@@ -24,7 +24,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
@@ -230,5 +234,4 @@ int main(int argc, char **argv)
 #endif
 } // main
 
-
-
+#endif

HMC/Mobius2p1f.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -195,5 +199,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1fEOFA.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #ifdef GRID_DEFAULT_PRECISION_DOUBLE
 #define MIXED_PRECISION
@@ -449,5 +453,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1fEOFA_F1.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #ifdef GRID_DEFAULT_PRECISION_DOUBLE
 #define MIXED_PRECISION
@@ -442,5 +446,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1fIDSDRGparityEOFA_40ID.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 using namespace Grid;
 
@@ -918,3 +922,5 @@ int main(int argc, char **argv) {
   return 0;
 #endif
 } // main
+
+#endif

HMC/Mobius2p1fIDSDRGparityEOFA_48ID.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 using namespace Grid;
 
@@ -873,3 +877,5 @@ int main(int argc, char **argv) {
   return 0;
 #endif
 } // main
+
+#endif

HMC/Mobius2p1fRHMC.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -193,5 +197,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_3level.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -512,5 +516,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_double.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -345,5 +349,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_mixed.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -516,5 +520,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_mshift.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -567,5 +571,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_RHMC.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -263,5 +267,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_RHMC_96I.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -417,5 +421,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_RHMC_96I_mixed.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -452,5 +456,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_EOFA_96I_hmc.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -462,5 +466,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_EOFA_96I_hmc_double.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 
 
@@ -264,5 +268,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/disable_examples_without_instantiations.h

@@ -0,0 +1,16 @@
+#include <Grid/Grid.h>
+#pragma once
+
+
+#ifndef ENABLE_FERMION_INSTANTIATIONS
+#include <iostream>
+
+int main(void) {
+  std::cout << "This build of Grid was configured to exclude fermion instantiations, "
+	    << "which this example relies on. "
+	    << "Please reconfigure and rebuild Grid with --enable-fermion-instantiations"
+	    << "to run this example."
+	    << std::endl;
+  return 1;
+}
+#endif

benchmarks/Benchmark_ITT.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace Grid;
@@ -731,3 +734,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_dwf.cc

@@ -20,6 +20,9 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -439,3 +442,4 @@ void Benchmark(int Ls, Coordinate Dirichlet,bool sloppy)
   GRID_ASSERT(norm2(src_e)<1.0e-4);
   GRID_ASSERT(norm2(src_o)<1.0e-4);
 }
+#endif

benchmarks/Benchmark_dwf_fp32.cc

@@ -20,6 +20,10 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -439,3 +443,5 @@ void Benchmark(int Ls, Coordinate Dirichlet,bool sloppy)
   GRID_ASSERT(norm2(src_e)<1.0e-4);
   GRID_ASSERT(norm2(src_o)<1.0e-4);
 }
+
+#endif

benchmarks/Benchmark_dwf_fp32_paranoid.cc

@@ -20,6 +20,9 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -385,3 +388,5 @@ int main (int argc, char ** argv)
   Grid_finalize();
   exit(0);
 }
+
+#endif

benchmarks/Benchmark_dwf_sweep.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -238,5 +241,4 @@ void benchDw(std::vector<int> & latt4, int Ls, int threads,int report )
   }
 }
 
-
-
+#endif

benchmarks/Benchmark_gparity.cc

@@ -1,3 +1,7 @@
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+
 #include <Grid/Grid.h>
 #include <sstream>
 using namespace std;
@@ -155,3 +159,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
+#endif

benchmarks/Benchmark_halo.cc

@@ -20,6 +20,9 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -129,3 +132,5 @@ int main (int argc, char ** argv)
   Grid_finalize();
   exit(0);
 }
+
+#endif

benchmarks/Benchmark_mooee.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -149,3 +152,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_schur.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -172,5 +175,4 @@ void benchDw(std::vector<int> & latt4, int Ls)
   //  Dw.Report();
 }
 
-
-
+#endif

benchmarks/Benchmark_staggered.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -110,3 +113,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_staggeredF.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -112,3 +115,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_usqcd.cc

@@ -26,6 +26,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+
 #include <Grid/Grid.h>
 #include <Grid/algorithms/blas/BatchedBlas.h>
 
@@ -873,7 +877,7 @@ int main (int argc, char ** argv)
   int do_su4=0;
   int do_memory=1;
   int do_comms =1;
-  int do_blas  =1;
+  int do_blas  =0;
   int do_dslash=1;
 
   int sel=4;
@@ -978,3 +982,5 @@ int main (int argc, char ** argv)
   Grid_finalize();
   fclose(FP);
 }
+
+#endif

benchmarks/Benchmark_wilson.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -258,3 +261,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_wilson_sweep.cc

@@ -19,6 +19,9 @@ Author: Richard Rollins <rprollins@users.noreply.github.com>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -161,3 +164,5 @@ void bench_wilson_eo (
   double flops = (single_site_flops * volume * ncall)/2.0;
   std::cout << flops/(t1-t0) << "\t\t";
 }
+
+#endif

benchmarks/disable_benchmarks_without_instantiations.h

@@ -0,0 +1,16 @@
+#include <Grid/Grid.h>
+
+#pragma once
+
+#ifndef ENABLE_FERMION_INSTANTIATIONS
+#include <iostream>
+
+int main(void) {
+  std::cout << "This build of Grid was configured to exclude fermion instantiations, "
+	    << "which this benchmark relies on. "
+	    << "Please reconfigure and rebuild Grid with --enable-fermion-instantiations"
+	    << "to run this benchmark."
+	    << std::endl;
+  return 1;
+}
+#endif

configure.ac

@@ -172,6 +172,12 @@ case ${ac_TRACING} in
 esac
 
 ############### fermions
+AC_ARG_ENABLE([fermion-instantiations],
+     [AS_HELP_STRING([--enable-fermion-instantiations=yes|no],[enable fermion instantiations])],
+     [ac_FERMION_INSTANTIATIONS=${enable_fermion_instantiations}], [ac_FERMION_INSTANTIATIONS=yes])
+
+AM_CONDITIONAL(BUILD_FERMION_INSTANTIATIONS, [ test "${ac_FERMION_INSTANTIATIONS}X" == "yesX" ])
+
 AC_ARG_ENABLE([fermion-reps],
      [AS_HELP_STRING([--enable-fermion-reps=yes|no],[enable extra fermion representation support])],
      [ac_FERMION_REPS=${enable_fermion_reps}], [ac_FERMION_REPS=yes])
@@ -194,6 +200,9 @@ AM_CONDITIONAL(BUILD_ZMOBIUS, [ test "${ac_ZMOBIUS}X" == "yesX" ])
 case ${ac_FERMION_REPS} in
    yes) AC_DEFINE([ENABLE_FERMION_REPS],[1],[non QCD fermion reps]);;
 esac
+case ${ac_FERMION_INSTANTIATIONS} in
+   yes) AC_DEFINE([ENABLE_FERMION_INSTANTIATIONS],[1],[enable fermions]);;
+esac
 case ${ac_GPARITY} in
    yes) AC_DEFINE([ENABLE_GPARITY],[1],[fermion actions with GPARITY BCs]);;
 esac
@@ -292,13 +301,14 @@ AC_ARG_ENABLE([accelerator],
 case ${ac_ACCELERATOR} in
     cuda)
       echo CUDA acceleration
-      LIBS="${LIBS} -lcuda -lcublas"
+      LIBS="${LIBS} -lcuda -lcublas -lcufft"
       AC_DEFINE([GRID_CUDA],[1],[Use CUDA offload]);;
     sycl)
       echo SYCL acceleration
       AC_DEFINE([GRID_SYCL],[1],[Use SYCL offload]);;
     hip)
       echo HIP acceleration
+      LIBS="${LIBS} -lhipblas -lrocblas -lhipfft"
       AC_DEFINE([GRID_HIP],[1],[Use HIP offload]);;
     none)
       echo NO acceleration    ;;

examples/Example_Mobius_spectrum.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -310,5 +313,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_arnoldi.cc

@@ -0,0 +1,430 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+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 */
+
+// copied here from Test_general_coarse_pvdagm.cc
+
+#include <cstdlib>
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+using namespace std;
+using namespace Grid;
+
+// Hermitize a DWF operator by squaring it
+template<class Matrix,class Field>
+class SquaredLinearOperator : public LinearOperatorBase<Field> {
+
+  public:
+  Matrix &_Mat;
+
+  public:
+    SquaredLinearOperator(Matrix &Mat): _Mat(Mat) {};
+
+    void OpDiag (const Field &in, Field &out) {    assert(0);  }
+    void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+    void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+    void Op     (const Field &in, Field &out){
+      // std::cout << "Op is overloaded as HermOp" << std::endl;
+      HermOp(in, out);
+    }
+    void AdjOp     (const Field &in, Field &out){
+      HermOp(in, out);
+    }
+    void _Op     (const Field &in, Field &out){
+      // std::cout << "Op: M "<<std::endl;
+      _Mat.M(in, out);
+    }
+    void _AdjOp     (const Field &in, Field &out){
+      // std::cout << "AdjOp: Mdag "<<std::endl;
+      _Mat.Mdag(in, out);
+    }
+    void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+    void HermOp(const Field &in, Field &out){
+      // std::cout << "HermOp: Mdag M Mdag M"<<std::endl;
+      Field tmp(in.Grid());
+      _Op(in,tmp);
+      _AdjOp(tmp,out);
+    }
+};
+
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    //    _Mat.M(in,tmp);
+    //    _PV.Mdag(tmp,out);
+    //    _PV.M(out,tmp);
+    //    _Mat.Mdag(tmp,out);
+    Op(in,tmp);
+    AdjOp(tmp,out);
+    //    std::cout << "HermOp done "<<norm2(out)<<std::endl;
+  }
+};
+
+template<class Matrix,class Field>
+class ShiftedPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  RealD shift;
+public:
+  ShiftedPVdagMLinearOperator(RealD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+};
+
+template<class Matrix, class Field>
+class ShiftedComplexPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  ComplexD shift;
+public:
+ShiftedComplexPVdagMLinearOperator(ComplexD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+  
+  void resetShift(ComplexD newShift) {
+    shift = newShift;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class MGPreconditioner : public LinearFunction< Lattice<Fobj> > {
+public:
+  using LinearFunction<Lattice<Fobj> >::operator();
+
+  typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::FineField    FineField;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
+  typedef LinearOperatorBase<FineField>                            FineOperator;
+  typedef LinearFunction    <FineField>                            FineSmoother;
+  typedef LinearOperatorBase<CoarseVector>                         CoarseOperator;
+  typedef LinearFunction    <CoarseVector>                         CoarseSolver;
+  Aggregates     & _Aggregates;
+  FineOperator   & _FineOperator;
+  FineSmoother   & _PreSmoother;
+  FineSmoother   & _PostSmoother;
+  CoarseOperator & _CoarseOperator;
+  CoarseSolver   & _CoarseSolve;
+
+  int    level;  void Level(int lv) {level = lv; };
+
+  MGPreconditioner(Aggregates &Agg,
+		   FineOperator &Fine,
+		   FineSmoother &PreSmoother,
+		   FineSmoother &PostSmoother,
+		   CoarseOperator &CoarseOperator_,
+		   CoarseSolver &CoarseSolve_)
+    : _Aggregates(Agg),
+      _FineOperator(Fine),
+      _PreSmoother(PreSmoother),
+      _PostSmoother(PostSmoother),
+      _CoarseOperator(CoarseOperator_),
+      _CoarseSolve(CoarseSolve_),
+      level(1)  {  }
+
+  virtual void operator()(const FineField &in, FineField & out) 
+  {
+    GridBase *CoarseGrid = _Aggregates.CoarseGrid;
+    //    auto CoarseGrid = _CoarseOperator.Grid();
+    CoarseVector Csrc(CoarseGrid);
+    CoarseVector Csol(CoarseGrid);
+    FineField vec1(in.Grid());
+    FineField vec2(in.Grid());
+
+    std::cout<<GridLogMessage << "Calling PreSmoother " <<std::endl;
+
+    //    std::cout<<GridLogMessage << "Calling PreSmoother input residual "<<norm2(in) <<std::endl;
+    double t;
+    // Fine Smoother
+    //    out = in;
+    out = Zero();
+    t=-usecond();
+    _PreSmoother(in,out);
+    t+=usecond();
+
+    std::cout<<GridLogMessage << "PreSmoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Update the residual
+    _FineOperator.Op(out,vec1);  sub(vec1, in ,vec1);   
+    //    std::cout<<GridLogMessage <<"Residual-1 now " <<norm2(vec1)<<std::endl;
+
+    // Fine to Coarse 
+    t=-usecond();
+    _Aggregates.ProjectToSubspace  (Csrc,vec1);
+    t+=usecond();
+    std::cout<<GridLogMessage << "Project to coarse took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse correction
+    t=-usecond();
+    Csol = Zero();
+    _CoarseSolve(Csrc,Csol);
+    //Csol=Zero();
+    t+=usecond();
+    std::cout<<GridLogMessage << "Coarse solve took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse to Fine
+    t=-usecond();  
+    //    _CoarseOperator.PromoteFromSubspace(_Aggregates,Csol,vec1);
+    _Aggregates.PromoteFromSubspace(Csol,vec1); 
+    add(out,out,vec1);
+    t+=usecond();
+    std::cout<<GridLogMessage << "Promote to this level took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Residual
+    _FineOperator.Op(out,vec1);  sub(vec1 ,in , vec1);  
+    //    std::cout<<GridLogMessage <<"Residual-2 now " <<norm2(vec1)<<std::endl;
+
+    // Fine Smoother
+    t=-usecond();
+    //    vec2=vec1;
+    vec2=Zero();
+    _PostSmoother(vec1,vec2);
+    t+=usecond();
+    std::cout<<GridLogMessage << "PostSmoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    add( out,out,vec2);
+    std::cout<<GridLogMessage << "Done " <<std::endl;
+  }
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=16;
+
+//   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  std::vector<int> lat_size {16, 16, 16, 32};
+  std::cout << "Lattice size: " << lat_size << std::endl;
+  GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(lat_size, 
+								          GridDefaultSimd(Nd,vComplex::Nsimd()),
+								          GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  // Construct a coarsened grid
+  // poare TODO: replace this with the following line?
+  Coordinate clatt = lat_size;
+//   Coordinate clatt = GridDefaultLatt();              // [PO] initial line before I edited it
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/2;
+    //    clatt[d] = clatt[d]/4;
+  }
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  LatticeFermion    src(FGrid); random(RNG5,src);
+  LatticeFermion result(FGrid); result=Zero();
+  LatticeFermion    ref(FGrid); ref=Zero();
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+//   std::string file("ckpoint_lat.4000");
+  std::string file("/Users/patrickoare/libraries/PETSc-Grid/ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  RealD mass=0.01;
+  RealD M5=1.8;
+
+  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionD Dpv(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
+
+  // const int nbasis = 20;            // size of approximate basis for low-mode space
+  const int nbasis = 3;            // size of approximate basis for low-mode space
+  const int cb = 0 ;
+  LatticeFermion prom(FGrid);
+
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNearestStencilGeometry5D geom(Coarse5d);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+
+  typedef PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM_t;
+  typedef ShiftedPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedPVdagM_t;
+  typedef ShiftedComplexPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedComplexPVdagM_t;
+  PVdagM_t PVdagM(Ddwf, Dpv);
+  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv);
+  SquaredLinearOperator<DomainWallFermionD, LatticeFermionD> Dsq (Ddwf);
+  NonHermitianLinearOperator<DomainWallFermionD, LatticeFermionD> DLinOp (Ddwf);
+
+  // PowerMethod<LatticeFermion> PM; PM(PVdagM, src);
+  int Nm = 10;
+  int Nk = 6;
+  // int Nm = 6;         // Nm = 6 case is acting really strangely... with Nm = 6 and Nm = 3 it zeros out the Hessenberg and also makes it imaginary?
+  // int Nk = 2;
+  // int Nk = Nm+1;     // if just running once
+  // int maxIter = 5;
+  // int maxIter = 1;
+  int maxIter = 3;
+  // int maxIter = 100;
+  int Nstop = 6;
+
+  Coordinate origin ({0,0,0,0});
+  auto tmpSrc = peekSite(src, origin);
+  std::cout << "[DEBUG] Source at origin = " <<  tmpSrc << std::endl;
+  LatticeFermion src2 = src;
+
+  // Run Lanczos and Arnoldi on a Hermitian matrix
+  // Arnoldi Arn (Dsq, FGrid, 1e-8, false);
+  // Arn(src, 1, Nm, -1);
+  Arnoldi Arn (Dsq, FGrid, 1e-8, EvalNormLarge);      // for comparison to Lanczos
+  // Arn(src, maxIter, Nm, Nk, Nstop);
+
+  // auto tmpSrcDup = peekSite(src, origin);
+  // std::cout << "[DEBUG] Source at origin = " <<  tmpSrcDup << std::endl;
+  // auto tmpSrc2Dup = peekSite(src2, origin);
+  // std::cout << "[DEBUG] Source2 at origin = " <<  tmpSrc2Dup << std::endl;
+
+  Arn(src, maxIter, Nm, Nk, Nstop);
+  std::cout << "Hessenberg mat for symmetric N = " << Nm << std::endl;
+  std::cout << Arn.getHessenbergMat() << std::endl;
+
+  // ImplicitlyRestartedLanczosHermOpTester<LatticeFermionD> SimpleTester (Dsq);
+  // ImplicitlyRestartedLanczos<LatticeFermionD> Lanc (Dsq, Dsq, SimpleTester, Nm, Nm, Nm, 1e-8, Nm);
+  int Nconv;
+  PlainHermOp DsqHermOp (Dsq);
+  // std::vector<RealD> levals (Nm+1); std::vector<LatticeFermionD> levecs (Nm+1, src);
+  // ImplicitlyRestartedLanczos<LatticeFermionD> Lanc (DsqHermOp, DsqHermOp, Nm, Nm, Nm + 1, 1e-8, Nm);
+  std::vector<RealD> levals (Nm+1); std::vector<LatticeFermionD> levecs (Nm, src);
+  ImplicitlyRestartedLanczos<LatticeFermionD> Lanc (DsqHermOp, DsqHermOp, Nstop, Nk, Nm, 1e-8, maxIter);
+  std::cout << GridLogMessage << "Calculating with Lanczos" << std::endl;
+
+  // auto tmpSrc1 = peekSite(src, origin);
+  // std::cout << "[DEBUG] Source at origin = " <<  tmpSrc1 << std::endl;
+  // auto tmpSrc2 = peekSite(src2, origin);
+  // std::cout << "[DEBUG] Source2 at origin = " <<  tmpSrc2 << std::endl;
+  // std::cout << "[DEBUG] Source norm2: " << norm2(src) << std::endl;
+
+  std::cout << "running Lanczos now" << std::endl;
+  Lanc.calc(levals, levecs, src2, Nconv);
+
+  std::cout<<GridLogMessage << "*******************************************" << std::endl;
+  std::cout<<GridLogMessage << "***************** RESULTS *****************" << std::endl;
+  std::cout<<GridLogMessage << "*******************************************" << std::endl;
+
+  std::cout << GridLogMessage << "Arnoldi eigenvalues: " << std::endl << Arn.getEvals() << std::endl;
+  std::cout << GridLogMessage << "Lanczos eigenvalues: " << std::endl << levals << std::endl;
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/Example_christoph.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -432,5 +435,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_gamma5_block_lanczos.cc

@@ -0,0 +1,188 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./examples/Example_gamma5_block_lanczos.cc
+
+    Copyright (C) 2026
+
+    Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+    γ5-Block Lanczos example for the Wilson Dirac operator.
+
+    Reads a gauge configuration from "config" (NERSC format) and Lanczos
+    parameters from "LanParams.xml".  Runs Gamma5BlockLanczos to
+    compute eigenvalues of D_W directly (not H_W = γ5 D_W).
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution
+    directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <cstdlib>
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+using namespace std;
+using namespace Grid;
+
+namespace Grid {
+
+struct LanczosParameters : Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+    RealD,   mass,
+    Integer, Nstop,
+    Integer, Nk,
+    Integer, Np,
+    Integer, maxIter,
+    Integer, reorthog,
+    Integer, verify,
+    Integer, ReadEvec,
+    RealD,   resid)
+
+  LanczosParameters()
+    : mass(-0.5), Nstop(10), Nk(20), maxIter(100),
+      reorthog(1), verify(0), ReadEvec(0), resid(1e-8)
+  {}
+
+  template<class ReaderClass>
+  LanczosParameters(Reader<ReaderClass>& r) { initialize(r); }
+
+  template<class ReaderClass>
+  void initialize(Reader<ReaderClass>& r) {
+    read(r, "LanczosParameters", *this);
+  }
+};
+
+} // namespace Grid
+
+template<class T>
+void writeField(T& in, std::string const& fname) {
+#ifdef HAVE_LIME
+  std::cout << GridLogMessage << "Writing to: " << fname << std::endl;
+  Grid::emptyUserRecord record;
+  Grid::ScidacWriter WR(in.Grid()->IsBoss());
+  WR.open(fname);
+  WR.writeScidacFieldRecord(in, record, 0);
+  WR.close();
+#endif
+}
+
+typedef WilsonFermionD WilsonOp;
+typedef typename WilsonFermionD::FermionField FermionField;
+
+int main(int argc, char** argv)
+{
+  Grid_init(&argc, &argv);
+
+  GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(
+      GridDefaultLatt(),
+      GridDefaultSimd(Nd, vComplex::Nsimd()),
+      GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  std::vector<int> seeds4({1, 2, 3, 4});
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+
+  // Read gauge configuration
+  LatticeGaugeField Umu(UGrid);
+  FieldMetaData header;
+  std::string configFile("config");
+  NerscIO::readConfiguration(Umu, header, configFile);
+  std::cout << GridLogMessage << "Loaded gauge configuration: " << configFile << std::endl;
+
+  // Read Lanczos parameters
+  LanczosParameters Params;
+  {
+    XmlReader rd("LanParams.xml");
+    read(rd, "LanczosParameters", Params);
+  }
+  std::cout << GridLogMessage << Params << std::endl;
+
+  // Build Wilson Dirac operator and wrap in a non-Hermitian linear operator
+  std::vector<Complex> boundary = {1, 1, 1, -1};
+  WilsonOp::ImplParams WilsonParams(boundary);
+  WilsonOp Dwilson(Umu, *UGrid, *UrbGrid, Params.mass, WilsonParams);
+  NonHermitianLinearOperator<WilsonOp, FermionField> DLinOp(Dwilson);
+
+  // γ5 functor: for 4D Wilson fermions γ5 is Gamma(Gamma5)
+  Gamma G5(Gamma::Algebra::Gamma5);
+  auto gamma5 = [&G5](const FermionField& in, FermionField& out) {
+    out = G5 * in;
+  };
+
+  // Starting vectors: two independent random vectors
+  FermionField src(UGrid), src2(UGrid);
+  random(RNG4, src);
+  random(RNG4, src2);
+  std::cout << GridLogMessage << "Using two random starting vectors" << std::endl;
+
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "*******************************************" << std::endl;
+  std::cout << GridLogMessage << " Running γ5-Block Lanczos" << std::endl;
+  std::cout << GridLogMessage << "   mass     = " << Params.mass     << std::endl;
+  std::cout << GridLogMessage << "   Nk       = " << Params.Nk       << std::endl;
+  std::cout << GridLogMessage << "   maxIter  = " << Params.maxIter  << std::endl;
+  std::cout << GridLogMessage << "   Nstop    = " << Params.Nstop    << std::endl;
+  std::cout << GridLogMessage << "   resid    = " << Params.resid    << std::endl;
+  std::cout << GridLogMessage << "   reorthog = " << Params.reorthog << std::endl;
+  std::cout << GridLogMessage << "   verify   = " << Params.verify   << std::endl;
+  std::cout << GridLogMessage << "*******************************************" << std::endl;
+  std::cout << GridLogMessage << std::endl;
+
+  Gamma5BlockLanczos<FermionField> G5BL(DLinOp, UGrid, gamma5, Params.resid);
+  G5BL.doEvalCheck = (Params.verify != 0);
+  G5BL.doVerify = (Params.verify != 0);
+//  G5BL(src, src2, Params.maxIter, Params.Nstop, Params.reorthog != 0);
+  G5BL.restart(src, src2, Params.maxIter, Params.Nk+Params.Np, Params.Nk, Params.Nstop, Params.reorthog != 0, EvalNormSmall);
+//  G5BL.implicitRestart(src, src2, Params.maxIter, Params.Nk+Params.Np, Params.Nk, Params.Nstop, Params.reorthog != 0, EvalNormSmall);
+  if (Params.verify ) G5BL.verify("after restart");
+
+  // Summary of results
+  Eigen::VectorXcd        evals     = G5BL.getEvals();
+  std::vector<RealD>      residuals = G5BL.getResiduals();
+  std::vector<FermionField> evecs   = G5BL.getEvecs();
+
+  int Nout = (int)evals.size();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "*******************************************" << std::endl;
+  std::cout << GridLogMessage << " γ5-Block Lanczos: " << Nout << " Ritz pairs" << std::endl;
+  std::cout << GridLogMessage << "*******************************************" << std::endl;
+  for (int i = 0; i < Nout; i++) {
+    std::cout << GridLogMessage
+              << "  [" << std::setw(3) << i << "]"
+              << "  lambda = " << evals(i)
+              << "  |res| = " << residuals[i] << std::endl;
+  }
+
+  // Write the first Nstop eigenvectors (in SCIDAC format when LIME is available)
+  int Nwrite = std::min((int)Params.Nstop, Nout);
+  for (int i = 0; i < Nwrite; i++) {
+    std::string fname = "./g5bl_evec_m" + std::to_string(Params.mass)
+                        + "_" + std::to_string(i);
+    writeField(evecs[i], fname);
+  }
+
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "Done" << std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/Example_krylov_coeffs.cc

@@ -0,0 +1,405 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+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 */
+
+// Tests code written to read off the Krylov coefficients
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+#include <Grid/algorithms/iterative/ConjugateGradient.h>
+
+using namespace std;
+using namespace Grid;
+
+// Hermitize a DWF operator by squaring it
+template<class Matrix,class Field>
+class SquaredLinearOperator : public LinearOperatorBase<Field> {
+
+  public:
+  Matrix &_Mat;
+
+  public:
+    SquaredLinearOperator(Matrix &Mat): _Mat(Mat) {};
+
+    void OpDiag (const Field &in, Field &out) {    assert(0);  }
+    void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+    void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+    void Op     (const Field &in, Field &out){
+      // std::cout << "Op is overloaded as HermOp" << std::endl;
+      HermOp(in, out);
+    }
+    void AdjOp     (const Field &in, Field &out){
+      HermOp(in, out);
+    }
+    void _Op     (const Field &in, Field &out){
+      // std::cout << "Op: M "<<std::endl;
+      _Mat.M(in, out);
+    }
+    void _AdjOp     (const Field &in, Field &out){
+      // std::cout << "AdjOp: Mdag "<<std::endl;
+      _Mat.Mdag(in, out);
+    }
+    void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+    void HermOp(const Field &in, Field &out){
+      // std::cout << "HermOp: Mdag M Mdag M"<<std::endl;
+      Field tmp(in.Grid());
+      _Op(in,tmp);
+      _AdjOp(tmp,out);
+    }
+};
+
+/**
+ * Computes the coefficients in the Krylov expansion for 1/D ~ \sum_{i=0}^N c_i D^i. 
+ * 
+ * Parameters
+ * ----------
+ * std::vector<double> &coeffs
+ *    Polynomial coeffients to return, with indexing order (c_0, c_1, c_2, ..., c_n). 
+ * LinearOperatorBase<FineField> &DiracOp
+ *    Dirac operator D. 
+ * FineField src
+ *    Source field b. 
+ * FineField psiStar
+ *    Output approximation for D^{-1} b coming from a Krylov method. 
+ * int N
+ *    Dimension of the polynomial approximation (Krylov space K_{N-1} = {b, Db, D^2 b, ..., D^{N-1} b}).
+ */
+void poly_coeffs(std::vector<ComplexD> &coeffs, LinearOperatorBase<LatticeFermion> &DiracOp, LatticeFermion src,
+                  LatticeFermion psiStar, GridCartesian* FGrid, int N, bool use_herm = false)
+{
+  // stdBasis = {b, Db, D^2 b, ..., D^N b}, kryBasis = {k0, k1, ..., kN}
+  std::vector<LatticeFermion> kryBasis;
+  Eigen::VectorXcd psiStarCoeffs (N);
+
+  // Normalize by 1 / ||src||; does not change the polynomial coefficients
+  double srcNorm   = 1 / std::sqrt(norm2(src));
+  kryBasis.push_back(srcNorm * src);                // normalized source
+  psiStar          = srcNorm * psiStar;
+  psiStarCoeffs(0) = innerProduct(kryBasis[0], psiStar);
+
+  // orthonormalize canonical Krylov basis {b, Db, D^2 b, ..., D^{N-1} b} <--> {k_i} and compute components <k_i | psi*>
+  LatticeFermion tmp (FGrid);
+  for (int i = 0; i < N - 1; i++) {               // construct ONB for {b, Db, ..., D^{i+1} b}
+    if (use_herm) {
+      DiracOp.HermOp(kryBasis.back(), tmp);         // tmp \in span{(D^\dag D)^{i+1} b} \oplus span{(D^\dag D)^i b, ..., D^\dag D b, b}
+    } else {
+      DiracOp.Op(kryBasis.back(), tmp);             // tmp \in span{D^{i+1} b} \oplus span{D^i b, ..., Db, b}
+    }
+
+    for (int j = 0; j < i+1; j++) {               // orthogonalize tmp with previous basis vectors
+      ComplexD coeff = innerProduct(kryBasis[j], tmp);      // <k_j | tmp>
+      tmp -= coeff * kryBasis[j];                           // subtract off |k_j><k_j | tmp>; now tmp is perp to |k_j>
+    }
+    double tmpNorm = 1 / std::sqrt(norm2(tmp));
+    kryBasis.push_back(
+      tmpNorm * tmp
+    );                                                      // normalize |k_i> and add to kryBasis
+    psiStarCoeffs(i+1) = innerProduct(kryBasis[i+1], psiStar);  // compute < k_i | psi* >
+  }
+
+  // To verify the basis is ONB
+  // for (int i = 0; i < N; i++) {
+  //   for (int j = 0; j < N; j++) {
+  //     std::cout << "<ki|kj> for (i,j) = (" << i << ", " << j << ") = "  << innerProduct(kryBasis[i], kryBasis[j]) << std::endl;
+  //   }
+  // }
+
+  // Compute change of basis matrix
+  LatticeFermion tmp2 (FGrid);
+  Eigen::MatrixXcd M = Eigen::MatrixXcd::Zero(N, N);
+  tmp = kryBasis[0];       // current Krylov vector; starts with tmp = src (normalized)
+  for (int i = 0; i < N; i++) {
+    for (int j = 0; j < i + 1; j++) {    // fill column with components of kryVec. Only need j <= i to get orthonormal components
+      M(j, i) = innerProduct(kryBasis[j], tmp);
+    }    
+    if (use_herm) {     // tmp --> D^\dag D(tmp)
+      DiracOp.HermOp(tmp, tmp2);
+      tmp = tmp2;
+    } else {      // tmp --> D(tmp). Note that DiracOp.Op(tmp, tmp) will cause a bug
+      DiracOp.Op(tmp, tmp2);
+      tmp = tmp2;
+    }
+  }
+  
+  // Compute M^{-1} @ psiStarCoeffs and copy to coeffs
+  Eigen::VectorXcd res (N);
+  res = M.inverse() * psiStarCoeffs;
+  for (int i = 0; i < N; i++) {
+    coeffs[i] = res(i);
+  }
+
+}
+
+// out file for poly coefficients (should it be complex?)
+// class PolynomialFile: Serializable {
+// public:
+//   GRID_SERIALIZABLE_CLASS_MEMBERS(OutputFile, std::vector< Real >, data);
+// };
+
+std::complex<double> poly_approx(std::complex<double> x, std::vector<std::complex<double>> coeffs) {
+  std::complex<double> px;
+  for (int i = 0; i < coeffs.size(); i++) {
+    px += coeffs[i] * std::pow(x, i);
+  }
+  return px;
+}
+
+/**
+ * Returns the approximation psi = \sum_i c_i D^i b resulting from a Krylov solver.
+ * 
+ * Parameters
+ * ----------
+ * LatticeFermion &psi
+ *    Approximation field, returned psi = \sum_i c_i D^i b.
+ * LatticeFermion src
+ *    Source b used to generate the Krylov space K_n(D, b).
+ * LinearOperatorBase<LatticeFermion> &Linop
+ *    Dirac operator used to generate the Krylov space K_n(D, b).
+ * std::vector<std::complex<double>> coeffs
+ *    Polynomial coefficients returned from the solver. 
+ */
+void krylovApprox(LatticeFermion &psi, LatticeFermion src, LinearOperatorBase<LatticeFermion> &Linop, std::vector<ComplexD> coeffs) {
+  psi = Zero();
+  LatticeFermion tmp (psi.Grid());
+  tmp = src;
+  LatticeFermion tmp2 (psi.Grid());
+  for (int i = 0; i < coeffs.size(); i++) {
+      psi = psi + coeffs[i] * tmp;
+      Linop.Op(tmp, tmp2);              // tmp = D*tmp
+      tmp = tmp2;
+  }
+}
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc, &argv);
+  const int Ls = 8;
+  std::vector<int> lat_size {16, 16, 16, 32};
+  std::cout << "Lattice size: " << lat_size << std::endl;
+  GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(lat_size, 
+								          GridDefaultSimd(Nd,vComplex::Nsimd()),
+								          GridDefaultMpi());
+
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  //////////////////////////////////////////////////////////////////////
+  // You can manage seeds however you like.
+  // Recommend SeedUniqueString.
+  //////////////////////////////////////////////////////////////////////
+  // std::vector<int> seeds4({1, 2, 3, 4}); 
+  // GridParallelRNG RNG4(UGrid);
+  // RNG4.SeedFixedIntegers(seeds4);
+
+  // std::vector<int> seeds5({1, 2, 3, 4, 5}); 
+  // GridParallelRNG RNG5(FGrid);
+  // RNG5.SeedFixedIntegers(seeds5);
+
+  // std::string outStrStem = "/Users/patrickoare/Dropbox (MIT)/research/multigrid/grid_out/";
+
+  LatticeGaugeField Umu(UGrid);
+  FieldMetaData header;
+  std::string file("/Users/patrickoare/libraries/PETSc-Grid/ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu, header, file);
+  
+  RealD mass=0.01;
+  RealD M5=1.8;
+  // RealD M5=1.0;
+  RealD b=1.5;// Scale factor b+c=2, b-c=1
+  RealD c=0.5;
+
+  // load in Dirac operators that we'll use; square it to Hermitize
+  // Dsq just needs to be a Hermitian operator so we can use CG on it
+  DomainWallFermionD Ddwf(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
+  SquaredLinearOperator<DomainWallFermionD, LatticeFermionD> Dsq (Ddwf);
+  NonHermitianLinearOperator<DomainWallFermionD, LatticeFermionD> DLinOp (Ddwf);
+
+  LatticeFermion src (FGrid); src = 1.0;                              // Source to use 
+  LatticeFermion psiCG (FGrid); psiCG = Zero();                       // Field to solve with for CG
+  LatticeFermion psiGCR (FGrid); psiGCR = Zero();                     // Field to solve with for GCR
+
+  std::cout << GridLogMessage << "*******************************************" << std::endl;
+  std::cout << GridLogMessage << "********** TESTING CG POLY COEFFS *********" << std::endl;
+  std::cout << GridLogMessage << "*******************************************" << std::endl << std::endl;
+
+  double tol = 1.0e-8;
+  int N = 5;           // max iterations (size of Krylov basis)
+
+  // GCR variables
+  int outer_iters = 1;                  // num restarts for GCR
+  TrivialPrecon<LatticeFermionD> prec;  // trivial preconditioner
+
+  ConjugateGradientPolynomial<LatticeFermion> CGP(tol, N, false);
+  CGP(Dsq, src, psiCG);
+
+  // Compute Krylov coeffs directly and compare
+  std::vector<ComplexD> cg_coeffs (N);
+  poly_coeffs(cg_coeffs, Dsq, src, psiCG, FGrid, N, true);
+
+  PolynomialFile PF;
+
+  // Use GCR solver, also get poly coeffs
+  std::vector<ComplexD> gcr_sym_coeffs (N);     // Can try N --> N + 3 to test to see if the last 3 comps are 0
+  PGCRPolynomial<LatticeFermionD> GCRPolySym(tol, outer_iters, Dsq, prec, N+1, N, PF);    // mmax sets the memory, note the last beta doesn't really matter for updating the polynomial
+  GCRPolySym(src, psiGCR);
+  // poly_coeffs(gcr_sym_coeffs, Dsq, src, psi, FGrid, N, true);
+  poly_coeffs(gcr_sym_coeffs, Dsq, src, psiGCR, FGrid, N, true);
+
+  std::cout << GridLogMessage << std::endl << "******** CG POLYNOMIAL COEFFICIENTS *******" << std::endl;
+  std::cout << GridLogMessage << CGP.polynomial << std::endl << std::endl;
+
+  std::cout << GridLogMessage << "****** DIRECT POLYNOMIAL COEFFICIENTS *****" << std::endl;
+  std::cout << GridLogMessage << cg_coeffs << std::endl << std::endl;
+
+  // TODO: try GCR with a Hermitian operator (Dsq)
+  std::cout << GridLogMessage << "****** GCR COEFFICIENTS *****" << std::endl;
+  std::cout << GridLogMessage << GCRPolySym.polynomial << std::endl << std::endl;
+
+  std::cout << GridLogMessage << "****** DIRECT GCR COEFFICIENTS *****" << std::endl;
+  std::cout << GridLogMessage << gcr_sym_coeffs << std::endl << std::endl;
+
+  // test how good the decomposition is
+  std::cout << "Testing fidelity of decomposition by computing ||psi* - sum_i c_i D^i b||^2!" << std::endl;
+  LatticeFermion psiPrime (FGrid);
+
+  // for CG
+  krylovApprox(psiPrime, src, Dsq, cg_coeffs);
+  std::cout << "CG with Dsq, ||psi - psiPrime||^2 = " << norm2(psiCG - psiPrime) << std::endl;
+
+  // for GCR with alpha / beta computation
+  krylovApprox(psiPrime, src, Dsq, GCRPolySym.polynomial);
+  std::cout << "GCR with Dsq, ||psi - psiPrime||^2 = " << norm2(psiGCR - psiPrime) << std::endl;
+
+  // for GCR with alpha / beta computation
+  krylovApprox(psiPrime, src, Dsq, gcr_sym_coeffs);
+  std::cout << "GCR direct with Dsq, ||psi - psiPrime||^2 = " << norm2(psiGCR - psiPrime) << std::endl;
+
+  
+  // std::vector<double> real_cg_diff (N);
+  // for (int i = 0; i < N; i++) { real_cg_diff[i] = std::abs(cg_coeffs[i].real() - CGP.polynomial[i]); }
+  // std::cout << GridLogMessage << "************* COEFF DIFFERENCE ************" << std::endl;
+  // std::cout << GridLogMessage << real_cg_diff << std::endl << std::endl;
+
+  // GCR polynomial reconstruction with Ddwf!
+  std::cout << GridLogMessage << "*******************************************" << std::endl;
+  std::cout << GridLogMessage << "********* TESTING GCR POLY COEFFS *********" << std::endl;
+  std::cout << GridLogMessage << "*******************************************" << std::endl << std::endl;
+
+  // re-init variables
+  src = 1.0;
+  src = (1 / std::sqrt(norm2(src))) * src;
+  psiGCR = Zero(); psiPrime = Zero();
+
+  // test GCR poly
+  PGCRPolynomial<LatticeFermionD> GCRPoly(tol, outer_iters, DLinOp, prec, N+1, N, PF);    // mmax sets the memory, note the last beta doesn't really matter for updating the polynomial
+  GCRPoly(src, psiGCR);
+
+  // Compute Krylov coeffs directly and compare
+  // N = 1;    // compare the N > 1 decomposition with the psi* resulting from N = 1
+  std::vector<ComplexD> gcr_coeffs (N);   // note N --> N + k should just give k coeffs that are 0; this works as intended
+  poly_coeffs(gcr_coeffs, DLinOp, src, psiGCR, FGrid, N, false);
+
+  std::cout << GridLogMessage << "******* GCR POLYNOMIAL COEFFICIENTS *******" << std::endl;
+  std::cout << GridLogMessage << GCRPoly.polynomial << std::endl << std::endl;
+
+  std::cout << GridLogMessage << "****** DIRECT POLYNOMIAL COEFFICIENTS *****" << std::endl;
+  std::cout << GridLogMessage << gcr_coeffs << std::endl << std::endl;
+
+  // test how good the decomposition is
+  std::cout << "Testing fidelity of decomposition by computing ||psi* - sum_i c_i D^i b||^2!" << std::endl;
+
+  // for GCR with alpha / beta computation
+  krylovApprox(psiPrime, src, DLinOp, GCRPoly.polynomial);
+  std::cout << "GCR with Dsq, ||psi - psiPrime||^2 = " << norm2(psiGCR - psiPrime) << std::endl;
+
+  // for GCR with alpha / beta computation
+  krylovApprox(psiPrime, src, DLinOp, gcr_coeffs);
+  std::cout << "GCR direct with Dsq, ||psi - psiPrime||^2 = " << norm2(psiGCR - psiPrime) << std::endl;
+
+  // TESTS TO DO THE N = 2 CASE DIRECTLY
+  /*
+  std::vector<std::complex<double>> alphas {
+    std::complex(0.244300601, 0.00013007545), 
+    std::complex(0.285370971, -0.000160704481)
+  };
+  std::complex<double> beta00 (-0.184661284, -6.52153945e-05);
+  LatticeFermion psi2 (FGrid);
+  LatticeFermion Dsrc (FGrid);
+  DLinOp.Op(src, Dsrc);
+  std::complex<double> c1 = alphas[0] + alphas[1] * (1. + beta00);
+  std::complex<double> c2 = -alphas[0] * alphas[1];
+  psi2 = c1 * src + c2 * Dsrc;
+
+  std::cout << "||b|| = " << norm2(src) << std::endl;
+  std::cout << "||Db|| = " << norm2(Dsrc) << std::endl;
+  // fail; so far this is giving something different than what's being computed in krylovApprox (idk how?)
+
+  std::cout << "c1 and c2 are: " << c1 << " and " << c2 << std::endl;
+  std::cout << "GCRPoly polynomial coeffs are (should equal c1 and c2): " << GCRPoly.polynomial << std::endl;
+  std::cout << "||GCRpsi - psi2||_2^2 = " << norm2(psiGCR - psi2) << std::endl;
+  // pass
+  
+  LatticeFermion src2 (FGrid);
+  src2 = 1.0;
+  src2 = (1 / std::sqrt(norm2(src2))) * src2;
+  std::cout << "||ones - src|| (to verify that src is the same throughout, should be 0) = " << norm2(src2 - src) << std::endl;
+  // pass
+
+  krylovApprox(psiPrime, src, DLinOp, GCRPoly.polynomial);
+  std::cout << "GCR with Dsq, ||psi2 - psiPrime||^2 = " << norm2(psi2 - psiPrime) << std::endl;
+
+  std::vector<ComplexD> psi2_coeffs (N);   // note N --> N + k should just give k coeffs that are 0; this works as intended
+  poly_coeffs(psi2_coeffs, DLinOp, src, psi2, FGrid, N, false);
+  krylovApprox(psiPrime, src, DLinOp, psi2_coeffs);
+  std::cout << "GCR direct with Dsq, ||psi - psiPrime||^2 = " << norm2(psi2 - psiPrime) << std::endl;
+  */
+
+  // std::complex z (10.0, 0.0);     // z = 10
+  // std::cout << GridLogMessage << "************* GCR POLY(z = 10) *************" << std::endl;
+  // std::cout << GridLogMessage << poly_approx(z, GCRPoly.polynomial) << std::endl;
+  // std::cout << GridLogMessage << "************ DIRECT POLY(z = 10) ***********" << std::endl;
+  // std::cout << GridLogMessage << poly_approx(z, gcr_coeffs) << std::endl;
+
+  // std::vector<std::complex<double>> gcr_diff (N);
+  // for (int i = 0; i < N; i++) { gcr_diff[i] = gcr_coeffs[i] - GCRPoly.polynomial[i]; }
+  // std::cout << GridLogMessage << "*********** GCR COEFF DIFFERENCE **********" << std::endl;
+  // std::cout << GridLogMessage << gcr_diff << std::endl << std::endl;
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/Example_krylov_schur.cc

@@ -0,0 +1,425 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./examples/Example_krylov_schur.cc
+
+    Copyright (C) 2023
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Patrick Oare <patrickoare@gmail.com>
+    Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+// copied here from Test_general_coarse_pvdagm.cc
+
+#include <cstdlib>
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+using namespace std;
+using namespace Grid;
+
+namespace Grid {
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass , 
+		  		RealD, mstep , 
+				Integer, Nstop,
+                                Integer, Nk,
+                                Integer, Np,
+                                Integer, ReadEvec,
+                                Integer, maxIter,
+                                Integer, Nblock,
+                                Integer, verify,
+		  		RealD, shift , 
+	  			RealD, resid,
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+
+  LanczosParameters() {
+    ////////////////////////////// Default values
+      mass = 0;
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass > 
+  void initialize(Reader<ReaderClass> &TheReader){
+//    std::cout << GridLogMessage << "Reading HMC\n";
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+  
+};
+
+}
+
+template <class T> void writeFile(T& in, std::string const fname){
+#if 1
+  // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+  std::cout << Grid::GridLogMessage << "Writes to: " << fname << std::endl;
+  Grid::emptyUserRecord record;
+  Grid::ScidacWriter WR(in.Grid()->IsBoss());
+  WR.open(fname);
+  WR.writeScidacFieldRecord(in,record,0);
+  WR.close();
+#endif
+  // What is the appropriate way to throw error?
+}
+
+
+typedef WilsonFermionD WilsonOp;
+typedef typename WilsonFermionD::FermionField FermionField;
+
+template<class Matrix,class Field>
+class InvertNonHermitianLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD _stp;
+public:
+  InvertNonHermitianLinearOperator(Matrix &Mat,RealD stp=1e-8): _Mat(Mat),_stp(stp){};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    assert(0);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+//    _Mat.MdirAll(in,out);
+    assert(0);
+  };
+  void Op     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _Mat.Mdag(in,tmp);
+    MdagMLinearOperator<Matrix,Field> HermOp(_Mat);
+    ConjugateGradient<Field> CG(_stp,10000);
+    CG(HermOp,tmp,out);
+//    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    _Mat.Mdag(in,out);
+//    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    assert(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    assert(0);
+  }
+};
+
+template<class Field>
+void testSchurFromHess(Arnoldi<Field>& Arn, Field& src, int Nlarge, int Nm, int Nk) {
+
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout << GridLogMessage << "Testing Schur reordering, Nm = " << Nm << ", Nk = " << Nk << std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+
+  std::cout << GridLogMessage << "Running Arnoldi for 1 iteration to get a Hessenberg." << std::endl;
+  Arn(src, 1, Nlarge, Nm, Nlarge);
+  Eigen::MatrixXcd Hess = Arn.getHessenbergMat();
+  std::cout << GridLogMessage << "Hessenberg for use: " << std::endl << Hess << std::endl;
+
+  ComplexSchurDecomposition schur (Hess, true);
+  bool isDecomposed = schur.checkDecomposition();
+  std::cout << "Schur decomp holds? " << isDecomposed << std::endl;
+
+  std::cout << GridLogMessage << "S = " << std::endl << schur.getMatrixS() << std::endl;
+  std::cout << GridLogMessage << "Swapping S(3, 3) with S(4, 4)" << std::endl;
+  schur.swapEvals(3);
+  std::cout << GridLogMessage << "S after swap = " << std::endl << schur.getMatrixS() << std::endl;
+  std::cout << "Schur decomp still holds? " << schur.checkDecomposition() << std::endl;
+
+  // Now move last diagonal element all the way to the front.
+  std::cout << GridLogMessage << "Moving last eval to front. S at start = " << std::endl << schur.getMatrixS() << std::endl;
+  for (int i = 0; i < Nk - 1; i++) {
+    int swapIdx = Nk - 2 - i;
+    schur.swapEvals(swapIdx);
+    std::cout << GridLogMessage << "S after swap of index " << swapIdx << " = " << std::endl << schur.getMatrixS() << std::endl;
+    std::cout << "Schur decomp still holds? " << schur.checkDecomposition() << std::endl;
+  }
+
+  std::cout << GridLogMessage << "Testing Schur reorder" << std::endl;
+  schur.schurReorder(Nk);
+  std::cout << GridLogMessage << "S after reorder = " << std::endl << schur.getMatrixS() << std::endl;
+  std::cout << "Schur decomp still holds? " << schur.checkDecomposition() << std::endl;
+
+}
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=16;
+
+//   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+//  std::vector<int> lat_size {32, 32, 32, 32};
+//  std::cout << "Lattice size: " << lat_size << std::endl;
+  GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), 
+								          GridDefaultSimd(Nd,vComplex::Nsimd()),
+								          GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+//  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+//  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+  GridCartesian         * FGrid   = UGrid;
+  GridRedBlackCartesian * FrbGrid = UrbGrid;
+
+  // Construct a coarsened grid
+  // poare TODO: replace this with the following line?
+  Coordinate clatt = GridDefaultLatt();
+//   Coordinate clatt = GridDefaultLatt();              // [PO] initial line before I edited it
+  for(int d=0;d<clatt.size();d++){
+  std::cout << GridLogMessage<< clatt[d] <<std::endl;
+    clatt[d] = clatt[d]/2;
+    //    clatt[d] = clatt[d]/4;
+  }
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  LatticeFermion result(FGrid); result=Zero();
+  LatticeFermion    ref(FGrid); ref=Zero();
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("config");
+//  std::string file("Users/patrickoare/libraries/PETSc-Grid/ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  LanczosParameters LanParams;
+  {
+    XmlReader  HMCrd("LanParams.xml");
+    read(HMCrd,"LanczosParameters",LanParams);
+  }
+
+  std::cout << GridLogMessage<< LanParams <<std::endl;
+  {
+    XmlWriter HMCwr("LanParams.xml.out");
+    write(HMCwr,"LanczosParameters",LanParams);
+  }
+
+
+  RealD mass=0.01;
+  RealD M5=1.8;
+
+  // PowerMethod<LatticeFermion> PM; PM(PVdagM, src);
+  int Nm = 50;
+  int Nk = 12; 
+  int Np = 38; 
+  // int Nk = Nm+1;     // if just running once
+  int maxIter = 10000;
+  int Nstop = 10;
+  RealD resid = 1.0e-5;
+
+  std::vector<Complex> boundary = {1,1,1,-1};
+  WilsonOp::ImplParams Params(boundary);
+
+//  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+//  DomainWallFermionD Dpv(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
+
+  mass=LanParams.mass;
+  std::cout << GridLogIRL<< "mass "<<mass<<std::endl;
+  WilsonOp WilsonOperator(Umu,*UGrid,*UrbGrid,mass,Params);
+
+  // const int nbasis = 20;            // size of approximate basis for low-mode space
+  const int nbasis = 3;            // size of approximate basis for low-mode space
+  const int cb = 0 ;
+  LatticeFermion prom(FGrid);
+
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNearestStencilGeometry5D geom(Coarse5d);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+
+//  typedef PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM_t;
+//  typedef ShiftedPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedPVdagM_t;
+//  typedef ShiftedComplexPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedComplexPVdagM_t;
+//  PVdagM_t PVdagM(Ddwf, Dpv);
+//  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv);
+//  SquaredLinearOperator<DomainWallFermionD, LatticeFermionD> Dsq (Ddwf);
+//  NonHermitianLinearOperator<DomainWallFermionD, LatticeFermionD> DLinOp (Ddwf);
+
+
+  NonHermitianLinearOperator<WilsonOp,FermionField> Dwilson(WilsonOperator); /// <-----
+  InvertNonHermitianLinearOperator<WilsonOp,FermionField> Iwilson(WilsonOperator); /// <-----
+  MdagMLinearOperator<WilsonOp,FermionField> HermOp(WilsonOperator); /// <-----
+  Gamma5HermitianLinearOperator <WilsonOp,LatticeFermion> HermOp2(WilsonOperator); /// <----
+
+  // PowerMethod<LatticeFermion> PM; PM(PVdagM, src);
+  resid=LanParams.resid;
+  Nstop=LanParams.Nstop;
+  Nk=LanParams.Nk;
+  Np=LanParams.Np;
+  maxIter=LanParams.maxIter;
+  Nm = Nk + Np;
+  int Nu=16;
+  std::vector<LatticeFermion> src(Nu,FGrid); 
+  for(int i=0;i<Nu;i++){
+     random(RNG5,src[i]);
+#if 0
+     LatticeFermion src_e(FrbGrid);
+     pickCheckerboard(Even, src_e, src[i]);
+     src_e=Zero();	
+     setCheckerboard(src[i],src_e);
+     std::cout << GridLogMessage << "src["<<i<<"] even sites zeroed " << std::endl;
+#endif
+  }
+
+  if(LanParams.ReadEvec) {
+    std::string evecs_file="evec_in";
+    std::cout << GridLogIRL<< "Reading evecs from "<<evecs_file<<std::endl;
+    Grid::emptyUserRecord record;
+    Grid::ScidacReader RD;
+    RD.open(evecs_file);
+    RD.readScidacFieldRecord(src[0],record);
+    RD.close();
+  }
+
+  Coordinate origin ({0,0,0,0});
+  auto tmpSrc = peekSite(src[0], origin);
+  std::cout << "[DEBUG] Source at origin = " <<  tmpSrc << std::endl;
+  LatticeFermion src2 = src[0];
+
+  // Run KrylovSchur / BlockKrylovSchur / TrueHarmonicBlockKrylovSchur on a Hermitian matrix
+    RealD shift=LanParams.shift;
+
+    // --use-blockKS : off -> plain (non-block) KrylovSchur
+    //                 on  -> block family (BlockKrylovSchur / TrueHarmonicBlockKrylovSchur)
+    // --use-harmonic: off -> no shift/harmonic Ritz extraction
+    //                 on, without --use-blockKS -> plain KrylovSchur with shifted (harmonic) Schur extraction
+    //                 on, with --use-blockKS    -> TrueHarmonicBlockKrylovSchur (true harmonic Ritz)
+    bool useBlockKS  = GridCmdOptionExists(argv, argv+argc, std::string("--use-blockKS"));
+    bool useHarmonic = GridCmdOptionExists(argv, argv+argc, std::string("--use-harmonic"));
+
+    std::vector<LatticeFermion> finalEvecs;
+
+    if (!useBlockKS) {
+      KrylovSchur KrySchur (Dwilson, UGrid, resid,EvalImNormSmall);
+      KrySchur.doEvalCheck=true;
+      if (useHarmonic) {
+        std::cout << GridLogMessage << "Running KrylovSchur (shifted/harmonic Ritz)" << std::endl;
+        KrySchur(src[0], maxIter, Nm, Nk, Nstop,&shift);
+      } else {
+        std::cout << GridLogMessage << "Running KrylovSchur" << std::endl;
+        KrySchur(src[0], maxIter, Nm, Nk, Nstop);
+      }
+      std::cout << GridLogMessage << "KrylovSchur evec.size= " << KrySchur.evecs.size()<< std::endl;
+      finalEvecs = KrySchur.evecs;
+    } else {
+      int Nblock=4;
+      Nblock=LanParams.Nblock;
+      bool if_verify=false;
+      if(LanParams.verify) if_verify=true;
+
+      auto gamma5Lambda = [](const FermionField& v, FermionField& out) {
+        Gamma g5(Gamma::Algebra::Gamma5);
+        out = g5 * v;
+      };
+
+      if (useHarmonic) {
+        std::cout << GridLogMessage << "Running TrueHarmonicBlockKrylovSchur" << std::endl;
+        TrueHarmonicBlockKrylovSchur<FermionField> KrySchur (Dwilson, UGrid, resid,shift,EvalImNormSmall);
+        KrySchur.doEvalCheck=true;
+        KrySchur.gamma5Func = gamma5Lambda;
+        KrySchur(src, maxIter, Nm, Nk, Nstop,Nblock,true,if_verify);
+        std::cout << GridLogMessage << "TrueHarmonicBlockKrylovSchur evec.size= " << KrySchur.evecs.size()<< std::endl;
+        finalEvecs = KrySchur.evecs;
+      } else {
+        std::cout << GridLogMessage << "Running BlockKrylovSchur" << std::endl;
+        BlockKrylovSchur<FermionField> KrySchur (Dwilson, UGrid, resid,EvalImNormSmall);
+        KrySchur.doEvalCheck=true;
+        KrySchur.gamma5Func = gamma5Lambda;
+        KrySchur(src, maxIter, Nm, Nk, Nstop,Nblock,true,if_verify);
+        std::cout << GridLogMessage << "BlockKrylovSchur evec.size= " << KrySchur.evecs.size()<< std::endl;
+        finalEvecs = KrySchur.evecs;
+      }
+    }
+
+  src[0]=finalEvecs[0];
+  std::cout << GridLogMessage << "finalEvecs= "<< finalEvecs.size() <<std::endl;
+  for (int i=1;i<Nk;i++) src[0]+=finalEvecs[i];
+  for (int i=0;i<Nstop;i++)
+  {
+	std::string evfile ("./evec_"+std::to_string(mass)+"_"+std::to_string(i));
+        auto evdensity = localInnerProduct(finalEvecs[i],finalEvecs[i] );
+        writeFile(evdensity,evfile);
+
+  }
+
+  {
+        std::string evfile ("./evec_"+std::to_string(mass)+"_sum");
+//        auto evdensity = localInnerProduct(evec[i],evec[i] );
+        writeFile(src[0],evfile);
+  }
+
+
+  /*
+  std::cout << GridLogMessage << "Running Arnoldi" << std::endl;
+  // Arnoldi Arn (Dsq, FGrid, 1e-8);
+  Arnoldi Arn (DLinOp, FGrid, 1e-8);
+  testSchurFromHess<LatticeFermion>(Arn, src, 10, 6, 4);
+
+  Arnoldi Arn2 (DLinOp, FGrid, 1e-8);
+  testSchurFromHess<LatticeFermion>(Arn2, src, 16, 12, 8);
+  */
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/Example_spec_kryschur.cc

@@ -0,0 +1,498 @@
+/*************************************************************************************
+
+    Runs the Krylov-Schur algorithm on a (pre-conditioned) domain-wall fermion operator 
+    to determine part of its spectrum. 
+
+    Usage : 
+      $ ./Example_spec_kryschur <Nm> <Nk> <maxiter> <Nstop> <inFile> <outDir> <?rf>
+
+      Nm = Maximum size of approximation subspace.
+      Nk = Size of truncation subspace
+      maxiter = Maximum number of iterations.
+      Nstop   = Stop when Nstop eigenvalues have converged. 
+      inFile  = Gauge configuration to read in.
+      outDir  = Directory to write output to.
+      rf      = (Optional) RitzFilter to sort with. Takes in any string in 
+                  {EvalNormSmall, EvalNormLarge, EvalReSmall, EvalReLarge, EvalImSmall, EvalImLarge}
+    
+    Output:
+      ${outDir}/evals.txt  = Contains all eigenvalues. Each line is formatted as `$idx $eval $ritz`, where:
+                              - $idx is the index of the eigenvalue.
+                              - $eval is the eigenvalue, formated as "(re,im)".
+                              - $ritz is the Ritz estimate of the eigenvalue (deviation from being a true eigenvalue)
+      ${outDir}/evec${idx} = Eigenvector $idx written out in SCIDAC format (if LIME is enabled).
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Patrick Oare <poare@bnl.edu>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+#include <cstdlib>
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+#include <Grid/parallelIO/IldgIOtypes.h>
+#include <Grid/parallelIO/IldgIO.h>
+
+using namespace std;
+using namespace Grid;
+
+namespace Grid {
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass ,
+		  		RealD, mstep ,
+				Integer, Nstop,
+                                Integer, Nk,
+                                Integer, Np,
+                                Integer, ReadEvec,
+	  			RealD, resid,
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+
+  LanczosParameters() {
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass >
+  void initialize(Reader<ReaderClass> &TheReader){
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+
+};
+
+}
+
+template <class T> void writeFile(T& in, std::string const fname){
+#ifdef HAVE_LIME
+  // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+  std::cout << Grid::GridLogMessage << "Writes to: " << fname << std::endl;
+  Grid::emptyUserRecord record;
+  Grid::ScidacWriter WR(in.Grid()->IsBoss());
+  WR.open(fname);
+  WR.writeScidacFieldRecord(in,record,0); // Lexico
+  WR.close();
+#endif
+}
+
+/**
+ * Writes the eigensystem of a Krylov Schur object to a directory.
+ *
+ * Parameters
+ * ----------
+ * std::string path
+ *    Directory to write to.
+ */
+template <class Field>
+void writeEigensystem(KrylovSchur<Field> KS, std::string outDir) {
+  int Nk = KS.getNk();
+  std::cout << GridLogMessage << "Writing output to directory: " << outDir << std::endl;
+
+  // Write evals
+  std::string evalPath = outDir + "/evals.txt";
+  std::ofstream fEval;
+  fEval.open(evalPath);
+  Eigen::VectorXcd evals = KS.getEvals();
+  std::vector<RealD> ritz  = KS.getRitzEstimates();
+  for (int i = 0; i < Nk; i++) {
+    // write eigenvalues and Ritz estimates
+    fEval << i << " " << evals(i) << " " << ritz[i];
+    if (i < Nk - 1) { fEval << "\n"; }
+  }
+  fEval.close();
+
+  // Write evecs (TODO: very heavy on storage costs! Don't write them all out)
+  // std::vector<Field> evecs = KS.getEvecs();
+  // for (int i = 0; i < Nk; i++) {
+  //   std::string fName = outDir + "/evec" + std::to_string(i);
+  //   writeFile(evecs[i], fName);     // using method from Grid/HMC/ComputeWilsonFlow.cc
+  // }
+}
+// Hermitize a DWF operator by squaring it
+template<class Matrix,class Field>
+class SquaredLinearOperator : public LinearOperatorBase<Field> {
+
+  public:
+  Matrix &_Mat;
+
+  public:
+    SquaredLinearOperator(Matrix &Mat): _Mat(Mat) {};
+
+    void OpDiag (const Field &in, Field &out) {    assert(0);  }
+    void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+    void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+    void Op     (const Field &in, Field &out){
+      // std::cout << "Op is overloaded as HermOp" << std::endl;
+      HermOp(in, out);
+    }
+    void AdjOp     (const Field &in, Field &out){
+      HermOp(in, out);
+    }
+    void _Op     (const Field &in, Field &out){
+      // std::cout << "Op: M "<<std::endl;
+      _Mat.M(in, out);
+    }
+    void _AdjOp     (const Field &in, Field &out){
+      // std::cout << "AdjOp: Mdag "<<std::endl;
+      _Mat.Mdag(in, out);
+    }
+    void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+    void HermOp(const Field &in, Field &out){
+      // std::cout << "HermOp: Mdag M Mdag M"<<std::endl;
+      Field tmp(in.Grid());
+      _Op(in,tmp);
+      _AdjOp(tmp,out);
+    }
+};
+
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    //    _Mat.M(in,tmp);
+    //    _PV.Mdag(tmp,out);
+    //    _PV.M(out,tmp);
+    //    _Mat.Mdag(tmp,out);
+    Op(in,tmp);
+    AdjOp(tmp,out);
+    //    std::cout << "HermOp done "<<norm2(out)<<std::endl;
+  }
+};
+
+template<class Matrix,class Field>
+class ShiftedPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  RealD shift;
+public:
+  ShiftedPVdagMLinearOperator(RealD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+};
+
+template<class Matrix, class Field>
+class ShiftedComplexPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  ComplexD shift;
+public:
+ShiftedComplexPVdagMLinearOperator(ComplexD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+  
+  void resetShift(ComplexD newShift) {
+    shift = newShift;
+  }
+};
+
+#if 0
+template<class Fobj,class CComplex,int nbasis>
+class MGPreconditioner : public LinearFunction< Lattice<Fobj> > {
+public:
+  using LinearFunction<Lattice<Fobj> >::operator();
+
+  typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::FineField    FineField;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
+  typedef LinearOperatorBase<FineField>                            FineOperator;
+  typedef LinearFunction    <FineField>                            FineSmoother;
+  typedef LinearOperatorBase<CoarseVector>                         CoarseOperator;
+  typedef LinearFunction    <CoarseVector>                         CoarseSolver;
+  Aggregates     & _Aggregates;
+  FineOperator   & _FineOperator;
+  FineSmoother   & _PreSmoother;
+  FineSmoother   & _PostSmoother;
+  CoarseOperator & _CoarseOperator;
+  CoarseSolver   & _CoarseSolve;
+
+  int    level;  void Level(int lv) {level = lv; };
+
+  MGPreconditioner(Aggregates &Agg,
+		   FineOperator &Fine,
+		   FineSmoother &PreSmoother,
+		   FineSmoother &PostSmoother,
+		   CoarseOperator &CoarseOperator_,
+		   CoarseSolver &CoarseSolve_)
+    : _Aggregates(Agg),
+      _FineOperator(Fine),
+      _PreSmoother(PreSmoother),
+      _PostSmoother(PostSmoother),
+      _CoarseOperator(CoarseOperator_),
+      _CoarseSolve(CoarseSolve_),
+      level(1)  {  }
+
+  virtual void operator()(const FineField &in, FineField & out) 
+  {
+    GridBase *CoarseGrid = _Aggregates.CoarseGrid;
+    //    auto CoarseGrid = _CoarseOperator.Grid();
+    CoarseVector Csrc(CoarseGrid);
+    CoarseVector Csol(CoarseGrid);
+    FineField vec1(in.Grid());
+    FineField vec2(in.Grid());
+
+    std::cout<<GridLogMessage << "Calling PreSmoother " <<std::endl;
+
+    //    std::cout<<GridLogMessage << "Calling PreSmoother input residual "<<norm2(in) <<std::endl;
+    double t;
+    // Fine Smoother
+    //    out = in;
+    out = Zero();
+    t=-usecond();
+    _PreSmoother(in,out);
+    t+=usecond();
+
+    std::cout<<GridLogMessage << "PreSmoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Update the residual
+    _FineOperator.Op(out,vec1);  sub(vec1, in ,vec1);   
+    //    std::cout<<GridLogMessage <<"Residual-1 now " <<norm2(vec1)<<std::endl;
+
+    // Fine to Coarse 
+    t=-usecond();
+    _Aggregates.ProjectToSubspace  (Csrc,vec1);
+    t+=usecond();
+    std::cout<<GridLogMessage << "Project to coarse took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse correction
+    t=-usecond();
+    Csol = Zero();
+    _CoarseSolve(Csrc,Csol);
+    //Csol=Zero();
+    t+=usecond();
+    std::cout<<GridLogMessage << "Coarse solve took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse to Fine
+    t=-usecond();  
+    //    _CoarseOperator.PromoteFromSubspace(_Aggregates,Csol,vec1);
+    _Aggregates.PromoteFromSubspace(Csol,vec1); 
+    add(out,out,vec1);
+    t+=usecond();
+    std::cout<<GridLogMessage << "Promote to this level took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Residual
+    _FineOperator.Op(out,vec1);  sub(vec1 ,in , vec1);  
+    //    std::cout<<GridLogMessage <<"Residual-2 now " <<norm2(vec1)<<std::endl;
+
+    // Fine Smoother
+    t=-usecond();
+    //    vec2=vec1;
+    vec2=Zero();
+    _PostSmoother(vec1,vec2);
+    t+=usecond();
+    std::cout<<GridLogMessage << "PostSmoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    add( out,out,vec2);
+    std::cout<<GridLogMessage << "Done " <<std::endl;
+  }
+};
+#endif
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  // Usage : $ ./Example_spec_kryschur <Nm> <Nk> <maaxiter> <Nstop> <inFile> <outDir>
+  std::string NmStr      = argv[1];
+  std::string NkStr      = argv[2];
+  std::string maxIterStr = argv[3];
+  std::string NstopStr   = argv[4];
+  std::string file       = argv[5];
+  std::string outDir     = argv[6];
+
+  RitzFilter RF;
+  if (argc == 8) {
+    std::string rf       = argv[7];
+    RF = selectRitzFilter(rf);
+  } else {
+    RF = EvalReSmall;
+  }
+  std::cout << "Sorting eigenvalues using " << rfToString(RF) << std::endl;
+
+  //const int Ls=16;
+  const int Ls = 8;
+
+//   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  //std::vector<int> lat_size {16, 16, 16, 32};
+  std::vector<int> lat_size {8, 8, 8, 8};
+  std::cout << "Lattice size: " << lat_size << std::endl;
+  GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(lat_size, 
+								          GridDefaultSimd(Nd,vComplex::Nsimd()),
+								          GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermion    src(FGrid); random(RNG5,src);
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  NerscIO::readConfiguration(Umu,header,file);
+
+  // RealD mass=0.01;
+  RealD mass=0.001;
+  RealD M5=1.8;
+
+  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionD Dpv(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+
+  typedef PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM_t;
+  typedef ShiftedPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedPVdagM_t;
+  typedef ShiftedComplexPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedComplexPVdagM_t;
+  PVdagM_t PVdagM(Ddwf, Dpv);
+  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv);
+  SquaredLinearOperator<DomainWallFermionD, LatticeFermionD> Dsq (Ddwf);
+  NonHermitianLinearOperator<DomainWallFermionD, LatticeFermionD> DLinOp (Ddwf);
+
+  int Nm = std::stoi(NmStr);
+  int Nk = std::stoi(NkStr);
+  int maxIter = std::stoi(maxIterStr);
+  int Nstop = std::stoi(NstopStr);
+
+  std::cout << GridLogMessage << "Runnning Krylov Schur. Nm = " << Nm << ", Nk = " << Nk << ", maxIter = " << maxIter 
+                  << ", Nstop = " << Nstop << std::endl;
+  
+  KrylovSchur KrySchur (PVdagM, FGrid, 1e-8, RF);      // use preconditioned PV^\dag D_{dwf}
+  // KrylovSchur KrySchur (DLinOp, FGrid, 1e-8, RF);         // use D_{dwf}
+  KrySchur(src, maxIter, Nm, Nk, Nstop);
+
+  std::cout<<GridLogMessage << "*******************************************" << std::endl;
+  std::cout<<GridLogMessage << "***************** RESULTS *****************" << std::endl;
+  std::cout<<GridLogMessage << "*******************************************" << std::endl;
+
+  std::cout << GridLogMessage << "Krylov Schur eigenvalues: " << std::endl << KrySchur.getEvals() << std::endl;
+
+  writeEigensystem(KrySchur, outDir);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/Example_wall_wall_3pt.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -535,5 +538,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_wall_wall_spectrum.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -429,5 +432,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_wilson_evecs.cc

@@ -0,0 +1,383 @@
+/*************************************************************************************
+
+    Script for studying the Wilson eigenvectors resulting from the Krylov-Schur process. 
+
+    Usage : 
+      $ ./Example_spec_kryschur <Nm> <Nk> <maxiter> <Nstop> <inFile> <outDir> <?rf>
+
+      Nm = Maximum size of approximation subspace.
+      Nk = Size of truncation subspace
+      maxiter = Maximum number of iterations.
+      Nstop   = Stop when Nstop eigenvalues have converged. 
+      inFile  = Gauge configuration to read in.
+      outDir  = Directory to write output to.
+      rf      = (Optional) RitzFilter to sort with. Takes in any string in 
+                  {EvalNormSmall, EvalNormLarge, EvalReSmall, EvalReLarge, EvalImSmall, EvalImLarge}
+    
+    Output:
+      ${outDir}/evals.txt  = Contains all eigenvalues. Each line is formatted as `$idx $eval $ritz`, where:
+                              - $idx is the index of the eigenvalue.
+                              - $eval is the eigenvalue, formated as "(re,im)".
+                              - $ritz is the Ritz estimate of the eigenvalue (deviation from being a true eigenvalue)
+      ${outDir}/evec${idx} = Eigenvector $idx written out in SCIDAC format (if LIME is enabled).
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Patrick Oare <poare@bnl.edu>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+#include <cstdlib>
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+#include <Grid/parallelIO/IldgIOtypes.h>
+#include <Grid/parallelIO/IldgIO.h>
+
+using namespace std;
+using namespace Grid;
+
+template <class T> void writeFile(T& in, std::string const fname){  
+  #ifdef HAVE_LIME
+    // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+    std::cout << Grid::GridLogMessage << "Writes to: " << fname << std::endl;
+    Grid::emptyUserRecord record;
+    Grid::ScidacWriter WR(in.Grid()->IsBoss());
+    WR.open(fname);
+    WR.writeScidacFieldRecord(in,record,0); // Lexico
+    WR.close();
+  #endif
+}
+
+template <class T> void readFile(T& out, std::string const fname){  
+  #ifdef HAVE_LIME
+    // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+    std::cout << Grid::GridLogMessage << "Reads at: " << fname << std::endl;
+    Grid::emptyUserRecord record;
+    // Grid::ScidacReader SR(out.Grid()->IsBoss());
+    Grid::ScidacReader SR;
+    SR.open(fname);
+    SR.readScidacFieldRecord(out, record);
+    SR.close();
+  #endif
+}
+
+/**
+ * Writes the eigensystem of a Krylov Schur object to a directory. 
+ * 
+ * Parameters
+ * ----------
+ * std::string path
+ *    Directory to write to. 
+ */
+template <class Field>
+void writeEigensystem(KrylovSchur<Field> KS, std::string outDir) {
+  int Nk = KS.getNk();
+  std::cout << GridLogMessage << "Writing output to directory: " << outDir << std::endl;
+  
+  // Write evals
+  std::string evalPath = outDir + "/evals.txt";
+  std::ofstream fEval;
+  fEval.open(evalPath);
+  Eigen::VectorXcd evals = KS.getEvals();
+  std::vector<RealD> ritz  = KS.getRitzEstimates();
+  for (int i = 0; i < Nk; i++) {
+    // write eigenvalues and Ritz estimates
+    fEval << i << " " << evals(i) << " " << ritz[i];
+    if (i < Nk - 1) { fEval << "\n"; }
+  }
+  fEval.close();
+  
+  // Write evecs
+  int Nevecs = Nk;          // don't write all of them
+  std::vector<Field> evecs = KS.getEvecs();
+  for (int i = 0; i < Nevecs; i++) {
+    std::string fName = outDir + "/evec" + std::to_string(i);
+    writeFile(evecs[i], fName);     // using method from Grid/HMC/ComputeWilsonFlow.cc
+  }
+}
+
+// Hermitize a DWF operator by squaring it
+template<class Matrix,class Field>
+class SquaredLinearOperator : public LinearOperatorBase<Field> {
+
+  public:
+  Matrix &_Mat;
+
+  public:
+    SquaredLinearOperator(Matrix &Mat): _Mat(Mat) {};
+
+    void OpDiag (const Field &in, Field &out) {    assert(0);  }
+    void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+    void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+    void Op     (const Field &in, Field &out){
+      // std::cout << "Op is overloaded as HermOp" << std::endl;
+      HermOp(in, out);
+    }
+    void AdjOp     (const Field &in, Field &out){
+      HermOp(in, out);
+    }
+    void _Op     (const Field &in, Field &out){
+      // std::cout << "Op: M "<<std::endl;
+      _Mat.M(in, out);
+    }
+    void _AdjOp     (const Field &in, Field &out){
+      // std::cout << "AdjOp: Mdag "<<std::endl;
+      _Mat.Mdag(in, out);
+    }
+    void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+    void HermOp(const Field &in, Field &out){
+      // std::cout << "HermOp: Mdag M Mdag M"<<std::endl;
+      Field tmp(in.Grid());
+      _Op(in,tmp);
+      _AdjOp(tmp,out);
+    }
+};
+
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    //    _Mat.M(in,tmp);
+    //    _PV.Mdag(tmp,out);
+    //    _PV.M(out,tmp);
+    //    _Mat.Mdag(tmp,out);
+    Op(in,tmp);
+    AdjOp(tmp,out);
+    //    std::cout << "HermOp done "<<norm2(out)<<std::endl;
+  }
+};
+
+template<class Matrix,class Field>
+class ShiftedPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  RealD shift;
+public:
+  ShiftedPVdagMLinearOperator(RealD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+};
+
+template<class Matrix, class Field>
+class ShiftedComplexPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  ComplexD shift;
+public:
+ShiftedComplexPVdagMLinearOperator(ComplexD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+  
+  void resetShift(ComplexD newShift) {
+    shift = newShift;
+  }
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  // Usage : $ ./Example_wilson_evecs ${inFile}
+  std::string file       = argv[1];
+
+  const int Ls=16;
+
+//   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  //std::vector<int> lat_size {16, 16, 16, 32};
+  std::vector<int> lat_size {32, 32, 32, 32};
+  std::cout << "Lattice size: " << lat_size << std::endl;
+  GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(lat_size, 
+								          GridDefaultSimd(Nd,vComplex::Nsimd()),
+								          GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  // GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  // GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+  GridCartesian * FGrid = UGrid;
+  GridRedBlackCartesian * FrbGrid = UrbGrid;
+
+  std::vector<int> seeds4({1,2,3,4});
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermion    src(FGrid); random(RNG4, src);
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  NerscIO::readConfiguration(Umu, header, file);
+
+  std::cout << GridLogMessage << "Loaded configuration" << std::endl;
+
+  // RealD mass = 0.01;
+  RealD M5 = 1.8;
+
+  // Wilson mass
+  RealD mass = -1.6;
+
+  std::cout << GridLogMessage << "masses specified" << std::endl;
+
+  std::vector<Complex> boundary = {1,1,1,-1};
+  WilsonFermionD::ImplParams Params(boundary);
+
+  // DomainWallFermionD Ddwf(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
+  // NonHermitianLinearOperator<DomainWallFermionD, LatticeFermionD> DLinOp (Ddwf);
+
+  // WilsonFermionD Dwilson(Umu, *FGrid, *FrbGrid, mass);
+  WilsonFermionD Dwilson(Umu, *UGrid, *UrbGrid, mass, Params);
+  NonHermitianLinearOperator<WilsonFermionD, LatticeFermionD> DLinOp (Dwilson);
+
+  std::cout << GridLogMessage << "Dirac operator defined" << std::endl;
+
+  std::string eigenPath = "/home/poare/lqcd/multigrid/spectra/32cube-rho0.124-tau4/U_smr_3.000000/Nm72_Nk24_8111835.aurora-pbs-0001.hostmgmt.cm.aurora.alcf.anl.gov/";
+
+  std::cout << GridLogMessage << "Loading eigenvalues" << std::endl;
+  std::ifstream evalFile(eigenPath + "evals.txt");
+  std::string str;
+  std::vector<ComplexD> evals;
+  while (std::getline(evalFile, str)) {
+      std::cout << GridLogMessage << "Reading line: " << str << std::endl;
+      int i1 = str.find("(") + 1;
+      int i2 = str.find(",") + 1;
+      int i3 = str.find(")");
+      std::cout << "i1,i2,i3 = " << i1 << "," << i2 << "," << i3 << std::endl;
+      std::string reStr = str.substr(i1, i2 - i1);
+      std::string imStr = str.substr(i2, i3 - i2);
+      std::cout << GridLogMessage << "Parsed re = " << reStr << " and im = " << imStr << std::endl;
+      // ComplexD z (std::stof(reStr), std::stof(imStr));
+      ComplexD z (std::stod(reStr), std::stod(imStr));
+      evals.push_back(z);
+  }
+  std::cout << GridLogMessage << "Eigenvalues: " << evals << std::endl;
+
+  int Nevecs = 24;
+  std::vector<LatticeFermion> evecs;
+  LatticeFermion evec (FGrid);
+  for (int i = 0; i < Nevecs; i++) {
+    std::string evecPath = eigenPath + "evec" + std::to_string(i);
+    readFile(evec, evecPath);
+    evecs.push_back(evec);
+  }
+  std::cout << GridLogMessage << "Evecs loaded" << std::endl;
+
+  // Compute < evec | D - \lambda | evec >
+  std::cout << GridLogMessage << "Testing eigenvectors" << std::endl;
+  LatticeFermion Devec (FGrid);
+  ComplexD ritz;
+  for (int i = 0; i < Nevecs; i++) {
+    Devec = Zero();
+    DLinOp.Op(evecs[i], Devec);
+    ritz = std::sqrt(norm2(Devec - evals[i] * evecs[i]));
+    std::cout << GridLogMessage << "i = " << i << ", || (D - lambda) |vi> || = " << ritz << std::endl;
+  }
+  // Eigen::MatrixXcd Dw_evecs;
+  // Dw_evecs = Eigen::MatrixXcd::Zero(Nevecs, Nevecs);
+  // for (int i = 0; i < Nevecs; i++) {
+  //   Linop.Op(evecs[i], Devec);
+  //   for (int j = 0; j < Nevecs; j++) {
+
+  //   }
+  // }
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/Example_wilson_spectrum.cc

@@ -0,0 +1,374 @@
+/*************************************************************************************
+
+    Runs the Krylov-Schur algorithm on a Wilson fermion operator to determine part of its spectrum. 
+
+    TODO rename this file: really is running the topology change jobs on Aurora. 
+
+    Usage : 
+      $ ./Example_spec_kryschur <Nm> <Nk> <maxiter> <Nstop> <inFile> <outDir> <?rf>
+
+      Nm = Maximum size of approximation subspace.
+      Nk = Size of truncation subspace
+      maxiter = Maximum number of iterations.
+      Nstop   = Stop when Nstop eigenvalues have converged. 
+      inFile  = Gauge configuration to read in.
+      outDir  = Directory to write output to.
+      rf      = (Optional) RitzFilter to sort with. Takes in any string in 
+                  {EvalNormSmall, EvalNormLarge, EvalReSmall, EvalReLarge, EvalImSmall, EvalImLarge}
+    
+    Output:
+      ${outDir}/evals.txt  = Contains all eigenvalues. Each line is formatted as `$idx $eval $ritz`, where:
+                              - $idx is the index of the eigenvalue.
+                              - $eval is the eigenvalue, formated as "(re,im)".
+                              - $ritz is the Ritz estimate of the eigenvalue (deviation from being a true eigenvalue)
+      ${outDir}/evec${idx} = Eigenvector $idx written out in SCIDAC format (if LIME is enabled).
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Patrick Oare <poare@bnl.edu>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+
+#include <cstdlib>
+
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+#include <Grid/parallelIO/IldgIOtypes.h>
+#include <Grid/parallelIO/IldgIO.h>
+
+using namespace std;
+using namespace Grid;
+
+template <class T> void writeFile(T& in, std::string const fname){  
+  #ifdef HAVE_LIME
+    // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+    std::cout << Grid::GridLogMessage << "Writes to: " << fname << std::endl;
+    Grid::emptyUserRecord record;
+    Grid::ScidacWriter WR(in.Grid()->IsBoss());
+    WR.open(fname);
+    WR.writeScidacFieldRecord(in,record,0); // Lexico
+    WR.close();
+  #endif
+}
+
+/**
+ * Writes the eigensystem of a Krylov Schur object to a directory. 
+ * 
+ * Parameters
+ * ----------
+ * std::string path
+ *    Directory to write to. 
+ */
+template <class Field>
+void writeEigensystem(KrylovSchur<Field> KS, std::string outDir) {
+  int Nk = KS.getNk();
+  std::cout << GridLogMessage << "Writing output to directory: " << outDir << std::endl;
+  
+  // Write evals
+  std::string evalPath = outDir + "/evals.txt";
+  std::ofstream fEval;
+  fEval.open(evalPath);
+  Eigen::VectorXcd evals = KS.getEvals();
+  std::vector<RealD> ritz  = KS.getRitzEstimates();
+  for (int i = 0; i < Nk; i++) {
+    // write eigenvalues and Ritz estimates
+    fEval << i << " " << evals(i) << " " << ritz[i];
+    if (i < Nk - 1) { fEval << "\n"; }
+  }
+  fEval.close();
+  
+  // Write evecs
+  int Nevecs = Nk;          // don't write all of them
+  std::vector<Field> evecs = KS.getEvecs();
+  for (int i = 0; i < Nevecs; i++) {
+    std::string fName = outDir + "/evec" + std::to_string(i);
+    writeFile(evecs[i], fName);     // using method from Grid/HMC/ComputeWilsonFlow.cc
+  }
+}
+
+// Hermitize a DWF operator by squaring it
+template<class Matrix,class Field>
+class SquaredLinearOperator : public LinearOperatorBase<Field> {
+
+  public:
+  Matrix &_Mat;
+
+  public:
+    SquaredLinearOperator(Matrix &Mat): _Mat(Mat) {};
+
+    void OpDiag (const Field &in, Field &out) {    assert(0);  }
+    void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+    void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+    void Op     (const Field &in, Field &out){
+      // std::cout << "Op is overloaded as HermOp" << std::endl;
+      HermOp(in, out);
+    }
+    void AdjOp     (const Field &in, Field &out){
+      HermOp(in, out);
+    }
+    void _Op     (const Field &in, Field &out){
+      // std::cout << "Op: M "<<std::endl;
+      _Mat.M(in, out);
+    }
+    void _AdjOp     (const Field &in, Field &out){
+      // std::cout << "AdjOp: Mdag "<<std::endl;
+      _Mat.Mdag(in, out);
+    }
+    void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+    void HermOp(const Field &in, Field &out){
+      // std::cout << "HermOp: Mdag M Mdag M"<<std::endl;
+      Field tmp(in.Grid());
+      _Op(in,tmp);
+      _AdjOp(tmp,out);
+    }
+};
+
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    //    _Mat.M(in,tmp);
+    //    _PV.Mdag(tmp,out);
+    //    _PV.M(out,tmp);
+    //    _Mat.Mdag(tmp,out);
+    Op(in,tmp);
+    AdjOp(tmp,out);
+    //    std::cout << "HermOp done "<<norm2(out)<<std::endl;
+  }
+};
+
+template<class Matrix,class Field>
+class ShiftedPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  RealD shift;
+public:
+  ShiftedPVdagMLinearOperator(RealD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+};
+
+template<class Matrix, class Field>
+class ShiftedComplexPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  ComplexD shift;
+public:
+ShiftedComplexPVdagMLinearOperator(ComplexD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
+  
+  void resetShift(ComplexD newShift) {
+    shift = newShift;
+  }
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  // Usage : $ ./Example_spec_kryschur <Nm> <Nk> <maaxiter> <Nstop> <inFile> <outDir>
+  std::string NmStr      = argv[1];
+  std::string NkStr      = argv[2];
+  std::string maxIterStr = argv[3];
+  std::string NstopStr   = argv[4];
+  std::string file       = argv[5];
+  std::string outDir     = argv[6];
+
+  // RitzFilter RF;
+  // if (argc == 8) {
+  //   std::string rf       = argv[7];
+  //   RF = selectRitzFilter(rf);
+  // } else {
+  //   RF = EvalReSmall;
+  // }
+  // RitzFilter RF;
+  std::string rf       = argv[7];
+  RitzFilter RF        = selectRitzFilter(rf);
+  std::cout << "Sorting eigenvalues using " << rfToString(RF) << std::endl;
+
+  const int Ls=16;
+
+//   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  //std::vector<int> lat_size {16, 16, 16, 32};
+  std::vector<int> lat_size {32, 32, 32, 32};
+  std::cout << "Lattice size: " << lat_size << std::endl;
+  GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(lat_size, 
+								          GridDefaultSimd(Nd,vComplex::Nsimd()),
+								          GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  // GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  // GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+  GridCartesian * FGrid = UGrid;
+  GridRedBlackCartesian * FrbGrid = UrbGrid;
+
+  std::vector<int> seeds4({1,2,3,4});
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermion    src(FGrid); random(RNG4, src);
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  NerscIO::readConfiguration(Umu, header, file);
+
+  std::cout << GridLogMessage << "Loaded configuration" << std::endl;
+
+  // RealD mass = 0.01;
+  RealD M5 = 1.8;
+
+  // Wilson mass
+  RealD mass = -1.6;
+
+  std::cout << GridLogMessage << "masses specified" << std::endl;
+
+  std::vector<Complex> boundary = {1,1,1,-1};
+  WilsonFermionD::ImplParams Params(boundary);
+
+  // DomainWallFermionD Ddwf(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
+  // NonHermitianLinearOperator<DomainWallFermionD, LatticeFermionD> DLinOp (Ddwf);
+
+  // WilsonFermionD Dwilson(Umu, *FGrid, *FrbGrid, mass);
+  WilsonFermionD Dwilson(Umu, *UGrid, *UrbGrid, mass, Params);
+  NonHermitianLinearOperator<WilsonFermionD, LatticeFermionD> DLinOp (Dwilson);
+
+  std::cout << GridLogMessage << "Dirac operator defined" << std::endl;
+
+  // Define PV^dag D (if we want)
+  // DomainWallFermionD Dpv(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, 1.0, M5);
+  // typedef PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM_t;
+  // PVdagM_t PVdagM(Ddwf, Dpv);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+
+  // SquaredLinearOperator<WilsonFermionD, LatticeFermionD> Dsq (DWilson);
+  // NonHermitianLinearOperator<WilsonFermionD, LatticeFermionD> DLinOp (DWilson);
+
+  int Nm = std::stoi(NmStr);
+  int Nk = std::stoi(NkStr);
+  int maxIter = std::stoi(maxIterStr);
+  int Nstop = std::stoi(NstopStr);
+
+  std::cout << GridLogMessage << "Runnning Krylov Schur. Nm = " << Nm << ", Nk = " << Nk << ", maxIter = " << maxIter 
+                  << ", Nstop = " << Nstop << std::endl;
+  
+  // KrylovSchur KrySchur (PVdagM, FGrid, 1e-8, RF);         // use PV^\dag M
+  KrylovSchur KrySchur (DLinOp, FGrid, 1e-8, RF);         // use Ddwf
+  KrySchur(src, maxIter, Nm, Nk, Nstop);
+
+  std::cout << GridLogMessage << "Checking eigensystem." << std::endl;
+  KrySchur.checkRitzEstimate();
+
+  std::cout<<GridLogMessage << "*******************************************" << std::endl;
+  std::cout<<GridLogMessage << "***************** RESULTS *****************" << std::endl;
+  std::cout<<GridLogMessage << "*******************************************" << std::endl;
+
+  std::cout << GridLogMessage << "Krylov Schur eigenvalues: " << std::endl << KrySchur.getEvals() << std::endl;
+
+  writeEigensystem(KrySchur, outDir);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}

examples/LanParams.xml

@@ -0,0 +1,22 @@
+<?xml version="1.0"?>
+<grid>
+  <LanczosParameters>
+    <mass>0</mass>
+    <mstep>-0.025</mstep>
+    <M5>1.8</M5>
+    <Ls>48</Ls>
+    <Nstop>100</Nstop>
+    <Nk>100</Nk>
+    <Np>100</Np>
+    <ReadEvec>0</ReadEvec>
+    <maxIter>1000</maxIter>
+    <reorthog>1</reorthog>
+    <Nblock>4</Nblock>
+    <verify>0</verify>
+    <shift>1.5</shift>
+    <resid>1e-8</resid>
+    <ChebyLow>1</ChebyLow>
+    <ChebyHigh>100</ChebyHigh>
+    <ChebyOrder>51</ChebyOrder>
+  </LanczosParameters>
+</grid>

examples/disable_examples_without_instantiations.h

@@ -0,0 +1,15 @@
+#include <Grid/Grid.h>
+#pragma once
+
+#ifndef ENABLE_FERMION_INSTANTIATIONS
+#include <iostream>
+
+int main(void) {
+  std::cout << "This build of Grid was configured to exclude fermion instantiations, "
+	    << "which this example relies on. "
+	    << "Please reconfigure and rebuild Grid with --enable-fermion-instantiations"
+	    << "to run this example."
+	    << std::endl;
+  return 1;
+}
+#endif

examples/fft5d.cc

@@ -0,0 +1,469 @@
+/*************************************************************************************
+
+  fft5d.cc — Fourier analysis of a time series of 4-D lattice scalar fields.
+
+  Designed for force-norm files from FTHMC (one RealD per site, SCIDAC or binary).
+  Assembles a (4+1)-D structure [trajectory][t][z][y][x] using Grid lattice fields
+  and performs three FFT analyses:
+
+    --fft spatial
+        4-D spatial FFT using Grid's FFT class (MPI+GPU parallel via Cshift/cufft).
+        Outputs shell-averaged P(|k|^2) averaged over trajectories, and a
+        per-mode table P(kt,kz,ky,kx).
+
+    --fft traj
+        1-D trajectory-axis FFT at each lattice site using FFTW3 locally at each
+        MPI rank.  Outputs site-averaged P(f_traj) and autocorrelation C(lag).
+        Results are combined with GlobalSumVector.
+
+    --fft all
+        Both of the above, plus a 2-D cross spectrum P(f_traj, |k_spatial|^2):
+        spatial FFT per trajectory, then trajectory FFTW on the momentum-space series.
+
+  Normalisations:
+    Spatial FFT  (volume V): P_spatial(k) = |F(k)|^2 / V^2
+      => (1/V) * sum_k P(k) = site mean-square per trajectory
+    Trajectory FFT (Ntraj):  P_traj(f) = |F(f)|^2 / Ntraj^2
+      => sum_f P(f) = site mean-square over trajectories
+
+  Build: add to examples/Make.inc (see bottom of this file), then make.
+
+  Usage (follows Grid conventions):
+    fft5d --grid Lx.Ly.Lz.Lt [--mpi Px.Py.Pz.Pt] \
+          [--fft spatial|traj|all] [--format scidac|binary] \
+          [--output PREFIX] file1 file2 ...
+
+*************************************************************************************/
+
+#include <Grid/Grid.h>
+// Grid's FFT.h uses cufft on CUDA builds; for the trajectory axis we also need
+// CPU-side FFTW3 (already linked via -lfftw3 in GRID_LIBS).
+#include <fftw3.h>
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+using namespace Grid;
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Global coordinate of a local site (osite, lane) on this MPI rank
+// ─────────────────────────────────────────────────────────────────────────────
+static void globalCoor(int osite, int lane, GridCartesian* g, Coordinate& gc)
+{
+    Coordinate oc, ic;
+    Lexicographic::CoorFromIndex(oc, osite, g->_rdimensions);
+    Lexicographic::CoorFromIndex(ic, lane,  g->_simd_layout);
+    for (int d = 0; d < Nd; d++)
+        gc[d] = g->_processor_coor[d] * g->_ldimensions[d]
+              + oc[d] * g->_simd_layout[d] + ic[d];
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Shell-average a power LatticeRealD over |k|^2 bins (MPI-aware via GlobalSumVector)
+// ─────────────────────────────────────────────────────────────────────────────
+static void shellAverage(const LatticeRealD& power, GridCartesian* g,
+                         double norm, std::ostream& ofs)
+{
+    const int Nsimd = vRealD::Nsimd();
+    Coordinate fdims = g->_fdimensions;
+    int maxk2 = 0;
+    for (int d = 0; d < Nd; d++) { int h = fdims[d]/2; maxk2 += h*h; }
+
+    std::vector<RealD> psum(maxk2+1, 0.0), cnt(maxk2+1, 0.0);
+    {
+        auto pv = power.View(CpuRead);
+        Coordinate gc(Nd);
+        for (int os = 0; os < (int)g->oSites(); os++) {
+            for (int lane = 0; lane < Nsimd; lane++) {
+                globalCoor(os, lane, g, gc);
+                int k2 = 0;
+                for (int d = 0; d < Nd; d++) {
+                    int kd = std::min(gc[d], fdims[d] - gc[d]);
+                    k2 += kd*kd;
+                }
+                psum[k2] += (RealD)extractLane(lane, pv[os]);
+                cnt [k2] += 1.0;
+            }
+        }
+    }
+    g->GlobalSumVector(psum.data(), (int)psum.size());
+    g->GlobalSumVector(cnt .data(), (int)cnt .size());
+
+    for (int k2 = 0; k2 <= maxk2; k2++)
+        if (cnt[k2] > 0)
+            ofs << k2 << "  " << std::sqrt((double)k2) << "  "
+                << (psum[k2] / cnt[k2]) * norm << "\n";
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Convert LatticeRealD → LatticeComplexD (zero imaginary part)
+// ─────────────────────────────────────────────────────────────────────────────
+static LatticeComplexD toComplex(const LatticeRealD& r)
+{
+    std::vector<RealD>    lr;
+    unvectorizeToLexOrdArray(lr, r);
+    std::vector<ComplexD> lc(lr.size());
+    for (size_t i = 0; i < lr.size(); i++) lc[i] = ComplexD(lr[i], 0.0);
+    LatticeComplexD c(r.Grid());
+    vectorizeFromLexOrdArray(lc, c);
+    return c;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// File readers
+// ─────────────────────────────────────────────────────────────────────────────
+static LatticeComplexD readScidac(const std::string& fname, GridCartesian* g)
+{
+    LatticeRealD field(g);
+    emptyUserRecord rec;
+    ScidacReader RD;
+    RD.open(fname);
+    RD.readScidacFieldRecord(field, rec);
+    RD.close();
+    return toComplex(field);
+}
+
+static LatticeComplexD readBinary(const std::string& fname, GridCartesian* g)
+{
+    // Raw IEEE doubles in lex order [x][y][z][t] written serially.
+    // Rank 0 reads the file, broadcasts to all ranks via GlobalSumVector.
+    Coordinate fdims = g->_fdimensions;
+    long vol4 = 1; for (int d = 0; d < Nd; d++) vol4 *= fdims[d];
+
+    std::vector<RealD> buf(vol4, 0.0);
+    if (g->IsBoss()) {
+        std::ifstream f(fname, std::ios::binary);
+        if (!f) throw std::runtime_error("Cannot open: " + fname);
+        f.read(reinterpret_cast<char*>(buf.data()), vol4 * sizeof(RealD));
+    }
+    g->GlobalSumVector(buf.data(), (int)vol4);   // broadcast from rank 0
+
+    LatticeRealD field(g);
+    Coordinate ldims = g->_ldimensions, pcoor = g->_processor_coor;
+    for (long ls = 0; ls < g->lSites(); ls++) {
+        Coordinate lc;
+        g->LocalIndexToLocalCoor(ls, lc);
+        long glex = 0, stride = 1;
+        for (int d = 0; d < Nd; d++) {
+            glex += (pcoor[d]*ldims[d] + lc[d]) * stride;
+            stride *= fdims[d];
+        }
+        pokeLocalSite(buf[glex], field, lc);
+    }
+    return toComplex(field);
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Load all files into a trajectory vector
+// ─────────────────────────────────────────────────────────────────────────────
+static std::vector<LatticeComplexD>
+loadFiles(const std::vector<std::string>& files, GridCartesian* g,
+          const std::string& fmt)
+{
+    std::vector<LatticeComplexD> traj;
+    traj.reserve(files.size());
+    for (int n = 0; n < (int)files.size(); n++) {
+        std::cout << GridLogMessage << "[" << n << "] reading " << files[n] << "\n";
+        if (fmt == "scidac") traj.push_back(readScidac(files[n], g));
+        else                 traj.push_back(readBinary (files[n], g));
+    }
+    return traj;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Spatial FFT analysis
+// ─────────────────────────────────────────────────────────────────────────────
+static void analyzeSpatial(const std::vector<LatticeComplexD>& traj,
+                            GridCartesian* g, const std::string& pfx)
+{
+    int  Ntraj = (int)traj.size();
+    long vol4  = 1; for (int d = 0; d < Nd; d++) vol4 *= g->_fdimensions[d];
+
+    FFT theFFT(g);
+    LatticeRealD pavg(g); pavg = Zero();
+    for (int n = 0; n < Ntraj; n++) {
+        LatticeComplexD fk(g);
+        theFFT.FFT_all_dim(fk, traj[n], FFT::forward);
+        // Evaluate product before applying real() — real() is not defined for
+        // unevaluated LatticeBinaryExpression.
+        LatticeComplexD fk_sq(g); fk_sq = conjugate(fk) * fk;
+        std::vector<ComplexD> lc; unvectorizeToLexOrdArray(lc, fk_sq);
+        std::vector<RealD> lr(lc.size());
+        for (size_t i = 0; i < lc.size(); i++) lr[i] = lc[i].real();
+        LatticeRealD pk(g); vectorizeFromLexOrdArray(lr, pk);
+        pavg += pk;
+    }
+    pavg *= (1.0 / Ntraj);
+
+    // Shell-averaged spectrum
+    {
+        std::ofstream fs(pfx + "_spatial_shell.dat");
+        fs << "# k2  |k|  P_shell_avg   (P = |F|^2 / V^2 / shell_count)\n";
+        shellAverage(pavg, g, 1.0 / ((double)vol4 * vol4), fs);
+        std::cout << GridLogMessage << "Written: " << pfx << "_spatial_shell.dat\n";
+    }
+
+    // Per-mode table (rank 0 only, via peekSite)
+    if (g->IsBoss()) {
+        std::ofstream fm(pfx + "_spatial_modes.dat");
+        fm << "# kt  kz  ky  kx  k2  |k|  P_traj_avg\n";
+        Coordinate fdims = g->_fdimensions, site(Nd);
+        double norm = 1.0 / ((double)vol4 * vol4);
+        for (site[3]=0; site[3]<fdims[3]; site[3]++) { int kt=std::min(site[3],fdims[3]-site[3]);
+        for (site[2]=0; site[2]<fdims[2]; site[2]++) { int kz=std::min(site[2],fdims[2]-site[2]);
+        for (site[1]=0; site[1]<fdims[1]; site[1]++) { int ky=std::min(site[1],fdims[1]-site[1]);
+        for (site[0]=0; site[0]<fdims[0]; site[0]++) { int kx=std::min(site[0],fdims[0]-site[0]);
+            RealD p; peekSite(p, pavg, site);
+            int k2 = kt*kt+kz*kz+ky*ky+kx*kx;
+            fm << kt<<" "<<kz<<" "<<ky<<" "<<kx<<" "<<k2<<" "
+               <<std::sqrt((double)k2)<<" "<<p*norm<<"\n";
+        }}}}
+        std::cout << GridLogMessage << "Written: " << pfx << "_spatial_modes.dat\n";
+    }
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Local FFTW trajectory FFT: each rank FFTs its own local sites independently
+// ─────────────────────────────────────────────────────────────────────────────
+static void trajFFT(const std::vector<LatticeComplexD>& in,
+                    std::vector<LatticeComplexD>& out)
+{
+    int    Ntraj  = (int)in.size();
+    GridBase* g   = in[0].Grid();
+    long   lsites = g->lSites();
+
+    // fftw_complex is double[2] — std::vector<double[2]> is not supported by nvcc.
+    // Use std::vector<double> with 2x size and reinterpret_cast.
+    std::vector<double> ibuf((long)Ntraj * lsites * 2, 0.0);
+    for (int n = 0; n < Ntraj; n++) {
+        std::vector<ComplexD> lc;
+        unvectorizeToLexOrdArray(lc, in[n]);
+        for (long s = 0; s < lsites; s++) {
+            ibuf[((long)n*lsites + s)*2 + 0] = lc[s].real();
+            ibuf[((long)n*lsites + s)*2 + 1] = lc[s].imag();
+        }
+    }
+
+    // lsites transforms of length Ntraj, stride=lsites, dist=1
+    std::vector<double> obuf((long)Ntraj * lsites * 2, 0.0);
+    fftw_complex* iptr = reinterpret_cast<fftw_complex*>(ibuf.data());
+    fftw_complex* optr = reinterpret_cast<fftw_complex*>(obuf.data());
+    int n1[1] = {Ntraj};
+    fftw_plan p = fftw_plan_many_dft(
+        1, n1, (int)lsites,
+        iptr, nullptr, (int)lsites, 1,
+        optr, nullptr, (int)lsites, 1,
+        FFTW_FORWARD, FFTW_ESTIMATE);
+    fftw_execute(p);
+    fftw_destroy_plan(p);
+
+    // vector::assign triggers _M_fill_assign which needs a default constructor;
+    // Lattice has none.  Use explicit emplace_back instead.
+    out.clear(); out.reserve(Ntraj);
+    for (int k = 0; k < Ntraj; k++) out.emplace_back(g);
+    for (int k = 0; k < Ntraj; k++) {
+        std::vector<ComplexD> lc(lsites);
+        for (long s = 0; s < lsites; s++)
+            lc[s] = ComplexD(obuf[((long)k*lsites+s)*2], obuf[((long)k*lsites+s)*2+1]);
+        vectorizeFromLexOrdArray(lc, out[k]);
+    }
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Trajectory-axis FFT analysis
+// ─────────────────────────────────────────────────────────────────────────────
+static void analyzeTraj(const std::vector<LatticeComplexD>& traj,
+                         GridCartesian* g, const std::string& pfx)
+{
+    int  Ntraj = (int)traj.size();
+    int  Nf    = Ntraj / 2 + 1;
+    long vol4  = 1; for (int d = 0; d < Nd; d++) vol4 *= g->_fdimensions[d];
+
+    std::vector<LatticeComplexD> ftraj;
+    trajFFT(traj, ftraj);
+
+    // P(k) = (1/vol4) * sum_sites |F_traj(k)|^2 / Ntraj^2
+    std::vector<RealD> Pavg(Ntraj, 0.0);
+    for (int k = 0; k < Ntraj; k++) {
+        LatticeComplexD tmp(g); tmp = conjugate(ftraj[k]) * ftraj[k];
+        std::vector<ComplexD> lc; unvectorizeToLexOrdArray(lc, tmp);
+        std::vector<RealD> lr(lc.size());
+        for (size_t i = 0; i < lc.size(); i++) lr[i] = lc[i].real();
+        LatticeRealD pk(g); vectorizeFromLexOrdArray(lr, pk);
+        RealD s = 0.0;
+        for (long ls = 0; ls < g->lSites(); ls++) {
+            Coordinate lc; g->LocalIndexToLocalCoor(ls, lc);
+            RealD v; peekLocalSite(v, pk, lc);
+            s += v;
+        }
+        g->GlobalSum(s);
+        Pavg[k] = s / ((double)vol4 * Ntraj * Ntraj);
+    }
+
+    {
+        std::ofstream f(pfx + "_traj_power.dat");
+        f << "# k  freq  P_avg   (sum_k P = site mean-square)\n";
+        for (int k = 0; k < Nf; k++)
+            f << k << "  " << (double)k/Ntraj << "  " << Pavg[k] << "\n";
+        std::cout << GridLogMessage << "Written: " << pfx << "_traj_power.dat\n";
+    }
+
+    // Autocorrelation via IFFT of the power spectrum
+    std::vector<double> Pc_buf(Nf * 2, 0.0);
+    for (int k = 0; k < Nf; k++) { Pc_buf[k*2] = Pavg[k]; Pc_buf[k*2+1] = 0.0; }
+    fftw_complex* Pc = reinterpret_cast<fftw_complex*>(Pc_buf.data());
+    std::vector<double> acorr(Ntraj, 0.0);
+    fftw_plan ip = fftw_plan_dft_c2r(1, &Ntraj, Pc, acorr.data(), FFTW_ESTIMATE);
+    fftw_execute(ip); fftw_destroy_plan(ip);
+    double c0 = acorr[0] / Ntraj;
+    {
+        std::ofstream f(pfx + "_traj_autocorr.dat");
+        f << "# lag  C(lag)  C(lag)/C(0)\n";
+        for (int lag = 0; lag < Ntraj/2; lag++) {
+            double c = acorr[lag] / Ntraj;
+            f << lag << "  " << c << "  " << (c0 > 0 ? c/c0 : 0.0) << "\n";
+        }
+        std::cout << GridLogMessage << "Written: " << pfx << "_traj_autocorr.dat\n";
+    }
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Full 5-D analysis: spatial FFT → trajectory FFT → 2-D cross spectrum
+// ─────────────────────────────────────────────────────────────────────────────
+static void analyzeAll5D(const std::vector<LatticeComplexD>& traj,
+                          GridCartesian* g, const std::string& pfx)
+{
+    int  Ntraj = (int)traj.size();
+    int  Nf    = Ntraj / 2 + 1;
+    long vol4  = 1; for (int d = 0; d < Nd; d++) vol4 *= g->_fdimensions[d];
+    Coordinate fdims = g->_fdimensions;
+
+    // Step 1: spatial FFT for each trajectory
+    FFT theFFT(g);
+    std::vector<LatticeComplexD> sfft(Ntraj, LatticeComplexD(g));
+    for (int n = 0; n < Ntraj; n++)
+        theFFT.FFT_all_dim(sfft[n], traj[n], FFT::forward);
+
+    // Step 2: trajectory FFTW on the momentum-space series
+    std::vector<LatticeComplexD> tfft;
+    trajFFT(sfft, tfft);
+
+    // Step 3: P(f_traj, |k_spatial|^2) shell-averaged
+    int maxk2 = 0;
+    for (int d = 0; d < Nd; d++) { int h = fdims[d]/2; maxk2 += h*h; }
+
+    long nb = (long)Nf * (maxk2+1);
+    std::vector<RealD> p2d(nb, 0.0), cnt2d(nb, 0.0);
+    const int Nsimd = vComplexD::Nsimd();
+    double norm = 1.0 / ((double)Ntraj * Ntraj * vol4 * vol4);
+
+    for (int kf = 0; kf < Nf; kf++) {
+        LatticeComplexD tmp(g); tmp = conjugate(tfft[kf]) * tfft[kf];
+        std::vector<ComplexD> lc2; unvectorizeToLexOrdArray(lc2, tmp);
+        std::vector<RealD> lr2(lc2.size());
+        for (size_t i = 0; i < lc2.size(); i++) lr2[i] = lc2[i].real();
+        LatticeRealD pk(g); vectorizeFromLexOrdArray(lr2, pk);
+        auto pv = pk.View(CpuRead);
+        Coordinate gc(Nd);
+        for (int os = 0; os < (int)g->oSites(); os++) {
+            for (int lane = 0; lane < Nsimd; lane++) {
+                globalCoor(os, lane, g, gc);
+                int k2 = 0;
+                for (int d = 0; d < Nd; d++) {
+                    int kd = std::min(gc[d], fdims[d] - gc[d]);
+                    k2 += kd*kd;
+                }
+                long b = (long)kf*(maxk2+1) + k2;
+                p2d [b] += (RealD)extractLane(lane, pv[os]);
+                cnt2d[b] += 1.0;
+            }
+        }
+    }
+    g->GlobalSumVector(p2d .data(), (int)nb);
+    g->GlobalSumVector(cnt2d.data(), (int)nb);
+
+    if (g->IsBoss()) {
+        std::ofstream f(pfx + "_5d_cross.dat");
+        f << "# k_traj  freq_traj  k2_spatial  |k_spatial|  P_avg\n";
+        for (int kf = 0; kf < Nf; kf++) {
+            double freq = (double)kf / Ntraj;
+            for (int k2 = 0; k2 <= maxk2; k2++) {
+                long b = (long)kf*(maxk2+1) + k2;
+                if (cnt2d[b] > 0)
+                    f << kf << "  " << freq << "  " << k2 << "  "
+                      << std::sqrt((double)k2) << "  "
+                      << (p2d[b]/cnt2d[b]) * norm << "\n";
+            }
+            f << "\n";   // blank line between kf slices for gnuplot pm3d
+        }
+        std::cout << GridLogMessage << "Written: " << pfx << "_5d_cross.dat\n";
+    }
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Main
+// ─────────────────────────────────────────────────────────────────────────────
+static void usage(const char* prog)
+{
+    std::cerr
+        << "Usage: " << prog << " --grid Lx.Ly.Lz.Lt [--mpi Px.Py.Pz.Pt]\n"
+        << "       [--fft spatial|traj|all] [--format scidac|binary]\n"
+        << "       [--output PREFIX] file1 [file2 ...]\n";
+    exit(1);
+}
+
+int main(int argc, char** argv)
+{
+    Grid_init(&argc, &argv);   // consumes --grid, --mpi, etc.
+
+    std::string fftMode = "all";
+    std::string fmt     = "scidac";
+    std::string pfx     = "fft5d";
+    std::vector<std::string> files;
+
+    for (int i = 1; i < argc; i++) {
+        std::string a = argv[i];
+        if      (a == "--fft"    && i+1 < argc) fftMode = argv[++i];
+        else if (a == "--format" && i+1 < argc) fmt     = argv[++i];
+        else if (a == "--output" && i+1 < argc) pfx     = argv[++i];
+        else if (!a.empty() && a[0] != '-')     files.push_back(a);
+        // unknown --flags skipped (may be Grid flags already consumed)
+    }
+
+    if (files.empty()) usage(argv[0]);
+
+    GridCartesian* grid = SpaceTimeGrid::makeFourDimGrid(
+        GridDefaultLatt(),
+        GridDefaultSimd(Nd, vComplexD::Nsimd()),
+        GridDefaultMpi());
+
+    Coordinate latt = GridDefaultLatt();
+    std::cout << GridLogMessage << "Lattice : "
+              << latt[0]<<"x"<<latt[1]<<"x"<<latt[2]<<"x"<<latt[3] << "\n"
+              << GridLogMessage << "Ntraj   : " << files.size() << "\n"
+              << GridLogMessage << "FFT     : " << fftMode << "\n"
+              << GridLogMessage << "Format  : " << fmt << "\n"
+              << GridLogMessage << "Output  : " << pfx << "\n";
+
+    auto traj = loadFiles(files, grid, fmt);
+
+    if (fftMode == "spatial" || fftMode == "all") analyzeSpatial(traj, grid, pfx);
+    if (fftMode == "traj"    || fftMode == "all") analyzeTraj   (traj, grid, pfx);
+    if (fftMode == "all")                         analyzeAll5D  (traj, grid, pfx);
+
+    Grid_finalize();
+    return 0;
+}
+
+/*
+  Add to examples/Make.inc to build:
+
+    bin_PROGRAMS += ... fft5d
+    fft5d_SOURCES = fft5d.cc
+    fft5d_LDADD   = $(top_builddir)/Grid/libGrid.a
+*/

systems/Aurora/config-command

@@ -10,7 +10,7 @@ export HDF5=/opt/cray/pe/hdf5/1.12.2.3/gnu/9.1
 	--disable-gparity \
 	--disable-fermion-reps \
 	--enable-shm=nvlink \
-	--enable-checksum-comms=yes \
+	--enable-checksum-comms=no \
 	--enable-log-views=yes \
 	--enable-accelerator=sycl \
 	--enable-accelerator-aware-mpi=no \

systems/Frontier/config-command

@@ -10,12 +10,11 @@ CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
 --disable-fermion-reps \
 --enable-simd=GPU \
 --with-gmp=$OLCF_GMP_ROOT \
---with-fftw=$FFTW_DIR/.. \
 --with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 --disable-fermion-reps \
 CXX=hipcc MPICXX=mpicxx \
 CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 " \
- LDFLAGS="-L/lib64 -L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lhipblas -lrocblas"
+ LDFLAGS="-L/lib64 -L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lhipblas -lrocblas -lhipfft"
 
 
 

tests/Test_cayley_even_odd_vec.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -273,8 +276,6 @@ void  TestWhat(What & Ddwf,
 
   err = phi-chi;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<< std::endl;
-
-  
 }
 
-
+#endif

tests/Test_compressed_lanczos_hot_start.cc

@@ -30,6 +30,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
  *  Reimplement the badly named "multigrid" lanczos as compressed Lanczos using the features 
  *  in Grid that were intended to be used to support blocked Aggregates, from
  */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/iterative/LocalCoherenceLanczos.h>
@@ -256,3 +259,4 @@ int main (int argc, char ** argv) {
   Grid_finalize();
 }
 
+#endif

tests/Test_dwf_dslash_repro.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -237,3 +240,5 @@ int main (int argc, char ** argv)
   
   Grid_finalize();
 }
+
+#endif

tests/Test_dwf_mixedcg_prec.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -222,3 +225,5 @@ int main (int argc, char ** argv)
   
   Grid_finalize();
 }
+
+#endif

tests/Test_dwf_mixedcg_prec_halfcomms.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -118,3 +121,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 #endif
+#endif

tests/Test_meson_field.cc

@@ -24,6 +24,8 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 
 See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
 
 #include <Grid/Grid.h>
 #include <Grid/qcd/utils/A2Autils.h>
@@ -157,3 +159,5 @@ int main(int argc, char *argv[])
   
   return EXIT_SUCCESS;
 }
+
+#endif

tests/core/Test_fft.cc

@@ -29,7 +29,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/Grid.h>
 
 using namespace Grid;
- ;
 
 int main (int argc, char ** argv)
 {
@@ -116,10 +115,10 @@ int main (int argc, char ** argv)
 
   Stilde=S;
   std::cout<<" Benchmarking FFT of LatticeSpinMatrix  "<<std::endl;
-  theFFT.FFT_dim(Stilde,S,0,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
-  theFFT.FFT_dim(Stilde,S,1,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
-  theFFT.FFT_dim(Stilde,S,2,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
-  theFFT.FFT_dim(Stilde,S,3,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,0,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,1,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,2,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,3,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<std::endl;
 
   SpinMatrixD Sp; 
   Sp = Zero(); Sp = Sp+cVol;
@@ -202,11 +201,16 @@ int main (int argc, char ** argv)
     FFT theFFT5(FGrid);
       
     theFFT5.FFT_dim(result5,tmp5,1,FFT::forward); tmp5 = result5;
+    std::cout<<"Fourier xformed Ddwf 1 "<<norm2(result5)<<std::endl;
     theFFT5.FFT_dim(result5,tmp5,2,FFT::forward); tmp5 = result5;
+    std::cout<<"Fourier xformed Ddwf 2 "<<norm2(result5)<<std::endl;
     theFFT5.FFT_dim(result5,tmp5,3,FFT::forward); tmp5 = result5;
-    theFFT5.FFT_dim(result5,tmp5,4,FFT::forward); result5 = result5*ComplexD(::sqrt(1.0/vol),0.0);
+    std::cout<<"Fourier xformed Ddwf 3 "<<norm2(result5)<<std::endl;
+    theFFT5.FFT_dim(result5,tmp5,4,FFT::forward); 
+    std::cout<<"Fourier xformed Ddwf 4 "<<norm2(result5)<<std::endl;
+    result5 = result5*ComplexD(::sqrt(1.0/vol),0.0);
     
-    std::cout<<"Fourier xformed Ddwf"<<std::endl;
+    std::cout<<"Fourier xformed Ddwf "<<norm2(result5)<<std::endl;
     
     tmp5 = src5;
     theFFT5.FFT_dim(src5_p,tmp5,1,FFT::forward); tmp5 = src5_p;
@@ -214,7 +218,7 @@ int main (int argc, char ** argv)
     theFFT5.FFT_dim(src5_p,tmp5,3,FFT::forward); tmp5 = src5_p;
     theFFT5.FFT_dim(src5_p,tmp5,4,FFT::forward); src5_p = src5_p*ComplexD(::sqrt(1.0/vol),0.0);
 
-    std::cout<<"Fourier xformed src5"<<std::endl;
+    std::cout<<"Fourier xformed src5"<< norm2(src5)<<" -> "<<norm2(src5_p)<<std::endl;
       
     /////////////////////////////////////////////////////////////////
     // work out the predicted from Fourier
@@ -251,7 +255,8 @@ int main (int argc, char ** argv)
       Kinetic = Kinetic + sin(kmu)*ci*(Gamma(Gmu[mu])*src5_p);
       
     }
-    
+    std::cout << " src5    "<<norm2(src5_p)<<std::endl;
+    std::cout << " Kinetic "<<norm2(Kinetic)<<std::endl;
     // NB implicit sum over mu
     //
     // 1-1/2 Dw = 1 - 1/2 ( eip+emip)
@@ -260,18 +265,23 @@ int main (int argc, char ** argv)
     //          = 2 sink/2 ink/2 = sk2
     
     W = one - M5 + sk2; 
+    std::cout << " W "<<norm2(W)<<std::endl;
     Kinetic = Kinetic + W * src5_p;
+
+    std::cout << " Kinetic "<<norm2(Kinetic)<<std::endl;
     
     LatticeCoordinate(scoor,sdir);
     
     tmp5 = Cshift(src5_p,sdir,+1);
     tmp5 = (tmp5 - G5*tmp5)*0.5;
     tmp5 = where(scoor==Integer(Ls-1),mass*tmp5,-tmp5);
+    std::cout << " tmp5 "<<norm2(tmp5)<<std::endl;
     Kinetic = Kinetic + tmp5;
     
     tmp5 = Cshift(src5_p,sdir,-1);
     tmp5 = (tmp5 + G5*tmp5)*0.5;
     tmp5 = where(scoor==Integer(0),mass*tmp5,-tmp5);
+    std::cout << " tmp5 "<<norm2(tmp5)<<std::endl;
     Kinetic = Kinetic + tmp5;
     
     std::cout<<"Momentum space Ddwf  "<< norm2(Kinetic)<<std::endl;
@@ -339,7 +349,7 @@ int main (int argc, char ** argv)
     Ddwf.Mdag(src5,tmp5);
     src5=tmp5;
     MdagMLinearOperator<DomainWallFermionD,LatticeFermionD> HermOp(Ddwf);
-    ConjugateGradient<LatticeFermionD> CG(1.0e-16,10000);
+    ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
     CG(HermOp,src5,result5);
     
     ////////////////////////////////////////////////////////////////////////
@@ -423,7 +433,7 @@ int main (int argc, char ** argv)
     Dov.Mdag(src5,tmp5);
     src5=tmp5;
     MdagMLinearOperator<OverlapWilsonCayleyTanhFermionD,LatticeFermionD> HermOp(Dov);
-    ConjugateGradient<LatticeFermionD> CG(1.0e-16,10000);
+    ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
     CG(HermOp,src5,result5);
     
     ////////////////////////////////////////////////////////////////////////

tests/core/Test_fft_prop.cc

@@ -0,0 +1,101 @@
+    /*************************************************************************************
+
+    grid` physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_cshift.cc
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace Grid;
+
+template<class LatticeObject>
+void bench(GridCartesian *grid, std::string name)
+{
+  LatticeComplexD       C(grid);
+  LatticeComplexD       coor(grid);
+
+  ComplexD ci(0.0,1.0);
+  Coordinate p({1,2,3,4});
+
+  Coordinate latt_size   = grid->_fdimensions;
+  std::cout<<"*************************************************"<<std::endl;
+  std::cout<<" Benchmarking FFT of "<<name<<" on plane wave    "<<std::endl;
+  std::cout<<"*************************************************"<<std::endl;
+  C=Zero();
+  for(int mu=0;mu<4;mu++){
+    RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
+    LatticeCoordinate(coor,mu);
+    C = C + (TwoPiL * p[mu]) * coor;
+  }
+  C = exp(C*ci);
+
+  LatticeObject S(grid);
+  LatticeObject Stilde(grid);
+
+  S=Zero();
+  S = S+C;
+
+  FFT theFFT(grid);
+
+  Stilde=S;
+  std::cout << " norm2(s) "<<norm2(Stilde)<<std::endl;
+  double tt= -usecond();
+  theFFT.FFT_dim(Stilde,Stilde,0,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<norm2(Stilde)<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,1,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<norm2(Stilde)<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,2,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<norm2(Stilde)<<std::endl;
+  theFFT.FFT_dim(Stilde,Stilde,3,FFT::forward); std::cout << theFFT.MFlops()<<" mflops "<<norm2(Stilde)<<std::endl;
+  tt+= usecond();
+
+  std::cout<<"*************************************************"<<std::endl;
+  std::cout<<" FFT of "<<latt_size <<" "<<name<<" took "<<tt/1.e6<<" s"<<std::endl;
+  std::cout<<"*************************************************"<<std::endl;
+
+}
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+
+  int vol = 1;
+  for(int d=0;d<latt_size.size();d++){
+    vol = vol * latt_size[d];
+  }
+  GridCartesian         GRID(latt_size,simd_layout,mpi_layout);
+
+
+  bench<LatticeComplexD>(&GRID,std::string("LatticeComplexD"));
+  bench<LatticeColourMatrixD>(&GRID,std::string("LatticeColourMatrixD"));
+  bench<LatticePropagatorD>(&GRID,std::string("LatticePropagatorD"));
+  
+  Grid_finalize();
+}

tests/core/Test_plaquette_stats.cc

@@ -0,0 +1,271 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: tests/core/Test_plaquette_stats.cc
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+/**
+ * Test_plaquette_stats
+ *
+ * Measure every plaquette Re Tr[U_mu(x) U_nu(x+mu) U_mu†(x+nu) U_nu†(x)] / Nc
+ * across all sites and all (mu,nu) planes and report max, min, and average.
+ *
+ * Usage:
+ *   ./Test_plaquette_stats [Grid options] [--file <nersc_config>] [--hot]
+ *                          [--threshold <val>]
+ *
+ *   --file <path>       Read gauge field from NERSC-format file
+ *   --hot               Use random (hot) SU(3) start (default: cold/unit start)
+ *   --threshold <val>   Print coordinates of every plaquette with Re Tr/Nc < val
+ *
+ * Grid size defaults to 8^3 x 16; override with --grid (e.g. --grid 4.4.4.8).
+ */
+
+#include <Grid/Grid.h>
+#include <limits>
+
+using namespace std;
+using namespace Grid;
+
+template<class T>
+static void writeFile(T& in, const std::string& fname)
+{
+#ifdef HAVE_LIME
+  std::cout << GridLogMessage << "Writing: " << fname << std::endl;
+  Grid::emptyUserRecord record;
+  Grid::ScidacWriter WR(in.Grid()->IsBoss());
+  WR.open(fname);
+  WR.writeScidacFieldRecord(in, record, 0);
+  WR.close();
+#endif
+}
+
+// Return the plane label string for output
+static std::string planeName(int mu, int nu)
+{
+  const char dirs[] = "xyzt";
+  std::string s;
+  s += dirs[mu];
+  s += dirs[nu];
+  return s;
+}
+
+int main(int argc, char** argv)
+{
+  Grid_init(&argc, &argv);
+
+  // ---- Lattice geometry ----
+  Coordinate latt_size   = GridDefaultLatt();
+  if (latt_size.size() == 0) {
+    latt_size = Coordinate(std::vector<int>{8, 8, 8, 16});
+  }
+  Coordinate simd_layout = GridDefaultSimd(Nd, vComplex::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+
+  GridCartesian grid(latt_size, simd_layout, mpi_layout);
+
+  std::cout << GridLogMessage << "Lattice: ";
+  for (int d = 0; d < Nd; d++)
+    std::cout << latt_size[d] << (d < Nd-1 ? "x" : "\n");
+
+  // ---- Gauge field ----
+  LatticeGaugeField Umu(&grid);
+
+  // Check for --file argument
+  std::string config_file = "";
+  for (int i = 1; i < argc - 1; i++) {
+    if (std::string(argv[i]) == "--file") {
+      config_file = argv[i+1];
+      break;
+    }
+  }
+
+  bool doHot = false;
+  for (int i = 1; i < argc; i++) {
+    if (std::string(argv[i]) == "--hot") { doHot = true; break; }
+  }
+
+  double threshold = std::numeric_limits<double>::quiet_NaN();
+  bool   doThresh  = false;
+  for (int i = 1; i < argc - 1; i++) {
+    if (std::string(argv[i]) == "--threshold") {
+      threshold = std::stod(argv[i+1]);
+      doThresh  = true;
+      break;
+    }
+  }
+
+  if (!config_file.empty()) {
+    std::cout << GridLogMessage << "Reading gauge field from " << config_file << std::endl;
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu, header, config_file);
+  } else {
+    std::vector<int> seeds({1, 2, 3, 4});
+    GridParallelRNG pRNG(&grid);
+    pRNG.SeedFixedIntegers(seeds);
+    if (doHot) {
+      std::cout << GridLogMessage << "Generating hot (random SU(3)) start" << std::endl;
+      SU<Nc>::HotConfiguration(pRNG, Umu);
+    } else {
+      std::cout << GridLogMessage << "Using cold (unit) gauge start" << std::endl;
+      SU<Nc>::ColdConfiguration(pRNG, Umu);
+    }
+  }
+
+  // ---- Extract link matrices ----
+  std::vector<LatticeColourMatrix> U(Nd, &grid);
+  for (int mu = 0; mu < Nd; mu++)
+    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+
+  // ---- Per-plane plaquette statistics ----
+  //
+  // For each (mu, nu) plane (mu > nu) compute
+  //   P_munu(x) = Re Tr[plaquette] / Nc
+  // then report max, min, mean over all sites.
+  //
+  // WilsonLoops::traceDirPlaquette gives Tr[U_mu U_nu U_mu† U_nu†] (complex).
+
+  double vol = grid.gSites();
+
+  // Accumulate site-average plaquette (sum over planes / Nplanes / Nc)
+  LatticeComplex plaq_all(&grid);
+  plaq_all = Zero();
+  int Nplanes = 0;
+
+  // Per-site running sum over planes (Re Tr/Nc, normalised per plane)
+  LatticeRealD plaq_sum(&grid);
+  plaq_sum = Zero();
+
+  std::cout << GridLogMessage
+            << "======== Per-plane plaquette statistics ========" << std::endl;
+  std::cout << GridLogMessage
+            << std::setw(6)  << "plane"
+            << std::setw(20) << "max"
+            << std::setw(20) << "min"
+            << std::setw(20) << "average"
+            << std::endl;
+
+  for (int mu = 1; mu < Nd; mu++) {
+    for (int nu = 0; nu < mu; nu++) {
+
+      // Per-site trace of plaquette in (mu,nu) plane
+      LatticeComplex sitePlaq(&grid);
+      ColourWilsonLoops::traceDirPlaquette(sitePlaq, U, mu, nu);
+
+      plaq_all = plaq_all + sitePlaq;
+      Nplanes++;
+
+      // --- global average via sum() ---
+      TComplex Tsum  = sum(sitePlaq);
+      ComplexD csum  = TensorRemove(Tsum);
+      RealD    avg   = csum.real() / vol / Nc;
+
+      // --- global max and min via unvectorize + GlobalMax/GlobalMin ---
+      std::vector<TComplex> sv;
+      unvectorizeToLexOrdArray(sv, sitePlaq);
+
+      RealD local_max = -1e38, local_min = 1e38;
+      for (auto& tc : sv) {
+        RealD r = TensorRemove(tc).real() / Nc;
+        if (r > local_max) local_max = r;
+        if (r < local_min) local_min = r;
+      }
+
+      // Reduce across MPI ranks (no GlobalMin; negate to use GlobalMax)
+      grid.GlobalMax(local_max);
+      local_min = -local_min;
+      grid.GlobalMax(local_min);
+      local_min = -local_min;
+
+      std::cout << GridLogMessage
+                << std::setw(6)  << planeName(mu, nu)
+                << std::setw(20) << std::setprecision(10) << local_max
+                << std::setw(20) << std::setprecision(10) << local_min
+                << std::setw(20) << std::setprecision(10) << avg
+                << std::endl;
+
+      if (doThresh) {
+        std::vector<RealD> vr(sv.size(), 0.0);
+        for (int s = 0; s < (int)sv.size(); s++) {
+          RealD val = TensorRemove(sv[s]).real() / Nc;
+          if (val < threshold) {
+            Coordinate lc(Nd), gc(Nd);
+            Lexicographic::CoorFromIndex(lc, s, grid._ldimensions);
+            for (int d = 0; d < Nd; d++)
+              gc[d] = grid._processor_coor[d] * grid._ldimensions[d] + lc[d];
+            std::cerr << "BELOW_THRESHOLD  plane=" << planeName(mu, nu)
+//                      << "  global=(" << grid._processor_coor[0] << "," << grid._processor_coor[1] << "," << grid._processor_coor[2] << "," << grid._processor_coor[3] << ")"
+//                      << "  dim=(" << grid._ldimensions[0] << "," << grid._ldimensions[1] << "," << grid._ldimensions[2] << "," << grid._ldimensions[3] << ")"
+                      << "  site=(" << gc[0] << "," << gc[1] << "," << gc[2] << "," << gc[3] << ")"
+                      << "  P=" << std::setprecision(10) << val << std::endl;
+            vr[s] = val;
+          }
+        }
+        LatticeRealD plaqRe(&grid);
+        vectorizeFromLexOrdArray(vr, plaqRe);
+        plaq_sum = plaq_sum + plaqRe;
+      }
+    }
+  }
+
+  // Write per-site sum of below-threshold plaquette values over all planes
+  if (doThresh)
+    writeFile(plaq_sum, "./plaq_sum");
+
+  // ---- Overall (averaged over all planes) statistics ----
+  // plaq_all = sum of Tr[...] over all 6 (mu,nu) planes
+  // Normalise to Re Tr / Nc per plane
+  plaq_all = plaq_all * (1.0 / Nc / Nplanes);
+
+  TComplex Tsum_all = sum(plaq_all);
+  RealD    avg_all  = TensorRemove(Tsum_all).real() / vol;
+
+  std::vector<TComplex> sv_all;
+  unvectorizeToLexOrdArray(sv_all, plaq_all);
+
+  RealD max_all = -1e38, min_all = 1e38;
+  for (auto& tc : sv_all) {
+    RealD r = TensorRemove(tc).real();
+    if (r > max_all) max_all = r;
+    if (r < min_all) min_all = r;
+  }
+  grid.GlobalMax(max_all);
+  min_all = -min_all;
+  grid.GlobalMax(min_all);
+  min_all = -min_all;
+
+  // Cross-check with built-in avgPlaquette
+  RealD avg_builtin = ColourWilsonLoops::avgPlaquette(Umu);
+
+  std::cout << GridLogMessage
+            << "======== Overall plaquette statistics (all planes) ========" << std::endl;
+  std::cout << GridLogMessage << "  max     = " << max_all  << std::endl;
+  std::cout << GridLogMessage << "  min     = " << min_all  << std::endl;
+  std::cout << GridLogMessage << "  average = " << avg_all  << std::endl;
+  std::cout << GridLogMessage << "  avgPlaquette (builtin check) = " << avg_builtin << std::endl;
+
+  Grid_finalize();
+  return 0;
+}