Reverted restart to inline in KrylovSchur

Making constructUR to use basisRotate()

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,54 @@
+name: Bug report
+description: Report a bug.
+title: "<insert title>"
+labels: [bug]
+
+body:
+  - type: markdown
+    attributes:
+      value: >
+        Thank you for taking the time to file a bug report.
+        Please check that the code is pointing to the HEAD of develop
+        or any commit in master which is tagged with a version number.
+
+  - type: textarea
+    attributes:
+      label: "Describe the issue:"
+      description: >
+        Describe the issue and any previous attempt to solve it.
+    validations:
+      required: true
+
+  - type: textarea
+    attributes:
+      label: "Code example:"
+      description: >
+        If relevant, show how to reproduce the issue using a minimal working
+        example.
+      placeholder: |
+        << your code here >>
+      render: shell
+    validations:
+      required: false
+
+  - type: textarea
+    attributes:
+      label: "Target platform:"
+      description: >
+        Give a description of the target platform (CPU, network, compiler).
+        Please give the full CPU part description, using for example
+        `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux)
+        or `sysctl machdep.cpu.brand_string` (macOS) and the full output
+        the `--version` option of your compiler.
+    validations:
+      required: true
+
+  - type: textarea
+    attributes:
+      label: "Configure options:"
+      description: >
+        Please give the exact configure command used and attach
+        `config.log`, `grid.config.summary` and the output of `make V=1`.
+      render: shell
+    validations:
+      required: true

.gitignore

@@ -1,3 +1,7 @@
+# Doxygen stuff
+html/*
+latex/*
+
 # Compiled Object files #
 #########################
 *.slo
@@ -88,6 +92,7 @@ Thumbs.db
 # build directory #
 ###################
 build*/*
+Documentation/_build
 
 # IDE related files #
 #####################

.travis.yml

@@ -1,61 +0,0 @@
-language: cpp
-
-cache:
-  directories:
-    - clang
-
-matrix:
-  include:
-    - os:        osx
-      osx_image: xcode8.3
-      compiler: clang
-      env: PREC=single
-    - os:        osx
-      osx_image: xcode8.3
-      compiler: clang
-      env: PREC=double
-      
-before_install:
-    - export GRIDDIR=`pwd`
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]] && [ ! -e clang/bin ]; then wget $CLANG_LINK; tar -xf `basename $CLANG_LINK`; mkdir clang; mv clang+*/* clang/; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export PATH="${GRIDDIR}/clang/bin:${PATH}"; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc openssl; fi
-    
-install:
-    - export CWD=`pwd`
-    - echo $CWD
-    - export CC=$CC$VERSION
-    - export CXX=$CXX$VERSION
-    - echo $PATH
-    - which autoconf
-    - autoconf  --version
-    - which automake
-    - automake  --version
-    - which $CC
-    - $CC  --version
-    - which $CXX
-    - $CXX --version
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export EXTRACONF='--with-openssl=/usr/local/opt/openssl'; fi
-    
-script:
-    - ./bootstrap.sh
-    - mkdir build
-    - cd build
-    - mkdir lime
-    - cd lime
-    - mkdir build
-    - cd build
-    - wget http://usqcd-software.github.io/downloads/c-lime/lime-1.3.2.tar.gz
-    - tar xf lime-1.3.2.tar.gz
-    - cd lime-1.3.2
-    - ./configure --prefix=$CWD/build/lime/install
-    - make -j4
-    - make install
-    - cd $CWD/build
-    - ../configure --enable-precision=$PREC --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
-    - make -j4 
-    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
-    - make check

BLAS_benchmark/compile-command

@@ -0,0 +1,2 @@
+
+mpicxx -qmkl=parallel -fsycl BatchBlasBench.cc -o BatchBlasBench -DGRID_SYCL

BLAS_benchmark/compile-command-frontier

@@ -0,0 +1,5 @@
+CXX=hipcc
+MPICXX=mpicxx 
+CXXFLAGS="-fPIC -I{$ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 -I/opt/cray/pe/mpich/8.1.28/ofi/gnu/12.3/include -DGRID_HIP"
+LDFLAGS="-L/lib64 -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64 -lhipblas -lrocblas -lmpi_gnu_123"
+hipcc $CXXFLAGS $LDFLAGS BatchBlasBench.cc -o BatchBlasBench

BLAS_benchmark/compile-command-sunspot

@@ -0,0 +1,2 @@
+
+mpicxx -qmkl=parallel -fsycl BatchBlasBench.cc -o BatchBlasBench -DGRID_SYCL

Grid/DisableWarnings.h

@@ -37,19 +37,31 @@ directory
 #endif
 
  //disables and intel compiler specific warning (in json.hpp)
+#ifdef __ICC
 #pragma warning disable 488  
+#endif
 
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
+
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+ //disables nvcc specific warning in json.hpp
+#pragma nv_diag_suppress unsigned_compare_with_zero
+#pragma nv_diag_suppress cast_to_qualified_type
+ //disables nvcc specific warning in many files
+#pragma nv_diag_suppress esa_on_defaulted_function_ignored
+#pragma nv_diag_suppress declared_but_not_referenced
+#pragma nv_diag_suppress extra_semicolon
+#else
+ //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
-
+#pragma diag_suppress declared_but_not_referenced
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon
-
-//Eigen only
+#endif
 #endif
 
 // Disable vectorisation in Eigen on the Power8/9 and PowerPC

Grid/GridCore.h

@@ -44,12 +44,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridStd.h>
 #include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
-#include <Grid/perfmon/PerfCount.h>
+//#include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
-#include <Grid/allocator/AlignedAllocator.h>
+#include <Grid/perfmon/Tracing.h>
+#include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
-#include <Grid/threads/Threads.h>
+#include <Grid/threads/ThreadReduction.h>
 #include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 
@@ -58,6 +59,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice.h>      
 #include <Grid/cshift/Cshift.h>       
 #include <Grid/stencil/Stencil.h>      
+#include <Grid/stencil/GeneralLocalStencil.h>      
 #include <Grid/parallelIO/BinaryIO.h>
 #include <Grid/algorithms/Algorithms.h>   
 NAMESPACE_CHECK(GridCore)

Grid/GridQCDcore.h

@@ -36,6 +36,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/qcd/QCD.h>
 #include <Grid/qcd/spin/Spin.h>
+#include <Grid/qcd/gparity/Gparity.h>
 #include <Grid/qcd/utils/Utils.h>
 #include <Grid/qcd/representations/Representations.h>
 NAMESPACE_CHECK(GridQCDCore);

Grid/GridStd.h

@@ -1,11 +1,20 @@
 #ifndef GRID_STD_H
 #define GRID_STD_H
 
+///////////////////
+// Grid config
+///////////////////
+#include "Config.h"
+
 ///////////////////
 // Std C++ dependencies
 ///////////////////
+#define _NBACKTRACE (256)
+extern void * Grid_backtrace_buffer[_NBACKTRACE];
+
 #include <cassert>
 #include <complex>
+#include <memory>
 #include <vector>
 #include <array>
 #include <string>
@@ -14,17 +23,48 @@
 #include <random>
 #include <functional>
 #include <stdio.h>
+#include <string.h>
 #include <stdlib.h>
+#include <unistd.h>
+#include <strings.h>
 #include <stdio.h>
 #include <signal.h>
 #include <ctime>
 #include <sys/time.h>
 #include <chrono>
 #include <zlib.h>
+#ifdef HAVE_EXECINFO_H
+#include <execinfo.h>
+#endif
 
-///////////////////
-// Grid config
-///////////////////
-#include "Config.h"
+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(" : ");							\
+    ASSLOG(#b);								\
+    ASSLOG(" : ");							\
+    int symbols = backtrace(Grid_backtrace_buffer,_NBACKTRACE);		\
+    backtrace_symbols_fd(Grid_backtrace_buffer,symbols,STDERR_FILENO);	\
+    GridAbort();							\
+  };
+#else
+#define GRID_ASSERT(b) if(!(b)) {					\
+    ASSLOG(" GRID_ASSERT failure: ");					\
+    ASSLOG(__FILE__);							\
+    ASSLOG(" : ");							\
+    ASSLOG(#b);								\
+    ASSLOG(" : ");							\
+    GridAbort();							\
+  };
+#endif
+
+
+#ifdef TOFU
+#undef GRID_COMMS_THREADS
+#endif
 #endif /* GRID_STD_H */

Grid/Grid_Eigen_Dense.h

@@ -14,25 +14,36 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+#pragma nv_diag_suppress code_is_unreachable
+#else
 #pragma diag_suppress code_is_unreachable
+#endif
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")
+#undef __CUDA_ARCH__
 #undef __NVCC__
 #undef __CUDACC__
-#undef __CUDA_ARCH__
 #define __NVCC__REDEFINE__
 #endif 
 
 /* SYCL save and restore compile environment*/
-#ifdef __SYCL_DEVICE_ONLY__  
+#ifdef GRID_SYCL
 #pragma push
 #pragma push_macro("__SYCL_DEVICE_ONLY__")
 #undef __SYCL_DEVICE_ONLY__
-#undef EIGEN_USE_SYCL
 #define EIGEN_DONT_VECTORIZE
+#undef EIGEN_USE_SYCL
+#define __SYCL__REDEFINE__
 #endif
 
+/* HIP save and restore compile environment*/
+#ifdef GRID_HIP
+#pragma push
+#pragma push_macro("__HIP_DEVICE_COMPILE__")
+#endif
+#define EIGEN_NO_HIP
 
 #include <Grid/Eigen/Dense>
 #include <Grid/Eigen/unsupported/CXX11/Tensor>
@@ -51,6 +62,12 @@
 #pragma pop
 #endif
 
+/*HIP restore*/
+#ifdef __HIP__REDEFINE__
+#pragma pop_macro("__HIP_DEVICE_COMPILE__")
+#pragma pop
+#endif
+
 #if defined __GNUC__
 #pragma GCC diagnostic pop
 #endif

Grid/Makefile.am

@@ -21,7 +21,8 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
-all: version-cache
+
+all: version-cache Version.h
 
 version-cache:
 	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
@@ -42,7 +43,7 @@ version-cache:
 	fi;\
 	rm -f vertmp
 
-Version.h:
+Version.h: version-cache
 	cp version-cache Version.h
 
 .PHONY: version-cache
@@ -53,6 +54,27 @@ Version.h:
 include Make.inc
 include Eigen.inc
 
+if BUILD_FERMION_INSTANTIATIONS
+  extra_sources+=$(WILS_FERMION_FILES)
+  extra_sources+=$(STAG_FERMION_FILES)
+if BUILD_ZMOBIUS
+    extra_sources+=$(ZWILS_FERMION_FILES)
+endif
+if BUILD_GPARITY
+    extra_sources+=$(GP_FERMION_FILES)
+endif
+if BUILD_FERMION_REPS
+    extra_sources+=$(ADJ_FERMION_FILES)
+    extra_sources+=$(TWOIND_FERMION_FILES)
+endif
+if BUILD_SP
+    extra_sources+=$(SP_FERMION_FILES)
+if BUILD_FERMION_REPS
+      extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
+endif
+endif
+
 lib_LIBRARIES = libGrid.a
 
 CCFILES += $(extra_sources)

Grid/Namespace.h

@@ -29,6 +29,7 @@ directory
 #pragma once
 
 #include <type_traits>
+#include <exception>
 #include <cassert>
 
 #define NAMESPACE_BEGIN(A) namespace A {
@@ -36,3 +37,7 @@ directory
 #define GRID_NAMESPACE_BEGIN NAMESPACE_BEGIN(Grid)
 #define GRID_NAMESPACE_END   NAMESPACE_END(Grid)
 #define NAMESPACE_CHECK(x) struct namespaceTEST##x {};  static_assert(std::is_same<namespaceTEST##x, ::namespaceTEST##x>::value,"Not in :: at"  ); 
+
+#define EXCEPTION_CHECK_BEGIN(A) try {
+#define EXCEPTION_CHECK_END(A)   } catch ( std::exception e ) { BACKTRACEFP(stderr); std::cerr << __PRETTY_FUNCTION__ << " : " <<__LINE__<< " Caught exception "<<e.what()<<std::endl; throw; }
+

Grid/algorithms/Algorithms.h

@@ -29,9 +29,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_H
 #define GRID_ALGORITHMS_H
 
+NAMESPACE_CHECK(blas);
+#include <Grid/algorithms/blas/BatchedBlas.h>
+
+NAMESPACE_CHECK(algorithms);
 #include <Grid/algorithms/SparseMatrix.h>
 #include <Grid/algorithms/LinearOperator.h>
 #include <Grid/algorithms/Preconditioner.h>
+NAMESPACE_CHECK(SparseMatrix);
 
 #include <Grid/algorithms/approx/Zolotarev.h>
 #include <Grid/algorithms/approx/Chebyshev.h>
@@ -41,15 +46,26 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/approx/Forecast.h>
 #include <Grid/algorithms/approx/RemezGeneral.h>
 #include <Grid/algorithms/approx/ZMobius.h>
-
-#include <Grid/algorithms/iterative/Deflation.h>
+NAMESPACE_CHECK(approx);
+#include <Grid/algorithms/deflation/Deflation.h>
+#include <Grid/algorithms/deflation/MultiRHSBlockProject.h>
+#include <Grid/algorithms/deflation/MultiRHSDeflation.h>
+#include <Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h>
+// Not really deflation, but useful
+#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);
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h>
 #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
@@ -60,9 +76,26 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #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/CoarsenedMatrix.h>
+#include <Grid/algorithms/iterative/AdefGeneric.h>
+#include <Grid/algorithms/iterative/AdefMrhs.h>
+NAMESPACE_CHECK(PowerMethod);
+#include <Grid/algorithms/multigrid/MultiGrid.h>
+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/CoarsenedMatrix.h

@@ -1,960 +0,0 @@
-    // blockZaxpy in bockPromote - 3s, 5%
-    // noncoalesced linalg in Preconditionoer ~ 3s 5%
-    // Lancos tuning or replace 10-20s ~ 25%, open ended
-    // setup tuning   5s  ~  8%
-    //    -- e.g. ordermin, orderstep tunables.
-    // MdagM path without norm in LinOp code.     few seconds
-
-    // Mdir calc blocking kernels
-    // Fuse kernels in blockMaskedInnerProduct
-    // preallocate Vectors in Cayley 5D ~ few percent few seconds
-
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/CoarsenedMatrix.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-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_ALGORITHM_COARSENED_MATRIX_H
-#define  GRID_ALGORITHM_COARSENED_MATRIX_H
-
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj,class CComplex>
-inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
-				    const Lattice<decltype(innerProduct(vobj(),vobj()))> &FineMask,
-				    const Lattice<vobj> &fineX,
-				    const Lattice<vobj> &fineY)
-{
-  typedef decltype(innerProduct(vobj(),vobj())) dotp;
-
-  GridBase *coarse(CoarseInner.Grid());
-  GridBase *fine  (fineX.Grid());
-
-  Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
-  Lattice<dotp> fine_inner_msk(fine);
-
-  // Multiply could be fused with innerProduct
-  // Single block sum kernel could do both masks.
-  fine_inner = localInnerProduct(fineX,fineY);
-  mult(fine_inner_msk, fine_inner,FineMask);
-  blockSum(CoarseInner,fine_inner_msk);
-}
-
-
-class Geometry {
-public:
-  int npoint;
-  std::vector<int> directions   ;
-  std::vector<int> displacements;
-
-  Geometry(int _d)  {
-    
-    int base = (_d==5) ? 1:0;
-
-    // make coarse grid stencil for 4d , not 5d
-    if ( _d==5 ) _d=4;
-
-    npoint = 2*_d+1;
-    directions.resize(npoint);
-    displacements.resize(npoint);
-    for(int d=0;d<_d;d++){
-      directions[d   ] = d+base;
-      directions[d+_d] = d+base;
-      displacements[d  ] = +1;
-      displacements[d+_d]= -1;
-    }
-    directions   [2*_d]=0;
-    displacements[2*_d]=0;
-      
-    //// report back
-    std::cout<<GridLogMessage<<"directions    :";
-    for(int d=0;d<npoint;d++) std::cout<< directions[d]<< " ";
-    std::cout<<std::endl;
-    std::cout<<GridLogMessage<<"displacements :";
-    for(int d=0;d<npoint;d++) std::cout<< displacements[d]<< " ";
-    std::cout<<std::endl;
-  }
-  
-  /*
-  // Original cleaner code
-  Geometry(int _d) : dimension(_d), npoint(2*_d+1), directions(npoint), displacements(npoint) {
-  for(int d=0;d<dimension;d++){
-  directions[2*d  ] = d;
-  directions[2*d+1] = d;
-  displacements[2*d  ] = +1;
-  displacements[2*d+1] = -1;
-  }
-  directions   [2*dimension]=0;
-  displacements[2*dimension]=0;
-  }
-  std::vector<int> GetDelta(int point) {
-  std::vector<int> delta(dimension,0);
-  delta[directions[point]] = displacements[point];
-  return delta;
-  };
-  */    
-
-};
-  
-template<class Fobj,class CComplex,int nbasis>
-class Aggregation   {
-public:
-  typedef iVector<CComplex,nbasis >             siteVector;
-  typedef Lattice<siteVector>                 CoarseVector;
-  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
-
-  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
-  typedef Lattice<Fobj >        FineField;
-
-  GridBase *CoarseGrid;
-  GridBase *FineGrid;
-  std::vector<Lattice<Fobj> > subspace;
-  int checkerboard;
-  int Checkerboard(void){return checkerboard;}
-  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
-    CoarseGrid(_CoarseGrid),
-    FineGrid(_FineGrid),
-    subspace(nbasis,_FineGrid),
-    checkerboard(_checkerboard)
-  {
-  };
-  
-  void Orthogonalise(void){
-    CoarseScalar InnerProd(CoarseGrid); 
-    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
-    blockOrthogonalise(InnerProd,subspace);
-    //    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 2"<<std::endl; // Really have to do twice? Yuck
-    //    blockOrthogonalise(InnerProd,subspace);
-    //      std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
-    //      CheckOrthogonal();
-  } 
-  void CheckOrthogonal(void){
-    CoarseVector iProj(CoarseGrid); 
-    CoarseVector eProj(CoarseGrid); 
-    for(int i=0;i<nbasis;i++){
-      blockProject(iProj,subspace[i],subspace);
-      eProj=Zero(); 
-      accelerator_for(ss, CoarseGrid->oSites(),1,{
-	eProj[ss](i)=CComplex(1.0);
-      });
-      eProj=eProj - iProj;
-      std::cout<<GridLogMessage<<"Orthog check error "<<i<<" " << norm2(eProj)<<std::endl;
-    }
-    std::cout<<GridLogMessage <<"CheckOrthog done"<<std::endl;
-  }
-  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
-    blockProject(CoarseVec,FineVec,subspace);
-  }
-  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
-    FineVec.Checkerboard() = subspace[0].Checkerboard();
-    blockPromote(CoarseVec,FineVec,subspace);
-  }
-  void CreateSubspaceRandom(GridParallelRNG &RNG){
-    for(int i=0;i<nbasis;i++){
-      random(RNG,subspace[i]);
-    }
-  }
-
-  virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
-
-    RealD scale;
-
-    ConjugateGradient<FineField> CG(1.0e-2,100,false);
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-
-    for(int b=0;b<nn;b++){
-	
-      subspace[b] = Zero();
-      gaussian(RNG,noise);
-      scale = std::pow(norm2(noise),-0.5); 
-      noise=noise*scale;
-	
-      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-      for(int i=0;i<1;i++){
-
-	CG(hermop,noise,subspace[b]);
-
-	noise = subspace[b];
-	scale = std::pow(norm2(noise),-0.5); 
-	noise=noise*scale;
-
-      }
-
-      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
-      subspace[b]   = noise;
-
-    }
-  }
-
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
-  // and this is the best I found
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-#if 1
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-    {
-      // Filter
-      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
-      Cheb(hermop,noise,Mn);
-      // normalise
-      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
-      subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
-      b++;
-    }
-
-    // Generate a full sequence of Chebyshevs
-    {
-      lo=filterlo;
-      noise=Mn;
-
-      FineField T0(FineGrid); T0 = noise;  
-      FineField T1(FineGrid); 
-      FineField T2(FineGrid);
-      FineField y(FineGrid);
-      
-      FineField *Tnm = &T0;
-      FineField *Tn  = &T1;
-      FineField *Tnp = &T2;
-
-      // Tn=T1 = (xscale M + mscale)in
-      RealD xscale = 2.0/(hi-lo);
-      RealD mscale = -(hi+lo)/(hi-lo);
-      hermop.HermOp(T0,y);
-      T1=y*xscale+noise*mscale;
-
-      for(int n=2;n<=ordermin+orderstep*(nn-2);n++){
-	
-	hermop.HermOp(*Tn,y);
-
-	auto y_v = y.View();
-	auto Tn_v = Tn->View();
-	auto Tnp_v = Tnp->View();
-	auto Tnm_v = Tnm->View();
-	const int Nsimd = CComplex::Nsimd();
-	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
-	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
-	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
-        });
-
-	// Possible more fine grained control is needed than a linear sweep,
-	// but huge productivity gain if this is simple algorithm and not a tunable
-	int m =1;
-	if ( n>=ordermin ) m=n-ordermin;
-	if ( (m%orderstep)==0 ) { 
-	  Mn=*Tnp;
-	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
-	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
-	  b++;
-	}
-
-	// Cycle pointers to avoid copies
-	FineField *swizzle = Tnm;
-	Tnm    =Tn;
-	Tn     =Tnp;
-	Tnp    =swizzle;
-	  
-      }
-    }
-    assert(b==nn);
-  }
-#endif
-#if 0
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-    FineField combined(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-#define FILTERb(llo,hhi,oorder)						\
-    {									\
-      Chebyshev<FineField> Cheb(llo,hhi,oorder);			\
-      Cheb(hermop,noise,Mn);						\
-      scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;			\
-      subspace[b]   = Mn;						\
-      hermop.Op(Mn,tmp);						\
-      std::cout<<GridLogMessage << oorder<< " Cheb filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
-      b++;								\
-    }									
-
-    //      JacobiPolynomial<FineField> Cheb(0.002,60.0,1500,-0.5,3.5);	\
-
-    RealD alpha=-0.8;
-    RealD beta =-0.8;
-#define FILTER(llo,hhi,oorder)						\
-    {									\
-      Chebyshev<FineField> Cheb(llo,hhi,oorder);			\
-      /* JacobiPolynomial<FineField> Cheb(0.0,60.0,oorder,alpha,beta);*/\
-      Cheb(hermop,noise,Mn);						\
-      scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;			\
-      subspace[b]   = Mn;						\
-      hermop.Op(Mn,tmp);						\
-      std::cout<<GridLogMessage << oorder<< "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
-      b++;								\
-    }									
-    
-#define FILTERc(llo,hhi,oorder)				\
-    {							\
-      Chebyshev<FineField> Cheb(llo,hhi,oorder);	\
-      Cheb(hermop,noise,combined);			\
-    }									
-
-    double node = 0.000;
-    FILTERb(lo,hi,orderfilter);// 0
-    //    FILTERc(node,hi,51);// 0
-    noise = Mn;
-    int base = 0;
-    int mult = 100;
-    FILTER(node,hi,base+1*mult);
-    FILTER(node,hi,base+2*mult);
-    FILTER(node,hi,base+3*mult);
-    FILTER(node,hi,base+4*mult);
-    FILTER(node,hi,base+5*mult);
-    FILTER(node,hi,base+6*mult);
-    FILTER(node,hi,base+7*mult);
-    FILTER(node,hi,base+8*mult);
-    FILTER(node,hi,base+9*mult);
-    FILTER(node,hi,base+10*mult);
-    FILTER(node,hi,base+11*mult);
-    FILTER(node,hi,base+12*mult);
-    FILTER(node,hi,base+13*mult);
-    FILTER(node,hi,base+14*mult);
-    FILTER(node,hi,base+15*mult);
-    assert(b==nn);
-  }
-#endif
-
-#if 0
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-    FineField combined(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-    {						
-      Chebyshev<FineField> JacobiPoly(0.005,60.,1500);
-      //      JacobiPolynomial<FineField> JacobiPoly(0.002,60.0,1500,-0.5,3.5);
-      //JacobiPolynomial<FineField> JacobiPoly(0.03,60.0,500,-0.5,3.5);
-      //      JacobiPolynomial<FineField> JacobiPoly(0.00,60.0,1000,-0.5,3.5);
-      JacobiPoly(hermop,noise,Mn);
-      scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;
-      subspace[b]   = Mn;
-      hermop.Op(Mn,tmp);
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; 
-      b++;
-      //      scale = std::pow(norm2(tmp),-0.5);     tmp=tmp*scale;
-      //      subspace[b]   = tmp;      b++;
-      //    }									
-    }									
-
-#define FILTER(lambda)						\
-    {								\
-      hermop.HermOp(subspace[0],tmp);				\
-      tmp = tmp - lambda *subspace[0];				\
-      scale = std::pow(norm2(tmp),-0.5);			\
-      tmp=tmp*scale;							\
-      subspace[b]   = tmp;						\
-      hermop.Op(subspace[b],tmp);					\
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
-      b++;								\
-    }									
-    //      scale = std::pow(norm2(tmp),-0.5);     tmp=tmp*scale;
-    //      subspace[b]   = tmp;      b++;
-    //    }									
-
-    FILTER(2.0e-5);
-    FILTER(2.0e-4);
-    FILTER(4.0e-4);
-    FILTER(8.0e-4);
-    FILTER(8.0e-4);
-
-    FILTER(2.0e-3);
-    FILTER(3.0e-3);
-    FILTER(4.0e-3);
-    FILTER(5.0e-3);
-    FILTER(6.0e-3);
-
-    FILTER(2.5e-3);
-    FILTER(3.5e-3);
-    FILTER(4.5e-3);
-    FILTER(5.5e-3);
-    FILTER(6.5e-3);
-
-    //    FILTER(6.0e-5);//6
-    //    FILTER(7.0e-5);//8
-    //    FILTER(8.0e-5);//9
-    //    FILTER(9.0e-5);//3
-
-    /*
-    //    FILTER(1.0e-4);//10
-    FILTER(2.0e-4);//11
-    //   FILTER(3.0e-4);//12
-    //    FILTER(4.0e-4);//13
-    FILTER(5.0e-4);//14
-
-    FILTER(6.0e-3);//4
-    FILTER(7.0e-4);//1
-    FILTER(8.0e-4);//7
-    FILTER(9.0e-4);//15
-    FILTER(1.0e-3);//2
-
-    FILTER(2.0e-3);//2
-    FILTER(3.0e-3);//2
-    FILTER(4.0e-3);//2
-    FILTER(5.0e-3);//2
-    FILTER(6.0e-3);//2
-
-    FILTER(7.0e-3);//2
-    FILTER(8.0e-3);//2
-    FILTER(1.0e-2);//2
-    */
-    std::cout << GridLogMessage <<"Jacobi filtering done" <<std::endl;
-    assert(b==nn);
-  }
-#endif
-
-
-};
-
-// Fine Object == (per site) type of fine field
-// nbasis      == number of deflation vectors
-template<class Fobj,class CComplex,int nbasis>
-class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
-public:
-    
-  typedef iVector<CComplex,nbasis >           siteVector;
-  typedef Lattice<CComplex >                  CoarseComplexField;
-  typedef Lattice<siteVector>                 CoarseVector;
-  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
-  typedef iMatrix<CComplex,nbasis >  Cobj;
-  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
-  typedef Lattice<Fobj >        FineField;
-
-  ////////////////////
-  // Data members
-  ////////////////////
-  Geometry         geom;
-  GridBase *       _grid; 
-  int hermitian;
-
-  CartesianStencil<siteVector,siteVector,int> Stencil; 
-
-  std::vector<CoarseMatrix> A;
-      
-  ///////////////////////
-  // Interface
-  ///////////////////////
-  GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know
-
-  void M (const CoarseVector &in, CoarseVector &out)
-  {
-    conformable(_grid,in.Grid());
-    conformable(in.Grid(),out.Grid());
-
-    SimpleCompressor<siteVector> compressor;
-
-    Stencil.HaloExchange(in,compressor);
-
-    auto in_v = in.View();
-    auto out_v = out.View();
-    typedef LatticeView<Cobj> Aview;
-
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
-    Aview *Aview_p = & AcceleratorViewContainer[0];
-
-    const int Nsimd = CComplex::Nsimd();
-    typedef decltype(coalescedRead(in_v[0])) calcVector;
-    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
-
-    int osites=Grid()->oSites();
-
-    accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
-      int ss = sss/nbasis;
-      int b  = sss%nbasis;
-      calcComplex res = Zero();
-      calcVector nbr;
-      int ptype;
-      StencilEntry *SE;
-
-      int lane=SIMTlane(Nsimd);
-      for(int point=0;point<geom.npoint;point++){
-
-	SE=Stencil.GetEntry(ptype,point,ss);
-	  
-	if(SE->_is_local) { 
-	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute,lane);
-	} else {
-	  nbr = coalescedRead(Stencil.CommBuf()[SE->_offset],lane);
-	}
-	synchronise();
-
-	for(int bb=0;bb<nbasis;bb++) {
-	  res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
-	}
-      }
-      coalescedWrite(out_v[ss](b),res,lane);
-    });
-  };
-
-  void Mdag (const CoarseVector &in, CoarseVector &out)
-  {
-    if(hermitian) {
-      // corresponds to Petrov-Galerkin coarsening
-      return M(in,out);
-    } else {
-      // corresponds to Galerkin coarsening
-      CoarseVector tmp(Grid());
-      G5C(tmp, in); 
-      M(tmp, out);
-      G5C(out, out);
-    }
-  };
-  void MdirComms(const CoarseVector &in)
-  {
-    SimpleCompressor<siteVector> compressor;
-    Stencil.HaloExchange(in,compressor);
-  }
-  void MdirCalc(const CoarseVector &in, CoarseVector &out, int point)
-  {
-    conformable(_grid,in.Grid());
-    conformable(_grid,out.Grid());
-
-    typedef LatticeView<Cobj> Aview;
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
-    Aview *Aview_p = & AcceleratorViewContainer[0];
-
-    auto out_v = out.View();
-    auto in_v  = in.View();
-
-    const int Nsimd = CComplex::Nsimd();
-    typedef decltype(coalescedRead(in_v[0])) calcVector;
-    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
-
-    accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
-      int ss = sss/nbasis;
-      int b  = sss%nbasis;
-      calcComplex res = Zero();
-      calcVector nbr;
-      int ptype;
-      StencilEntry *SE;
-
-      int lane=SIMTlane(Nsimd);
-      SE=Stencil.GetEntry(ptype,point,ss);
-	  
-      if(SE->_is_local) { 
-	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute,lane);
-      } else {
-	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset],lane);
-      }
-      synchronise();
-
-      for(int bb=0;bb<nbasis;bb++) {
-	res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
-      }
-      coalescedWrite(out_v[ss](b),res,lane);
-    });
-#if 0
-    accelerator_for(ss,Grid()->oSites(),1,{
-
-      siteVector res = Zero();
-      siteVector nbr;
-      int ptype;
-      StencilEntry *SE;
-      
-      SE=Stencil.GetEntry(ptype,point,ss);
-      
-      if(SE->_is_local&&SE->_permute) {
-	permute(nbr,in_v[SE->_offset],ptype);
-      } else if(SE->_is_local) {
-	nbr = in_v[SE->_offset];
-      } else {
-	nbr = Stencil.CommBuf()[SE->_offset];
-      }
-      synchronise();
-
-      res = res + Aview_p[point][ss]*nbr;
-      
-      out_v[ss]=res;
-    });
-#endif
-  }
-  void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
-  {
-    this->MdirComms(in);
-    int ndir=geom.npoint-1;
-    if ((out.size()!=ndir)&&(out.size()!=ndir+1)) { 
-      std::cout <<"MdirAll out size "<< out.size()<<std::endl;
-      std::cout <<"MdirAll ndir "<< ndir<<std::endl;
-      assert(0);
-    }
-    for(int p=0;p<ndir;p++){
-      MdirCalc(in,out[p],p);
-    }
-  };
-  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
-
-    this->MdirComms(in);
-
-    int ndim = in.Grid()->Nd();
-
-    //////////////
-    // 4D action like wilson
-    // 0+ => 0 
-    // 0- => 1
-    // 1+ => 2 
-    // 1- => 3
-    // etc..
-    //////////////
-    // 5D action like DWF
-    // 1+ => 0 
-    // 1- => 1
-    // 2+ => 2 
-    // 2- => 3
-    // etc..
-    auto point = [dir, disp, ndim](){
-      if(dir == 0 and disp == 0)
-	return 8;
-      else if ( ndim==4 ) { 
-	return (4 * dir + 1 - disp) / 2;
-      } else { 
-	return (4 * (dir-1) + 1 - disp) / 2;
-      }
-    }();
-
-    MdirCalc(in,out,point);
-
-  };
-
-  void Mdiag(const CoarseVector &in, CoarseVector &out)
-  {
-    int point=geom.npoint-1;
-    MdirCalc(in, out, point); // No comms
-  };
-
-  
- CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	: 
-
-    _grid(&CoarseGrid),
-    geom(CoarseGrid._ndimension),
-    hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-      A(geom.npoint,&CoarseGrid)
-  {
-  };
-
-  void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
-		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
-  {
-    typedef Lattice<typename Fobj::tensor_reduced> FineComplexField;
-    typedef typename Fobj::scalar_type scalar_type;
-
-    FineComplexField one(FineGrid); one=scalar_type(1.0,0.0);
-    FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0);
-
-    std::vector<FineComplexField> masks(geom.npoint,FineGrid);
-    FineComplexField imask(FineGrid); // contributions from within this block
-    FineComplexField omask(FineGrid); // contributions from outwith this block
-
-    FineComplexField evenmask(FineGrid);
-    FineComplexField oddmask(FineGrid); 
-
-    FineField     phi(FineGrid);
-    FineField     tmp(FineGrid);
-    FineField     zz(FineGrid); zz=Zero();
-    FineField    Mphi(FineGrid);
-    FineField    Mphie(FineGrid);
-    FineField    Mphio(FineGrid);
-    std::vector<FineField>     Mphi_p(geom.npoint,FineGrid);
-
-    Lattice<iScalar<vInteger> > coor (FineGrid);
-    Lattice<iScalar<vInteger> > bcoor(FineGrid);
-    Lattice<iScalar<vInteger> > bcb  (FineGrid); bcb = Zero();
-
-    CoarseVector iProj(Grid()); 
-    CoarseVector oProj(Grid()); 
-    CoarseVector SelfProj(Grid()); 
-    CoarseComplexField iZProj(Grid()); 
-    CoarseComplexField oZProj(Grid()); 
-
-    CoarseScalar InnerProd(Grid()); 
-
-    // Orthogonalise the subblocks over the basis
-    blockOrthogonalise(InnerProd,Subspace.subspace);
-
-    // Compute the matrix elements of linop between this orthonormal
-    // set of vectors.
-    int self_stencil=-1;
-    for(int p=0;p<geom.npoint;p++)
-    { 
-      int dir   = geom.directions[p];
-      int disp  = geom.displacements[p];
-      A[p]=Zero();
-      if( geom.displacements[p]==0){
-	self_stencil=p;
-      }
-
-      Integer block=(FineGrid->_rdimensions[dir])/(Grid()->_rdimensions[dir]);
-
-      LatticeCoordinate(coor,dir);
-
-      ///////////////////////////////////////////////////////
-      // Work out even and odd block checkerboarding for fast diagonal term
-      ///////////////////////////////////////////////////////
-      if ( disp==1 ) {
-	bcb   = bcb + div(coor,block);
-      }
-	
-      if ( disp==0 ) {
-	  masks[p]= Zero();
-      } else if ( disp==1 ) {
-	masks[p] = where(mod(coor,block)==(block-1),one,zero);
-      } else if ( disp==-1 ) {
-	masks[p] = where(mod(coor,block)==(Integer)0,one,zero);
-      }
-    }
-    evenmask = where(mod(bcb,2)==(Integer)0,one,zero);
-    oddmask  = one-evenmask;
-
-    assert(self_stencil!=-1);
-
-    for(int i=0;i<nbasis;i++){
-
-      phi=Subspace.subspace[i];
-
-      //      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl;
-      linop.OpDirAll(phi,Mphi_p);
-      linop.OpDiag  (phi,Mphi_p[geom.npoint-1]);
-
-      for(int p=0;p<geom.npoint;p++){ 
-
-	Mphi = Mphi_p[p];
-
-	int dir   = geom.directions[p];
-	int disp  = geom.displacements[p];
-
-	if ( (disp==-1) || (!hermitian ) ) {
-
-	  ////////////////////////////////////////////////////////////////////////
-	  // Pick out contributions coming from this cell and neighbour cell
-	  ////////////////////////////////////////////////////////////////////////
-	  omask = masks[p];
-	  imask = one-omask;
-	
-	  for(int j=0;j<nbasis;j++){
-	    
-	    blockMaskedInnerProduct(oZProj,omask,Subspace.subspace[j],Mphi);
-	    
-	    auto iZProj_v = iZProj.View() ;
-	    auto oZProj_v = oZProj.View() ;
-	    auto A_p     =  A[p].View();
-	    auto A_self  = A[self_stencil].View();
-
-	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
-
-	  }
-	}
-      }
-
-      ///////////////////////////////////////////
-      // Faster alternate self coupling.. use hermiticity to save 2x
-      ///////////////////////////////////////////
-      {
-	mult(tmp,phi,evenmask);  linop.Op(tmp,Mphie);
-	mult(tmp,phi,oddmask );  linop.Op(tmp,Mphio);
-
-	{
-	  auto tmp_      = tmp.View();
-	  auto evenmask_ = evenmask.View();
-	  auto oddmask_  =  oddmask.View();
-	  auto Mphie_    =  Mphie.View();
-	  auto Mphio_    =  Mphio.View();
-	  accelerator_for(ss, FineGrid->oSites(), Fobj::Nsimd(),{ 
-	      coalescedWrite(tmp_[ss],evenmask_(ss)*Mphie_(ss) + oddmask_(ss)*Mphio_(ss));
-	    });
-	}
-
-	blockProject(SelfProj,tmp,Subspace.subspace);
-
-	auto SelfProj_ = SelfProj.View();
-	auto A_self  = A[self_stencil].View();
-
-	accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{
-	  for(int j=0;j<nbasis;j++){
-	    coalescedWrite(A_self[ss](j,i), SelfProj_(ss)(j));
-	  }
-	});
-
-      }
-    }
-    if(hermitian) {
-      std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
-      ForceHermitian();
-    }
-      // AssertHermitian();
-      // ForceDiagonal();
-  }
-
-#if 0
-    ///////////////////////////
-    // test code worth preserving in if block
-    ///////////////////////////
-    std::cout<<GridLogMessage<< " Computed matrix elements "<< self_stencil <<std::endl;
-    for(int p=0;p<geom.npoint;p++){
-      std::cout<<GridLogMessage<< "A["<<p<<"]" << std::endl;
-      std::cout<<GridLogMessage<< A[p] << std::endl;
-    }
-    std::cout<<GridLogMessage<< " picking by block0 "<< self_stencil <<std::endl;
-
-    phi=Subspace.subspace[0];
-    std::vector<int> bc(FineGrid->_ndimension,0);
-
-    blockPick(Grid(),phi,tmp,bc);      // Pick out a block
-    linop.Op(tmp,Mphi);                // Apply big dop
-    blockProject(iProj,Mphi,Subspace.subspace); // project it and print it
-    std::cout<<GridLogMessage<< " Computed matrix elements from block zero only "<<std::endl;
-    std::cout<<GridLogMessage<< iProj <<std::endl;
-    std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
-#endif
-
-
-  void ForceHermitian(void) {
-    CoarseMatrix Diff  (Grid());
-    for(int p=0;p<geom.npoint;p++){
-      int dir   = geom.directions[p];
-      int disp  = geom.displacements[p];
-      if(disp==-1) {
-	// Find the opposite link
-	for(int pp=0;pp<geom.npoint;pp++){
-	  int dirp   = geom.directions[pp];
-	  int dispp  = geom.displacements[pp];
-	  if ( (dirp==dir) && (dispp==1) ){
-	    //	    Diff = adj(Cshift(A[p],dir,1)) - A[pp]; 
-	    //	    std::cout << GridLogMessage<<" Replacing stencil leg "<<pp<<" with leg "<<p<< " diff "<<norm2(Diff) <<std::endl;
-	    A[pp] = adj(Cshift(A[p],dir,1));
-	  }
-	}
-      }
-    }
-  }
-  void AssertHermitian(void) {
-    CoarseMatrix AA    (Grid());
-    CoarseMatrix AAc   (Grid());
-    CoarseMatrix Diff  (Grid());
-    for(int d=0;d<4;d++){
-	
-      int dd=d+1;
-      AAc = Cshift(A[2*d+1],dd,1);
-      AA  = A[2*d];
-	
-      Diff = AA - adj(AAc);
-
-      std::cout<<GridLogMessage<<"Norm diff dim "<<d<<" "<< norm2(Diff)<<std::endl;
-      std::cout<<GridLogMessage<<"Norm dim "<<d<<" "<< norm2(AA)<<std::endl;
-	  
-    }
-    Diff = A[8] - adj(A[8]);
-    std::cout<<GridLogMessage<<"Norm diff local "<< norm2(Diff)<<std::endl;
-    std::cout<<GridLogMessage<<"Norm local "<< norm2(A[8])<<std::endl;
-  }
-    
-};
-
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/FFT.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -29,96 +28,206 @@ 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
-#ifdef USE_MKL
+#if defined(USE_MKL) || defined(GRID_SYCL)
 #include <fftw/fftw3.h>
 #else
 #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;
@@ -128,31 +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);
   };
     
   ~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;
@@ -162,59 +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
-    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);
+    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);
-    auto pgbuf_v = pgbuf.View();
-
+    //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 assert(0);
-      
+    else GRID_ASSERT(0);
+
+    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];
@@ -228,61 +342,154 @@ public:
     }
       
     // Barrel shift and collect global pencil
-    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++) {
-      thread_for(idx, sgrid->lSites(),{
-          Coordinate cbuf(Nd);
-          sobj s;
-	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,result,cbuf);
-	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  //            cbuf[dim]+=p*L;
-	  pokeLocalSite(s,pgbuf,cbuf);
+      t_copy-=usecond();
+      autoView(r_v,result,AcceleratorRead);
+      accelerator_for(idx, grid->oSites(), vobj::Nsimd(), {
+#ifdef GRID_SIMT
+      {
+	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();
       
-    // 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;
-      
-    // writing out result
-    thread_for(idx,sgrid->lSites(),{
-	Coordinate clbuf(Nd), cgbuf(Nd);
-	sobj s;
-	sgrid->LocalIndexToLocalCoor(idx,clbuf);
-	cgbuf = clbuf;
-	cgbuf[dim] = clbuf[dim]+L*pc;
-	peekLocalSite(s,pgbuf,cgbuf);
-	pokeLocalSite(s,result,clbuf);
-    });
+    flops_call = 5.0*howmany*G*log2(G);
+    usec = t_fft;
+    flops= flops_call;
+
+    result = Zero();
+    
+    double t_insert = -usecond();
+    {
+      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;
-      
+
+    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/LinearOperator.h

@@ -52,6 +52,7 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
+  virtual ~LinearOperatorBase(){};
 };
 
 
@@ -63,7 +64,7 @@ public:
 //
 // I'm not entirely happy with implementation; to share the Schur code between herm and non-herm
 // while still having a "OpAndNorm" in the abstract base I had to implement it in both cases
-// with an assert trap in the non-herm. This isn't right; there must be a better C++ way to
+// with an GRID_ASSERT trap in the non-herm. This isn't right; there must be a better C++ way to
 // do it, but I fear it required multiple inheritance and mixed in abstract base classes
 /////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -102,6 +103,38 @@ public:
     _Mat.MdagM(in,out);
   }
 };
+template<class Matrix,class Field>
+class MMdagLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+public:
+  MMdagLinearOperator(Matrix &Mat): _Mat(Mat){};
+
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    _Mat.Mdiag(in,out);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    _Mat.M(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    _Mat.Mdag(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    _Mat.MMdag(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+    _Mat.MMdag(in,out);
+  }
+};
 
 ////////////////////////////////////////////////////////////////////
 // Construct herm op and shift it for mgrid smoother
@@ -115,22 +148,22 @@ public:
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
     _Mat.Mdiag(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
     _Mat.Mdir(in,out,dir,disp);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
   void Op     (const Field &in, Field &out){
     _Mat.M(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void AdjOp     (const Field &in, Field &out){
     _Mat.Mdag(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
     HermOp(in,out);
@@ -144,6 +177,44 @@ public:
   }
 };
 
+////////////////////////////////////////////////////////////////////
+// Create a shifted HermOp
+////////////////////////////////////////////////////////////////////
+template<class Field>
+class ShiftedHermOpLinearOperator : public LinearOperatorBase<Field> {
+  LinearOperatorBase<Field> &_Mat;
+  RealD _shift;
+public:
+  ShiftedHermOpLinearOperator(LinearOperatorBase<Field> &Mat,RealD shift): _Mat(Mat), _shift(shift){};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    GRID_ASSERT(0);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    GRID_ASSERT(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    GRID_ASSERT(0);
+  };
+  void Op     (const Field &in, Field &out){
+    HermOp(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    HermOp(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+    _Mat.HermOp(in,out);
+    out = out + _shift*in;
+  }
+};
+
+
 ////////////////////////////////////////////////////////////////////
 // Wrap an already herm matrix
 ////////////////////////////////////////////////////////////////////
@@ -200,10 +271,42 @@ public:
     _Mat.Mdag(in,out);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    assert(0);
+    GRID_ASSERT(0);
   }
   void HermOp(const Field &in, Field &out){
-    assert(0);
+    GRID_ASSERT(0);
+  }
+};
+template<class Matrix,class Field>
+class ShiftedNonHermitianLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD shift;
+public:
+  ShiftedNonHermitianLinearOperator(Matrix &Mat,RealD shft): _Mat(Mat),shift(shft){};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    _Mat.Mdiag(in,out);
+    out = out + shift*in;
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    _Mat.M(in,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){
+    GRID_ASSERT(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    GRID_ASSERT(0);
   }
 };
 
@@ -242,13 +345,13 @@ class SchurOperatorBase :  public LinearOperatorBase<Field> {
   }
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
-    assert(0); // must coarsen the unpreconditioned system
+    GRID_ASSERT(0); // must coarsen the unpreconditioned system
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
 };
 template<class Matrix,class Field>
@@ -344,10 +447,10 @@ class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field>
     MpcDag(tmp,out);
   }
   virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   virtual void HermOp(const Field& in, Field& out) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void Op(const Field& in, Field& out) {
     Mpc(in, out);
@@ -357,13 +460,13 @@ class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field>
   }
   // Support for coarsening to a multigrid
   void OpDiag(const Field& in, Field& out) {
-    assert(0); // must coarsen the unpreconditioned system
+    GRID_ASSERT(0); // must coarsen the unpreconditioned system
   }
   void OpDir(const Field& in, Field& out, int dir, int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll(const Field& in, std::vector<Field>& out){
-    assert(0);
+    GRID_ASSERT(0);
   };
 };
 
@@ -477,7 +580,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
  public:
   SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
   { 
-    assert( _Mat.isTrivialEE() );
+    GRID_ASSERT( _Mat.isTrivialEE() );
     mass = _Mat.Mass();
   }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
@@ -507,8 +610,8 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   virtual  void MpcDag   (const Field &in, Field &out){
     Mpc(in,out);
   }
-  virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
-    assert(0);// Never need with staggered
+  virtual void MpcDagMpc(const Field &in, Field &out) {
+    GRID_ASSERT(0);// Never need with staggered
   }
 };
 template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
@@ -520,16 +623,28 @@ template<class Field> class OperatorFunction {
 public:
   virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
   virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
-    assert(in.size()==out.size());
+    GRID_ASSERT(in.size()==out.size());
     for(int k=0;k<in.size();k++){
       (*this)(Linop,in[k],out[k]);
     }
   };
+  virtual ~OperatorFunction(){};
 };
 
 template<class Field> class LinearFunction {
 public:
   virtual void operator() (const Field &in, Field &out) = 0;
+
+  virtual void operator() (const std::vector<Field> &in, std::vector<Field> &out)
+  {
+    GRID_ASSERT(in.size() == out.size());
+
+    for (unsigned int i = 0; i < in.size(); ++i)
+    {
+      (*this)(in[i], out[i]);
+    }
+  }
+  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
@@ -575,6 +690,7 @@ class HermOpOperatorFunction : public OperatorFunction<Field> {
 template<typename Field>
 class PlainHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   LinearOperatorBase<Field> &_Linop;
       
   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
@@ -588,6 +704,7 @@ public:
 template<typename Field>
 class FunctionHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator(); 
   OperatorFunction<Field>   & _poly;
   LinearOperatorBase<Field> &_Linop;
       

Grid/algorithms/Preconditioner.h

@@ -30,13 +30,19 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Field> class Preconditioner :  public LinearFunction<Field> { 
+template<class Field> using Preconditioner =  LinearFunction<Field> ;
+
+/*
+template<class Field> class Preconditioner :  public LinearFunction<Field> {
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field & psi)=0;
 };
+*/
 
 template<class Field> class TrivialPrecon :  public Preconditioner<Field> { 
 public:
-  void operator()(const Field &src, Field & psi){
+  using Preconditioner<Field>::operator();
+  virtual void operator()(const Field &src, Field & psi){
     psi = src;
   }
   TrivialPrecon(void){};

Grid/algorithms/SparseMatrix.h

@@ -45,9 +45,15 @@ public:
     M(in,tmp);
     Mdag(tmp,out);
   }
+  virtual void  MMdag(const Field &in, Field &out) {
+    Field tmp (in.Grid());
+    Mdag(in,tmp);
+    M(tmp,out);
+  }
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;
+  virtual ~SparseMatrixBase() {};
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -72,7 +78,7 @@ public:
   virtual  void MeooeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeInvDag (const Field &in, Field &out)=0;
-
+  virtual ~CheckerBoardedSparseMatrixBase() {};
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/approx/Chebyshev.h

@@ -59,7 +59,7 @@ public:
     RealD diff = hi-lo;
     RealD delta = diff*1.0e-9;
     for (RealD x=lo; x<hi; x+=delta) {
-      delta*=1.1;
+      delta*=1.02;
       RealD f = approx(x);
       out<< x<<" "<<f<<std::endl;
     }
@@ -90,9 +90,8 @@ public:
     order=_order;
       
     if(order < 2) exit(-1);
-    Coeffs.resize(order);
-    Coeffs.assign(0.,order);
-    Coeffs[order-1] = 1.;
+    Coeffs.resize(order,0.0);
+    Coeffs[order-1] = 1.0;
   };
   
   // PB - more efficient low pass drops high modes above the low as 1/x uses all Chebyshev's.
@@ -132,6 +131,26 @@ public:
       Coeffs[j] = s * 2.0/order;
     }
   };
+  template<class functor>
+  void Init(RealD _lo,RealD _hi,int _order, functor & func)
+  {
+    lo=_lo;
+    hi=_hi;
+    order=_order;
+      
+    if(order < 2) exit(-1);
+    Coeffs.resize(order);
+    for(int j=0;j<order;j++){
+      RealD s=0;
+      for(int k=0;k<order;k++){
+	RealD y=std::cos(M_PI*(k+0.5)/order);
+	RealD x=0.5*(y*(hi-lo)+(hi+lo));
+	RealD f=func(x);
+	s=s+f*std::cos( j*M_PI*(k+0.5)/order );
+      }
+      Coeffs[j] = s * 2.0/order;
+    }
+  };
 
     
   void JacksonSmooth(void){
@@ -250,7 +269,9 @@ public:
     RealD xscale = 2.0/(hi-lo);
     RealD mscale = -(hi+lo)/(hi-lo);
     Linop.HermOp(T0,y);
+    grid->Barrier();
     axpby(T1,xscale,mscale,y,in);
+    grid->Barrier();
 
     // sum = .5 c[0] T0 + c[1] T1
     //    out = ()*T0 + Coeffs[1]*T1;
@@ -258,26 +279,12 @@ public:
     for(int n=2;n<order;n++){
 
       Linop.HermOp(*Tn,y);
-#if 0
-      auto y_v = y.View();
-      auto Tn_v = Tn->View();
-      auto Tnp_v = Tnp->View();
-      auto Tnm_v = Tnm->View();
-      constexpr int Nsimd = vector_type::Nsimd();
-      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
-	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
-	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
-      });
-      if ( Coeffs[n] != 0.0) {
-	axpy(out,Coeffs[n],*Tnp,out);
-      }
-#else
       axpby(y,xscale,mscale,y,(*Tn));
       axpby(*Tnp,2.0,-1.0,y,(*Tnm));
       if ( Coeffs[n] != 0.0) {
 	axpy(out,Coeffs[n],*Tnp,out);
       }
-#endif
+
       // Cycle pointers to avoid copies
       Field *swizzle = Tnm;
       Tnm    =Tn;

Grid/algorithms/approx/MultiShiftFunction.h

@@ -40,7 +40,7 @@ public:
   RealD norm;
   RealD lo,hi;
 
-  MultiShiftFunction(int n,RealD _lo,RealD _hi): poles(n), residues(n), lo(_lo), hi(_hi) {;};
+  MultiShiftFunction(int n,RealD _lo,RealD _hi): poles(n), residues(n), tolerances(n), lo(_lo), hi(_hi) {;};
   RealD approx(RealD x);
   void csv(std::ostream &out);
   void gnuplot(std::ostream &out);

Grid/algorithms/approx/Remez.cc

@@ -121,7 +121,7 @@ double AlgRemez::generateApprox(int num_degree, int den_degree,
   // Reallocate arrays, since degree has changed
   if (num_degree != n || den_degree != d) allocate(num_degree,den_degree);
 
-  assert(a_len<=SUM_MAX);
+  GRID_ASSERT(a_len<=SUM_MAX);
 
   step = new bigfloat[num_degree+den_degree+2];
 
@@ -151,9 +151,9 @@ double AlgRemez::generateApprox(int num_degree, int den_degree,
     equations();
     if (delta < tolerance) {
       std::cout<<"Delta too small, try increasing precision\n";
-      assert(0);
+      GRID_ASSERT(0);
     };    
-    assert( delta>= tolerance);
+    GRID_ASSERT( delta>= tolerance);
 
     search(step);
   }

Grid/algorithms/approx/Remez.h

@@ -134,7 +134,7 @@ class AlgRemez
   virtual ~AlgRemez();
 
   int getDegree(void){ 
-    assert(n==d);
+    GRID_ASSERT(n==d);
     return n;
   }
   // Reset the bounds of the approximation

Grid/algorithms/approx/RemezGeneral.cc

@@ -28,11 +28,11 @@ void AlgRemezGeneral::setupPolyProperties(int num_degree, int den_degree, PolyTy
   pow_n = num_degree;
   pow_d = den_degree;
 
-  if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) assert(0);
-  if(pow_n % 2 == 1 && num_type_in == PolyType::Even) assert(0);
+  if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) GRID_ASSERT(0);
+  if(pow_n % 2 == 1 && num_type_in == PolyType::Even) GRID_ASSERT(0);
 
-  if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) assert(0);
-  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) assert(0);
+  if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) GRID_ASSERT(0);
+  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) GRID_ASSERT(0);
 
   num_type = num_type_in;
   den_type = den_type_in;
@@ -112,9 +112,9 @@ double AlgRemezGeneral::generateApprox(const int num_degree, const int den_degre
     equations();
     if (delta < tolerance) {
       std::cout<<"Iteration " << iter-1 << " delta too small (" << delta << "<" << tolerance << "), try increasing precision\n";
-      assert(0);
+      GRID_ASSERT(0);
     };    
-    assert( delta>= tolerance );
+    GRID_ASSERT( delta>= tolerance );
 
     search();
   }
@@ -278,7 +278,7 @@ void AlgRemezGeneral::equations(){
       if(num_pows[j] != -1){ *aa++ = z; t++; }
       z *= x;
     }
-    assert(t == n+1);
+    GRID_ASSERT(t == n+1);
 
     z = (bigfloat)1l;
     t = 0;
@@ -286,7 +286,7 @@ void AlgRemezGeneral::equations(){
       if(den_pows[j] != -1){ *aa++ = -y * z; t++; }
       z *= x;
     }
-    assert(t == d);
+    GRID_ASSERT(t == d);
 
     B[i] = y * z;		// Right hand side vector
   }

Grid/algorithms/approx/RemezGeneral.h

@@ -106,7 +106,7 @@ class AlgRemezGeneral{
 		  bigfloat (*f)(bigfloat x, void *data), void *data);
 
   inline int getDegree(void) const{ 
-    assert(n==d);
+    GRID_ASSERT(n==d);
     return n;
   }
   // Reset the bounds of the approximation

Grid/algorithms/approx/ZMobius.cc

@@ -74,7 +74,7 @@ bigfloat epsilonMobius(bigfloat x, void* data){
 void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out,
 			 const std::vector<RealD> &omega_in, const int Ls_in,
 			 const RealD lambda_bound){
-  assert(omega_in.size() == Ls_in);
+  GRID_ASSERT(omega_in.size() == Ls_in);
   omega_out.resize(Ls_out);
 
   //Use the Remez algorithm to generate the appropriate rational polynomial

Grid/algorithms/approx/Zolotarev.cc

@@ -293,7 +293,7 @@ static void sncndnFK(INTERNAL_PRECISION u, INTERNAL_PRECISION k,
  * Set type = 0 for the Zolotarev approximation, which is zero at x = 0, and
  * type = 1 for the approximation which is infinite at x = 0. */
 
-zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) {
+zolotarev_data* zolotarev(ZOLO_PRECISION epsilon, int n, int type) {
   INTERNAL_PRECISION A, c, cp, kp, ksq, sn, cn, dn, Kp, Kj, z, z0, t, M, F,
     l, invlambda, xi, xisq, *tv, s, opl;
   int m, czero, ts;
@@ -375,12 +375,12 @@ zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) {
   construct_partfrac(d);
   construct_contfrac(d);
 
-  /* Converting everything to PRECISION for external use only */
+  /* Converting everything to ZOLO_PRECISION for external use only */
 
   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data));
-  zd -> A = (PRECISION) d -> A;
-  zd -> Delta = (PRECISION) d -> Delta;
-  zd -> epsilon = (PRECISION) d -> epsilon;
+  zd -> A = (ZOLO_PRECISION) d -> A;
+  zd -> Delta = (ZOLO_PRECISION) d -> Delta;
+  zd -> epsilon = (ZOLO_PRECISION) d -> epsilon;
   zd -> n = d -> n;
   zd -> type = d -> type;
   zd -> dn = d -> dn;
@@ -390,24 +390,24 @@ zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) {
   zd -> deg_num = d -> deg_num;
   zd -> deg_denom = d -> deg_denom;
 
-  zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION));
-  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m];
+  zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m];
   free(d -> a);
 
-  zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION));
-  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m];
+  zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m];
   free(d -> ap);
 
-  zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION));
-  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m];
+  zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m];
   free(d -> alpha);
 
-  zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION));
-  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m];
+  zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m];
   free(d -> beta);
 
-  zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION));
-  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m];
+  zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m];
   free(d -> gamma);
 
   free(d);
@@ -426,7 +426,7 @@ void zolotarev_free(zolotarev_data *zdata)
 }
 
 
-zolotarev_data* higham(PRECISION epsilon, int n) {
+zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) {
   INTERNAL_PRECISION A, M, c, cp, z, z0, t, epssq;
   int m, czero;
   zolotarev_data *zd;
@@ -481,9 +481,9 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
   /* Converting everything to PRECISION for external use only */
 
   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data));
-  zd -> A = (PRECISION) d -> A;
-  zd -> Delta = (PRECISION) d -> Delta;
-  zd -> epsilon = (PRECISION) d -> epsilon;
+  zd -> A = (ZOLO_PRECISION) d -> A;
+  zd -> Delta = (ZOLO_PRECISION) d -> Delta;
+  zd -> epsilon = (ZOLO_PRECISION) d -> epsilon;
   zd -> n = d -> n;
   zd -> type = d -> type;
   zd -> dn = d -> dn;
@@ -493,24 +493,24 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
   zd -> deg_num = d -> deg_num;
   zd -> deg_denom = d -> deg_denom;
 
-  zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION));
-  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m];
+  zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m];
   free(d -> a);
 
-  zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION));
-  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m];
+  zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m];
   free(d -> ap);
 
-  zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION));
-  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m];
+  zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m];
   free(d -> alpha);
 
-  zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION));
-  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m];
+  zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m];
   free(d -> beta);
 
-  zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION));
-  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m];
+  zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m];
   free(d -> gamma);
 
   free(d);
@@ -523,17 +523,17 @@ NAMESPACE_END(Grid);
 #ifdef TEST
 
 #undef ZERO
-#define ZERO ((PRECISION) 0)
+#define ZERO ((ZOLO_PRECISION) 0)
 #undef ONE
-#define ONE ((PRECISION) 1)
+#define ONE ((ZOLO_PRECISION) 1)
 #undef TWO
-#define TWO ((PRECISION) 2)
+#define TWO ((ZOLO_PRECISION) 2)
 
 /* Evaluate the rational approximation R(x) using the factored form */
 
-static PRECISION zolotarev_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION R;
+  ZOLO_PRECISION R;
 
   if (rdata -> type == 0) {
     R = rdata -> A * x;
@@ -551,9 +551,9 @@ static PRECISION zolotarev_eval(PRECISION x, zolotarev_data* rdata) {
 
 /* Evaluate the rational approximation R(x) using the partial fraction form */
 
-static PRECISION zolotarev_partfrac_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_partfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION R = rdata -> alpha[rdata -> da - 1];
+  ZOLO_PRECISION R = rdata -> alpha[rdata -> da - 1];
   for (m = 0; m < rdata -> dd; m++)
     R += rdata -> alpha[m] / (x * x - rdata -> ap[m]);
   if (rdata -> type == 1) R += rdata -> alpha[rdata -> dd] / (x * x);
@@ -568,18 +568,18 @@ static PRECISION zolotarev_partfrac_eval(PRECISION x, zolotarev_data* rdata) {
  * non-signalling overflow this will work correctly since 1/(1/0) = 1/INF = 0,
  * but with signalling overflow you will get an error message. */
 
-static PRECISION zolotarev_contfrac_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_contfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION R = rdata -> beta[0] * x;
+  ZOLO_PRECISION R = rdata -> beta[0] * x;
   for (m = 1; m < rdata -> db; m++) R = rdata -> beta[m] * x + ONE / R;
   return R;
 }    
 
 /* Evaluate the rational approximation R(x) using Cayley form */
 
-static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_cayley_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION T;
+  ZOLO_PRECISION T;
 
   T = rdata -> type == 0 ? ONE : -ONE;
   for (m = 0; m < rdata -> n; m++)
@@ -607,7 +607,7 @@ int main(int argc, char** argv) {
   int m, n, plotpts = 5000, type = 0;
   float eps, x, ypferr, ycferr, ycaylerr, maxypferr, maxycferr, maxycaylerr;
   zolotarev_data *rdata;
-  PRECISION y;
+  ZOLO_PRECISION y;
   FILE *plot_function, *plot_error, 
     *plot_partfrac, *plot_contfrac, *plot_cayley;
 
@@ -626,13 +626,13 @@ int main(int argc, char** argv) {
   }
 
   rdata = type == 2 
-    ? higham((PRECISION) eps, n) 
-    : zolotarev((PRECISION) eps, n, type);
+    ? higham((ZOLO_PRECISION) eps, n) 
+    : zolotarev((ZOLO_PRECISION) eps, n, type);
 
   printf("Zolotarev Test: R(epsilon = %g, n = %d, type = %d)\n\t" 
 	 STRINGIFY(VERSION) "\n\t" STRINGIFY(HVERSION)
 	 "\n\tINTERNAL_PRECISION = " STRINGIFY(INTERNAL_PRECISION)
-	 "\tPRECISION = " STRINGIFY(PRECISION)
+	 "\tZOLO_PRECISION = " STRINGIFY(ZOLO_PRECISION)
 	 "\n\n\tRational approximation of degree (%d,%d), %s at x = 0\n"
 	 "\tDelta = %g (maximum error)\n\n"
 	 "\tA = %g (overall factor)\n",
@@ -681,15 +681,15 @@ int main(int argc, char** argv) {
     x = 2.4 * (float) m / plotpts - 1.2;
     if (rdata -> type == 0 || fabs(x) * (float) plotpts > 1.0) {
       /* skip x = 0 for type 1, as R(0) is singular */
-      y = zolotarev_eval((PRECISION) x, rdata);
+      y = zolotarev_eval((ZOLO_PRECISION) x, rdata);
       fprintf(plot_function, "%g %g\n", x, (float) y);
       fprintf(plot_error, "%g %g\n",
 	      x, (float)((y - ((x > 0.0 ? ONE : -ONE))) / rdata -> Delta));
-      ypferr = (float)((zolotarev_partfrac_eval((PRECISION) x, rdata) - y)
+      ypferr = (float)((zolotarev_partfrac_eval((ZOLO_PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
-      ycferr = (float)((zolotarev_contfrac_eval((PRECISION) x, rdata) - y)
+      ycferr = (float)((zolotarev_contfrac_eval((ZOLO_PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
-      ycaylerr = (float)((zolotarev_cayley_eval((PRECISION) x, rdata) - y)
+      ycaylerr = (float)((zolotarev_cayley_eval((ZOLO_PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
       if (fabs(x) < 1.0 && fabs(x) > rdata -> epsilon) {
 	maxypferr = MAX(maxypferr, fabs(ypferr));

Grid/algorithms/approx/Zolotarev.h

@@ -9,10 +9,10 @@ NAMESPACE_BEGIN(Approx);
 #define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY>
 
 #ifndef ZOLOTAREV_INTERNAL
-#ifndef PRECISION
-#define PRECISION double
+#ifndef ZOLO_PRECISION
+#define ZOLO_PRECISION double
 #endif
-#define ZPRECISION PRECISION
+#define ZPRECISION ZOLO_PRECISION
 #define ZOLOTAREV_DATA zolotarev_data
 #endif
 
@@ -77,8 +77,8 @@ typedef struct {
  * zolotarev_data structure. The arguments must satisfy the constraints that
  * epsilon > 0, n > 0, and type = 0 or 1. */
 
-ZOLOTAREV_DATA* higham(PRECISION epsilon, int n) ;
-ZOLOTAREV_DATA* zolotarev(PRECISION epsilon, int n, int type);
+ZOLOTAREV_DATA* higham(ZOLO_PRECISION epsilon, int n) ;
+ZOLOTAREV_DATA* zolotarev(ZOLO_PRECISION epsilon, int n, int type);
 void zolotarev_free(zolotarev_data *zdata);
 #endif
 
@@ -86,3 +86,4 @@ void zolotarev_free(zolotarev_data *zdata);
 NAMESPACE_END(Approx);
 NAMESPACE_END(Grid);
 #endif
+

Grid/algorithms/blas/BatchedBlas.cc

@@ -0,0 +1,34 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@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 */
+#include <Grid/GridCore.h>
+#include <Grid/algorithms/blas/BatchedBlas.h>
+NAMESPACE_BEGIN(Grid);
+gridblasHandle_t GridBLAS::gridblasHandle;
+int              GridBLAS::gridblasInit;
+NAMESPACE_END(Grid);
+

Grid/algorithms/blas/MomentumProject.h

@@ -0,0 +1,300 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MomentumProject.h
+
+    Copyright (C) 2025
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+/* 
+   MultiMomProject
+
+   Import vectors -> nxyz x (ncomponent x nt)
+   Import complex phases -> nmom x nxy
+
+   apply = via (possibly batched) GEMM
+*/
+template<class Field, class ComplexField>
+class MomentumProject
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+
+  GridBase *grid;
+  uint64_t nmom;
+  uint64_t nxyz;
+  uint64_t nt;
+  uint64_t nbtw;
+  uint64_t words;
+
+  deviceVector<scalar> BLAS_V;      // 
+  deviceVector<scalar> BLAS_M;      // 
+  deviceVector<scalar> BLAS_P;      // 
+  
+  MomentumProject(){};
+ ~MomentumProject(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    grid=nullptr;
+    nmom=0;
+    nxyz=0;
+    nt=0;
+    nbtw=0;
+    words=0;
+    BLAS_V.resize(0);
+    BLAS_M.resize(0);
+    BLAS_P.resize(0);
+  }
+  void Allocate(int _nmom,GridBase *_grid)
+  {
+    grid=_grid;
+    Coordinate ldims = grid->LocalDimensions();
+
+    nmom=_nmom;
+    nt   = ldims[grid->Nd()-1];
+    nxyz = grid->lSites()/nt;
+    words = sizeof(scalar_object)/sizeof(scalar);
+    nbtw = nt * words;
+
+    BLAS_V.resize (nxyz * nt * words );
+    BLAS_M.resize (nmom * nxyz       );
+    BLAS_P.resize (nmom * nt * words );
+  }
+  void ImportMomenta(const std::vector <ComplexField> &momenta)
+  {
+    GRID_ASSERT(momenta.size()==nmom);
+    //    might as well just make the momenta here
+    typedef typename Field::vector_object vobj;
+
+    int nd = grid->_ndimension;
+
+    uint64_t sz = BLAS_M.size();
+
+    GRID_ASSERT(momenta.size()==nmom)
+    GRID_ASSERT(momenta[0].Grid()==grid);
+    GRID_ASSERT(sz = nxyz * nmom);
+    
+    Coordinate rdimensions = grid->_rdimensions;
+    Coordinate ldims       = grid->LocalDimensions();
+    int64_t osites         = grid->oSites();
+    Coordinate simd        = grid->_simd_layout;
+    const int Nsimd        = vobj::Nsimd();
+    uint64_t lwords        = words; // local variable for copy in to GPU
+    int64_t Nxyz = nxyz;
+    auto blasData_p  = &BLAS_M[0];
+    for(int m=0;m<momenta.size();m++){
+
+      autoView( Data   , momenta[m], AcceleratorRead);
+      auto Data_p  = &Data[0];
+
+      accelerator_for(xyz,nxyz,1,{
+	  //////////////////////////////////////////
+	  // isite -- map lane within buffer to lane within lattice
+	  ////////////////////////////////////////////
+	    Coordinate lcoor(nd,0);
+	    Lexicographic::CoorFromIndex(lcoor,xyz,ldims);
+	    
+	    Coordinate icoor(nd);
+	    Coordinate ocoor(nd);
+	    for (int d = 0; d < nd; d++) {
+	      icoor[d] = lcoor[d]/rdimensions[d];
+	      ocoor[d] = lcoor[d]%rdimensions[d];
+	    }
+	    int64_t osite;
+	    int64_t isite;
+	    Lexicographic::IndexFromCoor(ocoor,osite,rdimensions);
+	    Lexicographic::IndexFromCoor(icoor,isite,simd);
+	    
+	    // BLAS_M[nmom][slice_vol]
+	    // Fortran Column major BLAS layout is M_xyz,mom
+	    scalar data = extractLane(isite,Data[osite]);
+	    uint64_t idx = xyz+m*Nxyz;
+	    blasData_p[idx] = data;
+	});
+    }
+  }
+  void ImportVector(Field &vec)
+  {
+    typedef typename Field::vector_object vobj;
+
+    int nd = grid->_ndimension;
+
+    uint64_t sz = BLAS_V.size();
+
+    GRID_ASSERT(sz = nxyz * words * nt);
+    
+    Coordinate rdimensions = grid->_rdimensions;
+    Coordinate ldims= grid->LocalDimensions();
+    int64_t osites = grid->oSites();
+    Coordinate simd = grid->_simd_layout;
+    const int Nsimd = vobj::Nsimd();
+    uint64_t lwords= words; // local variable for copy in to GPU
+
+    auto blasData_p  = &BLAS_V[0];
+    autoView( Data   , vec, AcceleratorRead);
+    auto Data_p  = &Data[0];
+
+    int64_t nwords = words;// for capture
+    int64_t Nt     = nt;// for capture
+    
+    accelerator_for(sf,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  //////////////////////////////////////////
+	  // isite -- map lane within buffer to lane within lattice
+	  ////////////////////////////////////////////
+	    Coordinate lcoor(nd,0);
+	    Coordinate icoor(nd);
+	    Coordinate ocoor(nd);
+	    
+	    Lexicographic::CoorFromIndex(icoor,lane,simd);
+	    Lexicographic::CoorFromIndex(ocoor,sf,rdimensions);
+
+	  
+	    int64_t l_xyz = 0;
+	    for (int d = 0; d < nd; d++) {
+	      lcoor[d] = rdimensions[d]*icoor[d] + ocoor[d];
+	    }
+	    uint64_t l_t   = lcoor[nd-1];
+
+	    Coordinate xyz_coor = lcoor;
+	    xyz_coor[nd-1] =0;
+	    Lexicographic::IndexFromCoor(xyz_coor,l_xyz,ldims);
+
+	    
+	    scalar_object data = extractLane(lane,Data[sf]);
+	    scalar *data_words = (scalar *) &data;
+	    for(int w = 0 ; w < nwords; w++) {
+	      // BLAS_V[slice_vol][nt][words]
+	      // Fortran Column major BLAS layout is V_(t,w)_xyz
+	      uint64_t idx = w+l_t*nwords + l_xyz * nwords * Nt;
+	      blasData_p[idx] = data_words[w];
+	    }
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+	});
+  }
+  void ExportMomentumProjection(std::vector<typename Field::scalar_object> &projection)
+  {
+    projection.resize(nmom*nt);
+    acceleratorCopyFromDevice(&BLAS_P[0],(scalar *)&projection[0],BLAS_P.size()*sizeof(scalar));
+    // Could decide on a layout late?
+  }
+
+  // Row major layout "C" order:
+  // BLAS_V[slice_vol][nt][words]
+  // BLAS_M[nmom][slice_vol]
+  // BLAS_P[nmom][nt][words]
+  //
+  // Fortran Column major BLAS layout is V_(w,t)_xyz
+  // Fortran Column major BLAS layout is M_xyz,mom
+  // Fortran Column major BLAS layout is P_(w,t),mom
+  //
+  // Projected
+  //
+  // P = (V * M)_(w,t),mom
+  //
+  void Project(Field &data,std::vector< typename Field::scalar_object > & projected_gdata)
+  {
+    double t_import=0;
+    double t_export=0;
+    double t_gemm  =0;
+    double t_allreduce=0;
+    t_import-=usecond();
+    this->ImportVector(data);
+
+    std::vector< typename Field::scalar_object > projected_planes;
+
+    deviceVector<scalar *> Vd(1);
+    deviceVector<scalar *> Md(1);
+    deviceVector<scalar *> Pd(1);
+
+    scalar * Vh = & BLAS_V[0];
+    scalar * Mh = & BLAS_M[0];
+    scalar * Ph = & BLAS_P[0];
+
+    acceleratorPut(Vd[0],Vh);
+    acceleratorPut(Md[0],Mh);
+    acceleratorPut(Pd[0],Ph);
+    t_import+=usecond();
+
+    GridBLAS BLAS;
+
+    /////////////////////////////////////////
+    // P_im = VMmx . Vxi
+    /////////////////////////////////////////
+    t_gemm-=usecond();
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     words*nt,nmom,nxyz,
+		     scalar(1.0),
+		     Vd,
+		     Md,
+		     scalar(0.0),  // wipe out result
+		     Pd);
+    BLAS.synchronise();
+    t_gemm+=usecond();
+
+    t_export-=usecond();
+    ExportMomentumProjection(projected_planes); // resizes
+    t_export+=usecond();
+
+    /////////////////////////////////
+    // Reduce across MPI ranks
+    /////////////////////////////////
+    int nd = grid->Nd();
+    int gt = grid->GlobalDimensions()[nd-1];
+    int lt = grid->LocalDimensions()[nd-1];
+    projected_gdata.resize(gt*nmom);
+    for(int t=0;t<gt*nmom;t++){ // global Nt array with zeroes for stuff not on this node
+      projected_gdata[t]=Zero();
+    }
+    for(int t=0;t<lt;t++){
+    for(int m=0;m<nmom;m++){
+      int st = grid->LocalStarts()[nd-1];
+      projected_gdata[t+st + gt*m] = projected_planes[t+lt*m];
+    }}
+    t_allreduce-=usecond();
+    grid->GlobalSumVector((scalar *)&projected_gdata[0],gt*nmom*words);
+    t_allreduce+=usecond();
+
+    std::cout << GridLogPerformance<<" MomentumProject t_import  "<<t_import<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_export  "<<t_export<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_gemm    "<<t_gemm<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_reduce  "<<t_allreduce<<"us"<<std::endl;
+
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/Deflation.h

@@ -33,16 +33,19 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
 };
 template<class Field>
 class DoNothingGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) {  };
 };
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) { guess = src; };
 };
 
@@ -54,15 +57,24 @@ class DeflatedGuesser: public LinearFunction<Field> {
 private:
   const std::vector<Field> &evec;
   const std::vector<RealD> &eval;
+  const unsigned int       N;
 
 public:
+  using LinearFunction<Field>::operator();
 
-  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
+  DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
+  : DeflatedGuesser(_evec, _eval, _evec.size())
+  {}
+
+  DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N)
+  : evec(_evec), eval(_eval), N(_N)
+  {
+    GRID_ASSERT(evec.size()==eval.size());
+    GRID_ASSERT(N <= evec.size());
+  } 
 
   virtual void operator()(const Field &src,Field &guess) {
     guess = Zero();
-    assert(evec.size()==eval.size());
-    auto N = evec.size();
     for (int i=0;i<N;i++) {
       const Field& tmp = evec[i];
       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
@@ -79,6 +91,7 @@ private:
   const std::vector<RealD>       &eval_coarse;
 public:
   
+  using LinearFunction<FineField>::operator();
   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
 				const std::vector<CoarseField> &_evec_coarse,
 				const std::vector<RealD>       &_eval_coarse)
@@ -100,7 +113,42 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
-};
+
+  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
+    int Nevec = (int)evec_coarse.size();
+    int Nsrc = (int)src.size();
+    // make temp variables
+    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
+    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
+    //Preporcessing
+    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    guess_coarse[j] = Zero();
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockProject(src_coarse[j],src[j],subspace);
+    }
+    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
+    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
+    for (int i=0;i<Nevec;i++)
+    {
+      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
+      const CoarseField & tmp = evec_coarse[i];
+      for (int j=0;j<Nsrc;j++)
+      {
+        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
+      }
+    }
+    //postprocessing
+    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++) {
+      std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+      blockPromote(guess_coarse[j],guess[j],subspace);
+      guess[j].Checkerboard() = src[j].Checkerboard();
+    }
+  };
+
+  };
 
 
 

Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h

@@ -0,0 +1,376 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MultiRHSBlockCGLinalg.h
+
+    Copyright (C) 2024
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/* Need helper object for BLAS accelerated mrhs blockCG */
+template<class Field>
+class MultiRHSBlockCGLinalg
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+  typedef typename Field::vector_object vector_object;
+
+  deviceVector<scalar> BLAS_X;      // nrhs x vol -- the sources
+  deviceVector<scalar> BLAS_Y;      // nrhs x vol -- the result
+  deviceVector<scalar> BLAS_C;      // nrhs x nrhs -- the coefficients 
+  deviceVector<scalar> BLAS_Cred;   // nrhs x nrhs x oSites -- reduction buffer
+  deviceVector<scalar *> Xdip;
+  deviceVector<scalar *> Ydip;
+  deviceVector<scalar *> Cdip;
+  
+  MultiRHSBlockCGLinalg() {};
+  ~MultiRHSBlockCGLinalg(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    Xdip.resize(0);
+    Ydip.resize(0);
+    Cdip.resize(0);
+    BLAS_Cred.resize(0);
+    BLAS_C.resize(0);
+    BLAS_X.resize(0);
+    BLAS_Y.resize(0);
+  }
+  void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0)
+  {
+    std::vector<Field> Y_copy(AP.size(),AP[0].Grid());
+    for(int r=0;r<AP.size();r++){
+      Y_copy[r] = Y[r];
+    }
+    MulMatrix(AP,m,X);
+    for(int r=0;r<AP.size();r++){
+      AP[r] = scale*AP[r]+Y_copy[r];
+    }
+  }
+  void MulMatrix(std::vector<Field> &Y, Eigen::MatrixXcd &m , const std::vector<Field> &X)
+  {
+    typedef typename Field::scalar_type scomplex;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    int nrhs = Y.size();
+    grid  = X[0].Grid();
+    vol   = grid->lSites();
+    words = sizeof(scalar_object)/sizeof(scalar);
+    int64_t vw = vol * words;
+
+    RealD t0 = usecond();
+    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_C.resize(nrhs * nrhs);// cost free if size doesn't change
+    RealD t1 = usecond();
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources
+    /////////////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(x_v,X[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&x_v[0],&BLAS_X[offset],sizeof(scalar_object)*vol);
+    }
+
+    // Assumes Eigen storage contiguous
+    acceleratorCopyToDevice(&m(0,0),&BLAS_C[0],BLAS_C.size()*sizeof(scalar));
+    
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Xxr = [X1(x)][..][Xn(x)]
+   * Yxr = [Y1(x)][..][Ym(x)]
+   * Y = X . C
+   */
+    deviceVector<scalar *> Xd(1);
+    deviceVector<scalar *> Yd(1);
+    deviceVector<scalar *> Cd(1);
+
+    scalar * Xh = & BLAS_X[0];
+    scalar * Yh = & BLAS_Y[0];
+    scalar * Ch = & BLAS_C[0];
+
+    acceleratorPut(Xd[0],Xh);
+    acceleratorPut(Yd[0],Yh);
+    acceleratorPut(Cd[0],Ch);
+
+    RealD t2 = usecond();
+    GridBLAS BLAS;
+    /////////////////////////////////////////
+    // Y = X*C (transpose?)
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     vw,nrhs,nrhs,
+		     scalar(1.0),
+		     Xd,
+		     Cd,
+		     scalar(0.0),  // wipe out Y
+		     Yd);
+    BLAS.synchronise();
+    RealD t3 = usecond();
+
+    // Copy back Y = m X 
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(y_v,Y[r],AcceleratorWrite);
+      acceleratorCopyDeviceToDevice(&BLAS_Y[offset],&y_v[0],sizeof(scalar_object)*vol);
+    }    
+    RealD t4 = usecond();
+    std::cout <<GridLogPerformance << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix total "<< t4-t0<<" us"<<std::endl;
+  }
+  
+  void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y)
+  {
+#if 0    
+    int nrhs;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    nrhs = X.size();
+    GRID_ASSERT(X.size()==Y.size());
+    conformable(X[0],Y[0]);
+
+    grid  = X[0].Grid();
+    vol   = grid->lSites();
+    words = sizeof(scalar_object)/sizeof(scalar);
+    int64_t vw = vol * words;
+
+    RealD t0 = usecond();
+    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_C.resize(nrhs * nrhs);// cost free if size doesn't change
+    RealD t1 = usecond();
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources
+    /////////////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(x_v,X[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&x_v[0],&BLAS_X[offset],sizeof(scalar_object)*vol);
+      autoView(y_v,Y[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&y_v[0],&BLAS_Y[offset],sizeof(scalar_object)*vol);
+    }
+    RealD t2 = usecond();
+
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Xxr = [X1(x)][..][Xn(x)]
+   *
+   * Yxr = [Y1(x)][..][Ym(x)]
+   *
+   * C_rs = X^dag Y
+   */
+    deviceVector<scalar *> Xd(1);
+    deviceVector<scalar *> Yd(1);
+    deviceVector<scalar *> Cd(1);
+
+    scalar * Xh = & BLAS_X[0];
+    scalar * Yh = & BLAS_Y[0];
+    scalar * Ch = & BLAS_C[0];
+
+    acceleratorPut(Xd[0],Xh);
+    acceleratorPut(Yd[0],Yh);
+    acceleratorPut(Cd[0],Ch);
+
+    GridBLAS BLAS;
+
+    RealD t3 = usecond();
+    /////////////////////////////////////////
+    // C_rs = X^dag Y
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nrhs,nrhs,vw,
+		     ComplexD(1.0),
+		     Xd,
+		     Yd,
+		     ComplexD(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+    RealD t4 = usecond();
+
+    std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nrhs -- the coefficients 
+    acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));
+    grid->GlobalSumVector(&HOST_C[0],nrhs*nrhs);
+
+    RealD t5 = usecond();
+    for(int rr=0;rr<nrhs;rr++){
+      for(int r=0;r<nrhs;r++){
+	int off = r+nrhs*rr;
+	m(r,rr)=HOST_C[off];
+      }
+    }
+    RealD t6 = usecond();
+    uint64_t M=nrhs;
+    uint64_t N=nrhs;
+    uint64_t K=vw;
+    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K);
+    RealD flops = 8.0*M*N*K;
+    flops = flops/(t4-t3)/1.e3;
+    bytes = bytes/(t4-t3)/1.e3;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+#else
+    int nrhs;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    nrhs = X.size();
+    GRID_ASSERT(X.size()==Y.size());
+    conformable(X[0],Y[0]);
+
+    grid  = X[0].Grid();
+    int rd0 =  grid->_rdimensions[0] * grid->_rdimensions[1];
+    vol   = grid->oSites()/rd0;
+    words = rd0*sizeof(vector_object)/sizeof(scalar);
+    int64_t vw = vol * words;
+    GRID_ASSERT(vw == grid->lSites()*sizeof(scalar_object)/sizeof(scalar));
+
+    RealD t0 = usecond();
+    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Cred.resize(nrhs * nrhs * vol);// cost free if size doesn't change
+    RealD t1 = usecond();
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources -- layout batched BLAS ready
+    /////////////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      autoView(x_v,X[r],AcceleratorRead);
+      autoView(y_v,Y[r],AcceleratorRead);
+      scalar *from_x=(scalar *)&x_v[0];
+      scalar *from_y=(scalar *)&y_v[0];
+      scalar *BX = &BLAS_X[0];
+      scalar *BY = &BLAS_Y[0];
+      accelerator_for(ssw,vw,1,{
+	  uint64_t ss=ssw/words;
+	  uint64_t  w=ssw%words;
+	  uint64_t offset = w+r*words+ss*nrhs*words; // [ss][rhs][words]
+	  BX[offset] = from_x[ssw];
+	  BY[offset] = from_y[ssw];
+	});
+    }
+    RealD t2 = usecond();
+
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Xxr = [X1(x)][..][Xn(x)]
+   *
+   * Yxr = [Y1(x)][..][Ym(x)]
+   *
+   * C_rs = X^dag Y
+   */
+    Xdip.resize(vol);
+    Ydip.resize(vol);
+    Cdip.resize(vol);
+    std::vector<scalar *> Xh(vol);
+    std::vector<scalar *> Yh(vol);
+    std::vector<scalar *> Ch(vol);
+    for(uint64_t ss=0;ss<vol;ss++){
+
+      Xh[ss] = & BLAS_X[ss*nrhs*words];
+      Yh[ss] = & BLAS_Y[ss*nrhs*words];
+      Ch[ss] = & BLAS_Cred[ss*nrhs*nrhs];
+
+    }
+    acceleratorCopyToDevice(&Xh[0],&Xdip[0],vol*sizeof(scalar *));
+    acceleratorCopyToDevice(&Yh[0],&Ydip[0],vol*sizeof(scalar *));
+    acceleratorCopyToDevice(&Ch[0],&Cdip[0],vol*sizeof(scalar *));
+    
+    GridBLAS BLAS;
+
+    RealD t3 = usecond();
+    /////////////////////////////////////////
+    // C_rs = X^dag Y
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nrhs,nrhs,words,
+		     ComplexD(1.0),
+		     Xdip,
+		     Ydip,
+		     ComplexD(0.0),  // wipe out C
+		     Cdip);
+    BLAS.synchronise();
+    RealD t4 = usecond();
+
+    std::vector<scalar> HOST_C(BLAS_Cred.size());      // nrhs . nrhs -- the coefficients 
+    acceleratorCopyFromDevice(&BLAS_Cred[0],&HOST_C[0],BLAS_Cred.size()*sizeof(scalar));
+
+    RealD t5 = usecond();
+    m = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    for(int ss=0;ss<vol;ss++){
+      Eigen::Map<Eigen::MatrixXcd> eC((std::complex<double> *)&HOST_C[ss*nrhs*nrhs],nrhs,nrhs);
+      m = m + eC;
+    }
+    RealD t6l = usecond();
+    grid->GlobalSumVector((scalar *) &m(0,0),nrhs*nrhs);
+    RealD t6 = usecond();
+    uint64_t M=nrhs;
+    uint64_t N=nrhs;
+    uint64_t K=vw;
+    RealD xybytes = grid->lSites()*sizeof(scalar_object);
+    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K);
+    RealD flops = 8.0*M*N*K;
+    flops = flops/(t4-t3)/1.e3;
+    bytes = bytes/(t4-t3)/1.e3;
+    xybytes = 4*xybytes/(t2-t1)/1.e3;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+#endif
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSBlockProject.h

@@ -0,0 +1,513 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MultiRHSDeflation.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/* 
+   MultiRHS block projection
+
+   Import basis -> nblock x nbasis x  (block x internal) 
+   Import vector of fine lattice objects -> nblock x nrhs x (block x internal) 
+
+   => coarse_(nrhs x nbasis )^block = via batched GEMM
+
+//template<class vobj,class CComplex,int nbasis,class VLattice>
+//inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
+//			   const VLattice &fineData,
+//			   const VLattice &Basis)
+*/
+
+template<class Field>
+class MultiRHSBlockProject
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+  typedef Field Fermion;
+
+  int nbasis;
+  GridBase *coarse_grid;
+  GridBase *fine_grid;
+  uint64_t block_vol;
+  uint64_t fine_vol;
+  uint64_t coarse_vol;
+  uint64_t words;
+
+  // Row major layout "C" order:
+  // BLAS_V[coarse_vol][nbasis][block_vol][words]
+  // BLAS_F[coarse_vol][nrhs][block_vol][words]
+  // BLAS_C[coarse_vol][nrhs][nbasis]
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Vxb = [v1(x)][..][vn(x)] ... x coarse vol
+   *
+   * Fxr = [r1(x)][..][rm(x)] ... x coarse vol
+   *
+   * Block project:
+   * C_br = V^dag F x coarse vol
+   *
+   * Block promote:
+   * F_xr = Vxb Cbr x coarse_vol
+   */  
+  deviceVector<scalar> BLAS_V;      // words * block_vol * nbasis x coarse_vol 
+  deviceVector<scalar> BLAS_F;      // nrhs x fine_vol * words   -- the sources
+  deviceVector<scalar> BLAS_C;      // nrhs x coarse_vol * nbasis -- the coarse coeffs
+
+  RealD blasNorm2(deviceVector<scalar> &blas)
+  {
+    scalar ss(0.0);
+    std::vector<scalar> tmp(blas.size());
+    acceleratorCopyFromDevice(&blas[0],&tmp[0],blas.size()*sizeof(scalar));
+    for(int64_t s=0;s<blas.size();s++){
+      ss=ss+tmp[s]*adj(tmp[s]);
+    }
+    coarse_grid->GlobalSum(ss);
+    return real(ss);
+  }
+  
+  MultiRHSBlockProject(){};
+ ~MultiRHSBlockProject(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    nbasis=0;
+    coarse_grid=nullptr;
+    fine_grid=nullptr;
+    fine_vol=0;
+    block_vol=0;
+    coarse_vol=0;
+    words=0;
+    BLAS_V.resize(0);
+    BLAS_F.resize(0);
+    BLAS_C.resize(0);
+  }
+  void Allocate(int _nbasis,GridBase *_fgrid,GridBase *_cgrid)
+  {
+    nbasis=_nbasis;
+
+    fine_grid=_fgrid;
+    coarse_grid=_cgrid;
+
+    fine_vol   = fine_grid->lSites();
+    coarse_vol = coarse_grid->lSites();
+    block_vol = fine_vol/coarse_vol;
+    
+    words = sizeof(scalar_object)/sizeof(scalar);
+
+    BLAS_V.resize (fine_vol * words * nbasis );
+  }
+  void ImportFineGridVectors(std::vector <Field > &vecs, deviceVector<scalar> &blas)
+  {
+    int nvec = vecs.size();
+    typedef typename Field::vector_object vobj;
+    //    std::cout << GridLogMessage <<" BlockProjector importing "<<nvec<< " fine grid vectors" <<std::endl;
+
+    GRID_ASSERT(vecs[0].Grid()==fine_grid);
+
+    subdivides(coarse_grid,fine_grid); // require they map
+
+    int _ndimension = coarse_grid->_ndimension;
+    GRID_ASSERT(block_vol == fine_grid->oSites() / coarse_grid->oSites());
+    
+    Coordinate  block_r      (_ndimension);
+    for(int d=0 ; d<_ndimension;d++){
+      block_r[d] = fine_grid->_rdimensions[d] / coarse_grid->_rdimensions[d];
+    }
+
+    uint64_t sz = blas.size();
+
+    acceleratorMemSet(&blas[0],0,blas.size()*sizeof(scalar));
+
+    Coordinate fine_rdimensions = fine_grid->_rdimensions;
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    int64_t bv= block_vol;
+    for(int v=0;v<vecs.size();v++){
+
+      //      std::cout << " BlockProjector importing vector"<<v<<" "<<norm2(vecs[v])<<std::endl;
+      autoView( fineData   , vecs[v], AcceleratorRead);
+
+      auto blasData_p  = &blas[0];
+      auto fineData_p  = &fineData[0];
+
+      int64_t osites = fine_grid->oSites();
+
+      // loop over fine sites
+      const int Nsimd = vobj::Nsimd();
+      //      std::cout << "sz "<<sz<<std::endl;
+      //      std::cout << "prod "<<Nsimd * coarse_grid->oSites() * block_vol * nvec * words<<std::endl;
+      GRID_ASSERT(sz == Nsimd * coarse_grid->oSites() * block_vol * nvec * words);
+      uint64_t lwords= words; // local variable for copy in to GPU
+      accelerator_for(sf,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  // One thread per fine site
+	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_b(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  // Fine site to fine coor
+	  Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_r[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_r[d];
+	  
+	  int sc;// coarse site
+	  int sb;// block site
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_r);
+
+          scalar_object data = extractLane(lane,fineData[sf]);
+
+	  // BLAS layout address calculation
+	  // words * block_vol * nbasis x coarse_vol
+	  // coarse oSite x block vole x lanes
+	  int64_t site = (lane*osites + sc*bv)*nvec
+   	               + v*bv
+	               + sb;
+
+	  //	  GRID_ASSERT(site*lwords<sz);
+
+	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
+
+	  *ptr = data;
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+      //      std::cout << " import fine Blas norm "<<blasNorm2(blas)<<std::endl;
+      //      std::cout << " BlockProjector imported vector"<<v<<std::endl;
+    }
+  }
+  void ExportFineGridVectors(std::vector <Field> &vecs, deviceVector<scalar> &blas)
+  {
+    typedef typename Field::vector_object vobj;
+
+    int nvec = vecs.size();
+
+    GRID_ASSERT(vecs[0].Grid()==fine_grid);
+
+    subdivides(coarse_grid,fine_grid); // require they map
+
+    int _ndimension = coarse_grid->_ndimension;
+    GRID_ASSERT(block_vol == fine_grid->oSites() / coarse_grid->oSites());
+    
+    Coordinate  block_r      (_ndimension);
+    for(int d=0 ; d<_ndimension;d++){
+      block_r[d] = fine_grid->_rdimensions[d] / coarse_grid->_rdimensions[d];
+    }
+    Coordinate fine_rdimensions = fine_grid->_rdimensions;
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+
+    //    std::cout << " export fine Blas norm "<<blasNorm2(blas)<<std::endl;
+
+    int64_t bv= block_vol;
+    for(int v=0;v<vecs.size();v++){
+
+      autoView( fineData   , vecs[v], AcceleratorWrite);
+
+      auto blasData_p  = &blas[0];
+      auto fineData_p    = &fineData[0];
+
+      int64_t osites = fine_grid->oSites();
+      uint64_t lwords = words;
+      //      std::cout << " Nsimd is "<<vobj::Nsimd() << std::endl;
+      //      std::cout << " lwords is "<<lwords << std::endl;
+      //      std::cout << " sizeof(scalar_object) is "<<sizeof(scalar_object) << std::endl;
+      // loop over fine sites
+      accelerator_for(sf,osites,vobj::Nsimd(),{
+      
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(vobj::Nsimd()); // buffer lane
+#else
+	  for(int lane=0;lane<vobj::Nsimd();lane++) {
+#endif
+	  // One thread per fine site
+	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_b(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_r[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_r[d];
+	  
+	  int sc;
+	  int sb;
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_r);
+
+	  // BLAS layout address calculation
+	  // words * block_vol * nbasis x coarse_vol 	  
+	  int64_t site = (lane*osites + sc*bv)*nvec
+   	               + v*bv
+	               + sb;
+
+	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
+
+	  scalar_object data = *ptr;
+
+	  insertLane(lane,fineData[sf],data);
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+    }
+  }
+  template<class vobj>
+  void ImportCoarseGridVectors(std::vector <Lattice<vobj> > &vecs, deviceVector<scalar> &blas)
+  {
+    int nvec = vecs.size();
+    typedef typename vobj::scalar_object coarse_scalar_object;
+
+    //    std::cout << " BlockProjector importing "<<nvec<< " coarse grid vectors" <<std::endl;
+
+    GRID_ASSERT(vecs[0].Grid()==coarse_grid);
+
+    int _ndimension = coarse_grid->_ndimension;
+
+    uint64_t sz = blas.size();
+
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    
+    for(int v=0;v<vecs.size();v++){
+
+      //      std::cout << " BlockProjector importing coarse vector"<<v<<" "<<norm2(vecs[v])<<std::endl;
+      autoView( coarseData   , vecs[v], AcceleratorRead);
+
+      auto blasData_p  = &blas[0];
+      auto coarseData_p  = &coarseData[0];
+
+      int64_t osites = coarse_grid->oSites();
+
+      // loop over fine sites
+      const int Nsimd = vobj::Nsimd();
+      uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
+      GRID_ASSERT(cwords==nbasis);
+      
+      accelerator_for(sc,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+           // C_br per site
+	    int64_t blas_site = (lane*osites + sc)*nvec*cwords + v*cwords;
+	    
+	    coarse_scalar_object data = extractLane(lane,coarseData[sc]);
+
+	    coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
+
+	    *ptr = data;
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+      //      std::cout << " import coarsee Blas norm "<<blasNorm2(blas)<<std::endl;
+    }
+  }
+  template<class vobj>
+  void ExportCoarseGridVectors(std::vector <Lattice<vobj> > &vecs, deviceVector<scalar> &blas)
+  {
+    int nvec = vecs.size();
+    typedef typename vobj::scalar_object coarse_scalar_object;
+    //    std::cout << GridLogMessage<<" BlockProjector exporting "<<nvec<< " coarse grid vectors" <<std::endl;
+
+    GRID_ASSERT(vecs[0].Grid()==coarse_grid);
+
+    int _ndimension = coarse_grid->_ndimension;
+    
+    uint64_t sz = blas.size();
+
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    
+    //    std::cout << " export coarsee Blas norm "<<blasNorm2(blas)<<std::endl;
+    for(int v=0;v<vecs.size();v++){
+
+      //  std::cout << " BlockProjector exporting coarse vector"<<v<<std::endl;
+      autoView( coarseData   , vecs[v], AcceleratorWrite);
+
+      auto blasData_p  = &blas[0];
+      auto coarseData_p  = &coarseData[0];
+
+      int64_t osites = coarse_grid->oSites();
+
+      // loop over fine sites
+      const int Nsimd = vobj::Nsimd();
+      uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
+      GRID_ASSERT(cwords==nbasis);
+      
+      accelerator_for(sc,osites,Nsimd,{
+	  // Wrap in a macro "FOR_ALL_LANES(lane,{ ... });
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	    int64_t blas_site = (lane*osites + sc)*nvec*cwords + v*cwords;
+	    coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
+	    coarse_scalar_object data = *ptr;
+	    insertLane(lane,coarseData[sc],data);
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+    }
+  }
+  void ImportBasis(std::vector < Field > &vecs)
+  {
+    //    std::cout << " BlockProjector Import basis size "<<vecs.size()<<std::endl;
+    ImportFineGridVectors(vecs,BLAS_V);
+  }
+
+  template<class cobj>
+  void blockProject(std::vector<Field> &fine,std::vector< Lattice<cobj> > & coarse)
+  {
+    int nrhs=fine.size();
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    //    std::cout << "blockProject nbasis " <<nbasis<<" " << _nbasis<<std::endl;
+    GRID_ASSERT(nbasis==_nbasis);
+    
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources to same data layout
+    /////////////////////////////////////////////
+    //    std::cout << "BlockProject import fine"<<std::endl;
+    ImportFineGridVectors(fine,BLAS_F);
+    
+    deviceVector<scalar *> Vd(coarse_vol);
+    deviceVector<scalar *> Fd(coarse_vol);
+    deviceVector<scalar *> Cd(coarse_vol);
+
+    //    std::cout << "BlockProject pointers"<<std::endl;
+    for(int c=0;c<coarse_vol;c++){
+      // BLAS_V[coarse_vol][nbasis][block_vol][words]
+      // BLAS_F[coarse_vol][nrhs][block_vol][words]
+      // BLAS_C[coarse_vol][nrhs][nbasis]
+      scalar * Vh = & BLAS_V[c*nbasis*block_vol*words];
+      scalar * Fh = & BLAS_F[c*nrhs*block_vol*words];
+      scalar * Ch = & BLAS_C[c*nrhs*nbasis];
+
+      acceleratorPut(Vd[c],Vh);
+      acceleratorPut(Fd[c],Fh);
+      acceleratorPut(Cd[c],Ch);
+    }
+
+    GridBLAS BLAS;
+
+    //    std::cout << "BlockProject BLAS"<<std::endl;
+    int64_t vw = block_vol * words;
+    /////////////////////////////////////////
+    // C_br = V^dag R
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nbasis,nrhs,vw,
+		     scalar(1.0),
+		     Vd,
+		     Fd,
+		     scalar(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+    //    std::cout << "BlockProject done"<<std::endl;
+    ExportCoarseGridVectors(coarse, BLAS_C);
+    //    std::cout << "BlockProject done"<<std::endl;
+
+  }
+
+  template<class cobj>
+  void blockPromote(std::vector<Field> &fine,std::vector<Lattice<cobj> > & coarse)
+  {
+    int nrhs=fine.size();
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportCoarseGridVectors(coarse, BLAS_C);
+
+    GridBLAS BLAS;
+
+    deviceVector<scalar *> Vd(coarse_vol);
+    deviceVector<scalar *> Fd(coarse_vol);
+    deviceVector<scalar *> Cd(coarse_vol);
+
+    for(int c=0;c<coarse_vol;c++){
+      // BLAS_V[coarse_vol][nbasis][block_vol][words]
+      // BLAS_F[coarse_vol][nrhs][block_vol][words]
+      // BLAS_C[coarse_vol][nrhs][nbasis]
+      scalar * Vh = & BLAS_V[c*nbasis*block_vol*words];
+      scalar * Fh = & BLAS_F[c*nrhs*block_vol*words];
+      scalar * Ch = & BLAS_C[c*nrhs*nbasis];
+      acceleratorPut(Vd[c],Vh);
+      acceleratorPut(Fd[c],Fh);
+      acceleratorPut(Cd[c],Ch);
+    }
+
+    /////////////////////////////////////////
+    // Block promote:
+    // F_xr = Vxb Cbr (x coarse_vol)
+    /////////////////////////////////////////
+
+    int64_t vw = block_vol * words;
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     vw,nrhs,nbasis,
+		     scalar(1.0),
+		     Vd,
+		     Cd,
+		     scalar(0.0),  // wipe out C
+		     Fd);
+    BLAS.synchronise();
+    //    std::cout << " blas call done"<<std::endl;
+    
+    ExportFineGridVectors(fine, BLAS_F);
+    //    std::cout << " exported "<<std::endl;
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSDeflation.h

@@ -0,0 +1,233 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MultiRHSDeflation.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/* Need helper object for BLAS accelerated mrhs projection
+
+   i) MultiRHS Deflation
+
+   Import Evecs -> nev x vol x internal 
+   Import vector of Lattice objects -> nrhs x vol x internal
+   => Cij (nrhs x Nev) via GEMM.
+   => Guess  (nrhs x vol x internal)  = C x evecs (via GEMM)
+   Export
+
+   
+   ii) MultiRHS block projection
+
+   Import basis -> nblock x nbasis x  (block x internal) 
+   Import vector of fine lattice objects -> nblock x nrhs x (block x internal) 
+
+   => coarse_(nrhs x nbasis )^block = via batched GEMM
+
+   iii)   Alternate interface: 
+   Import higher dim Lattice object-> vol x nrhs layout
+   
+*/
+template<class Field>
+class MultiRHSDeflation
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+
+  int nev;
+  std::vector<RealD> eval;
+  GridBase *grid;
+  uint64_t vol;
+  uint64_t words;
+  
+  deviceVector<scalar> BLAS_E;      //  nev x vol -- the eigenbasis   (up to a 1/sqrt(lambda))
+  deviceVector<scalar> BLAS_R;      // nrhs x vol -- the sources
+  deviceVector<scalar> BLAS_G;      // nrhs x vol -- the guess
+  deviceVector<scalar> BLAS_C;      // nrhs x nev -- the coefficients 
+  
+  MultiRHSDeflation(){};
+  ~MultiRHSDeflation(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    nev=0;
+    grid=nullptr;
+    vol=0;
+    words=0;
+    BLAS_E.resize(0);
+    BLAS_R.resize(0);
+    BLAS_C.resize(0);
+    BLAS_G.resize(0);
+  }
+  void Allocate(int _nev,GridBase *_grid)
+  {
+    nev=_nev;
+    grid=_grid;
+    vol   = grid->lSites();
+    words = sizeof(scalar_object)/sizeof(scalar);
+    eval.resize(nev);
+    BLAS_E.resize (vol * words * nev );
+    std::cout << GridLogMessage << " Allocate for "<<nev<<" eigenvectors and volume "<<vol<<std::endl;
+  }
+  void ImportEigenVector(Field &evec,RealD &_eval, int ev)
+  {
+    //    std::cout << " ev " <<ev<<" eval "<<_eval<< std::endl;
+    GRID_ASSERT(ev<eval.size());
+    eval[ev] = _eval;
+
+    int64_t offset = ev*vol*words;
+    autoView(v,evec,AcceleratorRead);
+    acceleratorCopyDeviceToDevice(&v[0],&BLAS_E[offset],sizeof(scalar_object)*vol);
+
+  }
+  void ImportEigenBasis(std::vector<Field> &evec,std::vector<RealD> &_eval)
+  {
+    ImportEigenBasis(evec,_eval,0,evec.size());
+  }
+  // Could use to import a batch of eigenvectors
+  void ImportEigenBasis(std::vector<Field> &evec,std::vector<RealD> &_eval, int _ev0, int _nev)
+  {
+    GRID_ASSERT(_ev0+_nev<=evec.size());
+
+    Allocate(_nev,evec[0].Grid());
+    
+    // Imports a sub-batch of eigenvectors, _ev0, ..., _ev0+_nev-1
+    for(int e=0;e<nev;e++){
+      std::cout << "Importing eigenvector "<<e<<" evalue "<<_eval[_ev0+e]<<std::endl;
+      ImportEigenVector(evec[_ev0+e],_eval[_ev0+e],e);
+    }
+  }
+  void DeflateSources(std::vector<Field> &source,std::vector<Field> & guess)
+  {
+    int nrhs = source.size();
+    GRID_ASSERT(source.size()==guess.size());
+    GRID_ASSERT(grid == guess[0].Grid());
+    conformable(guess[0],source[0]);
+
+    int64_t vw = vol * words;
+
+    RealD t0 = usecond();
+    BLAS_R.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_G.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_C.resize(nev * nrhs);// cost free if size doesn't change
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources
+    /////////////////////////////////////////////
+    //    for(int r=0;r<nrhs;r++){
+    //      std::cout << " source["<<r<<"] = "<<norm2(source[r])<<std::endl;
+    //    }
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(v,source[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&v[0],&BLAS_R[offset],sizeof(scalar_object)*vol);
+    }
+
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Exe = [e1(x)][..][en(x)]
+   *
+   * Rxr = [r1(x)][..][rm(x)]
+   *
+   * C_er = E^dag R
+   * C_er = C_er / lambda_e 
+   * G_xr = Exe Cer
+   */
+    deviceVector<scalar *> Ed(1);
+    deviceVector<scalar *> Rd(1);
+    deviceVector<scalar *> Cd(1);
+    deviceVector<scalar *> Gd(1);
+
+    scalar * Eh = & BLAS_E[0];
+    scalar * Rh = & BLAS_R[0];
+    scalar * Ch = & BLAS_C[0];
+    scalar * Gh = & BLAS_G[0];
+
+    acceleratorPut(Ed[0],Eh);
+    acceleratorPut(Rd[0],Rh);
+    acceleratorPut(Cd[0],Ch);
+    acceleratorPut(Gd[0],Gh);
+
+    GridBLAS BLAS;
+
+    /////////////////////////////////////////
+    // C_er = E^dag R
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nev,nrhs,vw,
+		     scalar(1.0),
+		     Ed,
+		     Rd,
+		     scalar(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+
+    GRID_ASSERT(BLAS_C.size()==nev*nrhs);
+
+    std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nev -- the coefficients 
+    acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));
+    grid->GlobalSumVector(&HOST_C[0],nev*nrhs);
+    for(int e=0;e<nev;e++){
+      RealD lam(1.0/eval[e]);
+      for(int r=0;r<nrhs;r++){
+	int off = e+nev*r;
+	HOST_C[off]=HOST_C[off] * lam;
+	//	std::cout << "C["<<e<<"]["<<r<<"] ="<<HOST_C[off]<< " eval[e] "<<eval[e] <<std::endl;
+      }
+    }
+    acceleratorCopyToDevice(&HOST_C[0],&BLAS_C[0],BLAS_C.size()*sizeof(scalar));
+
+    
+    /////////////////////////////////////////
+    // Guess G_xr = Exe Cer
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+		     vw,nrhs,nev,
+		     scalar(1.0),
+		     Ed, // x . nev
+		     Cd, // nev . nrhs
+		     scalar(0.0),
+		     Gd);
+    BLAS.synchronise();
+
+    ///////////////////////////////////////
+    // Copy out the multirhs
+    ///////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(v,guess[r],AcceleratorWrite);
+      acceleratorCopyDeviceToDevice(&BLAS_G[offset],&v[0],sizeof(scalar_object)*vol);
+    }
+    RealD t1 = usecond();
+    std::cout << GridLogMessage << "MultiRHSDeflation for "<<nrhs<<" sources with "<<nev<<" eigenvectors took " << (t1-t0)/1e3 <<" ms"<<std::endl;
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/iterative/AdefGeneric.h

@@ -33,109 +33,111 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    * Script A = SolverMatrix 
    * Script P = Preconditioner
    *
-   * Deflation methods considered
-   *      -- Solve P A x = P b        [ like Luscher ]
-   * DEF-1        M P A x = M P b     [i.e. left precon]
-   * DEF-2        P^T M A x = P^T M b
-   * ADEF-1       Preconditioner = M P + Q      [ Q + M + M A Q]
-   * ADEF-2       Preconditioner = P^T M + Q
-   * BNN          Preconditioner = P^T M P + Q
-   * BNN2         Preconditioner = M P + P^TM +Q - M P A M 
-   * 
    * Implement ADEF-2
    *
    * Vstart = P^Tx + Qb
    * M1 = P^TM + Q
    * M2=M3=1
-   * Vout = x
    */
+NAMESPACE_BEGIN(Grid);
 
-// abstract base
-template<class Field, class CoarseField>
-class TwoLevelFlexiblePcg : public LinearFunction<Field>
+
+template<class Field>
+class TwoLevelCG : public LinearFunction<Field>
 {
  public:
-  int verbose;
   RealD   Tolerance;
   Integer MaxIterations;
-  const int mmax = 5;
   GridBase *grid;
-  GridBase *coarsegrid;
 
-  LinearOperatorBase<Field>   *_Linop
-  OperatorFunction<Field>     *_Smoother,
-  LinearFunction<CoarseField> *_CoarseSolver;
-
-  // Need somthing that knows how to get from Coarse to fine and back again
+  // Fine operator, Smoother, CoarseSolver
+  LinearOperatorBase<Field>   &_FineLinop;
+  LinearFunction<Field>   &_Smoother;
   
   // more most opertor functions
-  TwoLevelFlexiblePcg(RealD tol,
-		     Integer maxit,
-		     LinearOperatorBase<Field> *Linop,
-		     LinearOperatorBase<Field> *SmootherLinop,
-		     OperatorFunction<Field>   *Smoother,
-		     OperatorFunction<CoarseField>  CoarseLinop
-		     ) : 
+  TwoLevelCG(RealD tol,
+	     Integer maxit,
+	     LinearOperatorBase<Field>   &FineLinop,
+	     LinearFunction<Field>       &Smoother,
+	     GridBase *fine) : 
       Tolerance(tol), 
       MaxIterations(maxit),
-      _Linop(Linop),
-      _PreconditionerLinop(PrecLinop),
-      _Preconditioner(Preconditioner)
-  { 
-    verbose=0;
+      _FineLinop(FineLinop),
+      _Smoother(Smoother)
+  {
+    grid       = fine;
   };
-
-  // The Pcg routine is common to all, but the various matrices differ from derived 
-  // implementation to derived implmentation
-  void operator() (const Field &src, Field &psi){
-  void operator() (const Field &src, Field &psi){
-
-    psi.Checkerboard() = src.Checkerboard();
-    grid             = src.Grid();
-
+  
+  virtual void operator() (const Field &src, Field &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: fPcg starting single RHS"<<std::endl;
     RealD f;
     RealD rtzp,rtz,a,d,b;
     RealD rptzp;
-    RealD tn;
-    RealD guess = norm2(psi);
-    RealD ssq   = norm2(src);
-    RealD rsq   = ssq*Tolerance*Tolerance;
-    
+
     /////////////////////////////
     // Set up history vectors
     /////////////////////////////
-    std::vector<Field> p  (mmax,grid);
+    int mmax = 5;
+    std::cout << GridLogMessage<<"HDCG: fPcg allocating"<<std::endl;
+    std::vector<Field> p(mmax,grid);
     std::vector<Field> mmp(mmax,grid);
     std::vector<RealD> pAp(mmax);
-
-    Field x  (grid); x = psi;
-    Field z  (grid);
+    Field z(grid);
     Field tmp(grid);
-    Field r  (grid);
-    Field mu (grid);
-  
+    Field  mp (grid);
+    Field  r  (grid);
+    Field  mu (grid);
+    
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated"<<std::endl;
+    //Initial residual computation & set up
+    RealD guess   = norm2(x);
+    std::cout << GridLogMessage<<"HDCG: fPcg guess nrm "<<guess<<std::endl;
+    RealD src_nrm = norm2(src);
+    std::cout << GridLogMessage<<"HDCG: fPcg src nrm "<<src_nrm<<std::endl;
+    
+    if ( src_nrm == 0.0 ) {
+      std::cout << GridLogMessage<<"HDCG: fPcg given trivial source norm "<<src_nrm<<std::endl;
+      x=Zero();
+    }
+    RealD tn;
+    
+    GridStopWatch HDCGTimer;
+    HDCGTimer.Start();
     //////////////////////////
     // x0 = Vstart -- possibly modify guess
     //////////////////////////
-    x=src;
     Vstart(x,src);
-
+    
     // r0 = b -A x0
-    HermOp(x,mmp); // Shouldn't this be something else?
+    _FineLinop.HermOp(x,mmp[0]);
     axpy (r, -1.0,mmp[0], src);    // Recomputes r=src-Ax0
+    {
+      double n1 = norm2(x);
+      double n2 = norm2(mmp[0]);
+      double n3 = norm2(r);
+      std::cout<<GridLogMessage<<"x,vstart,r = "<<n1<<" "<<n2<<" "<<n3<<std::endl;
+    }
 
     //////////////////////////////////
     // Compute z = M1 x
     //////////////////////////////////
-    M1(r,z,tmp,mp,SmootherMirs);
+    PcgM1(r,z);
     rtzp =real(innerProduct(r,z));
-
+    
     ///////////////////////////////////////
     // Solve for Mss mu = P A z and set p = z-mu
-    // Def2: p = 1 - Q Az = Pright z 
+    // Def2 p = 1 - Q Az = Pright z
     // Other algos M2 is trivial
     ///////////////////////////////////////
-    M2(z,p[0]);
+    PcgM2(z,p[0]);
+
+    RealD ssq =  norm2(src);
+    RealD rsq =  ssq*Tolerance*Tolerance;
+
+    std::cout << GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" rsq "<<rsq<<"\n";
+
+    Field pp(grid);
 
     for (int k=0;k<=MaxIterations;k++){
     
@@ -143,31 +145,46 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       int peri_kp = (k+1) % mmax;
 
       rtz=rtzp;
-      d= M3(p[peri_k],mp,mmp[peri_k],tmp);
+      d= PcgM3(p[peri_k],mmp[peri_k]);
       a = rtz/d;
     
       // Memorise this
       pAp[peri_k] = d;
-
+      
       axpy(x,a,p[peri_k],x);
       RealD rn = axpy_norm(r,-a,mmp[peri_k],r);
 
       // Compute z = M x
-      M1(r,z,tmp,mp);
-
+      PcgM1(r,z);
+      
+      {
+	RealD n1,n2;
+	n1=norm2(r);
+	n2=norm2(z);
+	std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : vector r,z "<<n1<<" "<<n2<<"\n";
+      }
       rtzp =real(innerProduct(r,z));
+      std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : inner rtzp "<<rtzp<<"\n";
 
-      M2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+      //    PcgM2(z,p[0]);
+      PcgM2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+      
+      p[peri_kp]=mu;
 
-      p[peri_kp]=p[peri_k];
-
-      // Standard search direction  p -> z + b p    ; b = 
+      // Standard search direction  p -> z + b p    
       b = (rtzp)/rtz;
-
+      
       int northog;
+      // k=zero  <=> peri_kp=1;        northog = 1
+      // k=1     <=> peri_kp=2;        northog = 2
+      // ...               ...                  ...
+      // k=mmax-2<=> peri_kp=mmax-1;   northog = mmax-1
+      // k=mmax-1<=> peri_kp=0;        northog = 1
+
       //    northog     = (peri_kp==0)?1:peri_kp; // This is the fCG(mmax) algorithm
       northog     = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
     
+      std::cout<<GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : orthogonalising to last "<<northog<<" vectors\n";
       for(int back=0; back < northog; back++){
 	int peri_back = (k-back)%mmax;
 	RealD pbApk= real(innerProduct(mmp[peri_back],p[peri_kp]));
@@ -176,75 +193,324 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       }
 
       RealD rrn=sqrt(rn/ssq);
-      std::cout<<GridLogMessage<<"TwoLevelfPcg: k= "<<k<<" residual = "<<rrn<<std::endl;
+      RealD rtn=sqrt(rtz/ssq);
+      RealD rtnp=sqrt(rtzp/ssq);
+
+      std::cout<<GridLogMessage<<"HDCG: fPcg k= "<<k<<" residual = "<<rrn<<"\n";
 
       // Stopping condition
       if ( rn <= rsq ) { 
 
-	HermOp(x,mmp); // Shouldn't this be something else?
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	
+	_FineLinop.HermOp(x,mmp[0]);			  
 	axpy(tmp,-1.0,src,mmp[0]);
 	
-	RealD psinorm = sqrt(norm2(x));
-	RealD srcnorm = sqrt(norm2(src));
-	RealD tmpnorm = sqrt(norm2(tmp));
-	RealD true_residual = tmpnorm/srcnorm;
-	std::cout<<GridLogMessage<<"TwoLevelfPcg:   true residual is "<<true_residual<<std::endl;
-	std::cout<<GridLogMessage<<"TwoLevelfPcg: target residual was"<<Tolerance<<std::endl;
-	return k;
+	RealD  mmpnorm = sqrt(norm2(mmp[0]));
+	RealD  xnorm   = sqrt(norm2(x));
+	RealD  srcnorm = sqrt(norm2(src));
+	RealD  tmpnorm = sqrt(norm2(tmp));
+	RealD  true_residual = tmpnorm/srcnorm;
+	std::cout<<GridLogMessage
+	       <<"HDCG: true residual is "<<true_residual
+	       <<" solution "<<xnorm
+	       <<" source "<<srcnorm
+	       <<" mmp "<<mmpnorm	  
+	       <<std::endl;
+      
+	return;
       }
+
     }
-    // Non-convergence
-    assert(0);
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    RealD  xnorm   = sqrt(norm2(x));
+    RealD  srcnorm = sqrt(norm2(src));
+    std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
   }
 
+
+
+  virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    std::vector<RealD> f(nrhs);
+    std::vector<RealD> rtzp(nrhs);
+    std::vector<RealD> rtz(nrhs);
+    std::vector<RealD> a(nrhs);
+    std::vector<RealD> d(nrhs);
+    std::vector<RealD> b(nrhs);
+    std::vector<RealD> rptzp(nrhs);
+    /////////////////////////////
+    // Set up history vectors
+    /////////////////////////////
+    int mmax = 3;
+    std::cout << GridLogMessage<<"HDCG: fPcg allocating"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<Field> > p(nrhs);   for(int r=0;r<nrhs;r++)  p[r].resize(mmax,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated p"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<Field> > mmp(nrhs); for(int r=0;r<nrhs;r++) mmp[r].resize(mmax,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated mmp"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<RealD> > pAp(nrhs); for(int r=0;r<nrhs;r++) pAp[r].resize(mmax);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated pAp"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<Field> z(nrhs,grid);
+    std::vector<Field>  mp (nrhs,grid);
+    std::vector<Field>  r  (nrhs,grid);
+    std::vector<Field>  mu (nrhs,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated z,mp,r,mu"<<std::endl;
+    src[0].Grid()->Barrier();
+
+    //Initial residual computation & set up
+    std::vector<RealD> src_nrm(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      src_nrm[rhs]=norm2(src[rhs]);
+      GRID_ASSERT(src_nrm[rhs]!=0.0);
+    }
+    std::vector<RealD> tn(nrhs);
+
+    GridStopWatch HDCGTimer;
+    HDCGTimer.Start();
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    Vstart(x,src);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      // r0 = b -A x0
+      _FineLinop.HermOp(x[rhs],mmp[rhs][0]);
+      axpy (r[rhs], -1.0,mmp[rhs][0], src[rhs]);    // Recomputes r=src-Ax0
+    }
+
+    //////////////////////////////////
+    // Compute z = M1 x
+    //////////////////////////////////
+    // This needs a multiRHS version for acceleration
+    PcgM1(r,z);
+
+    std::vector<RealD> ssq(nrhs);
+    std::vector<RealD> rsq(nrhs);
+    std::vector<Field> pp(nrhs,grid);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+      p[rhs][0]=z[rhs];
+      ssq[rhs]=norm2(src[rhs]);
+      rsq[rhs]=  ssq[rhs]*Tolerance*Tolerance;
+      std::cout << GridLogMessage<<"mrhs HDCG: "<<rhs<<" k=0 residual "<<rtzp[rhs]<<" rsq "<<rsq[rhs]<<"\n";
+    }
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+    
+      int peri_k  = k % mmax;
+      int peri_kp = (k+1) % mmax;
+
+      for(int rhs=0;rhs<nrhs;rhs++){
+	rtz[rhs]=rtzp[rhs];
+	d[rhs]= PcgM3(p[rhs][peri_k],mmp[rhs][peri_k]);
+	a[rhs] = rtz[rhs]/d[rhs];
+    
+	// Memorise this
+	pAp[rhs][peri_k] = d[rhs];
+
+	axpy(x[rhs],a[rhs],p[rhs][peri_k],x[rhs]);
+	rn[rhs] = axpy_norm(r[rhs],-a[rhs],mmp[rhs][peri_k],r[rhs]);
+      }
+
+      // Compute z = M x (for *all* RHS)
+      PcgM1(r,z);
+      std::cout << GridLogMessage<<"HDCG::fPcg M1 complete"<<std::endl;
+      grid->Barrier();
+      
+      RealD max_rn=0.0;
+      for(int rhs=0;rhs<nrhs;rhs++){
+
+	rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+
+	std::cout << GridLogMessage<<"HDCG::fPcg rhs"<<rhs<<" iteration "<<k<<" : inner rtzp "<<rtzp[rhs]<<"\n";
+	
+	mu[rhs]=z[rhs];
+
+	p[rhs][peri_kp]=mu[rhs];
+
+	// Standard search direction p == z + b p 
+	b[rhs] = (rtzp[rhs])/rtz[rhs];
+
+	int northog = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
+	std::cout<<GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : orthogonalising to last "<<northog<<" vectors\n";
+	for(int back=0; back < northog; back++){
+	  int peri_back = (k-back)%mmax;
+	  RealD pbApk= real(innerProduct(mmp[rhs][peri_back],p[rhs][peri_kp]));
+	  RealD beta = -pbApk/pAp[rhs][peri_back];
+	  axpy(p[rhs][peri_kp],beta,p[rhs][peri_back],p[rhs][peri_kp]);
+	}
+
+	RealD rrn=sqrt(rn[rhs]/ssq[rhs]);
+	RealD rtn=sqrt(rtz[rhs]/ssq[rhs]);
+	RealD rtnp=sqrt(rtzp[rhs]/ssq[rhs]);
+	
+	std::cout<<GridLogMessage<<"HDCG: rhs "<<rhs<<"fPcg k= "<<k<<" residual = "<<rrn<<"\n";
+	if ( rrn > max_rn ) max_rn = rrn;
+      }
+
+      // Stopping condition based on worst case
+      if ( max_rn <= Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+	  _FineLinop.HermOp(x[rhs],mmp[rhs][0]);			  
+	  Field tmp(grid);
+	  axpy(tmp,-1.0,src[rhs],mmp[rhs][0]);
+      
+	  RealD  mmpnorm = sqrt(norm2(mmp[rhs][0]));
+	  RealD  xnorm   = sqrt(norm2(x[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(tmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<" mmp "<<mmpnorm	  
+		   <<std::endl;
+	}
+	return;
+      }
+      
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    for(int rhs=0;rhs<nrhs;rhs++){
+      RealD  xnorm   = sqrt(norm2(x[rhs]));
+      RealD  srcnorm = sqrt(norm2(src[rhs]));
+      std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+    }
+  }
+  
+
  public:
 
-  virtual void M(Field & in,Field & out,Field & tmp) {
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out)
+  {
+    std::cout << "PcgM1 default (cheat) mrhs version"<<std::endl;
+    for(int rhs=0;rhs<in.size();rhs++){
+      this->PcgM1(in[rhs],out[rhs]);
+    }
+  }
+  virtual void PcgM1(Field & in, Field & out)     =0;
+  virtual void Vstart(std::vector<Field> & x,std::vector<Field> & src)
+  {
+    std::cout << "Vstart default (cheat) mrhs version"<<std::endl;
+    for(int rhs=0;rhs<x.size();rhs++){
+      this->Vstart(x[rhs],src[rhs]);
+    }
+  }
+  virtual void Vstart(Field & x,const Field & src)=0;
 
+  virtual void PcgM2(const Field & in, Field & out) {
+    out=in;
   }
 
-  virtual void M1(Field & in, Field & out) {// the smoother
+  virtual RealD PcgM3(const Field & p, Field & mmp){
+    RealD dd;
+    _FineLinop.HermOp(p,mmp);
+    ComplexD dot = innerProduct(p,mmp);
+    dd=real(dot);
+    return dd;
+  }
 
+  /////////////////////////////////////////////////////////////////////
+  // Only Def1 has non-trivial Vout.
+  /////////////////////////////////////////////////////////////////////
+
+};
+  
+template<class Field, class CoarseField, class Aggregation>
+class TwoLevelADEF2 : public TwoLevelCG<Field>
+{
+ public:
+  ///////////////////////////////////////////////////////////////////////////////////
+  // Need something that knows how to get from Coarse to fine and back again
+  //  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
+  //  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+  ///////////////////////////////////////////////////////////////////////////////////
+  GridBase *coarsegrid;
+  Aggregation &_Aggregates;                    
+  LinearFunction<CoarseField> &_CoarseSolver;
+  LinearFunction<CoarseField> &_CoarseSolverPrecise;
+  ///////////////////////////////////////////////////////////////////////////////////
+  
+  // more most opertor functions
+  TwoLevelADEF2(RealD tol,
+		Integer maxit,
+		LinearOperatorBase<Field>    &FineLinop,
+		LinearFunction<Field>        &Smoother,
+		LinearFunction<CoarseField>  &CoarseSolver,
+		LinearFunction<CoarseField>  &CoarseSolverPrecise,
+		Aggregation &Aggregates
+		) :
+      TwoLevelCG<Field>(tol,maxit,FineLinop,Smoother,Aggregates.FineGrid),
+      _CoarseSolver(CoarseSolver),
+      _CoarseSolverPrecise(CoarseSolverPrecise),
+      _Aggregates(Aggregates)
+  {
+    coarsegrid = Aggregates.CoarseGrid;
+  };
+
+  virtual void PcgM1(Field & in, Field & out)
+  {
+    GRID_TRACE("MultiGridPreconditioner ");
     // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
-    Field tmp(grid);
-    Field Min(grid);
 
-    PcgM(in,Min); // Smoother call
+    Field tmp(this->grid);
+    Field Min(this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
 
-    HermOp(Min,out);
+    GridStopWatch SmootherTimer;
+    GridStopWatch MatrixTimer;
+    SmootherTimer.Start();
+    this->_Smoother(in,Min);
+    SmootherTimer.Stop();
+
+    MatrixTimer.Start();
+    this->_FineLinop.HermOp(Min,out);
+    MatrixTimer.Stop();
     axpy(tmp,-1.0,out,in);          // tmp  = in - A Min
 
-    ProjectToSubspace(tmp,PleftProj);     
-    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
-    PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    GridStopWatch ProjTimer;
+    GridStopWatch CoarseTimer;
+    GridStopWatch PromTimer;
+    ProjTimer.Start();
+    this->_Aggregates.ProjectToSubspace(PleftProj,tmp);     
+    ProjTimer.Stop();
+    CoarseTimer.Start();
+    this->_CoarseSolver(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
+    CoarseTimer.Stop();
+    PromTimer.Start();
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    PromTimer.Stop();
+    std::cout << GridLogPerformance << "PcgM1 breakdown "<<std::endl;
+    std::cout << GridLogPerformance << "\tSmoother   " << SmootherTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tProj       " << ProjTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tCoarse     " << CoarseTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tProm       " << PromTimer.Elapsed() <<std::endl;
+
     axpy(out,1.0,Min,tmp); // Min+tmp
   }
 
-  virtual void M2(const Field & in, Field & out) {
-    out=in;
-    // Must override for Def2 only
-    //  case PcgDef2:
-    //    Pright(in,out);
-    //    break;
-  }
-
-  virtual RealD M3(const Field & p, Field & mmp){
-    double d,dd;
-    HermOpAndNorm(p,mmp,d,dd);
-    return dd;
-    // Must override for Def1 only
-    //  case PcgDef1:
-    //    d=linop_d->Mprec(p,mmp,tmp,0,1);// Dag no
-    //      linop_d->Mprec(mmp,mp,tmp,1);// Dag yes
-    //    Pleft(mp,mmp);
-    //    d=real(linop_d->inner(p,mmp));
-  }
-
-  virtual void VstartDef2(Field & xconst Field & src){
-    //case PcgDef2:
-    //case PcgAdef2: 
-    //case PcgAdef2f:
-    //case PcgV11f:
+  virtual void Vstart(Field & x,const Field & src)
+  {
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart "<<std::endl;
     ///////////////////////////////////
     // Choose x_0 such that 
     // x_0 = guess +  (A_ss^inv) r_s = guess + Ass_inv [src -Aguess]
@@ -256,142 +522,78 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
     //                   = src_s - (A guess)_s - src_s  + (A guess)_s 
     //                   = 0 
     ///////////////////////////////////
-    Field r(grid);
-    Field mmp(grid);
-    
-    HermOp(x,mmp);
-    axpy (r, -1.0, mmp, src);        // r_{-1} = src - A x
-    ProjectToSubspace(r,PleftProj);     
-    ApplyInverseCG(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    PromoteFromSubspace(PleftMss_proj,mmp);  
-    x=x+mmp;
+    Field r(this->grid);
+    Field mmp(this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
+
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart projecting "<<std::endl;
+    this->_Aggregates.ProjectToSubspace(PleftProj,src);     
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart coarse solve "<<std::endl;
+    this->_CoarseSolverPrecise(PleftProj,PleftMss_proj); // Ass^{-1} r_s
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart promote "<<std::endl;
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,x);  
 
   }
 
+};
+
+  
+template<class Field>
+class TwoLevelADEF1defl : public TwoLevelCG<Field>
+{
+public:
+  const std::vector<Field> &evec;
+  const std::vector<RealD> &eval;
+  
+  TwoLevelADEF1defl(RealD tol,
+		   Integer maxit,
+		   LinearOperatorBase<Field>   &FineLinop,
+		   LinearFunction<Field>   &Smoother,
+		   std::vector<Field> &_evec,
+		   std::vector<RealD> &_eval) : 
+    TwoLevelCG<Field>(tol,maxit,FineLinop,Smoother,_evec[0].Grid()),
+    evec(_evec),
+    eval(_eval)
+  {};
+
+  // Can just inherit existing M2
+  // Can just inherit existing M3
+
+  // Simple vstart - do nothing
   virtual void Vstart(Field & x,const Field & src){
-    return;
+    x=src; // Could apply Q
+  };
+
+  // Override PcgM1
+  virtual void PcgM1(Field & in, Field & out)
+  {
+    GRID_TRACE("EvecPreconditioner ");
+    int N=evec.size();
+    Field Pin(this->grid);
+    Field Qin(this->grid);
+
+    //MP  + Q = M(1-AQ) + Q = M
+    // // If we are eigenvector deflating in coarse space
+    // // Q   = Sum_i |phi_i> 1/lambda_i <phi_i|
+    // // A Q = Sum_i |phi_i> <phi_i|
+    // // M(1-AQ) = M(1-proj) + Q
+    Qin.Checkerboard()=in.Checkerboard();
+    Qin = Zero();
+    Pin = in;
+    for (int i=0;i<N;i++) {
+      const Field& tmp = evec[i];
+      auto ip = TensorRemove(innerProduct(tmp,in));
+      axpy(Qin, ip / eval[i],tmp,Qin);
+      axpy(Pin, -ip ,tmp,Pin);
+    }
+
+    this->_Smoother(Pin,out);
+
+    out = out + Qin;
   }
+};
 
-  /////////////////////////////////////////////////////////////////////
-  // Only Def1 has non-trivial Vout. Override in Def1
-  /////////////////////////////////////////////////////////////////////
-  virtual void   Vout  (Field & in, Field & out,Field & src){
-    out = in;
-    //case PcgDef1:
-    //    //Qb + PT x
-    //    ProjectToSubspace(src,PleftProj);     
-    //    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    //    PromoteFromSubspace(PleftMss_proj,tmp);  
-    //    
-    //    Pright(in,out);
-    //    
-    //    linop_d->axpy(out,tmp,out,1.0);
-    //    break;
-  }
+NAMESPACE_END(Grid);
 
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // Pright and Pleft are common to all implementations
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  virtual void Pright(Field & in,Field & out){
-    // P_R  = [ 1              0 ] 
-    //        [ -Mss^-1 Msb    0 ] 
-    Field in_sbar(grid);
-
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
-    axpy(in_sbar,-1.0,out,in);       // in_sbar = in - in_s 
-
-    HermOp(in_sbar,out);
-    ProjectToSubspace(out,PleftProj);           // Mssbar in_sbar  (project)
-
-    ApplyInverse     (PleftProj,PleftMss_proj); // Mss^{-1} Mssbar 
-    PromoteFromSubspace(PleftMss_proj,out);     // 
-
-    axpy(out,-1.0,out,in_sbar);     // in_sbar - Mss^{-1} Mssbar in_sbar
-  }
-  virtual void Pleft (Field & in,Field & out){
-    // P_L  = [ 1  -Mbs Mss^-1] 
-    //        [ 0   0         ] 
-    Field in_sbar(grid);
-    Field    tmp2(grid);
-    Field    Mtmp(grid);
-
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
-    axpy(in_sbar,-1.0,out,in);      // in_sbar = in - in_s
-
-    ApplyInverse(PleftProj,PleftMss_proj); // Mss^{-1} in_s
-    PromoteFromSubspace(PleftMss_proj,out);
-
-    HermOp(out,Mtmp);
-
-    ProjectToSubspace(Mtmp,PleftProj);      // Msbar s Mss^{-1}
-    PromoteFromSubspace(PleftProj,tmp2);
-
-    axpy(out,-1.0,tmp2,Mtmp);
-    axpy(out,-1.0,out,in_sbar);     // in_sbar - Msbars Mss^{-1} in_s
-  }
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgADef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp){
-
-  } 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp){
-
-  }
-  virtual void M2(Field & in, Field & out){
-
-  }
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp){
-
-  }
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp){
-
-  }
-}
-/*
-template<class Field>
-class TwoLevelFlexiblePcgAD : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef1 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-  virtual void   Vout  (Field & in, Field & out,Field & src,Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgV11: public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-*/
 #endif

Grid/algorithms/iterative/AdefMrhs.h

@@ -0,0 +1,734 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/AdefGeneric.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#pragma once
+
+
+  /*
+   * Compared to Tang-2009:  P=Pleft. P^T = PRight Q=MssInv. 
+   * Script A = SolverMatrix 
+   * Script P = Preconditioner
+   *
+   * Implement ADEF-2
+   *
+   * Vstart = P^Tx + Qb
+   * M1 = P^TM + Q
+   * M2=M3=1
+   */
+NAMESPACE_BEGIN(Grid);
+
+
+template<class Field>
+class TwoLevelCGmrhs
+{
+ public:
+  RealD   Tolerance;
+  Integer MaxIterations;
+  GridBase *grid;
+
+  // Fine operator, Smoother, CoarseSolver
+  LinearOperatorBase<Field>   &_FineLinop;
+  LinearFunction<Field>   &_Smoother;
+  MultiRHSBlockCGLinalg<Field> _BlockCGLinalg;
+
+  GridStopWatch ProjectTimer;
+  GridStopWatch PromoteTimer;
+  GridStopWatch DeflateTimer;
+  GridStopWatch CoarseTimer;
+  GridStopWatch FineTimer;
+  GridStopWatch SmoothTimer;
+  GridStopWatch InsertTimer;
+
+  /*
+    Field rrr;
+  Field sss;
+  Field qqq;
+  Field zzz;
+  */  
+  // more most opertor functions
+  TwoLevelCGmrhs(RealD tol,
+		 Integer maxit,
+		 LinearOperatorBase<Field>   &FineLinop,
+		 LinearFunction<Field>       &Smoother,
+		 GridBase *fine) : 
+    Tolerance(tol), 
+    MaxIterations(maxit),
+    _FineLinop(FineLinop),
+    _Smoother(Smoother)
+    /*
+    rrr(fine),
+    sss(fine),
+    qqq(fine),
+    zzz(fine)
+*/
+  {
+    grid       = fine;
+  };
+  
+  // Vector case
+  virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    SolveSingleSystem(src,x);
+	// SolvePrecBlockCG(src,x);
+  }
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Thin QR factorisation (google it)
+////////////////////////////////////////////////////////////////////////////////////////////////////
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  //Dimensions
+  // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
+  //
+  // Rdag R = m_rr = Herm = L L^dag        <-- Cholesky decomposition (LLT routine in Eigen)
+  //
+  //   Q  C = R => Q = R C^{-1}
+  //
+  // Want  Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock} 
+  //
+  // Set C = L^{dag}, and then Q^dag Q = ident 
+  //
+  // Checks:
+  // Cdag C = Rdag R ; passes.
+  // QdagQ  = 1      ; passes
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  void ThinQRfact (Eigen::MatrixXcd &m_zz,
+		   Eigen::MatrixXcd &C,
+		   Eigen::MatrixXcd &Cinv,
+		   std::vector<Field> &  Q,
+		   std::vector<Field> & MQ,
+		   const std::vector<Field> & Z,
+		   const std::vector<Field> & MZ)
+  {
+    RealD t0=usecond();
+    _BlockCGLinalg.InnerProductMatrix(m_zz,MZ,Z);
+    RealD t1=usecond();
+
+    m_zz = 0.5*(m_zz+m_zz.adjoint());
+    
+    Eigen::MatrixXcd L    = m_zz.llt().matrixL(); 
+    
+    C    = L.adjoint();
+    Cinv = C.inverse();
+    
+    RealD t3=usecond();
+    _BlockCGLinalg.MulMatrix( Q,Cinv,Z);
+    _BlockCGLinalg.MulMatrix(MQ,Cinv,MZ);
+    RealD t4=usecond();
+    std::cout << " ThinQRfact IP    :"<< t1-t0<<" us"<<std::endl;
+    std::cout << " ThinQRfact Eigen :"<< t3-t1<<" us"<<std::endl;
+    std::cout << " ThinQRfact MulMat:"<< t4-t3<<" us"<<std::endl;
+  }
+
+  virtual void SolvePrecBlockCG (std::vector<Field> &src, std::vector<Field> &X)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPrecBlockcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    //    std::vector<RealD> f(nrhs);
+    //    std::vector<RealD> rtzp(nrhs);
+    //    std::vector<RealD> rtz(nrhs);
+    //    std::vector<RealD> a(nrhs);
+    //    std::vector<RealD> d(nrhs);
+    //    std::vector<RealD> b(nrhs);
+    //    std::vector<RealD> rptzp(nrhs);
+
+    ////////////////////////////////////////////
+    //Initial residual computation & set up
+    ////////////////////////////////////////////
+    std::vector<RealD> ssq(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++){
+      ssq[rhs]=norm2(src[rhs]); GRID_ASSERT(ssq[rhs]!=0.0);
+    }      
+
+    ///////////////////////////
+    // Fields -- eliminate duplicates between fPcg and block cg
+    ///////////////////////////
+    std::vector<Field> Mtmp(nrhs,grid);
+    std::vector<Field> tmp(nrhs,grid);
+    std::vector<Field>   Z(nrhs,grid); // Rename Z to R
+    std::vector<Field>  MZ(nrhs,grid); // Rename MZ to Z
+    std::vector<Field>   Q(nrhs,grid); // 
+    std::vector<Field>  MQ(nrhs,grid); // Rename to P
+    std::vector<Field>   D(nrhs,grid);
+    std::vector<Field>  AD(nrhs,grid);
+    
+    /************************************************************************
+     * Preconditioned Block conjugate gradient rQ
+     * Generalise Sebastien Birk Thesis, after Dubrulle 2001.
+     * Introduce preconditioning following Saad Ch9
+     ************************************************************************
+     * Dimensions:
+     *
+     *   X,B etc... ==(Nferm x nrhs)
+     *  Matrix A==(Nferm x Nferm)
+     *  
+     * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
+     * QC => Thin QR factorisation (google it)
+     *
+     * R = B-AX
+     * Z = Mi R
+     * QC = Z
+     * D = Q 
+     * for k: 
+     *   R  = AD
+     *   Z  = Mi R
+     *   M  = [D^dag R]^{-1}
+     *   X  = X + D M C
+     *   QS = Q - Z.M
+     *   D  = Q + D S^dag
+     *   C  = S C
+     */
+    Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+    Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+    Eigen::MatrixXcd m_zz     = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    
+    Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    
+    Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+    Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+
+    GridStopWatch HDCGTimer;
+
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    Vstart(X,src);
+
+    //////////////////////////
+    // R = B-AX
+    //////////////////////////
+    for(int rhs=0;rhs<nrhs;rhs++){
+      // r0 = b -A x0
+      _FineLinop.HermOp(X[rhs],tmp[rhs]);
+      axpy (Z[rhs], -1.0,tmp[rhs], src[rhs]);    // Computes R=Z=src - A X0
+    }
+
+    //////////////////////////////////
+    // Compute MZ = M1 Z = M1 B - M1 A x0
+    //////////////////////////////////
+    PcgM1(Z,MZ);  
+
+    //////////////////////////////////
+    // QC = Z
+    //////////////////////////////////
+    ThinQRfact (m_zz, m_C, m_Cinv, Q, MQ, Z, MZ);
+
+    //////////////////////////////////
+    // D=MQ
+    //////////////////////////////////
+    for(int b=0;b<nrhs;b++) D[b]=MQ[b]; // LLT rotation of the MZ basis of search dirs
+
+    std::cout << GridLogMessage<<"PrecBlockCGrQ vec computed initial residual and QR fact " <<std::endl;
+
+    ProjectTimer.Reset();
+    PromoteTimer.Reset();
+    DeflateTimer.Reset();
+    CoarseTimer.Reset();
+    SmoothTimer.Reset();
+    FineTimer.Reset();
+    InsertTimer.Reset();
+
+    GridStopWatch M1Timer;
+    GridStopWatch M2Timer;
+    GridStopWatch M3Timer;
+    GridStopWatch LinalgTimer;
+    GridStopWatch InnerProdTimer;
+
+    HDCGTimer.Start();
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+
+      ////////////////////
+      // Z  = AD
+      ////////////////////
+      M3Timer.Start();
+      for(int b=0;b<nrhs;b++) _FineLinop.HermOp(D[b], Z[b]);      
+      M3Timer.Stop();
+
+      ////////////////////
+      // MZ  = M1 Z <==== the Multigrid preconditioner
+      ////////////////////
+      M1Timer.Start();
+      PcgM1(Z,MZ);
+      M1Timer.Stop();
+
+      FineTimer.Start();
+      ////////////////////
+      // M  = [D^dag Z]^{-1} = (<Ddag MZ>_M)^{-1} inner prod, generalising Saad derivation of Precon CG
+      ////////////////////
+      InnerProdTimer.Start();
+      _BlockCGLinalg.InnerProductMatrix(m_DZ,D,Z);
+      InnerProdTimer.Stop();
+      m_M       = m_DZ.inverse();
+
+      ///////////////////////////
+      // X  = X + D MC
+      ///////////////////////////
+      m_tmp     = m_M * m_C;
+      LinalgTimer.Start();
+      _BlockCGLinalg.MaddMatrix(X,m_tmp, D,X);     // D are the search directions and X takes the updates 
+      LinalgTimer.Stop();
+
+      ///////////////////////////
+      // QS = Q - M Z
+      // (MQ) S = MQ - M (M1Z)
+      ///////////////////////////
+      LinalgTimer.Start();
+      _BlockCGLinalg.MaddMatrix(tmp ,m_M, Z, Q,-1.0);
+      _BlockCGLinalg.MaddMatrix(Mtmp,m_M,MZ,MQ,-1.0);
+      ThinQRfact (m_zz, m_S, m_Sinv, Q, MQ, tmp, Mtmp);
+      LinalgTimer.Stop();
+
+      ////////////////////////////
+      // D  = MQ + D S^dag
+      ////////////////////////////
+      m_tmp = m_S.adjoint();
+      LinalgTimer.Start();
+      _BlockCGLinalg.MaddMatrix(D,m_tmp,D,MQ);
+      LinalgTimer.Stop();
+
+      ////////////////////////////
+      // C  = S C
+      ////////////////////////////
+      m_C = m_S*m_C;
+      
+      ////////////////////////////
+      // convergence monitor
+      ////////////////////////////
+      m_rr = m_C.adjoint() * m_C;
+      
+      FineTimer.Stop();
+
+      RealD max_resid=0;
+      RealD rrsum=0;
+      RealD sssum=0;
+      RealD rr;
+
+      for(int b=0;b<nrhs;b++) {
+	rrsum+=real(m_rr(b,b));
+	sssum+=ssq[b];
+	rr = real(m_rr(b,b))/ssq[b];
+	if ( rr > max_resid ) max_resid = rr;
+      }
+      std::cout << GridLogMessage <<
+	  "\t Prec BlockCGrQ Iteration "<<k<<" ave resid "<< std::sqrt(rrsum/sssum) << " max "<< std::sqrt(max_resid) <<std::endl;
+
+
+      if ( max_resid < Tolerance*Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Linalg  "<<LinalgTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : fine H  "<<M3Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : prec M1 "<<M1Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"**** M1 breakdown:"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Project "<<ProjectTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Promote "<<PromoteTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Deflate "<<DeflateTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Coarse  "<<CoarseTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Fine    "<<FineTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Smooth  "<<SmoothTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Insert  "<<InsertTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+
+	  _FineLinop.HermOp(X[rhs],tmp[rhs]);			  
+
+	  Field mytmp(grid);
+	  axpy(mytmp,-1.0,src[rhs],tmp[rhs]);
+      
+	  RealD  xnorm   = sqrt(norm2(X[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(mytmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<std::endl;
+	}
+	return;
+      }
+      
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: PrecBlockCGrQ not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    GRID_ASSERT(0);
+  }
+
+  virtual void SolveSingleSystem (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    std::vector<RealD> f(nrhs);
+    std::vector<RealD> rtzp(nrhs);
+    std::vector<RealD> rtz(nrhs);
+    std::vector<RealD> a(nrhs);
+    std::vector<RealD> d(nrhs);
+    std::vector<RealD> b(nrhs);
+    std::vector<RealD> rptzp(nrhs);
+    /////////////////////////////
+    // Set up history vectors
+    /////////////////////////////
+    int mmax = 3;
+
+    std::vector<std::vector<Field> > p(nrhs);   for(int r=0;r<nrhs;r++)  p[r].resize(mmax,grid);
+    std::vector<std::vector<Field> > mmp(nrhs); for(int r=0;r<nrhs;r++) mmp[r].resize(mmax,grid);
+    std::vector<std::vector<RealD> > pAp(nrhs); for(int r=0;r<nrhs;r++) pAp[r].resize(mmax);
+
+    std::vector<Field> z(nrhs,grid);
+    std::vector<Field>  mp (nrhs,grid);
+    std::vector<Field>  r  (nrhs,grid);
+    std::vector<Field>  mu (nrhs,grid);
+
+    //Initial residual computation & set up
+    std::vector<RealD> src_nrm(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      src_nrm[rhs]=norm2(src[rhs]);
+      GRID_ASSERT(src_nrm[rhs]!=0.0);
+    }
+    std::vector<RealD> tn(nrhs);
+
+    GridStopWatch HDCGTimer;
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    Vstart(x,src);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      // r0 = b -A x0
+      _FineLinop.HermOp(x[rhs],mmp[rhs][0]);
+      axpy (r[rhs], -1.0,mmp[rhs][0], src[rhs]);    // Recomputes r=src-Ax0
+    }
+
+    //////////////////////////////////
+    // Compute z = M1 x
+    //////////////////////////////////
+    // This needs a multiRHS version for acceleration
+    PcgM1(r,z);
+
+    std::vector<RealD> ssq(nrhs);
+    std::vector<RealD> rsq(nrhs);
+    std::vector<Field> pp(nrhs,grid);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+      p[rhs][0]=z[rhs];
+      ssq[rhs]=norm2(src[rhs]);
+      rsq[rhs]=  ssq[rhs]*Tolerance*Tolerance;
+      //      std::cout << GridLogMessage<<"mrhs HDCG: "<<rhs<<" k=0 residual "<<rtzp[rhs]<<" rsq "<<rsq[rhs]<<"\n";
+    }
+
+    ProjectTimer.Reset();
+    PromoteTimer.Reset();
+    DeflateTimer.Reset();
+    CoarseTimer.Reset();
+    SmoothTimer.Reset();
+    FineTimer.Reset();
+    InsertTimer.Reset();
+
+    GridStopWatch M1Timer;
+    GridStopWatch M2Timer;
+    GridStopWatch M3Timer;
+    GridStopWatch LinalgTimer;
+
+    HDCGTimer.Start();
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+    
+      int peri_k  = k % mmax;
+      int peri_kp = (k+1) % mmax;
+
+      for(int rhs=0;rhs<nrhs;rhs++){
+	rtz[rhs]=rtzp[rhs];
+	M3Timer.Start();
+	d[rhs]= PcgM3(p[rhs][peri_k],mmp[rhs][peri_k]);
+	M3Timer.Stop();
+	a[rhs] = rtz[rhs]/d[rhs];
+
+	LinalgTimer.Start();
+	// Memorise this
+	pAp[rhs][peri_k] = d[rhs];
+
+	axpy(x[rhs],a[rhs],p[rhs][peri_k],x[rhs]);
+	rn[rhs] = axpy_norm(r[rhs],-a[rhs],mmp[rhs][peri_k],r[rhs]);
+	LinalgTimer.Stop();
+      }
+
+      // Compute z = M x (for *all* RHS)
+      M1Timer.Start();
+      PcgM1(r,z);
+      M1Timer.Stop();
+      
+      RealD max_rn=0.0;
+      LinalgTimer.Start();
+      for(int rhs=0;rhs<nrhs;rhs++){
+
+	rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+
+	//	std::cout << GridLogMessage<<"HDCG::fPcg rhs"<<rhs<<" iteration "<<k<<" : inner rtzp "<<rtzp[rhs]<<"\n";
+	mu[rhs]=z[rhs];
+
+	p[rhs][peri_kp]=mu[rhs];
+
+	// Standard search direction p == z + b p 
+	b[rhs] = (rtzp[rhs])/rtz[rhs];
+
+	int northog = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
+	for(int back=0; back < northog; back++){
+	  int peri_back = (k-back)%mmax;
+	  RealD pbApk= real(innerProduct(mmp[rhs][peri_back],p[rhs][peri_kp]));
+	  RealD beta = -pbApk/pAp[rhs][peri_back];
+	  axpy(p[rhs][peri_kp],beta,p[rhs][peri_back],p[rhs][peri_kp]);
+	}
+
+	RealD rrn=sqrt(rn[rhs]/ssq[rhs]);
+	RealD rtn=sqrt(rtz[rhs]/ssq[rhs]);
+	RealD rtnp=sqrt(rtzp[rhs]/ssq[rhs]);
+	
+	std::cout<<GridLogMessage<<"HDCG:fPcg rhs "<<rhs<<" k= "<<k<<" residual = "<<rrn<<"\n";
+	if ( rrn > max_rn ) max_rn = rrn;
+      }
+      LinalgTimer.Stop();
+
+      // Stopping condition based on worst case
+      if ( max_rn <= Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Linalg  "<<LinalgTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : fine M3 "<<M3Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : prec M1 "<<M1Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"**** M1 breakdown:"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Project "<<ProjectTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Promote "<<PromoteTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Deflate "<<DeflateTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Coarse  "<<CoarseTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Fine    "<<FineTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Smooth  "<<SmoothTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Insert  "<<InsertTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+	  _FineLinop.HermOp(x[rhs],mmp[rhs][0]);			  
+	  Field tmp(grid);
+	  axpy(tmp,-1.0,src[rhs],mmp[rhs][0]);
+      
+	  RealD  mmpnorm = sqrt(norm2(mmp[rhs][0]));
+	  RealD  xnorm   = sqrt(norm2(x[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(tmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<" mmp "<<mmpnorm	  
+		   <<std::endl;
+	}
+	return;
+      }
+      
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    for(int rhs=0;rhs<nrhs;rhs++){
+      RealD  xnorm   = sqrt(norm2(x[rhs]));
+      RealD  srcnorm = sqrt(norm2(src[rhs]));
+      std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+    }
+  }
+  
+
+ public:
+
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out) = 0;
+  virtual void Vstart(std::vector<Field> & x,std::vector<Field> & src) = 0;
+  virtual void PcgM2(const Field & in, Field & out) {
+    out=in;
+  }
+
+  virtual RealD PcgM3(const Field & p, Field & mmp){
+    RealD dd;
+    _FineLinop.HermOp(p,mmp);
+    ComplexD dot = innerProduct(p,mmp);
+    dd=real(dot);
+    return dd;
+  }
+
+};
+
+template<class Field, class CoarseField>
+class TwoLevelADEF2mrhs : public TwoLevelCGmrhs<Field>
+{
+public:
+  GridBase *coarsegrid;
+  GridBase *coarsegridmrhs;
+  LinearFunction<CoarseField> &_CoarseSolverMrhs;
+  LinearFunction<CoarseField> &_CoarseSolverPreciseMrhs;
+  MultiRHSBlockProject<Field>    &_Projector;
+  MultiRHSDeflation<CoarseField> &_Deflator;
+
+  
+  TwoLevelADEF2mrhs(RealD tol,
+		    Integer maxit,
+		    LinearOperatorBase<Field>    &FineLinop,
+		    LinearFunction<Field>        &Smoother,
+		    LinearFunction<CoarseField>  &CoarseSolverMrhs,
+		    LinearFunction<CoarseField>  &CoarseSolverPreciseMrhs,
+		    MultiRHSBlockProject<Field>    &Projector,
+		    MultiRHSDeflation<CoarseField> &Deflator,
+		    GridBase *_coarsemrhsgrid) :
+    TwoLevelCGmrhs<Field>(tol, maxit,FineLinop,Smoother,Projector.fine_grid),
+    _CoarseSolverMrhs(CoarseSolverMrhs),
+    _CoarseSolverPreciseMrhs(CoarseSolverPreciseMrhs),
+    _Projector(Projector),
+    _Deflator(Deflator)
+  {
+    coarsegrid = Projector.coarse_grid;
+    coarsegridmrhs = _coarsemrhsgrid;// Thi could be in projector
+  };
+
+  // Override Vstart
+  virtual void Vstart(std::vector<Field> & x,std::vector<Field> & src)
+  {
+    int nrhs=x.size();
+    ///////////////////////////////////
+    // Choose x_0 such that 
+    // x_0 = guess +  (A_ss^inv) r_s = guess + Ass_inv [src -Aguess]
+    //                               = [1 - Ass_inv A] Guess + Assinv src
+    //                               = P^T guess + Assinv src 
+    //                               = Vstart  [Tang notation]
+    // This gives:
+    // W^T (src - A x_0) = src_s - A guess_s - r_s
+    //                   = src_s - (A guess)_s - src_s  + (A guess)_s 
+    //                   = 0 
+    ///////////////////////////////////
+    std::vector<CoarseField> PleftProj(nrhs,this->coarsegrid);
+    std::vector<CoarseField> PleftMss_proj(nrhs,this->coarsegrid);
+    CoarseField PleftProjMrhs(this->coarsegridmrhs);
+    CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
+
+    this->_Projector.blockProject(src,PleftProj);
+    this->_Deflator.DeflateSources(PleftProj,PleftMss_proj);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      InsertSliceFast(PleftProj[rhs],PleftProjMrhs,rhs,0);
+      InsertSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0); // the guess
+    }
+    
+    this->_CoarseSolverPreciseMrhs(PleftProjMrhs,PleftMss_projMrhs); // Ass^{-1} r_s
+
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      ExtractSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0);
+    }
+    this->_Projector.blockPromote(x,PleftMss_proj);
+  }
+
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out){
+
+    int nrhs=in.size();
+
+    // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
+    std::vector<Field> tmp(nrhs,this->grid);
+    std::vector<Field> Min(nrhs,this->grid);
+
+    std::vector<CoarseField> PleftProj(nrhs,this->coarsegrid);
+    std::vector<CoarseField> PleftMss_proj(nrhs,this->coarsegrid);
+
+    CoarseField PleftProjMrhs(this->coarsegridmrhs);
+    CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
+
+    //    this->rrr=in[0];
+
+#undef SMOOTHER_BLOCK_SOLVE
+#if SMOOTHER_BLOCK_SOLVE
+    this->SmoothTimer.Start();
+    this->_Smoother(in,Min);
+    this->SmoothTimer.Stop();
+#else
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      this->SmoothTimer.Start();
+      this->_Smoother(in[rhs],Min[rhs]);
+      this->SmoothTimer.Stop();
+    }
+#endif
+    //    this->sss=Min[0];
+    
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      
+      this->FineTimer.Start();
+      this->_FineLinop.HermOp(Min[rhs],out[rhs]);
+      axpy(tmp[rhs],-1.0,out[rhs],in[rhs]);          // resid  = in - A Min
+      this->FineTimer.Stop();
+
+    }
+
+    this->ProjectTimer.Start();
+    this->_Projector.blockProject(tmp,PleftProj);
+    this->ProjectTimer.Stop();
+    this->DeflateTimer.Start();
+    this->_Deflator.DeflateSources(PleftProj,PleftMss_proj);
+    this->DeflateTimer.Stop();
+    this->InsertTimer.Start();
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      InsertSliceFast(PleftProj[rhs],PleftProjMrhs,rhs,0);
+      InsertSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0); // the guess
+    }
+    this->InsertTimer.Stop();
+
+    this->CoarseTimer.Start();
+    this->_CoarseSolverMrhs(PleftProjMrhs,PleftMss_projMrhs); // Ass^{-1} [in - A Min]_s
+    this->CoarseTimer.Stop();
+
+    this->InsertTimer.Start();
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      ExtractSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0);
+    }
+    this->InsertTimer.Stop();
+    this->PromoteTimer.Start();
+    this->_Projector.blockPromote(tmp,PleftMss_proj);// tmp= Q[in - A Min]  
+    this->PromoteTimer.Stop();
+    this->FineTimer.Start();
+    //    this->qqq=tmp[0];
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      axpy(out[rhs],1.0,Min[rhs],tmp[rhs]); // Min+tmp
+    }
+    //    this->zzz=out[0];
+    this->FineTimer.Stop();
+  }
+};
+
+
+NAMESPACE_END(Grid);
+
+

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/BiCGSTAB.h

@@ -47,7 +47,7 @@ class BiCGSTAB : public OperatorFunction<Field>
   public:
     using OperatorFunction<Field>::operator();
     
-    bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+    bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                              // Defaults true.
     RealD Tolerance;
     Integer MaxIterations;
@@ -77,7 +77,7 @@ class BiCGSTAB : public OperatorFunction<Field>
 
       // Initial residual computation & set up
       RealD guess = norm2(psi);
-      assert(std::isnan(guess) == 0);
+      GRID_ASSERT(std::isnan(guess) == 0);
     
       Linop.Op(psi, v);
       b = norm2(v);
@@ -122,12 +122,14 @@ class BiCGSTAB : public OperatorFunction<Field>
 
         LinearCombTimer.Start();
         bo = beta * omega;
-        auto p_v = p.View();
-        auto r_v = r.View();
-        auto v_v = v.View();
-        accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
-        });
+	{
+	  autoView( p_v , p, AcceleratorWrite);
+	  autoView( r_v , r, AcceleratorRead);
+	  autoView( v_v , v, AcceleratorRead);
+	  accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
+	    });
+	}
         LinearCombTimer.Stop();
         LinalgTimer.Stop();
 
@@ -142,16 +144,20 @@ class BiCGSTAB : public OperatorFunction<Field>
         alpha = rho / Calpha.real();
 
         LinearCombTimer.Start();
-        auto h_v = h.View();
-        auto psi_v = psi.View();
-        accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
-        });
-        
-        auto s_v = s.View();
-        accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
-        });
+	{
+	  autoView( p_v , p, AcceleratorRead);
+	  autoView( r_v , r, AcceleratorRead);
+	  autoView( v_v , v, AcceleratorRead);
+	  autoView( psi_v,psi, AcceleratorRead);
+	  autoView( h_v  ,  h, AcceleratorWrite);
+	  autoView( s_v  ,  s, AcceleratorWrite);
+	  accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
+	    });
+	  accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
+ 	  });
+        }
         LinearCombTimer.Stop();
         LinalgTimer.Stop();
 
@@ -166,13 +172,19 @@ class BiCGSTAB : public OperatorFunction<Field>
         omega = Comega.real() / norm2(t);
 
         LinearCombTimer.Start();
-        auto t_v = t.View();
-        accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
-          coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
-          coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
-        });
+	{
+	  autoView( psi_v,psi, AcceleratorWrite);
+	  autoView( r_v , r, AcceleratorWrite);
+	  autoView( h_v , h, AcceleratorRead);
+	  autoView( s_v , s, AcceleratorRead);
+	  autoView( t_v , t, AcceleratorRead);
+	  accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
+	      coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
+	      coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
+	    });
+	}
         LinearCombTimer.Stop();
-
+	
         cp = norm2(r);
         LinalgTimer.Stop();
 
@@ -202,7 +214,7 @@ class BiCGSTAB : public OperatorFunction<Field>
           std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() << std::endl;
           std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() << std::endl;
 
-          if(ErrorOnNoConverge){ assert(true_residual / Tolerance < 10000.0); }
+          if(ErrorOnNoConverge){ GRID_ASSERT(true_residual / Tolerance < 10000.0); }
 
           IterationsToComplete = k;	
 
@@ -212,7 +224,7 @@ class BiCGSTAB : public OperatorFunction<Field>
       
       std::cout << GridLogMessage << "BiCGSTAB did NOT converge" << std::endl;
 
-      if(ErrorOnNoConverge){ assert(0); }
+      if(ErrorOnNoConverge){ GRID_ASSERT(0); }
       IterationsToComplete = k;
     }
 };

Grid/algorithms/iterative/BiCGSTABMixedPrec.h

@@ -36,7 +36,8 @@ NAMESPACE_BEGIN(Grid);
 template<class FieldD, class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
 class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD> 
 {
-  public:                                                
+  public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -31,6 +31,58 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+template<class Field>
+void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
+  typedef typename Field::scalar_type scomplex;
+  int Nblock = X.size();
+  for(int b=0;b<Nblock;b++){
+  for(int bp=0;bp<Nblock;bp++) {
+    m(b,bp) = innerProduct(X[b],Y[bp]);  
+  }}
+}
+template<class Field>
+void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
+  // Should make this cache friendly with site outermost, parallel_for
+  // Deal with case AP aliases with either Y or X
+  //
+  //Could pack "X" and "AP" into a Nblock x Volume dense array.
+  // AP(Nrhs x vol) = Y(Nrhs x vol) + scale * m(nrhs x nrhs) * X(nrhs*vol)
+  typedef typename Field::scalar_type scomplex;
+  int Nblock = AP.size();
+  std::vector<Field> tmp(Nblock,X[0]);
+  for(int b=0;b<Nblock;b++){
+    tmp[b]   = Y[b];
+    for(int bp=0;bp<Nblock;bp++) {
+      tmp[b] = tmp[b] +scomplex(scale*m(bp,b))*X[bp]; 
+    }
+  }
+  for(int b=0;b<Nblock;b++){
+    AP[b] = tmp[b];
+  }
+}
+template<class Field>
+void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
+  // Should make this cache friendly with site outermost, parallel_for
+  typedef typename Field::scalar_type scomplex;
+  int Nblock = AP.size();
+  for(int b=0;b<Nblock;b++){
+    AP[b] = Zero();
+    for(int bp=0;bp<Nblock;bp++) {
+      AP[b] += scomplex(m(bp,b))*X[bp]; 
+    }
+  }
+}
+template<class Field>
+double normv(const std::vector<Field> &P){
+  int Nblock = P.size();
+  double nn = 0.0;
+  for(int b=0;b<Nblock;b++) {
+    nn+=norm2(P[b]);
+  }
+  return nn;
+}
+
+
 enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
 
 //////////////////////////////////////////////////////////////////////////
@@ -46,7 +98,7 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
   int Nblock;
 
   BlockCGtype CGtype;
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -87,10 +139,19 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   sliceInnerProductMatrix(m_rr,R,R,Orthog);
 
   // Force manifest hermitian to avoid rounding related
+  /*
+  int rank=m_rr.rows();
+  for(int r=0;r<rank;r++){
+  for(int s=0;s<rank;s++){
+    std::cout << "QR m_rr["<<r<<","<<s<<"] "<<m_rr(r,s)<<std::endl;
+  }}
+  */
   m_rr = 0.5*(m_rr+m_rr.adjoint());
 
   Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
 
+//  ComplexD det = L.determinant();
+//  std::cout << " Det m_rr "<<det<<std::endl;
   C    = L.adjoint();
   Cinv = C.inverse();
   ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -110,11 +171,20 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
 		 const std::vector<Field> & R)
 {
   InnerProductMatrix(m_rr,R,R);
-
+  /*
+  int rank=m_rr.rows();
+  for(int r=0;r<rank;r++){
+  for(int s=0;s<rank;s++){
+    std::cout << "QRvec m_rr["<<r<<","<<s<<"] "<<m_rr(r,s)<<std::endl;
+  }}
+  */
   m_rr = 0.5*(m_rr+m_rr.adjoint());
 
   Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
 
+  //  ComplexD det = L.determinant();
+  //  std::cout << " Det m_rr "<<det<<std::endl;
+
   C    = L.adjoint();
   Cinv = C.inverse();
 
@@ -131,7 +201,7 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
   } else if (CGtype == CGmultiRHS ) {
     CGmultiRHSsolve(Linop,Src,Psi);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Field> &Src, std::vector<Field> &Psi) 
@@ -139,7 +209,7 @@ virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Fiel
   if ( CGtype == BlockCGrQVec ) {
     BlockCGrQsolveVec(Linop,Src,Psi);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 
@@ -186,12 +256,13 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   sliceNorm(ssq,B,Orthog);
   RealD sssum=0;
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+  for(int b=0;b<Nblock;b++) std::cout << "src["<<b<<"]" << ssq[b] <<std::endl;
 
   sliceNorm(residuals,B,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   sliceNorm(residuals,X,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   /************************************************************************
    * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
@@ -221,6 +292,9 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   Linop.HermOp(X, AD);
   tmp = B - AD;  
 
+  sliceNorm(residuals,tmp,Orthog);
+  for(int b=0;b<Nblock;b++) std::cout << "res["<<b<<"]" << residuals[b] <<std::endl;
+  
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
   D=Q;
 
@@ -236,6 +310,8 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   GridStopWatch SolverTimer;
   SolverTimer.Start();
 
+  RealD max_resid=0;
+
   int k;
   for (k = 1; k <= MaxIterations; k++) {
 
@@ -280,7 +356,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
      */
     m_rr = m_C.adjoint() * m_C;
 
-    RealD max_resid=0;
+    max_resid=0;
     RealD rrsum=0;
     RealD rr;
 
@@ -322,9 +398,11 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     }
 
   }
-  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge "<<k<<" / "<<MaxIterations
+	    <<" residual "<< std::sqrt(max_resid)<< std::endl;
+
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 //////////////////////////////////////////////////////////////////////////
@@ -360,10 +438,10 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
 
   sliceNorm(residuals,Src,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   sliceNorm(residuals,Psi,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   // Initial search dir is guess
   Linop.HermOp(Psi, AP);
@@ -462,47 +540,10 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   }
   std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 
-void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
-  for(int b=0;b<Nblock;b++){
-  for(int bp=0;bp<Nblock;bp++) {
-    m(b,bp) = innerProduct(X[b],Y[bp]);  
-  }}
-}
-void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
-  // Should make this cache friendly with site outermost, parallel_for
-  // Deal with case AP aliases with either Y or X
-  std::vector<Field> tmp(Nblock,X[0]);
-  for(int b=0;b<Nblock;b++){
-    tmp[b]   = Y[b];
-    for(int bp=0;bp<Nblock;bp++) {
-      tmp[b] = tmp[b] + scomplex(scale*m(bp,b))*X[bp]; 
-    }
-  }
-  for(int b=0;b<Nblock;b++){
-    AP[b] = tmp[b];
-  }
-}
-void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
-  // Should make this cache friendly with site outermost, parallel_for
-  for(int b=0;b<Nblock;b++){
-    AP[b] = Zero();
-    for(int bp=0;bp<Nblock;bp++) {
-      AP[b] += scomplex(m(bp,b))*X[bp]; 
-    }
-  }
-}
-double normv(const std::vector<Field> &P){
-  double nn = 0.0;
-  for(int b=0;b<Nblock;b++) {
-    nn+=norm2(P[b]);
-  }
-  return nn;
-}
-
 ////////////////////////////////////////////////////////////////////////////
 // BlockCGrQvec implementation:
 //--------------------------
@@ -513,7 +554,7 @@ double normv(const std::vector<Field> &P){
 void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field> &B, std::vector<Field> &X) 
 {
   Nblock = B.size();
-  assert(Nblock == X.size());
+  GRID_ASSERT(Nblock == X.size());
 
   std::cout<<GridLogMessage<<" Block Conjugate Gradient Vec rQ : Nblock "<<Nblock<<std::endl;
 
@@ -549,13 +590,14 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
 
   RealD sssum=0;
   for(int b=0;b<Nblock;b++){ ssq[b] = norm2(B[b]);}
+  for(int b=0;b<Nblock;b++){ std::cout << "ssq["<<b<<"] "<<ssq[b]<<std::endl;}
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
 
   for(int b=0;b<Nblock;b++){ residuals[b] = norm2(B[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   for(int b=0;b<Nblock;b++){ residuals[b] = norm2(X[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   /************************************************************************
    * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
@@ -585,6 +627,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   for(int b=0;b<Nblock;b++) {
     Linop.HermOp(X[b], AD[b]);
     tmp[b] = B[b] - AD[b];  
+    std::cout << "r0["<<b<<"] "<<norm2(tmp[b])<<std::endl;
   }
 
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
@@ -688,7 +731,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   }
   std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 

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/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when CAGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -82,7 +82,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -137,7 +137,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
     std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -185,7 +185,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/ConjugateGradient.h

@@ -38,13 +38,14 @@ NAMESPACE_BEGIN(Grid);
 // single input vec, single output vec.
 /////////////////////////////////////////////////////////////
 
+
 template <class Field>
 class ConjugateGradient : public OperatorFunction<Field> {
 public:
 
   using OperatorFunction<Field>::operator();
-
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -54,10 +55,26 @@ public:
   ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
     : Tolerance(tol),
       MaxIterations(maxit),
-      ErrorOnNoConverge(err_on_no_conv){};
+      ErrorOnNoConverge(err_on_no_conv)
+  {};
 
-  void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
+  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("ConjugateGradient");
+    GridStopWatch PreambleTimer;
+    GridStopWatch ConstructTimer;
+    GridStopWatch NormTimer;
+    GridStopWatch AssignTimer;
+    PreambleTimer.Start();
     psi.Checkerboard() = src.Checkerboard();
 
     conformable(psi, src);
@@ -65,22 +82,32 @@ public:
     RealD cp, c, a, d, b, ssq, qq;
     //RealD b_pred;
 
-    Field p(src);
-    Field mmp(src);
-    Field r(src);
+    // Was doing copies
+    ConstructTimer.Start();
+    Field p  (src.Grid());
+    Field mmp(src.Grid());
+    Field r  (src.Grid());
+    ConstructTimer.Stop();
 
     // Initial residual computation & set up
-    RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
-    
-    Linop.HermOpAndNorm(psi, mmp, d, b);
-    
-    r = src - mmp;
-    p = r;
-
-    a = norm2(p);
-    cp = a;
+    NormTimer.Start();
     ssq = norm2(src);
+    RealD guess = norm2(psi);
+    NormTimer.Stop();
+    GRID_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;
+      p = r;
+      a = norm2(p);
+    }
+    cp = a;
+    AssignTimer.Stop();
 
     // Handle trivial case of zero src
     if (ssq == 0.){
@@ -110,6 +137,7 @@ public:
     std::cout << GridLogIterative << std::setprecision(8)
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
+    PreambleTimer.Stop();
     GridStopWatch LinalgTimer;
     GridStopWatch InnerTimer;
     GridStopWatch AxpyNormTimer;
@@ -117,9 +145,13 @@ public:
     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();
@@ -140,43 +172,57 @@ public:
       b = cp / c;
 
       LinearCombTimer.Start();
-      auto psi_v = psi.View();
-      auto p_v   = p.View();
-      auto r_v   = r.View();
-      accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
-	  coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
-	  coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
-      });
+      {
+	autoView( psi_v , psi, AcceleratorWrite);
+	autoView( p_v   , p,   AcceleratorWrite);
+	autoView( r_v   , r,   AcceleratorWrite);
+	accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+	    coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+	    coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+	});
+      }
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
+      LogIteration(k,a,b);
 
-      std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+      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;
 
-        std::cout << GridLogIterative << "Time breakdown "<<std::endl;
-	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+	//	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;
 
-        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
+
+        if (ErrorOnNoConverge) GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
 	IterationsToComplete = k;	
 	TrueResidual = true_residual;
@@ -185,17 +231,143 @@ public:
       }
     }
     // Failed. Calculate true residual before giving up                                                         
-    Linop.HermOpAndNorm(psi, mmp, d, qq);
-    p = mmp - src;
+    // Linop.HermOpAndNorm(psi, mmp, d, qq);
+    //    p = mmp - src;
+    //TrueResidual = sqrt(norm2(p)/ssq);
+    //    TrueResidual = 1;
 
-    TrueResidual = sqrt(norm2(p)/ssq);
+    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;
 
-    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
-
-    if (ErrorOnNoConverge) assert(0);
+    if (ErrorOnNoConverge) GRID_ASSERT(0);
     IterationsToComplete = k;
 
   }
 };
+
+
+template <class Field>
+class ConjugateGradientPolynomial : public ConjugateGradient<Field> {
+public:
+  // Optionally record the CG polynomial
+  std::vector<double> ak;
+  std::vector<double> bk;
+  std::vector<double> poly_p;
+  std::vector<double> poly_r;
+  std::vector<double> poly_Ap;
+  std::vector<double> polynomial;
+
+public:
+  ConjugateGradientPolynomial(RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : ConjugateGradient<Field>(tol,maxit,err_on_no_conv)
+  { };
+  void PolyHermOp(LinearOperatorBase<Field> &Linop, 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;
+      Linop.HermOp(tmp,AtoN);
+      psi = psi + polynomial[n]*AtoN;
+    }
+  }
+  void CGsequenceHermOp(LinearOperatorBase<Field> &Linop, const Field &src, Field &x)
+  {
+    Field Ap(src.Grid());
+    Field r(src.Grid());
+    Field p(src.Grid());
+    p=src;
+    r=src;
+    x=Zero();
+    x.Checkerboard()=src.Checkerboard();
+    for(int k=0;k<ak.size();k++){
+      x = x + ak[k]*p;
+      Linop.HermOp(p,Ap);
+      r = r - ak[k] * Ap;
+      p = r + bk[k] * p;
+    }
+  }
+  void Solve(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
+  {
+    psi=Zero();
+    this->operator ()(Linop,src,psi);
+  }
+  virtual void LogBegin(void)
+  {
+    std::cout << "ConjugageGradientPolynomial::LogBegin() "<<std::endl;
+    ak.resize(0);
+    bk.resize(0);
+    polynomial.resize(0);
+    poly_Ap.resize(0);
+    poly_Ap.resize(0);
+    poly_p.resize(1);
+    poly_r.resize(1);
+    poly_p[0]=1.0;
+    poly_r[0]=1.0;
+  };
+  virtual void LogIteration(int k,RealD a,RealD b)
+  {
+    // With zero guess,
+    // p = r = src
+    //
+    // iterate:
+    //   x =  x + a p
+    //   r =  r - a A p
+    //   p =  r + b p
+    //
+    // [0]
+    // r = x
+    // p = x
+    // Ap=0
+    //
+    // [1]
+    // Ap = A x + 0  ==> shift poly P right by 1 and add 0.
+    // x  = x + a p  ==> add polynomials term by term 
+    // r  = r - a A p  ==> add polynomials term by term
+    // p  = r + b p  ==> add polynomials term by term
+    //
+    std::cout << "ConjugageGradientPolynomial::LogIteration() "<<k<<std::endl;
+    ak.push_back(a);
+    bk.push_back(b);
+    //  Ap= right_shift(p)
+    poly_Ap.resize(k+1);
+    poly_Ap[0]=0.0;
+    for(int i=0;i<k;i++){
+      poly_Ap[i+1]=poly_p[i];
+    }
+
+    //  x = x + a p
+    polynomial.resize(k);
+    polynomial[k-1]=0.0;
+    for(int i=0;i<k;i++){
+      polynomial[i] = polynomial[i] + a * poly_p[i];
+    }
+    
+    //  r = r - a Ap
+    //  p = r + b p
+    poly_r.resize(k+1);
+    poly_p.resize(k+1);
+    poly_r[k] = poly_p[k] = 0.0;
+    for(int i=0;i<k+1;i++){
+      poly_r[i] = poly_r[i] - a * poly_Ap[i];
+      poly_p[i] = poly_r[i] + b * poly_p[i];
+    }
+  }
+};
+
 NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -35,7 +35,8 @@ NAMESPACE_BEGIN(Grid);
     typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
-  public:                                                
+  public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;
@@ -48,6 +49,7 @@ NAMESPACE_BEGIN(Grid);
     Integer TotalInnerIterations; //Number of inner CG iterations
     Integer TotalOuterIterations; //Number of restarts
     Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
+    RealD TrueResidual;
 
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
@@ -67,6 +69,7 @@ NAMESPACE_BEGIN(Grid);
     }
   
   void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::cout << GridLogMessage << "MixedPrecisionConjugateGradient: Starting mixed precision CG with outer tolerance " << Tolerance << " and inner tolerance " << InnerTolerance << std::endl;
     TotalInnerIterations = 0;
 	
     GridStopWatch TotalTimer;
@@ -96,6 +99,7 @@ NAMESPACE_BEGIN(Grid);
     FieldF sol_f(SinglePrecGrid);
     sol_f.Checkerboard() = cb;
     
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting initial inner CG with tolerance " << inner_tol << std::endl;
     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
     CG_f.ErrorOnNoConverge = false;
 
@@ -104,22 +108,25 @@ NAMESPACE_BEGIN(Grid);
     GridStopWatch PrecChangeTimer;
     
     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
-      
+
+    precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
+    
     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
       RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
-      
+      std::cout<<GridLogMessage<<" rsd norm "<<norm<<std::endl;
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
 
       if(norm < OuterLoopNormMult * stop){
 	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
 	break;
       }
-      while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+      while(norm * inner_tol * inner_tol < stop*1.01) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d);
+      precisionChange(src_f, src_d, pc_wk_dp_to_sp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -129,6 +136,7 @@ NAMESPACE_BEGIN(Grid);
 	(*guesser)(src_f, sol_f);
 
       //Inner CG
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " << outer_iter << " starting inner CG with tolerance " << inner_tol << std::endl;
       CG_f.Tolerance = inner_tol;
       InnerCGtimer.Start();
       CG_f(Linop_f, src_f, sol_f);
@@ -137,7 +145,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f);
+      precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);
@@ -149,6 +157,7 @@ NAMESPACE_BEGIN(Grid);
     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
     CG_d(Linop_d, src_d_in, sol_d);
     TotalFinalStepIterations = CG_d.IterationsToComplete;
+    TrueResidual = CG_d.TrueResidual;
 
     TotalTimer.Stop();
     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -0,0 +1,213 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
+
+    Copyright (C) 2015
+
+    Author: Raoul Hodgson <raoul.hodgson@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_MIXED_PREC_BATCHED_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+
+NAMESPACE_BEGIN(Grid);
+
+//Mixed precision restarted defect correction CG
+template<class FieldD,class FieldF, 
+  typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+  typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
+public:
+  using LinearFunction<FieldD>::operator();
+  RealD   Tolerance;
+  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
+  Integer MaxInnerIterations;
+  Integer MaxOuterIterations;
+  Integer MaxPatchupIterations;
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+  LinearOperatorBase<FieldF> &Linop_f;
+  LinearOperatorBase<FieldD> &Linop_d;
+
+  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+  LinearFunction<FieldF> *guesser;
+  bool updateResidual;
+  
+  MixedPrecisionConjugateGradientBatched(RealD tol, 
+          Integer maxinnerit, 
+          Integer maxouterit, 
+          Integer maxpatchit,
+          GridBase* _sp_grid, 
+          LinearOperatorBase<FieldF> &_Linop_f, 
+          LinearOperatorBase<FieldD> &_Linop_d,
+          bool _updateResidual=true) :
+    Linop_f(_Linop_f), Linop_d(_Linop_d),
+    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
+    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
+
+  void useGuesser(LinearFunction<FieldF> &g){
+    guesser = &g;
+  }
+  
+  void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::vector<FieldD> srcs_d_in{src_d_in};
+    std::vector<FieldD> sols_d{sol_d};
+
+    (*this)(srcs_d_in,sols_d);
+
+    sol_d = sols_d[0];
+  }
+
+  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
+    GRID_ASSERT(src_d_in.size() == sol_d.size());
+    int NBatch = src_d_in.size();
+
+    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
+
+    Integer TotalOuterIterations = 0; //Number of restarts
+    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
+    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
+  
+    GridStopWatch TotalTimer;
+    TotalTimer.Start();
+
+    GridStopWatch InnerCGtimer;
+    GridStopWatch PrecChangeTimer;
+    
+    int cb = src_d_in[0].Checkerboard();
+    
+    std::vector<RealD> src_norm;
+    std::vector<RealD> norm;
+    std::vector<RealD> stop;
+    
+    GridBase* DoublePrecGrid = src_d_in[0].Grid();
+    FieldD tmp_d(DoublePrecGrid);
+    tmp_d.Checkerboard() = cb;
+    
+    FieldD tmp2_d(DoublePrecGrid);
+    tmp2_d.Checkerboard() = cb;
+
+    std::vector<FieldD> src_d;
+    std::vector<FieldF> src_f;
+    std::vector<FieldF> sol_f;
+
+    for (int i=0; i<NBatch; i++) {
+      sol_d[i].Checkerboard() = cb;
+
+      src_norm.push_back(norm2(src_d_in[i]));
+      norm.push_back(0.);
+      stop.push_back(src_norm[i] * Tolerance*Tolerance);
+
+      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
+
+      src_f.push_back(SinglePrecGrid);
+      src_f[i].Checkerboard() = cb;
+
+      sol_f.push_back(SinglePrecGrid);
+      sol_f[i].Checkerboard() = cb;
+    }
+    
+    RealD inner_tol = InnerTolerance;
+    
+    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+    CG_f.ErrorOnNoConverge = false;
+    
+    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      std::cout << GridLogMessage << std::endl;
+      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      
+      bool allConverged = true;
+      
+      for (int i=0; i<NBatch; i++) {
+        //Compute double precision rsd and also new RHS vector.
+        Linop_d.HermOp(sol_d[i], tmp_d);
+        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        
+        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
+
+        PrecChangeTimer.Start();
+        precisionChange(src_f[i], src_d[i]);
+        PrecChangeTimer.Stop();
+        
+        sol_f[i] = Zero();
+      
+        if(norm[i] > OuterLoopNormMult * stop[i]) {
+          allConverged = false;
+        }
+      }
+      if (allConverged) break;
+
+      if (updateResidual) {
+        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
+        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
+        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+        CG_f.Tolerance = inner_tol;
+      }
+
+      //Optionally improve inner solver guess (eg using known eigenvectors)
+      if(guesser != NULL) {
+        (*guesser)(src_f, sol_f);
+      }
+
+      for (int i=0; i<NBatch; i++) {
+        //Inner CG
+        InnerCGtimer.Start();
+        CG_f(Linop_f, src_f[i], sol_f[i]);
+        InnerCGtimer.Stop();
+        TotalInnerIterations[i] += CG_f.IterationsToComplete;
+        
+        //Convert sol back to double and add to double prec solution
+        PrecChangeTimer.Start();
+        precisionChange(tmp_d, sol_f[i]);
+        PrecChangeTimer.Stop();
+        
+        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
+      }
+
+    }
+    
+    //Final trial CG
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
+    
+    for (int i=0; i<NBatch; i++) {
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
+      CG_d(Linop_d, src_d_in[i], sol_d[i]);
+      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
+    }
+
+    TotalTimer.Stop();
+
+    std::cout << GridLogMessage << std::endl;
+    for (int i=0; i<NBatch; i++) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
+    }
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -44,7 +44,7 @@ public:
 
   using OperatorFunction<Field>::operator();
 
-  RealD   Tolerance;
+  //  RealD   Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -52,7 +52,7 @@ public:
   MultiShiftFunction shifts;
   std::vector<RealD> TrueResidualShift;
 
-  ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
+  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
     MaxIterations(maxit),
     shifts(_shifts)
   { 
@@ -84,6 +84,7 @@ public:
 
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
+    GRID_TRACE("ConjugateGradientMultiShift");
   
     GridBase *grid = src.Grid();
   
@@ -97,15 +98,15 @@ public:
     std::vector<RealD> alpha(nshift,1.0);
     std::vector<Field>   ps(nshift,grid);// Search directions
 
-    assert(psi.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(psi.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    // remove dynamic sized arrays on stack; 2d is a pain with vector
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -121,7 +122,7 @@ public:
   
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -143,7 +144,7 @@ public:
     for(int s=0;s<nshift;s++){
       rsq[s] = cp * mresidual[s] * mresidual[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShift: shift "<<s
-	       <<" target resid "<<rsq[s]<<std::endl;
+	       <<" target resid^2 "<<rsq[s]<<std::endl;
       ps[s] = src;
     }
     // r and p for primary
@@ -182,6 +183,9 @@ public:
     for(int s=0;s<nshift;s++) {
       axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
     }
+
+    std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
+    
   
   ///////////////////////////////////////
   // Timers
@@ -321,8 +325,8 @@ public:
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	
@@ -334,7 +338,7 @@ public:
     }
     // ugly hack
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
+    //  GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -0,0 +1,373 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+//For safety the a final regular CG is applied to clean up if necessary
+
+//PB Pure single, then double fixup
+
+template<class FieldD, class FieldF,
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class ConjugateGradientMultiShiftMixedPrecCleanup : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterationsMshift;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrecCleanup(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+    MaxIterations(20000)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GRID_TRACE("ConjugateGradientMultiShiftMixedPrecCleanup");
+    GridBase *DoublePrecGrid = src_d.Grid();
+
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldF> ps_f (nshift, SinglePrecGrid);// Search directions (single precision)
+    std::vector<FieldF> psi_f(nshift, SinglePrecGrid);// solutions (single precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldF r_f(SinglePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    GRID_ASSERT(psi_d.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<RealD>  rsqf(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF p_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      GRID_ASSERT( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	psi_f[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      rsqf[s] =rsq[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      //      ps_d[s] = src_d;
+      precisionChange(ps_f[s],src_d);
+    }
+    // r and p for primary
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
+    
+    //MdagM+m[0]
+    precisionChange(p_f,p_d);
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    precisionChange(tmp_d,mmp_f);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    //    GRID_ASSERT(norm2(tmp_d)< 1.0e-4);
+
+    axpy(mmp_d,mass[0],p_d,mmp_d);
+    RealD rn = norm2(p_d);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_d,b,mmp_d,r_d);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+      precisionChange(psi_f[s],psi_d[s]);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterationsMshift;k++){    
+
+      a = c /cp;
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChange(r_f, r_d);
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_f[s],a,ps_f[s],r_f);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_f[s],z[s][iz],as,r_f,ps_f[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      cp=c;
+      PrecChangeTimer.Start();
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f);
+      MatrixTimer.Stop();  
+      PrecChangeTimer.Start();
+      precisionChange(mmp_d, mmp_f); // From Float to Double
+      PrecChangeTimer.Stop();
+
+      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_d,mass[0],p_d,mmp_d);
+      RealD rn = norm2(p_d);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update single precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_f[ss],-bs[s]*alpha[s],ps_f[s],psi_f[ss]);
+	}
+      }
+      c = axpy_norm(r_d,b,mmp_d,r_d);
+      AXPYTimer.Stop();
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsqf[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged || k == MaxIterationsMshift-1){
+
+	SolverTimer.Stop();
+
+	for(int s=0;s<nshift;s++){
+	  precisionChange(psi_d[s],psi_f[s]);
+	}
+
+	
+	if ( all_converged ){
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: All shifts have converged iteration "<<k<<std::endl;
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Not all shifts have converged iteration "<<k<<std::endl;
+	}
+	
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrecCleanup: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+   
+    }
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    GRID_ASSERT(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -0,0 +1,416 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@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_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+    
+//For safety the a final regular CG is applied to clean up if necessary
+
+//Linop to add shift to input linop, used in cleanup CG
+namespace ConjugateGradientMultiShiftMixedPrecSupport{
+template<typename Field>
+class ShiftedLinop: public LinearOperatorBase<Field>{
+public:
+  LinearOperatorBase<Field> &linop_base;
+  RealD shift;
+
+  ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
+
+  void OpDiag (const Field &in, Field &out){ GRID_ASSERT(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ GRID_ASSERT(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ GRID_ASSERT(0); }
+  
+  void Op     (const Field &in, Field &out){ GRID_ASSERT(0); }
+  void AdjOp  (const Field &in, Field &out){ GRID_ASSERT(0); }
+
+  void HermOp(const Field &in, Field &out){
+    linop_base.HermOp(in, out);
+    axpy(out, shift, in, out);
+  }    
+
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+};
+};
+
+
+template<class FieldD, class FieldF,
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterationsMshift;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+    MaxIterations(20000)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
+    GridBase *DoublePrecGrid = src_d.Grid();
+
+    precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid);
+    
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    GRID_ASSERT(psi_d.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<RealD>  rsqf(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF p_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      GRID_ASSERT( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      rsqf[s] =rsq[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      ps_d[s] = src_d;
+    }
+    // r and p for primary
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
+    
+    //MdagM+m[0]
+    precisionChange(p_f, p_d, pc_wk_d_to_s);
+
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    precisionChange(tmp_d, mmp_f, pc_wk_s_to_d);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    GRID_ASSERT(norm2(tmp_d)< 1.0);
+
+    axpy(mmp_d,mass[0],p_d,mmp_d);
+    RealD rn = norm2(p_d);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_d,b,mmp_d,r_d);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterationsMshift;k++){    
+
+      a = c /cp;
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_d[s],a,ps_d[s],r_d);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+
+      cp=c;
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f);
+      MatrixTimer.Stop();  
+
+      PrecChangeTimer.Start();
+      precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_d,mass[0],p_d,mmp_d);
+      AXPYTimer.Stop();
+      RealD rn = norm2(p_d);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update double precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
+	}
+      }
+
+      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
+      c = axpy_norm(r_d,b,mmp_d,r_d);
+
+      AXPYTimer.Stop();
+
+      if(k % ReliableUpdateFreq == 0){
+	RealD c_old = c;
+	//Replace r with true residual
+	MatrixTimer.Start();  
+	Linop_d.HermOp(psi_d[0],mmp_d); 
+	MatrixTimer.Stop();  
+
+	AXPYTimer.Start();
+	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
+
+	c = axpy_norm(r_d, -1.0, mmp_d, src_d);
+	AXPYTimer.Stop();
+
+	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_old <<" with |r|^2 = "<<c<<std::endl;
+      }
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsqf[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged || k == MaxIterationsMshift-1){
+
+	SolverTimer.Stop();
+
+	if ( all_converged ){
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
+	}
+	
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+   
+    }
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    GRID_ASSERT(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -35,7 +35,7 @@ template<class FieldD,class FieldF,
 	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
 class ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
 public:
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
   // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -48,7 +48,7 @@ public:
   LinearOperatorBase<FieldF> &Linop_f;
   LinearOperatorBase<FieldD> &Linop_d;
   GridBase* SinglePrecGrid;
-  RealD Delta; //reliable update parameter
+  RealD Delta; //reliable update parameter. A reliable update is performed when the residual drops by a factor of Delta relative to its value at the last update
 
   //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
   LinearOperatorBase<FieldF> *Linop_fallback;
@@ -65,7 +65,9 @@ public:
       ErrorOnNoConverge(err_on_no_conv),
       DoFinalCleanup(true),
       Linop_fallback(NULL)
-  {};
+  {
+    GRID_ASSERT(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
+  };
 
   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
     Linop_fallback = &_Linop_fallback;
@@ -73,6 +75,7 @@ public:
   }
     
   void operator()(const FieldD &src, FieldD &psi) {
+    GRID_TRACE("ConjugateGradientReliableUpdate");
     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
     bool using_fallback = false;
       
@@ -87,7 +90,7 @@ public:
 
     // Initial residual computation & set up
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
     
     Linop_d.HermOpAndNorm(psi, mmp, d, b);
     
@@ -115,9 +118,12 @@ public:
     }
 
     //Single prec initialization
+    precisionChangeWorkspace pc_wk_sp_to_dp(src.Grid(), SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, src.Grid());
+    
     FieldF r_f(SinglePrecGrid);
     r_f.Checkerboard() = r.Checkerboard();
-    precisionChange(r_f, r);
+    precisionChange(r_f, r, pc_wk_dp_to_sp);
 
     FieldF psi_f(r_f);
     psi_f = Zero();
@@ -133,7 +139,8 @@ public:
     GridStopWatch LinalgTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
-
+    GridStopWatch PrecChangeTimer;
+    
     SolverTimer.Start();
     int k = 0;
     int l = 0;
@@ -172,7 +179,9 @@ public:
       // Stopping condition
       if (cp <= rsq) {
 	//Although not written in the paper, I assume that I have to add on the final solution
-	precisionChange(mmp, psi_f);
+	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
+	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 	
 	
@@ -193,7 +202,10 @@ public:
 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPrecChange " << PrecChangeTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPrecChange avg time " << PrecChangeTimer.Elapsed()/(2*l+1) <<std::endl;
 
+	
 	IterationsToComplete = k;	
 	ReliableUpdatesPerformed = l;
 	  
@@ -205,7 +217,7 @@ public:
 	  CG(Linop_d,src,psi);
 	  IterationsToCleanup = CG.IterationsToComplete;
 	}
-	else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+	else if (ErrorOnNoConverge) GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
 	return;
@@ -213,14 +225,21 @@ public:
       else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
 		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
-	precisionChange(mmp, psi_f);
+	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
+	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 
+	MatrixTimer.Start();
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	MatrixTimer.Stop();
+	
 	r = src - mmp;
 
 	psi_f = Zero();
-	precisionChange(r_f, r);
+	PrecChangeTimer.Start();
+	precisionChange(r_f, r, pc_wk_dp_to_sp);
+	PrecChangeTimer.Stop();
 	cp = norm2(r);
 	MaxResidSinceLastRelUp = cp;
 
@@ -244,7 +263,7 @@ public:
     std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
 	      << std::endl;
       
-    if (ErrorOnNoConverge) assert(0);
+    if (ErrorOnNoConverge) GRID_ASSERT(0);
     IterationsToComplete = k;
     ReliableUpdatesPerformed = l;      
   }    

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/ConjugateResidual.h

@@ -106,7 +106,7 @@ public:
     }
 
     std::cout<<GridLogMessage<<"ConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 };
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when FCAGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -87,7 +87,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -144,7 +144,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
     std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -191,7 +191,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when FGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -85,7 +85,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -142,7 +142,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
     std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -189,7 +189,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

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/GeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when GMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -80,7 +80,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -135,7 +135,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
     std::cout << GridLogMessage << "GeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -181,7 +181,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

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:
@@ -79,14 +91,16 @@ template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public Imp
     RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
 
     std::cout.precision(13);
-    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
-	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
-	     <<std::endl;
 
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
 
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<" target " << eresid*eresid << " conv " <<conv
+	     <<std::endl;
+
     return conv;
   }
 };
@@ -117,9 +131,10 @@ class ImplicitlyRestartedLanczos {
   /////////////////////////
   // Constructor
   /////////////////////////
-  
-public:       
+ public:
+  IRLeigsort EigSort;
 
+  
   //////////////////////////////////////////////////////////////////
   // PAB:
   //////////////////////////////////////////////////////////////////
@@ -152,6 +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,
@@ -168,6 +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)  { };
 
   ////////////////////////////////
@@ -209,7 +226,7 @@ until convergence
   void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
   {
     GridBase *grid = src.Grid();
-    assert(grid == evec[0].Grid());
+    GRID_ASSERT(grid == evec[0].Grid());
     
     //    GridLogIRL.TimingMode(1);
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
@@ -229,7 +246,7 @@ until convergence
     }
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
 	
-    assert(Nm <= evec.size() && Nm <= eval.size());
+    GRID_ASSERT(Nm <= evec.size() && Nm <= eval.size());
     
     // quickly get an idea of the largest eigenvalue to more properly normalize the residuum
     RealD evalMaxApprox = 0.0;
@@ -243,9 +260,10 @@ until convergence
 	_HermOp(src_n,tmp);
 	//	std::cout << GridLogMessage<< tmp<<std::endl; exit(0);
 	//	std::cout << GridLogIRL << " _HermOp " << norm2(tmp) << std::endl;
-	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+//	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vnum = real(innerProduct(tmp,tmp)); // HermOp^2.
 	RealD vden = norm2(src_n);
-	RealD na = vnum/vden;
+	RealD na = std::sqrt(vnum/vden);
 	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
 	  i=_MAX_ITER_IRL_MEVAPP_;
 	evalMaxApprox = na;
@@ -253,6 +271,7 @@ until convergence
 	src_n = tmp;
       }
     }
+    std::cout << GridLogIRL << " Final evalMaxApprox  " << evalMaxApprox << std::endl;
 	
     std::vector<RealD> lme(Nm);  
     std::vector<RealD> lme2(Nm);
@@ -312,8 +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 ;
@@ -329,11 +352,12 @@ 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;
 
-      assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
+      GRID_ASSERT(k2<Nm);      GRID_ASSERT(k2<Nm);      GRID_ASSERT(k1>0);
 
       basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
       std::cout<<GridLogIRL <<"basisRotated  by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
@@ -371,7 +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);	    
@@ -419,14 +444,15 @@ until convergence
 	}
       }
 
-      if ( Nconv < Nstop )
+      if ( Nconv < Nstop ) {
 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-
+	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
+      }
       eval=eval2;
       
       //Keep only converged
-      eval.resize(Nconv);// Nstop?
-      evec.resize(Nconv,grid);// Nstop?
+      eval.resize(Nstop);// was Nconv
+      evec.resize(Nstop,grid);// was Nconv
       basisSortInPlace(evec,eval,reverse);
       
     }
@@ -456,15 +482,15 @@ until convergence
 	    std::vector<Field>& evec,
 	    Field& w,int Nm,int k)
   {
-    std::cout<<GridLogIRL << "Lanczos step " <<k<<std::endl;
+    std::cout<<GridLogDebug << "Lanczos step " <<k<<std::endl;
     const RealD tiny = 1.0e-20;
-    assert( k< Nm );
+    GRID_ASSERT( k< Nm );
 
     GridStopWatch gsw_op,gsw_o;
 
     Field& evec_k = evec[k];
 
-    _PolyOp(evec_k,w);    std::cout<<GridLogIRL << "PolyOp" <<std::endl;
+    _PolyOp(evec_k,w);    std::cout<<GridLogDebug << "PolyOp" <<std::endl;
 
     if(k>0) w -= lme[k-1] * evec[k-1];
 
@@ -479,18 +505,18 @@ until convergence
     lme[k] = beta;
 
     if ( (k>0) && ( (k % orth_period) == 0 )) {
-      std::cout<<GridLogIRL << "Orthogonalising " <<k<<std::endl;
+      std::cout<<GridLogDebug << "Orthogonalising " <<k<<std::endl;
       orthogonalize(w,evec,k); // orthonormalise
-      std::cout<<GridLogIRL << "Orthogonalised " <<k<<std::endl;
+      std::cout<<GridLogDebug << "Orthogonalised " <<k<<std::endl;
     }
 
     if(k < Nm-1) evec[k+1] = w;
 
-    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+    std::cout<<GridLogIRL << "Lanczos step alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
     if ( beta < tiny ) 
       std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
 
-    std::cout<<GridLogIRL << "Lanczos step complete " <<k<<std::endl;
+    std::cout<<GridLogDebug << "Lanczos step complete " <<k<<std::endl;
   }
 
   void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
@@ -592,7 +618,7 @@ until convergence
     }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
       diagonalize_Eigen(lmd,lme,Nk,Nm,Qt,grid);
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
 
@@ -682,7 +708,7 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
     }
   }
 #else 
-  assert(0);
+  GRID_ASSERT(0);
 #endif
 }
 

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/LocalCoherenceLanczos.h

@@ -44,6 +44,7 @@ public:
 				  int, MinRes);    // Must restart
 };
 
+//This class is the input parameter class for some testing programs
 struct LocalCoherenceLanczosParams : Serializable {
 public:
   GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
@@ -67,6 +68,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -78,7 +80,7 @@ public:
   ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) : 
     _Linop(linop), subspace(_subspace)
   {  
-    assert(subspace.size() >0);
+    GRID_ASSERT(subspace.size() >0);
   };
 
   void operator()(const CoarseField& in, CoarseField& out) {
@@ -97,6 +99,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -143,16 +146,24 @@ public:
   LinearOperatorBase<FineField> &_Linop;
   RealD                             _coarse_relax_tol;
   std::vector<FineField>        &_subspace;
+
+  int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
+                                //As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
+                                //To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
+                                //out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
+                                //NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
   
   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
 					   OperatorFunction<FineField>   &smoother,
 					   LinearOperatorBase<FineField> &Linop,
 					   std::vector<FineField>        &subspace,
-					   RealD coarse_relax_tol=5.0e3) 
+					   RealD coarse_relax_tol=5.0e3,
+					   int largestEvalIdxForReport=-1) 
     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
-      _coarse_relax_tol(coarse_relax_tol)  
+      _coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
   {    };
 
+  //evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
   int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
     CoarseField v(B);
@@ -175,12 +186,26 @@ public:
 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
 	     <<std::endl;
 
+    if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
+      std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
+      RealD tmp_eval;
+      ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
+    }
+    
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
     return conv;
   }
-  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+
+  //This function is called at the end of the coarse grid Lanczos. It promotes the coarse eigenvector 'B' to the fine grid,
+  //applies a smoother to the result then computes the computes the *fine grid* eigenvalue (output as 'eval').
+
+  //evalMaxApprox should be the approximation of the largest eval of the fine Hermop. However when this function is called by IRL it actually passes the largest eval of the *Chebyshev* operator (as this is the max approx used for the TestConvergence above)
+  //As the largest eval of the Chebyshev is typically several orders of magnitude larger this makes the convergence test pass even when it should not.
+  //We therefore ignore evalMaxApprox here and use a value of 1.0 (note this value is already used by TestCoarse)
+  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)  
   {
+    evalMaxApprox = 1.0; //cf above
     GridBase *FineGrid = _subspace[0].Grid();    
     int checkerboard   = _subspace[0].Checkerboard();
     FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
@@ -199,13 +224,13 @@ public:
     eval   = vnum/vden;
     fv -= eval*fB;
     RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
-
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
+    
     std::cout.precision(13);
     std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
 	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv << " target " << eresid*eresid
 	     <<std::endl;
-    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
     if( (vv<eresid*eresid) ) return 1;
     return 0;
   }
@@ -283,6 +308,10 @@ public:
     evals_coarse.resize(0);
   };
 
+  //The block inner product is the inner product on the fine grid locally summed over the blocks
+  //to give a Lattice<Scalar> on the coarse grid. This function orthnormalizes the fine-grid subspace
+  //vectors under the block inner product. This step must be performed after computing the fine grid
+  //eigenvectors and before computing the coarse grid eigenvectors.    
   void Orthogonalise(void ) {
     CoarseScalar InnerProd(_CoarseGrid);
     std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
@@ -317,19 +346,21 @@ public:
 
   void testFine(RealD resid) 
   {
-    assert(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
+    GRID_ASSERT(evals_fine.size() == nbasis);
+    GRID_ASSERT(subspace.size() == nbasis);
     PlainHermOp<FineField>    Op(_FineOp);
     ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
     for(int k=0;k<nbasis;k++){
-      assert(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
+      GRID_ASSERT(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
     }
   }
 
+  //While this method serves to check the coarse eigenvectors, it also recomputes the eigenvalues from the smoothed reconstructed eigenvectors
+  //hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
   {
-    assert(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
+    GRID_ASSERT(evals_fine.size() == nbasis);
+    GRID_ASSERT(subspace.size() == nbasis);
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
@@ -349,7 +380,7 @@ public:
   void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
 		RealD MaxIt, RealD betastp, int MinRes)
   {
-    assert(nbasis<=Nm);
+    GRID_ASSERT(nbasis<=Nm);
     Chebyshev<FineField>      Cheby(cheby_parms);
     FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
     PlainHermOp<FineField>    Op(_FineOp);
@@ -369,40 +400,54 @@ public:
     IRL.calc(evals_fine,subspace,src,Nconv,false);
     
     // Shrink down to number saved
-    assert(Nstop>=nbasis);
-    assert(Nconv>=nbasis);
+    GRID_ASSERT(Nstop>=nbasis);
+    GRID_ASSERT(Nconv>=nbasis);
     evals_fine.resize(nbasis);
     subspace.resize(nbasis,_FineGrid);
   }
+
+
+  //cheby_op: Parameters of the fine grid Chebyshev polynomial used for the Lanczos acceleration
+  //cheby_smooth: Parameters of a separate Chebyshev polynomial used after the Lanczos has completed to smooth out high frequency noise in the reconstructed fine grid eigenvectors prior to computing the eigenvalue
+  //relax: Reconstructed eigenvectors (post smoothing) are naturally not as precise as true eigenvectors. This factor acts as a multiplier on the stopping condition when determining whether the results satisfy the user provided stopping condition
   void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
 		  int Nstop, int Nk, int Nm,RealD resid, 
 		  RealD MaxIt, RealD betastp, int MinRes)
   {
-    Chebyshev<FineField>                          Cheby(cheby_op);
-    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace);
-    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
+    Chebyshev<FineField>                          Cheby(cheby_op); //Chebyshev of fine operator on fine grid
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace); //Fine operator on coarse grid with intermediate fine grid conversion
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); //Chebyshev of fine operator on coarse grid with intermediate fine grid conversion
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
 
-    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1); 
 
     evals_coarse.resize(Nm);
     evec_coarse.resize(Nm,_CoarseGrid);
 
     CoarseField src(_CoarseGrid);     src=1.0; 
 
+    //Note the "tester" here is also responsible for generating the fine grid eigenvalues which are output into the "evals_coarse" array
     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
     int Nconv=0;
     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
-    assert(Nconv>=Nstop);
+    GRID_ASSERT(Nconv>=Nstop);
     evals_coarse.resize(Nstop);
     evec_coarse.resize (Nstop,_CoarseGrid);
     for (int i=0;i<Nstop;i++){
       std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl;
     }
   }
+
+  //Get the fine eigenvector 'i' by reconstruction
+  void getFineEvecEval(FineField &evec, RealD &eval, const int i) const{
+    blockPromote(evec_coarse[i],evec,subspace);  
+    eval = evals_coarse[i];
+  }
+    
+    
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/MinimalResidual.h

@@ -35,7 +35,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
+  bool ErrorOnNoConverge; // throw an GRID_ASSERT when the MR fails to converge.
                           // Defaults true.
   RealD   Tolerance;
   Integer MaxIterations;
@@ -59,7 +59,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
 
     // Initial residual computation & set up
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD ssq = norm2(src);
     RealD rsq = Tolerance * Tolerance * ssq;
@@ -136,7 +136,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
         std::cout << GridLogMessage << "MR Time elapsed: Linalg  " << LinalgTimer.Elapsed() << std::endl;
 
         if (ErrorOnNoConverge)
-          assert(true_residual / Tolerance < 10000.0);
+          GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
         IterationsToComplete = k;
 
@@ -148,7 +148,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
               << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
 
     IterationsToComplete = k;
   }

Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h

@@ -37,7 +37,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
 
   using OperatorFunction<FieldD>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when MPFGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -91,7 +91,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -150,7 +150,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
     std::cout << GridLogMessage << "MPFGMRES did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi, RealD rsq) {
@@ -197,7 +197,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/NormalEquations.h

@@ -33,7 +33,7 @@ NAMESPACE_BEGIN(Grid);
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
 // Take a matrix and form an NE solver calling a Herm solver
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class NormalEquations {
+template<class Field> class NormalEquations : public LinearFunction<Field>{
 private:
   SparseMatrixBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;
@@ -60,7 +60,33 @@ public:
   }     
 };
 
-template<class Field> class HPDSolver {
+template<class Field> class NormalResidual : public LinearFunction<Field>{
+private:
+  SparseMatrixBase<Field> & _Matrix;
+  OperatorFunction<Field> & _HermitianSolver;
+  LinearFunction<Field>   & _Guess;
+public:
+
+  /////////////////////////////////////////////////////
+  // Wrap the usual normal equations trick
+  /////////////////////////////////////////////////////
+ NormalResidual(SparseMatrixBase<Field> &Matrix, OperatorFunction<Field> &HermitianSolver,
+		 LinearFunction<Field> &Guess) 
+   :  _Matrix(Matrix), _HermitianSolver(HermitianSolver), _Guess(Guess) {}; 
+
+  void operator() (const Field &in, Field &out){
+ 
+    Field res(in.Grid());
+    Field tmp(in.Grid());
+
+    MMdagLinearOperator<SparseMatrixBase<Field>,Field> MMdagOp(_Matrix);
+    _Guess(in,res);
+    _HermitianSolver(MMdagOp,in,res);  // M Mdag res = in ;
+    _Matrix.Mdag(res,out);             // out = Mdag res
+  }     
+};
+
+template<class Field> class HPDSolver : public LinearFunction<Field> {
 private:
   LinearOperatorBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;
@@ -78,13 +104,13 @@ public:
   void operator() (const Field &in, Field &out){
  
     _Guess(in,out);
-    _HermitianSolver(_Matrix,in,out);  // Mdag M out = Mdag in
+    _HermitianSolver(_Matrix,in,out);  //M out = in
 
   }     
 };
 
 
-template<class Field> class MdagMSolver {
+template<class Field> class MdagMSolver : public LinearFunction<Field> {
 private:
   SparseMatrixBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;

Grid/algorithms/iterative/PowerMethod.h

@@ -20,7 +20,7 @@ template<class Field> class PowerMethod
     RealD evalMaxApprox = 0.0; 
     auto src_n = src; 
     auto tmp = src; 
-    const int _MAX_ITER_EST_ = 50; 
+    const int _MAX_ITER_EST_ = 200; 
 
     for (int i=0;i<_MAX_ITER_EST_;i++) { 
       
@@ -29,17 +29,18 @@ template<class Field> class PowerMethod
       RealD vnum = real(innerProduct(src_n,tmp)); // HermOp. 
       RealD vden = norm2(src_n); 
       RealD na = vnum/vden; 
+
+      std::cout << GridLogMessage << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
       
-      if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) { 
- 	evalMaxApprox = na; 
-	std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
- 	return evalMaxApprox; 
-      } 
+      //      if ( (fabs(evalMaxApprox/na - 1.0) < 0.0001) || (i==_MAX_ITER_EST_-1) ) { 
+	// 	evalMaxApprox = na; 
+	// 	return evalMaxApprox; 
+      //      } 
       evalMaxApprox = na; 
       src_n = tmp;
     }
-    assert(0);
-    return 0;
+    std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+    return evalMaxApprox;
   }
 };
 }

Grid/algorithms/iterative/PowerSpectrum.h

@@ -0,0 +1,76 @@
+#pragma once
+namespace Grid {
+
+class Band
+{
+  RealD lo, hi;
+public:
+  Band(RealD _lo,RealD _hi)
+  {
+    lo=_lo;
+    hi=_hi;
+  }
+  RealD operator() (RealD x){
+    if ( x>lo && x<hi ){
+      return 1.0;
+    } else {
+      return 0.0;
+    }
+  }
+};
+
+class PowerSpectrum
+{ 
+ public: 
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  std::vector<RealD> ranges;
+  std::vector<int> order;
+  
+  PowerSpectrum(  std::vector<RealD> &bins, std::vector<int> &_order ) : ranges(bins), order(_order)  { };
+
+  template<class Field>
+  RealD operator()(LinearOperatorBase<Field> &HermOp, const Field &src) 
+  { 
+    GridBase *grid = src.Grid(); 
+    int N=ranges.size();
+    RealD hi = ranges[N-1];
+
+    RealD lo_band = 0.0;
+    RealD hi_band;
+    RealD nn=norm2(src);
+    RealD ss=0.0;
+
+    Field tmp = src;
+
+    for(int b=0;b<N;b++){
+      hi_band = ranges[b];
+      Band Notch(lo_band,hi_band);
+      
+      Chebyshev<Field> polynomial;
+      polynomial.Init(0.0,hi,order[b],Notch);
+      polynomial.JacksonSmooth();
+
+      polynomial(HermOp,src,tmp) ;
+
+      RealD p=norm2(tmp);
+      ss=ss+p;
+      std::cout << GridLogMessage << " PowerSpectrum Band["<<lo_band<<","<<hi_band<<"] power "<<norm2(tmp)/nn<<std::endl;
+      
+      lo_band=hi_band;
+    }
+    std::cout << GridLogMessage << " PowerSpectrum total power "<<ss/nn<<std::endl;
+    std::cout << GridLogMessage << " PowerSpectrum total power (unnormalised) "<<nn<<std::endl;
+
+    return 0;
+  };
+};
+  
+}

Grid/algorithms/iterative/PrecConjugateResidual.h

@@ -112,7 +112,7 @@ public:
     }
 
     std::cout<<GridLogMessage<<"PrecConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 };
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
 template<class Field>
 class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
 public:                                                
-
+  using LinearFunction<Field>::operator();
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;
@@ -118,7 +118,7 @@ public:
 
     }
     GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    assert(0);
+    //    GRID_ASSERT(0);
   }
 
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
@@ -221,7 +221,7 @@ public:
       int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
       for(int back=0;back<northog;back++){
 
-	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+	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];
@@ -231,7 +231,7 @@ public:
       qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
     }
-    assert(0); // never reached
+    GRID_ASSERT(0); // never reached
     return cp;
   }
 };

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -0,0 +1,450 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/PrecGeneralisedConjugateResidual.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_PREC_GCR_NON_HERM_H
+#define GRID_PREC_GCR_NON_HERM_H
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+//VPGCR Abe and Zhang, 2005.
+//INTERNATIONAL JOURNAL OF NUMERICAL ANALYSIS AND MODELING
+//Computing and Information Volume 2, Number 2, Pages 147-161
+//NB. Likely not original reference since they are focussing on a preconditioner variant.
+//    but VPGCR was nicely written up in their paper
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+
+#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" " 
+
+template<class Field>
+class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
+public:                                                
+  using LinearFunction<Field>::operator();
+  RealD   Tolerance;
+  Integer MaxIterations;
+  int verbose;
+  int mmax;
+  int nstep;
+  int steps;
+  int level;
+  GridStopWatch PrecTimer;
+  GridStopWatch MatTimer;
+  GridStopWatch LinalgTimer;
+
+  LinearFunction<Field>     &Preconditioner;
+  LinearOperatorBase<Field> &Linop;
+
+  void Level(int lv) { level=lv; };
+
+  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax, int _nstep) : 
+    Tolerance(tol), 
+    MaxIterations(maxit),
+    Linop(_Linop),
+    Preconditioner(Prec),
+    mmax(_mmax),
+    nstep(_nstep)         // 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();
+    RealD cp, ssq,rsq;
+    ssq=norm2(src);
+    rsq=Tolerance*Tolerance*ssq;
+      
+    Field r(src.Grid());
+
+    PrecTimer.Reset();
+    MatTimer.Reset();
+    LinalgTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    steps=0;
+    for(int k=0;k<MaxIterations;k++){
+
+      cp=GCRnStep(src,psi,rsq);
+
+      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
+
+      if(cp<rsq) {
+        SolverTimer.Stop();
+
+        Linop.Op(psi,r);
+        axpy(r,-1.0,src,r);
+        RealD tr = norm2(r);
+        GCRLogLevel<<"PGCR: Converged on iteration " <<steps
+          << " computed residual "<<sqrt(cp/ssq)
+          << " true residual "    <<sqrt(tr/ssq)
+          << " target "           <<Tolerance <<std::endl;
+
+        GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
+        return;
+      }
+
+    }
+    GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
+    //    GRID_ASSERT(0);
+  }
+
+  RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
+
+    RealD cp;
+    ComplexD a, b;
+    //    ComplexD zAz;
+    RealD zAAz;
+    ComplexD rq;
+
+    GridBase *grid = src.Grid();
+
+    Field r(grid);
+    Field z(grid);
+    Field tmp(grid);
+    Field ttmp(grid);
+    Field Az(grid);
+
+    ////////////////////////////////
+    // history for flexible orthog
+    ////////////////////////////////
+    std::vector<Field> q(mmax,grid);        // 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;
+
+    //////////////////////////////////
+    // initial guess x0 is taken as nonzero.
+    // r0=src-A x0 = src
+    //////////////////////////////////
+    MatTimer.Start();
+    Linop.Op(psi,Az);
+    //    zAz = innerProduct(Az,psi);
+    zAAz= norm2(Az);
+    MatTimer.Stop();
+    
+
+    LinalgTimer.Start();
+    r=src-Az;
+    LinalgTimer.Stop();
+    GCRLogLevel<< "PGCR true residual r = src - A psi   "<< norm2(r) <<std::endl;
+
+    this->LogBegin();       // initialize polynomial GCR if needed (TODO think about placement of this)
+    
+    /////////////////////
+    // p = Prec(r)
+    /////////////////////
+
+    PrecTimer.Start();
+    Preconditioner(r,z);
+    PrecTimer.Stop();
+
+    MatTimer.Start();
+    Linop.Op(z,Az);
+    MatTimer.Stop();
+
+    LinalgTimer.Start();
+
+    //    zAz = innerProduct(Az,psi);
+    zAAz= norm2(Az);
+
+    //p[0],q[0],qq[0] 
+    p[0]= z;
+    q[0]= Az;
+    qq[0]= zAAz;
+
+    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;     // 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];              // compute alpha_j
+
+      all_alphas.push_back(a);
+
+      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();
+
+      // LogIterationA(k + 1, a);
+
+      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();
+      Preconditioner(r,z);// solve Az = r
+      PrecTimer.Stop();
+
+      MatTimer.Start();
+      Linop.Op(z,Az);
+      MatTimer.Stop();
+      //      zAz = innerProduct(Az,psi);
+      zAAz= norm2(Az);
+
+      LinalgTimer.Start();
+
+      q[peri_kp]=Az;
+      p[peri_kp]=z;
+
+      // 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];
+        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/QuasiMinimalResidual.h

@@ -79,7 +79,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
 
     LinOp.Op(x,r); r = b - r;
 
-    assert(normb> 0.0);
+    GRID_ASSERT(normb> 0.0);
 
     resid = norm2(r)/normb;
     if (resid <= Tolerance) {
@@ -105,8 +105,8 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
     for (int i = 1; i <= MaxIterations; i++) {
 
       // Breakdown tests
-      assert( rho != 0.0);
-      assert( xi  != 0.0);
+      GRID_ASSERT( rho != 0.0);
+      GRID_ASSERT( xi  != 0.0);
 
       v = (1. / rho) * v_tld;
       y = (1. / rho) * y;
@@ -134,10 +134,10 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       ep=Zep.real();
       std::cout << "Zep "<<Zep <<std::endl;
       // Complex Audit
-      assert(abs(ep)>0);
+      GRID_ASSERT(abs(ep)>0);
 
       beta = ep / delta;
-      assert(abs(beta)>0);
+      GRID_ASSERT(abs(beta)>0);
 
       v_tld = p_tld - beta * v;
       y = v_tld;
@@ -158,7 +158,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       std::cout << "theta "<<theta<<std::endl;
       std::cout << "gamma "<<gamma<<std::endl;
 
-      assert(abs(gamma)> 0.0);
+      GRID_ASSERT(abs(gamma)> 0.0);
 
       eta = -eta * rho_1 * gamma* gamma / (beta * gamma_1 * gamma_1);
 
@@ -178,7 +178,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       }
       std::cout << "Iteration "<<i<<" resid " << resid<<std::endl;
     }
-    assert(0);
+    GRID_ASSERT(0);
     return;                            // no convergence
   }
 #else

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/SchurRedBlack.h

@@ -132,6 +132,31 @@ namespace Grid {
       (*this)(_Matrix,in,out,guess);
     }
 
+    void RedBlackSource(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &src_o) 
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      Field tmp(grid);
+      int nblock = in.size();
+      for(int b=0;b<nblock;b++){
+	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+      }
+    }
+    // James can write his own deflated guesser
+    // with optimised code for the inner products
+    //    RedBlackSolveSplitGrid();
+    //    RedBlackSolve(_Matrix,src_o,sol_o); 
+
+    void RedBlackSolution(Matrix &_Matrix, const std::vector<Field> &in, const std::vector<Field> &sol_o, std::vector<Field> &out)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      Field tmp(grid);
+      int nblock = in.size();
+      for(int b=0;b<nblock;b++) {
+	pickCheckerboard(Even,tmp,in[b]);
+	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]);
+      }
+    }
+
     template<class Guesser>
     void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess) 
     {
@@ -150,24 +175,29 @@ namespace Grid {
       ////////////////////////////////////////////////
       // Prepare RedBlack source
       ////////////////////////////////////////////////
-      for(int b=0;b<nblock;b++){
-	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
-      }
+      RedBlackSource(_Matrix,in,src_o);
+	//      for(int b=0;b<nblock;b++){
+	//	RedBlackSource(_Matrix,in[b],tmp,src_o[b]);
+	//      }
+      
       ////////////////////////////////////////////////
       // Make the guesses
       ////////////////////////////////////////////////
       if ( subGuess ) guess_save.resize(nblock,grid);
 
-      for(int b=0;b<nblock;b++){
-        if(useSolnAsInitGuess) {
+      
+      if(useSolnAsInitGuess) {
+        for(int b=0;b<nblock;b++){
           pickCheckerboard(Odd, sol_o[b], out[b]);
-        } else {
-          guess(src_o[b],sol_o[b]); 
         }
+      } else {
+        guess(src_o, sol_o); 
+      }
 
-	if ( subGuess ) { 
-	  guess_save[b] = sol_o[b];
-	}
+	    if ( subGuess ) { 
+        for(int b=0;b<nblock;b++){
+          guess_save[b] = sol_o[b];
+        }
       }
       //////////////////////////////////////////////////////////////
       // Call the block solver
@@ -297,9 +327,9 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
     }
@@ -317,17 +347,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.Checkerboard()   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o,tmp);        GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      src_e = src_e-tmp;               GRID_ASSERT(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     }
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
     {
       SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  GRID_ASSERT(sol_o.Checkerboard()==Odd);
     };
     virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
     {
@@ -366,13 +396,13 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
 
     }
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -386,17 +416,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.Checkerboard()   ==Even);
-      src_e_i = src_e-tmp;               assert(  src_e_i.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o,tmp);          GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      src_e_i = src_e-tmp;               GRID_ASSERT(  src_e_i.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e_i,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     }
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
     {
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  GRID_ASSERT(sol_o.Checkerboard()==Odd);
     };
     virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
     {
@@ -431,9 +461,9 @@ namespace Grid {
         /////////////////////////////////////////////////////
         // src_o = Mdag * (source_o - Moe MeeInv source_e)
         /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
+        _Matrix.MooeeInv(src_e, tmp);   GRID_ASSERT(   tmp.Checkerboard() == Even );
+        _Matrix.Meooe   (tmp, Mtmp);    GRID_ASSERT(  Mtmp.Checkerboard() == Odd  );     
+        src_o -= Mtmp;                  GRID_ASSERT( src_o.Checkerboard() == Odd  );     
       }
       
       virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
@@ -448,18 +478,18 @@ namespace Grid {
         ///////////////////////////////////////////////////
         // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
         ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o, tmp);         assert(     tmp.Checkerboard() == Even );
-        src_e_i = src_e - tmp;             assert( src_e_i.Checkerboard() == Even );
-        _Matrix.MooeeInv(src_e_i, sol_e);  assert(   sol_e.Checkerboard() == Even );
+        _Matrix.Meooe(sol_o, tmp);         GRID_ASSERT(     tmp.Checkerboard() == Even );
+        src_e_i = src_e - tmp;             GRID_ASSERT( src_e_i.Checkerboard() == Even );
+        _Matrix.MooeeInv(src_e_i, sol_e);  GRID_ASSERT(   sol_e.Checkerboard() == Even );
        
-        setCheckerboard(sol, sol_e); assert( sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o); assert( sol_o.Checkerboard() == Odd  );
+        setCheckerboard(sol, sol_e); GRID_ASSERT( sol_e.Checkerboard() == Even );
+        setCheckerboard(sol, sol_o); GRID_ASSERT( sol_o.Checkerboard() == Odd  );
       }
 
       virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
       {
         NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  assert(sol_o.Checkerboard() == Odd);
+        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  GRID_ASSERT(sol_o.Checkerboard() == Odd);
       }
 
       virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o, std::vector<Field>& sol_o)
@@ -469,6 +499,87 @@ namespace Grid {
       }
   };
 
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Site diagonal is identity, left preconditioned by Mee^inv
+  // ( 1 - Mee^inv Meo Moo^inv Moe ) phi = Mee_inv ( Mee - Meo Moo^inv Moe Mee^inv  ) phi =  Mee_inv eta
+  //
+  // Solve:
+  // ( 1 - Mee^inv Meo Moo^inv Moe )^dag ( 1 - Mee^inv Meo Moo^inv Moe ) phi = ( 1 - Mee^inv Meo Moo^inv Moe )^dag  Mee_inv eta
+  //
+  // Old notation e<->o
+  //
+  // Left precon by Moo^-1
+  //  b) (Doo^{dag} M_oo^-dag) (Moo^-1 Doo) psi_o =  [ (D_oo)^dag M_oo^-dag ] Moo^-1 L^{-1}  eta_o
+  //                                   eta_o'     = (D_oo)^dag  M_oo^-dag Moo^-1 (eta_o - Moe Mee^{-1} eta_e)
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagOneSolve : public SchurRedBlackBase<Field> {
+  public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagOneSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+      const bool _solnAsInitGuess = false)  
+    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
+
+    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurDiagOneOperator<Matrix,Field> _HermOpEO(_Matrix);
+      
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd ,src_o,src);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mpcdag *MooeeInv * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      Mtmp=src_o-Mtmp;                 
+      _Matrix.MooeeInv(Mtmp,tmp);      GRID_ASSERT( tmp.Checkerboard() ==Odd);     
+      
+      // get the right MpcDag
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
+    }
+
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field   tmp(grid);
+      Field   sol_e(grid);
+
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);    GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      tmp = src_e-tmp;             GRID_ASSERT(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+     
+      setCheckerboard(sol,sol_e);  GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o);  GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
+    };
+
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
+    {
+      SchurDiagOneOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);
+    };
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
+    {
+      SchurDiagOneOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
+    }
+  };
+
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
   // Site diagonal is identity, right preconditioned by Mee^inv
   // ( 1 - Meo Moo^inv Moe Mee^inv  ) phi =( 1 - Meo Moo^inv Moe Mee^inv  ) Mee psi =  = eta  = eta
@@ -501,12 +612,12 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
     }
 
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -527,12 +638,12 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o_i,tmp);    GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      tmp = src_e-tmp;               GRID_ASSERT(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e);    assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e);    GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o_i);  GRID_ASSERT(  sol_o_i.Checkerboard() ==Odd );
     };
 
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
@@ -573,9 +684,9 @@ namespace Grid {
         /////////////////////////////////////////////////////
         // src_o = Mdag * (source_o - Moe MeeInv source_e)
         /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
+        _Matrix.MooeeInv(src_e, tmp);   GRID_ASSERT(   tmp.Checkerboard() == Even );
+        _Matrix.Meooe   (tmp, Mtmp);    GRID_ASSERT(  Mtmp.Checkerboard() == Odd  );     
+        src_o -= Mtmp;                  GRID_ASSERT( src_o.Checkerboard() == Odd  );     
       }
 
       virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
@@ -596,12 +707,12 @@ namespace Grid {
         ///////////////////////////////////////////////////
         // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
         ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o_i, tmp);    assert(   tmp.Checkerboard() == Even );
-        tmp = src_e - tmp;              assert( src_e.Checkerboard() == Even );
-        _Matrix.MooeeInv(tmp, sol_e);   assert( sol_e.Checkerboard() == Even );
+        _Matrix.Meooe(sol_o_i, tmp);    GRID_ASSERT(   tmp.Checkerboard() == Even );
+        tmp = src_e - tmp;              GRID_ASSERT( src_e.Checkerboard() == Even );
+        _Matrix.MooeeInv(tmp, sol_e);   GRID_ASSERT( sol_e.Checkerboard() == Even );
        
-        setCheckerboard(sol, sol_e);    assert(   sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o_i);  assert( sol_o_i.Checkerboard() == Odd  );
+        setCheckerboard(sol, sol_e);    GRID_ASSERT(   sol_e.Checkerboard() == Even );
+        setCheckerboard(sol, sol_o_i);  GRID_ASSERT( sol_o_i.Checkerboard() == Odd  );
       };
 
       virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)

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

@@ -0,0 +1,611 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/Aggregates.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+
+NAMESPACE_BEGIN(Grid);
+
+inline RealD AggregatePowerLaw(RealD x)
+{
+  //  return std::pow(x,-4);
+  //  return std::pow(x,-3);
+  return std::pow(x,-5);
+}
+
+template<class Fobj,class CComplex,int nbasis>
+class Aggregation {
+public:
+  constexpr int Nbasis(void) { return nbasis; };
+  
+  typedef iVector<CComplex,nbasis >             siteVector;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+
+  GridBase *CoarseGrid;
+  GridBase *FineGrid;
+  std::vector<Lattice<Fobj> > subspace;
+  int checkerboard;
+  int Checkerboard(void){return checkerboard;}
+  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
+    CoarseGrid(_CoarseGrid),
+    FineGrid(_FineGrid),
+    subspace(nbasis,_FineGrid),
+    checkerboard(_checkerboard)
+  {
+  };
+  
+  
+  void Orthogonalise(void){
+    CoarseScalar InnerProd(CoarseGrid); 
+    //    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
+    blockOrthogonalise(InnerProd,subspace);
+  } 
+  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
+    blockProject(CoarseVec,FineVec,subspace);
+  }
+  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+    FineVec.Checkerboard() = subspace[0].Checkerboard();
+    blockPromote(CoarseVec,FineVec,subspace);
+  }
+
+  virtual void CreateSubspaceRandom(GridParallelRNG  &RNG) {
+    int nn=nbasis;
+    RealD scale;
+    FineField noise(FineGrid);
+    for(int b=0;b<nn;b++){
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      subspace[b] = noise;
+    }
+  }
+  virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis)
+  {
+
+    RealD scale;
+
+    ConjugateGradient<FineField> CG(1.0e-4,2000,false);
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+      for(int i=0;i<4;i++){
+
+	CG(hermop,noise,subspace[b]);
+
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
+  virtual void CreateSubspaceGCR(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &DiracOp,int nn=nbasis)
+  {
+    RealD scale;
+
+    TrivialPrecon<FineField> simple_fine;
+    //    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);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
+
+      for(int i=0;i<3;i++){
+	//  void operator() (const Field &src, Field &psi){
+#if 1
+	if (i==0)std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	src = noise;
+	guess=Zero();
+	GCR(src,guess);
+	subspace[b] = guess;
+#else
+	if (i==0)std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	src=Zero();
+	guess = noise;
+	GCR(src,guess);
+	subspace[b] = guess;
+#endif
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      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;
+
+    }
+  }
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
+  // and this is the best I found
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+
+  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo,
+				       int orderfilter,
+				       int ordermin,
+				       int orderstep,
+				       double filterlo
+				       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" Chebyshev subspace pass-1 : ord "<<orderfilter<<" ["<<lo<<","<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pass-2 : nbasis"<<nn<<" min "
+	      <<ordermin<<" step "<<orderstep
+	      <<" lo"<<filterlo<<std::endl;
+
+    // Initial matrix element
+    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      ComplexD ip;
+      // Filter
+      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+
+      hermop.Op(Mn,tmp);
+      ip= innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+      hermop.AdjOp(Mn,tmp); 
+      ip = innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+      b++;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    {
+      lo=filterlo;
+      noise=Mn;
+
+      FineField T0(FineGrid); T0 = noise;  
+      FineField T1(FineGrid); 
+      FineField T2(FineGrid);
+      FineField y(FineGrid);
+      
+      FineField *Tnm = &T0;
+      FineField *Tn  = &T1;
+      FineField *Tnp = &T2;
+
+      // Tn=T1 = (xscale M + mscale)in
+      RealD xscale = 2.0/(hi-lo);
+      RealD mscale = -(hi+lo)/(hi-lo);
+      hermop.HermOp(T0,y);
+      T1=y*xscale+noise*mscale;
+
+      for(int n=2;n<=ordermin+orderstep*(nn-2);n++){
+	
+	hermop.HermOp(*Tn,y);
+
+	autoView( y_v , y, AcceleratorWrite);
+	autoView( Tn_v , (*Tn), AcceleratorWrite);
+	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
+	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
+	const int Nsimd = CComplex::Nsimd();
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
+	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
+	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
+        });
+
+	// Possible more fine grained control is needed than a linear sweep,
+	// but huge productivity gain if this is simple algorithm and not a tunable
+	int m =1;
+	if ( n>=ordermin ) m=n-ordermin;
+	if ( (m%orderstep)==0 ) { 
+	  Mn=*Tnp;
+	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
+	  subspace[b] = Mn;
+
+
+	  ComplexD ip;
+
+	  hermop.Op(Mn,tmp);
+	  ip= innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+	  hermop.AdjOp(Mn,tmp); 
+	  ip = innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+	  
+	  b++;
+	}
+
+	// Cycle pointers to avoid copies
+	FineField *swizzle = Tnm;
+	Tnm    =Tn;
+	Tn     =Tnp;
+	Tnp    =swizzle;
+	  
+      }
+    }
+    GRID_ASSERT(b==nn);
+  }
+
+
+  virtual void CreateSubspacePolyCheby(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo1,
+				       int orderfilter,
+				       double lo2,
+				       int orderstep)
+  {
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" CreateSubspacePolyCheby "<<std::endl;
+    // Initial matrix element
+    hermop.Op(noise,Mn);
+    std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      // Filter
+      std::cout << GridLogMessage << "Cheby "<<lo1<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo1,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|n> "<<norm2(Mn)<<std::endl;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    for(int n=1;n<nn;n++){
+      std::cout << GridLogMessage << "Cheby "<<lo2<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo2,hi,orderstep);
+      Cheb(hermop,subspace[n-1],Mn);
+
+      for(int m=0;m<n;m++){
+	ComplexD c = innerProduct(subspace[m],Mn);
+	Mn = Mn - c*subspace[m];
+      }
+      
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5);
+      Mn=Mn*scale;
+      
+      subspace[n]=Mn;
+      
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|n> "<<norm2(Mn)<<std::endl;
+
+    }
+  }
+
+  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo,
+				       int orderfilter
+				       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise : ord "<<orderfilter<<" ["<<lo<<","<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise  : nbasis "<<nn<<std::endl;
+
+
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+      // Filter
+      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+
+      // Refine
+      Chebyshev<FineField> PowerLaw(lo,hi,1000,AggregatePowerLaw);
+      noise = Mn;
+      PowerLaw(hermop,noise,Mn);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+
+      // normalise
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+
+  }
+
+  virtual void CreateSubspaceChebyshevPowerLaw(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					       int nn,
+					       double hi,
+					       int orderfilter
+					       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise : ord "<<orderfilter<<" [0,"<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise  : nbasis "<<nn<<std::endl;
+
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+      // Filter
+      Chebyshev<FineField> Cheb(0.0,hi,orderfilter,AggregatePowerLaw);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+
+  }
+  virtual void CreateSubspaceChebyshevNew(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					  double hi
+					  ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+      // Filter
+      //#opt2(x) =  acheb(x,3,90,300)* acheb(x,1,90,50) * acheb(x,0.5,90,200) * acheb(x,0.05,90,400) * acheb(x,0.01,90,1500)
+      /*266
+      Chebyshev<FineField> Cheb1(3.0,hi,300);
+      Chebyshev<FineField> Cheb2(1.0,hi,50);
+      Chebyshev<FineField> Cheb3(0.5,hi,300);
+      Chebyshev<FineField> Cheb4(0.05,hi,500);
+      Chebyshev<FineField> Cheb5(0.01,hi,2000);
+      */
+      /* 242 */
+      /*
+      Chebyshev<FineField> Cheb3(0.1,hi,300);
+      Chebyshev<FineField> Cheb2(0.02,hi,1000);
+      Chebyshev<FineField> Cheb1(0.003,hi,2000);
+      8?
+      */
+      /* How many??
+      */
+      Chebyshev<FineField> Cheb2(0.001,hi,2500); // 169 iters on HDCG after refine
+      Chebyshev<FineField> Cheb1(0.02,hi,600);
+
+      //      Chebyshev<FineField> Cheb2(0.001,hi,1500);
+      //      Chebyshev<FineField> Cheb1(0.02,hi,600);
+      Cheb1(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb1 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      Cheb2(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb2 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      //      Cheb3(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      //      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb3 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      //      Cheb4(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      //      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb4 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      //      Cheb5(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      //      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb5 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      subspace[b]   = noise;
+      hermop.Op(subspace[b],tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<< " norm " << norm2(noise)<<std::endl;
+    }
+
+  }
+
+  virtual void CreateSubspaceMultishift(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					double Lo,double tol,int maxit)
+  {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Multishift subspace : Lo "<<Lo<<std::endl;
+
+    // Filter
+    // [ 1/6(x+Lo)  - 1/2(x+2Lo) + 1/2(x+3Lo)  -1/6(x+4Lo) = Lo^3 /[ (x+1Lo)(x+2Lo)(x+3Lo)(x+4Lo) ]
+    //
+    // 1/(x+Lo)  - 1/(x+2 Lo)
+    double epsilon      = Lo/3;
+    std::vector<RealD> alpha({1.0/6.0,-1.0/2.0,1.0/2.0,-1.0/6.0});
+    std::vector<RealD> shifts({Lo,Lo+epsilon,Lo+2*epsilon,Lo+3*epsilon});
+    std::vector<RealD> tols({tol,tol,tol,tol});
+    std::cout << "sizes "<<alpha.size()<<" "<<shifts.size()<<" "<<tols.size()<<std::endl;
+
+    MultiShiftFunction msf(4,0.0,95.0);
+    std::cout << "msf constructed "<<std::endl;
+    msf.poles=shifts;
+    msf.residues=alpha;
+    msf.tolerances=tols;
+    msf.norm=0.0;
+    msf.order=alpha.size();
+    ConjugateGradientMultiShift<FineField> MSCG(maxit,msf);
+    
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+      MSCG(hermop,noise,Mn);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+    }
+
+  }
+  virtual void RefineSubspace(LinearOperatorBase<FineField> &hermop,
+			      double Lo,double tol,int maxit)
+  {
+    FineField tmp(FineGrid);
+    for(int b =0;b<nbasis;b++)
+    {
+      ConjugateGradient<FineField>  CGsloppy(tol,maxit,false);
+      ShiftedHermOpLinearOperator<FineField> ShiftedFineHermOp(hermop,Lo);
+      tmp=Zero();
+      CGsloppy(hermop,subspace[b],tmp);
+      RealD scale = std::pow(norm2(tmp),-0.5); 	tmp=tmp*scale;
+      subspace[b]=tmp;
+      hermop.Op(subspace[b],tmp);
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+  }
+  virtual void RefineSubspaceHDCG(LinearOperatorBase<FineField> &hermop,
+				  TwoLevelADEF2mrhs<FineField,CoarseVector> & theHDCG,
+				  int nrhs)
+  {
+    std::vector<FineField> src_mrhs(nrhs,FineGrid);
+    std::vector<FineField> res_mrhs(nrhs,FineGrid);
+    FineField tmp(FineGrid);
+    for(int b =0;b<nbasis;b+=nrhs)
+    {
+      tmp = subspace[b];
+      RealD scale = std::pow(norm2(tmp),-0.5); 	tmp=tmp*scale;
+      subspace[b] =tmp;
+      hermop.Op(subspace[b],tmp);
+      std::cout<<GridLogMessage << "before filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+      for(int r=0;r<MIN(nbasis-b,nrhs);r++){
+	src_mrhs[r] = subspace[b+r];
+      }
+      for(int r=0;r<nrhs;r++){
+	res_mrhs[r] = Zero();
+      }
+      theHDCG(src_mrhs,res_mrhs);
+
+      for(int r=0;r<MIN(nbasis-b,nrhs);r++){
+	tmp = res_mrhs[r];
+	RealD scale = std::pow(norm2(tmp),-0.5); tmp=tmp*scale;
+	subspace[b+r]=tmp;
+      }
+      hermop.Op(subspace[b],tmp);
+      std::cout<<GridLogMessage << "after filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+  }
+
+  
+  
+};
+NAMESPACE_END(Grid);
+

Grid/algorithms/multigrid/CoarsenedMatrix.h

@@ -0,0 +1,837 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/CoarsenedMatrix.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+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_ALGORITHM_COARSENED_MATRIX_H
+#define  GRID_ALGORITHM_COARSENED_MATRIX_H
+
+#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
+
+NAMESPACE_BEGIN(Grid);
+
+template<class vobj,class CComplex>
+inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
+				    const Lattice<decltype(innerProduct(vobj(),vobj()))> &FineMask,
+				    const Lattice<vobj> &fineX,
+				    const Lattice<vobj> &fineY)
+{
+  typedef decltype(innerProduct(vobj(),vobj())) dotp;
+
+  GridBase *coarse(CoarseInner.Grid());
+  GridBase *fine  (fineX.Grid());
+
+  Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
+  Lattice<dotp> fine_inner_msk(fine);
+
+  // Multiply could be fused with innerProduct
+  // Single block sum kernel could do both masks.
+  fine_inner = localInnerProduct(fineX,fineY);
+  mult(fine_inner_msk, fine_inner,FineMask);
+  blockSum(CoarseInner,fine_inner_msk);
+}
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class CoarsenedMatrix : public CheckerBoardedSparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+    
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef Lattice<CComplex >                  CoarseComplexField;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+  typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+  typedef CoarseVector FermionField;
+
+  // enrich interface, use default implementation as in FermionOperator ///////
+  void Dminus(CoarseVector const& in, CoarseVector& out) { out = in; }
+  void DminusDag(CoarseVector const& in, CoarseVector& out) { out = in; }
+  void ImportPhysicalFermionSource(CoarseVector const& input, CoarseVector& imported) { imported = input; }
+  void ImportUnphysicalFermion(CoarseVector const& input, CoarseVector& imported) { imported = input; }
+  void ExportPhysicalFermionSolution(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
+  void ExportPhysicalFermionSource(CoarseVector const& solution, CoarseVector& exported) { exported = solution; };
+
+  ////////////////////
+  // Data members
+  ////////////////////
+  Geometry         geom;
+  GridBase *       _grid; 
+  GridBase*        _cbgrid;
+  int hermitian;
+
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> Stencil; 
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilEven;
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilOdd;
+
+  std::vector<CoarseMatrix> A;
+  std::vector<CoarseMatrix> Aeven;
+  std::vector<CoarseMatrix> Aodd;
+
+  CoarseMatrix AselfInv;
+  CoarseMatrix AselfInvEven;
+  CoarseMatrix AselfInvOdd;
+
+  deviceVector<RealD> dag_factor;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know
+  GridBase * RedBlackGrid()     { return _cbgrid; };
+
+  int ConstEE() { return 0; }
+
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    conformable(_grid,in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    SimpleCompressor<siteVector> compressor;
+
+    Stencil.HaloExchange(in,compressor);
+    autoView( in_v , in, AcceleratorRead);
+    autoView( out_v , out, AcceleratorWrite);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+      
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
+    Aview *Aview_p = & AcceleratorViewContainer[0];
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    int osites=Grid()->oSites();
+
+    accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
+      int ss = sss/nbasis;
+      int b  = sss%nbasis;
+      calcComplex res = Zero();
+      calcVector nbr;
+      int ptype;
+      StencilEntry *SE;
+
+      for(int point=0;point<npoint;point++){
+
+	SE=Stencil_v.GetEntry(ptype,point,ss);
+	  
+	if(SE->_is_local) { 
+	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+	} else {
+	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
+	}
+	acceleratorSynchronise();
+
+	for(int bb=0;bb<nbasis;bb++) {
+	  res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+	}
+      }
+      coalescedWrite(out_v[ss](b),res);
+      });
+
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+  };
+
+  void Mdag (const CoarseVector &in, CoarseVector &out)
+  {
+    if(hermitian) {
+      // corresponds to Petrov-Galerkin coarsening
+      return M(in,out);
+    } else {
+      // corresponds to Galerkin coarsening
+      return MdagNonHermitian(in, out);
+    }
+  };
+
+  void MdagNonHermitian(const CoarseVector &in, CoarseVector &out)
+  {
+    conformable(_grid,in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    SimpleCompressor<siteVector> compressor;
+
+    Stencil.HaloExchange(in,compressor);
+    autoView( in_v , in, AcceleratorRead);
+    autoView( out_v , out, AcceleratorWrite);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+
+
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
+    Aview *Aview_p = & AcceleratorViewContainer[0];
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    int osites=Grid()->oSites();
+
+    deviceVector<int> points(geom.npoint);
+    for(int p=0; p<geom.npoint; p++) { 
+      acceleratorPut(points[p],geom.points_dagger[p]);
+    }
+    auto points_p = &points[0];
+
+    RealD* dag_factor_p = &dag_factor[0];
+
+    accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
+      int ss = sss/nbasis;
+      int b  = sss%nbasis;
+      calcComplex res = Zero();
+      calcVector nbr;
+      int ptype;
+      StencilEntry *SE;
+
+      for(int p=0;p<npoint;p++){
+        int point = points_p[p];
+
+	SE=Stencil_v.GetEntry(ptype,point,ss);
+
+	if(SE->_is_local) {
+	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+	} else {
+	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
+	}
+	acceleratorSynchronise();
+
+	for(int bb=0;bb<nbasis;bb++) {
+	  res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+	}
+      }
+      coalescedWrite(out_v[ss](b),res);
+      });
+
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+  }
+
+  void MdirComms(const CoarseVector &in)
+  {
+    SimpleCompressor<siteVector> compressor;
+    Stencil.HaloExchange(in,compressor);
+  }
+  void MdirCalc(const CoarseVector &in, CoarseVector &out, int point)
+  {
+    conformable(_grid,in.Grid());
+    conformable(_grid,out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    typedef LatticeView<Cobj> Aview;
+
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
+    Aview *Aview_p = & AcceleratorViewContainer[0];
+
+    autoView( out_v , out, AcceleratorWrite);
+    autoView( in_v  , in, AcceleratorRead);
+    autoView( Stencil_v  , Stencil, AcceleratorRead);
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
+      int ss = sss/nbasis;
+      int b  = sss%nbasis;
+      calcComplex res = Zero();
+      calcVector nbr;
+      int ptype;
+      StencilEntry *SE;
+
+      SE=Stencil_v.GetEntry(ptype,point,ss);
+	  
+      if(SE->_is_local) { 
+	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+      } else {
+	nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]);
+      }
+      acceleratorSynchronise();
+
+      for(int bb=0;bb<nbasis;bb++) {
+	res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+      }
+      coalescedWrite(out_v[ss](b),res);
+    });
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+  }
+  void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
+  {
+    this->MdirComms(in);
+    int ndir=geom.npoint-1;
+    if ((out.size()!=ndir)&&(out.size()!=ndir+1)) { 
+      std::cout <<"MdirAll out size "<< out.size()<<std::endl;
+      std::cout <<"MdirAll ndir "<< ndir<<std::endl;
+      GRID_ASSERT(0);
+    }
+    for(int p=0;p<ndir;p++){
+      MdirCalc(in,out[p],p);
+    }
+  };
+  void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
+
+    this->MdirComms(in);
+
+    MdirCalc(in,out,geom.point(dir,disp));
+  };
+
+  void Mdiag(const CoarseVector &in, CoarseVector &out)
+  {
+    int point=geom.npoint-1;
+    MdirCalc(in, out, point); // No comms
+  };
+
+  void Mooee(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerNo, InverseNo);
+  }
+
+  void MooeeInv(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerNo, InverseYes);
+  }
+
+  void MooeeDag(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerYes, InverseNo);
+  }
+
+  void MooeeInvDag(const CoarseVector &in, CoarseVector &out) {
+    MooeeInternal(in, out, DaggerYes, InverseYes);
+  }
+
+  void Meooe(const CoarseVector &in, CoarseVector &out) {
+    if(in.Checkerboard() == Odd) {
+      DhopEO(in, out, DaggerNo);
+    } else {
+      DhopOE(in, out, DaggerNo);
+    }
+  }
+
+  void MeooeDag(const CoarseVector &in, CoarseVector &out) {
+    if(in.Checkerboard() == Odd) {
+      DhopEO(in, out, DaggerYes);
+    } else {
+      DhopOE(in, out, DaggerYes);
+    }
+  }
+
+  void Dhop(const CoarseVector &in, CoarseVector &out, int dag) {
+    conformable(in.Grid(), _grid); // verifies full grid
+    conformable(in.Grid(), out.Grid());
+
+    out.Checkerboard() = in.Checkerboard();
+
+    DhopInternal(Stencil, A, in, out, dag);
+  }
+
+  void DhopOE(const CoarseVector &in, CoarseVector &out, int dag) {
+    conformable(in.Grid(), _cbgrid);    // verifies half grid
+    conformable(in.Grid(), out.Grid()); // drops the cb check
+
+    GRID_ASSERT(in.Checkerboard() == Even);
+    out.Checkerboard() = Odd;
+
+    DhopInternal(StencilEven, Aodd, in, out, dag);
+  }
+
+  void DhopEO(const CoarseVector &in, CoarseVector &out, int dag) {
+    conformable(in.Grid(), _cbgrid);    // verifies half grid
+    conformable(in.Grid(), out.Grid()); // drops the cb check
+
+    GRID_ASSERT(in.Checkerboard() == Odd);
+    out.Checkerboard() = Even;
+
+    DhopInternal(StencilOdd, Aeven, in, out, dag);
+  }
+
+  void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) {
+    out.Checkerboard() = in.Checkerboard();
+    GRID_ASSERT(in.Checkerboard() == Odd || in.Checkerboard() == Even);
+
+    CoarseMatrix *Aself = nullptr;
+    if(in.Grid()->_isCheckerBoarded) {
+      if(in.Checkerboard() == Odd) {
+        Aself = (inv) ? &AselfInvOdd : &Aodd[geom.npoint-1];
+        DselfInternal(StencilOdd, *Aself, in, out, dag);
+      } else {
+        Aself = (inv) ? &AselfInvEven : &Aeven[geom.npoint-1];
+        DselfInternal(StencilEven, *Aself, in, out, dag);
+      }
+    } else {
+      Aself = (inv) ? &AselfInv : &A[geom.npoint-1];
+      DselfInternal(Stencil, *Aself, in, out, dag);
+    }
+    GRID_ASSERT(Aself != nullptr);
+  }
+
+  void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
+                       const CoarseVector &in, CoarseVector &out, int dag) {
+    int point = geom.npoint-1;
+    autoView( out_v, out, AcceleratorWrite);
+    autoView( in_v,  in,  AcceleratorRead);
+    autoView( st_v,  st,  AcceleratorRead);
+    autoView( a_v,   a,   AcceleratorRead);
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    RealD* dag_factor_p = &dag_factor[0];
+
+    if(dag) {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        SE=st_v.GetEntry(ptype,point,ss);
+
+        if(SE->_is_local) {
+          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+        } else {
+          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+        }
+        acceleratorSynchronise();
+
+        for(int bb=0;bb<nbasis;bb++) {
+          res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(a_v[ss](b,bb))*nbr(bb);
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    } else {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        SE=st_v.GetEntry(ptype,point,ss);
+
+        if(SE->_is_local) {
+          nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+        } else {
+          nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+        }
+        acceleratorSynchronise();
+
+        for(int bb=0;bb<nbasis;bb++) {
+          res = res + coalescedRead(a_v[ss](b,bb))*nbr(bb);
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    }
+  }
+
+  void DhopInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, std::vector<CoarseMatrix> &a,
+                    const CoarseVector &in, CoarseVector &out, int dag) {
+    SimpleCompressor<siteVector> compressor;
+
+    st.HaloExchange(in,compressor);
+    autoView( in_v,  in,  AcceleratorRead);
+    autoView( out_v, out, AcceleratorWrite);
+    autoView( st_v , st,  AcceleratorRead);
+    typedef LatticeView<Cobj> Aview;
+
+    // determine in what order we need the points
+    int npoint = geom.npoint-1;
+    deviceVector<int> points(npoint);
+    for(int p=0; p<npoint; p++) {
+      int val = (dag && !hermitian) ? geom.points_dagger[p] : p;
+      acceleratorPut(points[p], val);
+    }
+    auto points_p = &points[0];
+
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = a[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
+    Aview *Aview_p = & AcceleratorViewContainer[0];
+
+    const int Nsimd = CComplex::Nsimd();
+    typedef decltype(coalescedRead(in_v[0])) calcVector;
+    typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+
+    RealD* dag_factor_p = &dag_factor[0];
+
+    if(dag) {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        for(int p=0;p<npoint;p++){
+          int point = points_p[p];
+          SE=st_v.GetEntry(ptype,point,ss);
+
+          if(SE->_is_local) {
+            nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+          } else {
+            nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+          }
+          acceleratorSynchronise();
+
+          for(int bb=0;bb<nbasis;bb++) {
+            res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+          }
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    } else {
+      accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, {
+        int ss = sss/nbasis;
+        int b  = sss%nbasis;
+        calcComplex res = Zero();
+        calcVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        for(int p=0;p<npoint;p++){
+          int point = points_p[p];
+          SE=st_v.GetEntry(ptype,point,ss);
+
+          if(SE->_is_local) {
+            nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
+          } else {
+            nbr = coalescedRead(st_v.CommBuf()[SE->_offset]);
+          }
+          acceleratorSynchronise();
+
+          for(int bb=0;bb<nbasis;bb++) {
+            res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+          }
+        }
+        coalescedWrite(out_v[ss](b),res);
+      });
+    }
+
+    for(int p=0;p<npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+  }
+  
+  CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
+    _grid(&CoarseGrid),
+    _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
+    geom(CoarseGrid._ndimension),
+    hermitian(hermitian_),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements),
+    A(geom.npoint,&CoarseGrid),
+    Aeven(geom.npoint,_cbgrid),
+    Aodd(geom.npoint,_cbgrid),
+    AselfInv(&CoarseGrid),
+    AselfInvEven(_cbgrid),
+    AselfInvOdd(_cbgrid),
+    dag_factor(nbasis*nbasis)
+  {
+    fillFactor();
+  };
+
+  CoarsenedMatrix(GridCartesian &CoarseGrid, GridRedBlackCartesian &CoarseRBGrid, int hermitian_=0) 	:
+
+    _grid(&CoarseGrid),
+    _cbgrid(&CoarseRBGrid),
+    geom(CoarseGrid._ndimension),
+    hermitian(hermitian_),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements),
+    A(geom.npoint,&CoarseGrid),
+    Aeven(geom.npoint,&CoarseRBGrid),
+    Aodd(geom.npoint,&CoarseRBGrid),
+    AselfInv(&CoarseGrid),
+    AselfInvEven(&CoarseRBGrid),
+    AselfInvOdd(&CoarseRBGrid),
+    dag_factor(nbasis*nbasis)
+  {
+    fillFactor();
+  };
+
+  void fillFactor() {
+    Eigen::MatrixXd dag_factor_eigen = Eigen::MatrixXd::Ones(nbasis, nbasis);
+    if(!hermitian) {
+      const int nb = nbasis/2;
+      dag_factor_eigen.block(0,nb,nb,nb) *= -1.0;
+      dag_factor_eigen.block(nb,0,nb,nb) *= -1.0;
+    }
+
+    // GPU readable prefactor
+    std::vector<RealD> h_dag_factor(nbasis*nbasis);
+    thread_for(i, nbasis*nbasis, {
+      int j = i/nbasis;
+      int k = i%nbasis;
+      h_dag_factor[i] = dag_factor_eigen(j, k);
+    });
+    acceleratorCopyToDevice(&h_dag_factor[0],&dag_factor[0],dag_factor.size()*sizeof(RealD));
+  }
+
+  void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    typedef Lattice<typename Fobj::tensor_reduced> FineComplexField;
+    typedef typename Fobj::scalar_type scalar_type;
+
+    std::cout << GridLogMessage<< "CoarsenMatrix "<< std::endl;
+
+    FineComplexField one(FineGrid); one=scalar_type(1.0,0.0);
+    FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0);
+
+    std::vector<FineComplexField> masks(geom.npoint,FineGrid);
+    FineComplexField imask(FineGrid); // contributions from within this block
+    FineComplexField omask(FineGrid); // contributions from outwith this block
+
+    FineComplexField evenmask(FineGrid);
+    FineComplexField oddmask(FineGrid); 
+
+    FineField     phi(FineGrid);
+    FineField     tmp(FineGrid);
+    FineField     zz(FineGrid); zz=Zero();
+    FineField    Mphi(FineGrid);
+    FineField    Mphie(FineGrid);
+    FineField    Mphio(FineGrid);
+    std::vector<FineField>     Mphi_p(geom.npoint,FineGrid);
+
+    Lattice<iScalar<vInteger> > coor (FineGrid);
+    Lattice<iScalar<vInteger> > bcoor(FineGrid);
+    Lattice<iScalar<vInteger> > bcb  (FineGrid); bcb = Zero();
+
+    CoarseVector iProj(Grid()); 
+    CoarseVector oProj(Grid()); 
+    CoarseVector SelfProj(Grid()); 
+    CoarseComplexField iZProj(Grid()); 
+    CoarseComplexField oZProj(Grid()); 
+
+    CoarseScalar InnerProd(Grid()); 
+
+    std::cout << GridLogMessage<< "CoarsenMatrix Orthog "<< std::endl;
+    // Orthogonalise the subblocks over the basis
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    // Compute the matrix elements of linop between this orthonormal
+    // set of vectors.
+    std::cout << GridLogMessage<< "CoarsenMatrix masks "<< std::endl;
+    int self_stencil=-1;
+    for(int p=0;p<geom.npoint;p++)
+    { 
+      int dir   = geom.directions[p];
+      int disp  = geom.displacements[p];
+      A[p]=Zero();
+      if( geom.displacements[p]==0){
+	self_stencil=p;
+      }
+
+      Integer block=(FineGrid->_rdimensions[dir])/(Grid()->_rdimensions[dir]);
+
+      LatticeCoordinate(coor,dir);
+
+      ///////////////////////////////////////////////////////
+      // Work out even and odd block checkerboarding for fast diagonal term
+      ///////////////////////////////////////////////////////
+      if ( disp==1 ) {
+	bcb   = bcb + div(coor,block);
+      }
+	
+      if ( disp==0 ) {
+	  masks[p]= Zero();
+      } else if ( disp==1 ) {
+	masks[p] = where(mod(coor,block)==(block-1),one,zero);
+      } else if ( disp==-1 ) {
+	masks[p] = where(mod(coor,block)==(Integer)0,one,zero);
+      }
+    }
+    evenmask = where(mod(bcb,2)==(Integer)0,one,zero);
+    oddmask  = one-evenmask;
+
+    GRID_ASSERT(self_stencil!=-1);
+
+    for(int i=0;i<nbasis;i++){
+
+      phi=Subspace.subspace[i];
+
+      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl;
+      linop.OpDirAll(phi,Mphi_p);
+      linop.OpDiag  (phi,Mphi_p[geom.npoint-1]);
+
+      for(int p=0;p<geom.npoint;p++){ 
+
+	Mphi = Mphi_p[p];
+
+	int dir   = geom.directions[p];
+	int disp  = geom.displacements[p];
+
+	if ( (disp==-1) || (!hermitian ) ) {
+
+	  ////////////////////////////////////////////////////////////////////////
+	  // Pick out contributions coming from this cell and neighbour cell
+	  ////////////////////////////////////////////////////////////////////////
+	  omask = masks[p];
+	  imask = one-omask;
+	
+	  for(int j=0;j<nbasis;j++){
+	    
+	    blockMaskedInnerProduct(oZProj,omask,Subspace.subspace[j],Mphi);
+	    
+	    autoView( iZProj_v , iZProj, AcceleratorRead) ;
+	    autoView( oZProj_v , oZProj, AcceleratorRead) ;
+	    autoView( A_p     ,  A[p], AcceleratorWrite);
+	    autoView( A_self  , A[self_stencil], AcceleratorWrite);
+
+	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
+	    if ( hermitian && (disp==-1) ) {
+	      for(int pp=0;pp<geom.npoint;pp++){// Find the opposite link and set <j|A|i> = <i|A|j>*
+		int dirp   = geom.directions[pp];
+		int dispp  = geom.displacements[pp];
+		if ( (dirp==dir) && (dispp==1) ){
+		  auto sft = conjugate(Cshift(oZProj,dir,1));
+		  autoView( sft_v    ,  sft  , AcceleratorWrite);
+		  autoView( A_pp     ,  A[pp], AcceleratorWrite);
+		  accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_pp[ss](i,j),sft_v(ss)); });
+		}
+	      }
+	    }
+
+	  }
+	}
+      }
+
+      ///////////////////////////////////////////
+      // Faster alternate self coupling.. use hermiticity to save 2x
+      ///////////////////////////////////////////
+      {
+	mult(tmp,phi,evenmask);  linop.Op(tmp,Mphie);
+	mult(tmp,phi,oddmask );  linop.Op(tmp,Mphio);
+
+	{
+	  autoView( tmp_      , tmp, AcceleratorWrite);
+	  autoView( evenmask_ , evenmask, AcceleratorRead);
+	  autoView( oddmask_  ,  oddmask, AcceleratorRead);
+	  autoView( Mphie_    ,  Mphie, AcceleratorRead);
+	  autoView( Mphio_    ,  Mphio, AcceleratorRead);
+	  accelerator_for(ss, FineGrid->oSites(), Fobj::Nsimd(),{ 
+	      coalescedWrite(tmp_[ss],evenmask_(ss)*Mphie_(ss) + oddmask_(ss)*Mphio_(ss));
+	    });
+	}
+
+	blockProject(SelfProj,tmp,Subspace.subspace);
+
+	autoView( SelfProj_ , SelfProj, AcceleratorRead);
+	autoView( A_self  , A[self_stencil], AcceleratorWrite);
+
+	accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{
+	  for(int j=0;j<nbasis;j++){
+	    coalescedWrite(A_self[ss](j,i), SelfProj_(ss)(j));
+	  }
+	});
+
+      }
+    }
+    if(hermitian) {
+      std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
+    }
+
+    InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl;
+    FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl;
+  }
+
+  void InvertSelfStencilLink() {
+    std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl;
+    int localVolume = Grid()->lSites();
+
+    typedef typename Cobj::scalar_object scalar_object;
+
+    autoView(Aself_v,    A[geom.npoint-1], CpuRead);
+    autoView(AselfInv_v, AselfInv,         CpuWrite);
+    thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke
+      Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis);
+      Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis);
+
+      scalar_object selfLink    = Zero();
+      scalar_object selfLinkInv = Zero();
+
+      Coordinate lcoor;
+
+      Grid()->LocalIndexToLocalCoor(site, lcoor);
+      peekLocalSite(selfLink, Aself_v, lcoor);
+
+      for (int i = 0; i < nbasis; ++i)
+        for (int j = 0; j < nbasis; ++j)
+          selfLinkEigen(i, j) = static_cast<ComplexD>(TensorRemove(selfLink(i, j)));
+
+      selfLinkInvEigen = selfLinkEigen.inverse();
+
+      for(int i = 0; i < nbasis; ++i)
+        for(int j = 0; j < nbasis; ++j)
+          selfLinkInv(i, j) = selfLinkInvEigen(i, j);
+
+      pokeLocalSite(selfLinkInv, AselfInv_v, lcoor);
+    });
+  }
+
+  void FillHalfCbs() {
+    std::cout << GridLogDebug << "CoarsenedMatrix::FillHalfCbs" << std::endl;
+    for(int p = 0; p < geom.npoint; ++p) {
+      pickCheckerboard(Even, Aeven[p], A[p]);
+      pickCheckerboard(Odd, Aodd[p], A[p]);
+    }
+    pickCheckerboard(Even, AselfInvEven, AselfInv);
+    pickCheckerboard(Odd, AselfInvOdd, AselfInv);
+  }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -0,0 +1,629 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrix.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
+
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+NAMESPACE_BEGIN(Grid);
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class GeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  typedef Lattice<iScalar<CComplex> >         CoarseComplexField;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+  typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef iVector<CComplex,nbasis >  Cvec;
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+  typedef Lattice<CComplex >    FineComplexField;
+  typedef CoarseVector Field;
+  ////////////////////
+  // Data members
+  ////////////////////
+  int hermitian;
+  GridBase      *       _FineGrid; 
+  GridCartesian *       _CoarseGrid; 
+  NonLocalStencilGeometry &geom;
+  PaddedCell Cell;
+  GeneralLocalStencil Stencil;
+  
+  std::vector<CoarseMatrix> _A;
+  std::vector<CoarseMatrix> _Adag;
+  std::vector<CoarseVector> MultTemporaries;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase      * Grid(void)           { return _CoarseGrid; };   // this is all the linalg routines need to know
+  GridBase      * FineGrid(void)       { return _FineGrid; };   // this is all the linalg routines need to know
+  GridCartesian * CoarseGrid(void)     { return _CoarseGrid; };   // this is all the linalg routines need to know
+
+  /*  void ShiftMatrix(RealD shift)
+  {
+    int Nd=_FineGrid->Nd(); 
+    Coordinate zero_shift(Nd,0);
+    for(int p=0;p<geom.npoint;p++){
+      if ( zero_shift==geom.shifts[p] ) {
+	_A[p] = _A[p]+shift;
+	//	_Adag[p] = _Adag[p]+shift;
+      }
+    }    
+  }
+  void ProjectNearestNeighbour(RealD shift, GeneralCoarseOp &CopyMe)
+  {
+    int nfound=0;
+    std::cout << GridLogMessage <<"GeneralCoarsenedMatrix::ProjectNearestNeighbour "<< CopyMe._A[0].Grid()<<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      for(int pp=0;pp<CopyMe.geom.npoint;pp++){
+ 	// Search for the same relative shift
+	// Avoids brutal handling of Grid pointers
+	if ( CopyMe.geom.shifts[pp]==geom.shifts[p] ) {
+	  _A[p] = CopyMe.Cell.Extract(CopyMe._A[pp]);
+	  //	  _Adag[p] = CopyMe.Cell.Extract(CopyMe._Adag[pp]);
+	  nfound++;
+	}
+      }
+    }
+    GRID_ASSERT(nfound==geom.npoint);
+    ExchangeCoarseLinks();
+  }
+  */
+  
+  GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid)
+    : geom(_geom),
+      _FineGrid(FineGrid),
+      _CoarseGrid(CoarseGrid),
+      hermitian(1),
+      Cell(_geom.Depth(),_CoarseGrid),
+      Stencil(Cell.grids.back(),geom.shifts)
+  {
+    {
+      int npoint = _geom.npoint;
+    }
+    _A.resize(geom.npoint,CoarseGrid);
+    //    _Adag.resize(geom.npoint,CoarseGrid);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    Mult(_A,in,out);
+  }
+  void Mdag (const CoarseVector &in, CoarseVector &out)
+  {
+    GRID_ASSERT(hermitian);
+    Mult(_A,in,out);
+    //    if ( hermitian ) M(in,out);
+    //    else Mult(_Adag,in,out);
+  }
+  void Mult (std::vector<CoarseMatrix> &A,const CoarseVector &in, CoarseVector &out)
+  {
+    RealD tviews=0;    RealD ttot=0;    RealD tmult=0;   RealD texch=0;    RealD text=0; RealD ttemps=0; RealD tcopy=0;
+    RealD tmult2=0;
+
+    ttot=-usecond();
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+    CoarseVector tin=in;
+
+    texch-=usecond();
+    CoarseVector pin = Cell.ExchangePeriodic(tin);
+    texch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+    
+    int64_t osites=pin.Grid()->oSites();
+
+    RealD flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    RealD bytes = 1.0*osites*sizeof(siteMatrix)*npoint
+                + 2.0*osites*sizeof(siteVector)*npoint;
+      
+    {
+      tviews-=usecond();
+      autoView( in_v , pin, AcceleratorRead);
+      autoView( out_v , pout, AcceleratorWriteDiscard);
+      autoView( Stencil_v  , Stencil, AcceleratorRead);
+      tviews+=usecond();
+
+      // Static and prereserve to keep UVM region live and not resized across multiple calls
+      ttemps-=usecond();
+      MultTemporaries.resize(npoint,pin.Grid());       
+      ttemps+=usecond();
+      std::vector<Aview> AcceleratorViewContainer_h;
+      std::vector<Vview> AcceleratorVecViewContainer_h; 
+
+      tviews-=usecond();
+      for(int p=0;p<npoint;p++) {
+	AcceleratorViewContainer_h.push_back(      A[p].View(AcceleratorRead));
+	AcceleratorVecViewContainer_h.push_back(MultTemporaries[p].View(AcceleratorWrite));
+      }
+      tviews+=usecond();
+
+      static deviceVector<Aview> AcceleratorViewContainer; AcceleratorViewContainer.resize(npoint);
+      static deviceVector<Vview> AcceleratorVecViewContainer; AcceleratorVecViewContainer.resize(npoint); 
+      
+      auto Aview_p = &AcceleratorViewContainer[0];
+      auto Vview_p = &AcceleratorVecViewContainer[0];
+      tcopy-=usecond();
+      acceleratorCopyToDevice(&AcceleratorViewContainer_h[0],&AcceleratorViewContainer[0],npoint *sizeof(Aview));
+      acceleratorCopyToDevice(&AcceleratorVecViewContainer_h[0],&AcceleratorVecViewContainer[0],npoint *sizeof(Vview));
+      tcopy+=usecond();
+
+      tmult-=usecond();
+      accelerator_for(spb, osites*nbasis*npoint, Nsimd, {
+	  typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+	  int32_t ss   = spb/(nbasis*npoint);
+	  int32_t bp   = spb%(nbasis*npoint);
+	  int32_t point= bp/nbasis;
+	  int32_t b    = bp%nbasis;
+	  auto SE  = Stencil_v.GetEntry(point,ss);
+	  auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
+	  auto res = coalescedRead(Aview_p[point][ss](0,b))*nbr(0);
+	  for(int bb=1;bb<nbasis;bb++) {
+	    res = res + coalescedRead(Aview_p[point][ss](bb,b))*nbr(bb);
+	  }
+	  coalescedWrite(Vview_p[point][ss](b),res);
+      });
+      tmult2-=usecond();
+      accelerator_for(sb, osites*nbasis, Nsimd, {
+	  int ss = sb/nbasis;
+	  int b  = sb%nbasis;
+	  auto res = coalescedRead(Vview_p[0][ss](b));
+	  for(int point=1;point<npoint;point++){
+	    res = res + coalescedRead(Vview_p[point][ss](b));
+	  }
+	  coalescedWrite(out_v[ss](b),res);
+      });
+      tmult2+=usecond();
+      tmult+=usecond();
+      for(int p=0;p<npoint;p++) {
+	AcceleratorViewContainer_h[p].ViewClose();
+	AcceleratorVecViewContainer_h[p].ViewClose();
+      }
+    }
+
+    text-=usecond();
+    out = Cell.Extract(pout);
+    text+=usecond();
+    ttot+=usecond();
+    
+    std::cout << GridLogPerformance<<"Coarse 1rhs Mult Aviews "<<tviews<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult exch "<<texch<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<" of which mult2  "<<tmult2<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult ext  "<<text<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult temps "<<ttemps<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult copy  "<<tcopy<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult tot  "<<ttot<<" us"<<std::endl;
+    //    std::cout << GridLogPerformance<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Kernel flops "<< flops<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Kernel bytes/s "<< bytes/tmult<<" MB/s"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+
+  };
+  
+  void PopulateAdag(void)
+  {
+    for(int64_t bidx=0;bidx<CoarseGrid()->gSites() ;bidx++){
+      Coordinate bcoor;
+      CoarseGrid()->GlobalIndexToGlobalCoor(bidx,bcoor);
+      
+      for(int p=0;p<geom.npoint;p++){
+	Coordinate scoor = bcoor;
+	for(int mu=0;mu<bcoor.size();mu++){
+	  int L = CoarseGrid()->GlobalDimensions()[mu];
+	  scoor[mu] = (bcoor[mu] - geom.shifts[p][mu] + L) % L; // Modulo arithmetic
+	}
+	// Flip to poke/peekLocalSite and not too bad
+	auto link = peekSite(_A[p],scoor);
+	int pp = geom.Reverse(p);
+	pokeSite(adj(link),_Adag[pp],bcoor);
+      }
+    }
+  }
+  /////////////////////////////////////////////////////////////
+  // 
+  // A) Only reduced flops option is to use a padded cell of depth 4
+  // and apply MpcDagMpc in the padded cell.
+  //
+  // Makes for ONE application of MpcDagMpc per vector instead of 30 or 80.
+  // With the effective cell size around (B+8)^4 perhaps 12^4/4^4 ratio
+  // Cost is 81x more, same as stencil size.
+  //
+  // But: can eliminate comms and do as local dirichlet.
+  //
+  // Local exchange gauge field once.
+  // Apply to all vectors, local only computation.
+  // Must exchange ghost subcells in reverse process of PaddedCell to take inner products
+  //
+  // B) Can reduce cost: pad by 1, apply Deo      (4^4+6^4+8^4+8^4 )/ (4x 4^4)
+  //                     pad by 2, apply Doe
+  //                     pad by 3, apply Deo
+  //                     then break out 8x directions; cost is ~10x MpcDagMpc per vector
+  //
+  // => almost factor of 10 in setup cost, excluding data rearrangement
+  //
+  // Intermediates -- ignore the corner terms, leave approximate and force Hermitian
+  // Intermediates -- pad by 2 and apply 1+8+24 = 33 times.
+  /////////////////////////////////////////////////////////////
+
+    //////////////////////////////////////////////////////////
+    // BFM HDCG style approach: Solve a system of equations to get Aij
+    //////////////////////////////////////////////////////////
+    /*
+     *     Here, k,l index which possible shift within the 3^Nd "ball" connected by MdagM.
+     *
+     *     conj(phases[block]) proj[k][ block*Nvec+j ] =  \sum_ball  e^{i q_k . delta} < phi_{block,j} | MdagM | phi_{(block+delta),i} > 
+     *                                                 =  \sum_ball e^{iqk.delta} A_ji
+     *
+     *     Must invert matrix M_k,l = e^[i q_k . delta_l]
+     *
+     *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+     */
+#if 0
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
+    GridBase *grid = FineGrid();
+
+    RealD tproj=0.0;
+    RealD teigen=0.0;
+    RealD tmat=0.0;
+    RealD tphase=0.0;
+    RealD tinv=0.0;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid()); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    const int npoint = geom.npoint;
+      
+    Coordinate clatt = CoarseGrid()->GlobalDimensions();
+    int Nd = CoarseGrid()->Nd();
+
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    teigen-=usecond();
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom.shifts[k][mu]*geom.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+    teigen+=usecond();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    CoarseVector coarseInner(CoarseGrid());
+
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid());
+    std::vector<CoarseVector>          FT(npoint,CoarseGrid());
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+	/////////////////////////////////////////////////////
+	// Stick a phase on every block
+	/////////////////////////////////////////////////////
+	tphase-=usecond();
+	CoarseComplexField coor(CoarseGrid());
+	CoarseComplexField pha(CoarseGrid());	pha=Zero();
+	for(int mu=0;mu<Nd;mu++){
+	  LatticeCoordinate(coor,mu);
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  pha = pha + (TwoPiL * geom.shifts[p][mu]) * coor;
+	}
+	pha  =exp(pha*ci);
+	phaV=Zero();
+	blockZAXPY(phaV,pha,Subspace.subspace[i],phaV);
+	tphase+=usecond();
+
+	/////////////////////////////////////////////////////////////////////
+	// Multiple phased subspace vector by matrix and project to subspace
+	// Remove local bulk phase to leave relative phases
+	/////////////////////////////////////////////////////////////////////
+	tmat-=usecond();
+	linop.Op(phaV,MphaV);
+	tmat+=usecond();
+
+	tproj-=usecond();
+	blockProject(coarseInner,MphaV,Subspace.subspace);
+	coarseInner = conjugate(pha) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+	tproj+=usecond();
+
+      }
+
+      tinv-=usecond();
+      for(int k=0;k<npoint;k++){
+	FT[k] = Zero();
+	for(int l=0;l<npoint;l++){
+	  FT[k]= FT[k]+ invMkl(l,k)*ComputeProj[l];
+	}
+      
+	int osites=CoarseGrid()->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT[k], AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    // Only needed if nonhermitian
+    if ( ! hermitian ) {
+      //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      //      PopulateAdag();
+    }
+
+    // Need to write something to populate Adag from A
+    ExchangeCoarseLinks();
+    std::cout << GridLogMessage<<"CoarsenOperator eigen  "<<teigen<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+  }
+#else
+  //////////////////////////////////////////////////////////////////////
+  // Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    CoarsenOperator(linop,Subspace,Subspace);
+  }
+  //////////////////////////////////////////////////////////////////////
+  // Petrov - Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & U,
+		       Aggregation<Fobj,CComplex,nbasis> & V)
+  {
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
+    GridBase *grid = FineGrid();
+
+    RealD tproj=0.0;
+    RealD teigen=0.0;
+    RealD tmat=0.0;
+    RealD tphase=0.0;
+    RealD tphaseBZ=0.0;
+    RealD tinv=0.0;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid()); 
+    blockOrthogonalise(InnerProd,V.subspace);
+    blockOrthogonalise(InnerProd,U.subspace);
+
+    const int npoint = geom.npoint;
+      
+    Coordinate clatt = CoarseGrid()->GlobalDimensions();
+    int Nd = CoarseGrid()->Nd();
+
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    teigen-=usecond();
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom.shifts[k][mu]*geom.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+    teigen+=usecond();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid());
+    
+    CoarseVector coarseInner(CoarseGrid());
+    
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    tphase=-usecond();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid());
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+      
+    }
+    tphase+=usecond();
+    
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid());
+    std::vector<CoarseVector>          FT(npoint,CoarseGrid());
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+	tphaseBZ-=usecond();
+	phaV = phaF[p]*V.subspace[i];
+	tphaseBZ+=usecond();
+
+	/////////////////////////////////////////////////////////////////////
+	// Multiple phased subspace vector by matrix and project to subspace
+	// Remove local bulk phase to leave relative phases
+	/////////////////////////////////////////////////////////////////////
+	tmat-=usecond();
+	linop.Op(phaV,MphaV);
+	tmat+=usecond();
+	//	std::cout << i << " " <<p << " MphaV "<<norm2(MphaV)<<" "<<norm2(phaV)<<std::endl;
+
+	tproj-=usecond();
+	blockProject(coarseInner,MphaV,U.subspace);
+	coarseInner = conjugate(pha[p]) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+	tproj+=usecond();
+	//	std::cout << i << " " <<p << " ComputeProj "<<norm2(ComputeProj[p])<<std::endl;
+
+      }
+
+      tinv-=usecond();
+      for(int k=0;k<npoint;k++){
+	FT[k] = Zero();
+	for(int l=0;l<npoint;l++){
+	  FT[k]= FT[k]+ invMkl(l,k)*ComputeProj[l];
+	}
+      
+	int osites=CoarseGrid()->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT[k], AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    // Only needed if nonhermitian
+    if ( ! hermitian ) {
+      //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      //      PopulateAdag();
+    }
+
+    for(int p=0;p<geom.npoint;p++){
+      std::cout << " _A["<<p<<"] "<<norm2(_A[p])<<std::endl;
+    }
+
+    // Need to write something to populate Adag from A
+    ExchangeCoarseLinks();
+    std::cout << GridLogMessage<<"CoarsenOperator eigen  "<<teigen<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phaseBZ "<<tphaseBZ<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+  }
+#endif  
+  void ExchangeCoarseLinks(void){
+    for(int p=0;p<geom.npoint;p++){
+      _A[p] = Cell.ExchangePeriodic(_A[p]);
+      //      _Adag[p]= Cell.ExchangePeriodic(_Adag[p]);
+    }
+  }
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
+  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
+  virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
+};
+
+
+  
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h

@@ -0,0 +1,729 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrixMultiRHS.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class MultiGeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+  typedef typename CComplex::scalar_object SComplex;
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef MultiGeneralCoarsenedMatrix<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
+
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  typedef iVector<SComplex,nbasis >           calcVector;
+  typedef iMatrix<SComplex,nbasis >           calcMatrix;
+  typedef Lattice<iScalar<CComplex> >         CoarseComplexField;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+  typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef iVector<CComplex,nbasis >  Cvec;
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+  typedef Lattice<CComplex >    FineComplexField;
+  typedef CoarseVector Field;
+
+  ////////////////////
+  // Data members
+  ////////////////////
+  GridCartesian *       _CoarseGridMulti; 
+  NonLocalStencilGeometry geom;
+  NonLocalStencilGeometry geom_srhs;
+  PaddedCell Cell;
+  GeneralLocalStencil Stencil;
+
+  deviceVector<calcVector> BLAS_B;
+  deviceVector<calcVector> BLAS_C;
+  std::vector<deviceVector<calcMatrix> > BLAS_A;
+
+  std::vector<deviceVector<ComplexD *> > BLAS_AP;
+  std::vector<deviceVector<ComplexD *> > BLAS_BP;
+  deviceVector<ComplexD *>               BLAS_CP;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase      * Grid(void)           { return _CoarseGridMulti; };   // this is all the linalg routines need to know
+  GridCartesian * CoarseGrid(void)     { return _CoarseGridMulti; };   // this is all the linalg routines need to know
+
+  // Can be used to do I/O on the operator matrices externally
+  void SetMatrix (int p,CoarseMatrix & A)
+  {
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
+    GridtoBLAS(A[p],BLAS_A[p]);
+  }
+  void GetMatrix (int p,CoarseMatrix & A)
+  {
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
+    BLAStoGrid(A[p],BLAS_A[p]);
+  }
+  void CopyMatrix (GeneralCoarseOp &_Op)
+  {
+    for(int p=0;p<geom.npoint;p++){
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      //Unpadded
+      GridtoBLAS(Aup,BLAS_A[p]);
+    }
+  }
+  /*
+  void CheckMatrix (GeneralCoarseOp &_Op)
+  {
+    std::cout <<"************* Checking the little direc operator mRHS"<<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      //Unpadded
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      auto Ack = Aup;
+      BLAStoGrid(Ack,BLAS_A[p]);
+      std::cout << p<<" Ack "<<norm2(Ack)<<std::endl;
+      std::cout << p<<" Aup "<<norm2(Aup)<<std::endl;
+    }
+    std::cout <<"************* "<<std::endl;
+  }
+  */
+  
+  MultiGeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGridMulti) :
+    _CoarseGridMulti(CoarseGridMulti),
+    geom_srhs(_geom),
+    geom(_CoarseGridMulti,_geom.hops,_geom.skip+1),
+    Cell(geom.Depth(),_CoarseGridMulti),
+    Stencil(Cell.grids.back(),geom.shifts) // padded cell stencil
+  {
+    int32_t padded_sites   = Cell.grids.back()->lSites();
+    int32_t unpadded_sites = CoarseGridMulti->lSites();
+    
+    int32_t nrhs  = CoarseGridMulti->FullDimensions()[0];  // # RHS
+    int32_t orhs  = nrhs/CComplex::Nsimd();
+
+    padded_sites   = padded_sites/nrhs;
+    unpadded_sites = unpadded_sites/nrhs;
+    
+    /////////////////////////////////////////////////
+    // Device data vector storage
+    /////////////////////////////////////////////////
+    BLAS_A.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_A[p].resize (unpadded_sites); // no ghost zone, npoint elements
+    }
+    
+    BLAS_B.resize(nrhs *padded_sites);   // includes ghost zone
+    BLAS_C.resize(nrhs *unpadded_sites); // no ghost zone
+    BLAS_AP.resize(geom.npoint);
+    BLAS_BP.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_AP[p].resize(unpadded_sites);
+      BLAS_BP[p].resize(unpadded_sites);
+    }
+    BLAS_CP.resize(unpadded_sites);
+
+    /////////////////////////////////////////////////
+    // Pointers to data
+    /////////////////////////////////////////////////
+
+    // Site identity mapping for A
+    for(int p=0;p<geom.npoint;p++){
+      for(int ss=0;ss<unpadded_sites;ss++){
+	ComplexD *ptr = (ComplexD *)&BLAS_A[p][ss];
+	acceleratorPut(BLAS_AP[p][ss],ptr);
+      }
+    }
+    // Site identity mapping for C
+    for(int ss=0;ss<unpadded_sites;ss++){
+      ComplexD *ptr = (ComplexD *)&BLAS_C[ss*nrhs];
+      acceleratorPut(BLAS_CP[ss],ptr);
+    }
+
+    // Neighbour table is more complicated
+    int32_t j=0; // Interior point counter (unpadded)
+    for(int32_t s=0;s<padded_sites;s++){ // 4 volume, padded
+      int ghost_zone=0;
+      for(int32_t point = 0 ; point < geom.npoint; point++){
+	int i=s*orhs*geom.npoint+point;
+	if( Stencil._entries[i]._wrap ) { // stencil is indexed by the oSite of the CoarseGridMulti, hence orhs factor
+	  ghost_zone=1; // If general stencil wrapped in any direction, wrap=1
+	}
+      }
+
+      if( ghost_zone==0) {
+	for(int32_t point = 0 ; point < geom.npoint; point++){
+	  int i=s*orhs*geom.npoint+point;
+ 	  int32_t nbr = Stencil._entries[i]._offset*CComplex::Nsimd(); // oSite -> lSite
+	  GRID_ASSERT(nbr<BLAS_B.size());
+	  ComplexD * ptr = (ComplexD *)&BLAS_B[nbr];
+	  acceleratorPut(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
+	}
+	j++;
+      }
+    }
+    GRID_ASSERT(j==unpadded_sites);
+  }
+  template<class vobj> void GridtoBLAS(const Lattice<vobj> &from,deviceVector<typename vobj::scalar_object> &to)
+  {
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Fg = from.Grid();
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
+  int nd = Fg->_ndimension;
+
+  to.resize(Fg->lSites());
+
+  Coordinate LocalLatt = Fg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+
+  autoView(from_v,from,AcceleratorRead);
+  auto to_v = &to[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate from_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	from_coor[i] = base[i];
+      }
+      int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      scalar_type* to = (scalar_type *)&to_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	to[w] = stmp;
+      }
+    });
+  }    
+  template<class vobj> void BLAStoGrid(Lattice<vobj> &grid,deviceVector<typename vobj::scalar_object> &in)
+  {
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Tg = grid.Grid();
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
+  int nd = Tg->_ndimension;
+  
+  GRID_ASSERT(in.size()==Tg->lSites());
+
+  Coordinate LocalLatt = Tg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
+
+  autoView(to_v,grid,AcceleratorWrite);
+  auto from_v = &in[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate to_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	to_coor[i] = base[i];
+      }
+      int to_oidx = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      scalar_type* from = (scalar_type *)&from_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp=from[w];
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  }
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace,
+		       GridBase *CoarseGrid)
+  {
+#if 0
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrixMrhs "<< std::endl;
+
+    GridBase *grid = Subspace.FineGrid;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    const int npoint = geom_srhs.npoint;
+
+    Coordinate clatt = CoarseGrid->GlobalDimensions();
+    int Nd = CoarseGrid->Nd();
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom_srhs.shifts[k][mu]*geom_srhs.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid);
+    
+    CoarseVector coarseInner(CoarseGrid);
+    
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid);
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom_srhs.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);	
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+    }
+
+    // Could save on temporary storage here
+    std::vector<CoarseMatrix> _A;
+    _A.resize(geom_srhs.npoint,CoarseGrid);
+
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid);
+    CoarseVector          FT(CoarseGrid);
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+
+	phaV = phaF[p]*Subspace.subspace[i];
+
+	/////////////////////////////////////////////////////////////////////
+	// Multiple phased subspace vector by matrix and project to subspace
+	// Remove local bulk phase to leave relative phases
+	/////////////////////////////////////////////////////////////////////
+	linop.Op(phaV,MphaV);
+
+	// Fixme, could use batched block projector here
+	blockProject(coarseInner,MphaV,Subspace.subspace);
+
+	coarseInner = conjugate(pha[p]) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+      }
+
+      // Could do this with a block promote or similar BLAS call via the MultiRHSBlockProjector with a const matrix.
+      for(int k=0;k<npoint;k++){
+
+	FT = Zero();
+	for(int l=0;l<npoint;l++){
+	  FT= FT+ invMkl(l,k)*ComputeProj[l];
+	}
+      
+	int osites=CoarseGrid->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT, AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+    }
+
+    // Only needed if nonhermitian
+    //    if ( ! hermitian ) {
+    //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+    //      PopulateAdag();
+    //    }
+    // Need to write something to populate Adag from A
+
+    for(int p=0;p<geom_srhs.npoint;p++){
+      GridtoBLAS(_A[p],BLAS_A[p]);
+    }
+    /*
+Grid : Message : 11698.730546 s : CoarsenOperator eigen  1334 us
+Grid : Message : 11698.730563 s : CoarsenOperator phase  34729 us
+Grid : Message : 11698.730565 s : CoarsenOperator phaseBZ 2423814 us
+Grid : Message : 11698.730566 s : CoarsenOperator mat    127890998 us
+Grid : Message : 11698.730567 s : CoarsenOperator proj   515840840 us
+Grid : Message : 11698.730568 s : CoarsenOperator inv    103948313 us
+Takes 600s to compute matrix elements, DOMINATED by the block project.
+Easy to speed up with the batched block project.
+Store npoint vectors, get npoint x Nbasis block projection, and 81 fold faster.
+
+// Block project below taks to 240s
+Grid : Message : 328.193418 s : CoarsenOperator phase      38338 us
+Grid : Message : 328.193434 s : CoarsenOperator phaseBZ  1711226 us
+Grid : Message : 328.193436 s : CoarsenOperator mat    122213270 us
+//Grid : Message : 328.193438 s : CoarsenOperator proj   1181154 us <-- this is mistimed
+//Grid : Message : 11698.730568 s : CoarsenOperator inv  103948313 us <-- Cut this ~10x if lucky by loop fusion
+     */
+#else
+    RealD tproj=0.0;
+    RealD tmat=0.0;
+    RealD tphase=0.0;
+    RealD tphaseBZ=0.0;
+    RealD tinv=0.0;
+
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrixMrhs "<< std::endl;
+
+    GridBase *grid = Subspace.FineGrid;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+
+    MultiRHSBlockProject<Lattice<Fobj> >    Projector;
+    Projector.Allocate(nbasis,grid,CoarseGrid);
+    Projector.ImportBasis(Subspace.subspace);
+    
+    const int npoint = geom_srhs.npoint;
+
+    Coordinate clatt = CoarseGrid->GlobalDimensions();
+    int Nd = CoarseGrid->Nd();
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom_srhs.shifts[k][mu]*geom_srhs.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid);
+    
+    CoarseVector coarseInner(CoarseGrid);
+    
+    tphase=-usecond();
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid);
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom_srhs.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);	
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+    }
+    tphase+=usecond();
+
+    // Could save on temporary storage here
+    std::vector<CoarseMatrix> _A;
+    _A.resize(geom_srhs.npoint,CoarseGrid);
+
+    // Count use small chunks than npoint == 81 and save memory
+    int batch = 9;
+    std::vector<FineField>    _MphaV(batch,grid);
+    std::vector<CoarseVector> TmpProj(batch,CoarseGrid);
+
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid);
+    CoarseVector          FT(CoarseGrid);
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+
+      //      std::cout << GridLogMessage << " phasing the fine vector "<<std::endl;
+      // Fixme : do this in batches
+      for(int p=0;p<npoint;p+=batch){ // Loop over momenta in npoint
+
+	for(int b=0;b<MIN(batch,npoint-p);b++){
+	  tphaseBZ-=usecond();
+	  phaV = phaF[p+b]*Subspace.subspace[i];
+	  tphaseBZ+=usecond();
+
+	  /////////////////////////////////////////////////////////////////////
+	  // Multiple phased subspace vector by matrix and project to subspace
+	  // Remove local bulk phase to leave relative phases
+	  /////////////////////////////////////////////////////////////////////
+	  // Memory footprint was an issue
+	  tmat-=usecond();
+	  linop.Op(phaV,MphaV);
+	  _MphaV[b] = MphaV;
+	  tmat+=usecond();
+	}      
+
+	//	std::cout << GridLogMessage << " Calling block project "<<std::endl;
+	tproj-=usecond();
+	Projector.blockProject(_MphaV,TmpProj);
+	tproj+=usecond();
+	
+	//	std::cout << GridLogMessage << " conj phasing the coarse vectors "<<std::endl;
+	for(int b=0;b<MIN(batch,npoint-p);b++){
+	  ComputeProj[p+b] = conjugate(pha[p+b])*TmpProj[b];
+	}
+      }
+
+      // Could do this with a block promote or similar BLAS call via the MultiRHSBlockProjector with a const matrix.
+      
+      // std::cout << GridLogMessage << " Starting FT inv "<<std::endl;
+      tinv-=usecond();
+      for(int k=0;k<npoint;k++){
+	FT = Zero();
+	// 81 kernel calls as many ComputeProj vectors
+	// Could fuse with a vector of views, but ugly
+	// Could unroll the expression and run fewer kernels -- much more attractive
+	// Could also do non blocking.
+#if 0	
+	for(int l=0;l<npoint;l++){
+	  FT= FT+ invMkl(l,k)*ComputeProj[l];
+	}
+#else
+	const int radix = 9;
+	int ll;
+	for(ll=0;ll+radix-1<npoint;ll+=radix){
+	  // When ll = npoint-radix, ll+radix-1 = npoint-1, and we do it all.
+	  FT = FT 
+	    + invMkl(ll+0,k)*ComputeProj[ll+0]
+	    + invMkl(ll+1,k)*ComputeProj[ll+1]
+	    + invMkl(ll+2,k)*ComputeProj[ll+2]
+	    + invMkl(ll+3,k)*ComputeProj[ll+3]
+	    + invMkl(ll+4,k)*ComputeProj[ll+4]
+	    + invMkl(ll+5,k)*ComputeProj[ll+5]
+	    + invMkl(ll+6,k)*ComputeProj[ll+6]
+	    + invMkl(ll+7,k)*ComputeProj[ll+7]
+	    + invMkl(ll+8,k)*ComputeProj[ll+8];
+	}
+	for(int l=ll;l<npoint;l++){
+	  FT= FT+ invMkl(l,k)*ComputeProj[l];
+	}
+#endif
+      
+	// 1 kernel call -- must be cheaper
+	int osites=CoarseGrid->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT, AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    // Only needed if nonhermitian
+    //    if ( ! hermitian ) {
+    //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+    //      PopulateAdag();
+    //    }
+    // Need to write something to populate Adag from A
+    //    std::cout << GridLogMessage << " Calling GridtoBLAS "<<std::endl;
+    for(int p=0;p<geom_srhs.npoint;p++){
+      GridtoBLAS(_A[p],BLAS_A[p]);
+    }
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phaseBZ "<<tphaseBZ<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+#endif
+  }
+  void Mdag(const CoarseVector &in, CoarseVector &out)
+  {
+    this->M(in,out);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    //    std::cout << GridLogMessage << "New Mrhs coarse"<<std::endl;
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    RealD t_tot;
+    RealD t_exch;
+    RealD t_GtoB;
+    RealD t_BtoG;
+    RealD t_mult;
+
+    t_tot=-usecond();
+    CoarseVector tin=in;
+    t_exch=-usecond();
+    CoarseVector pin = Cell.ExchangePeriodic(tin); //padded input
+    t_exch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef calcMatrix* Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+
+    int64_t nrhs  =pin.Grid()->GlobalDimensions()[0];
+    GRID_ASSERT(nrhs>=1);
+
+    RealD flops,bytes;
+    int64_t osites=in.Grid()->oSites(); // unpadded
+    int64_t unpadded_vol = CoarseGrid()->lSites()/nrhs;
+    
+    flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    bytes = 1.0*osites*sizeof(siteMatrix)*npoint/pin.Grid()->GlobalDimensions()[0]
+          + 2.0*osites*sizeof(siteVector)*npoint;
+    
+
+    t_GtoB=-usecond();
+    GridtoBLAS(pin,BLAS_B);
+    t_GtoB+=usecond();
+
+    GridBLAS BLAS;
+
+    t_mult=-usecond();
+    for(int p=0;p<geom.npoint;p++){
+      RealD c = 1.0;
+      if (p==0) c = 0.0;
+      ComplexD beta(c);
+
+      BLAS.gemmBatched(nbasis,nrhs,nbasis,
+		       ComplexD(1.0),
+		       BLAS_AP[p], 
+		       BLAS_BP[p], 
+		       ComplexD(c), 
+		       BLAS_CP);
+    }
+    BLAS.synchronise();
+    t_mult+=usecond();
+
+    t_BtoG=-usecond();
+    BLAStoGrid(out,BLAS_C);
+    t_BtoG+=usecond();
+    t_tot+=usecond();
+    /*
+    std::cout << GridLogMessage << "New Mrhs coarse DONE "<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult exch "<<t_exch<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult mult "<<t_mult<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult GtoB  "<<t_GtoB<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult BtoG  "<<t_BtoG<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult tot  "<<t_tot<<" us"<<std::endl;
+    */
+    //    std::cout << GridLogMessage<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel flops "<< flops<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel flop/s "<< flops/t_mult<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel bytes/s "<< bytes/t_mult/1000<<" GB/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse overall flops/s "<< flops/t_tot<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+  };
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
+  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
+  virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
+};
+  
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/Geometry.h

@@ -0,0 +1,238 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrix.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/////////////////////////////////////////////////////////////////
+// Geometry class in cartesian case
+/////////////////////////////////////////////////////////////////
+
+class Geometry {
+public:
+  int npoint;
+  int base;
+  std::vector<int> directions   ;
+  std::vector<int> displacements;
+  std::vector<int> points_dagger;
+
+  Geometry(int _d)  {
+    
+    base = (_d==5) ? 1:0;
+
+    // make coarse grid stencil for 4d , not 5d
+    if ( _d==5 ) _d=4;
+
+    npoint = 2*_d+1;
+    directions.resize(npoint);
+    displacements.resize(npoint);
+    points_dagger.resize(npoint);
+    for(int d=0;d<_d;d++){
+      directions[d   ] = d+base;
+      directions[d+_d] = d+base;
+      displacements[d  ] = +1;
+      displacements[d+_d]= -1;
+      points_dagger[d   ] = d+_d;
+      points_dagger[d+_d] = d;
+    }
+    directions   [2*_d]=0;
+    displacements[2*_d]=0;
+    points_dagger[2*_d]=2*_d;
+  }
+
+  int point(int dir, int disp) {
+    GRID_ASSERT(disp == -1 || disp == 0 || disp == 1);
+    GRID_ASSERT(base+0 <= dir && dir < base+4);
+
+    // directions faster index = new indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  1  2  3  0  1  2  3  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  2  3  4  1  2  3  4  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+
+    // displacements faster index = old indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  0  1  1  2  2  3  3  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  1  2  2  3  3  4  4  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+
+    if(dir == 0 and disp == 0)
+      return 8;
+    else // New indexing
+      return (1 - disp) / 2 * 4 + dir - base;
+    // else // Old indexing
+    //   return (4 * (dir - base) + 1 - disp) / 2;
+  }
+};
+
+/////////////////////////////////////////////////////////////////
+// Less local equivalent of Geometry class in cartesian case
+/////////////////////////////////////////////////////////////////
+class NonLocalStencilGeometry {
+public:
+  //  int depth;
+  int skip;
+  int hops;
+  int npoint;
+  std::vector<Coordinate> shifts;
+  Coordinate stencil_size;
+  Coordinate stencil_lo;
+  Coordinate stencil_hi;
+  GridCartesian *grid;
+  GridCartesian *Grid() {return grid;};
+  int Depth(void){return 1;};   // Ghost zone depth
+  int Hops(void){return hops;}; // # of hops=> level of corner fill in in stencil
+  int DimSkip(void){return skip;};
+
+  virtual ~NonLocalStencilGeometry() {};
+
+  int  Reverse(int point)
+  {
+    int Nd = Grid()->Nd();
+    Coordinate shft = shifts[point];
+    Coordinate rev(Nd);
+    for(int mu=0;mu<Nd;mu++) rev[mu]= -shft[mu];
+    for(int p=0;p<npoint;p++){
+      if(rev==shifts[p]){
+	return p;
+      }
+    }
+    GRID_ASSERT(0);
+    return -1;
+  }
+  void BuildShifts(void)
+  {
+    this->shifts.resize(0);
+    int Nd = this->grid->Nd();
+
+    int dd = this->DimSkip();
+    for(int s0=this->stencil_lo[dd+0];s0<=this->stencil_hi[dd+0];s0++){
+    for(int s1=this->stencil_lo[dd+1];s1<=this->stencil_hi[dd+1];s1++){
+    for(int s2=this->stencil_lo[dd+2];s2<=this->stencil_hi[dd+2];s2++){
+    for(int s3=this->stencil_lo[dd+3];s3<=this->stencil_hi[dd+3];s3++){
+      Coordinate sft(Nd,0);
+      sft[dd+0] = s0;
+      sft[dd+1] = s1;
+      sft[dd+2] = s2;
+      sft[dd+3] = s3;
+      int nhops = abs(s0)+abs(s1)+abs(s2)+abs(s3);
+      if(nhops<=this->hops) this->shifts.push_back(sft);
+    }}}}
+    this->npoint = this->shifts.size();
+    std::cout << GridLogMessage << "NonLocalStencilGeometry has "<< this->npoint << " terms in stencil "<<std::endl;
+  }
+  
+  NonLocalStencilGeometry(GridCartesian *_coarse_grid,int _hops,int _skip) : grid(_coarse_grid), hops(_hops), skip(_skip)
+  {
+    Coordinate latt = grid->GlobalDimensions();
+    stencil_size.resize(grid->Nd());
+    stencil_lo.resize(grid->Nd());
+    stencil_hi.resize(grid->Nd());
+    for(int d=0;d<grid->Nd();d++){
+     if ( latt[d] == 1 ) {
+      stencil_lo[d] = 0;
+      stencil_hi[d] = 0;
+      stencil_size[d]= 1;
+     } else if ( latt[d] == 2 ) {
+      stencil_lo[d] = -1;
+      stencil_hi[d] = 0;
+      stencil_size[d]= 2;
+     } else if ( latt[d] > 2 ) {
+       stencil_lo[d] = -1;
+       stencil_hi[d] =  1;
+       stencil_size[d]= 3;
+     }
+    }
+    this->BuildShifts();
+  };
+
+};
+
+// Need to worry about red-black now
+class NonLocalStencilGeometry4D : public NonLocalStencilGeometry {
+public:
+  virtual int DerivedDimSkip(void) { return 0;};
+  NonLocalStencilGeometry4D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,0) { };
+  virtual ~NonLocalStencilGeometry4D() {};
+};
+class NonLocalStencilGeometry5D : public NonLocalStencilGeometry {
+public:
+  virtual int DerivedDimSkip(void) { return 1; }; 
+  NonLocalStencilGeometry5D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,1)  { };
+  virtual ~NonLocalStencilGeometry5D() {};
+};
+/*
+ * Bunch of different options classes
+ */
+class NextToNextToNextToNearestStencilGeometry4D : public NonLocalStencilGeometry4D {
+public:
+  NextToNextToNextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,4)
+  {
+  };
+};
+class NextToNextToNextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNextToNextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,4)
+  {
+  };
+};
+class NextToNearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,2)
+  {
+  };
+};
+class NextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,2)
+  {
+  };
+};
+class NearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,1)
+  {
+  };
+};
+class NearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,1)
+  {
+  };
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/MultiGrid.h

@@ -0,0 +1,34 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Grid/algorithms/multigrid/MultiGrid.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@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 */
+#pragma once
+
+#include <Grid/algorithms/multigrid/Aggregates.h>
+#include <Grid/algorithms/multigrid/Geometry.h>
+#include <Grid/algorithms/multigrid/CoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h>

Grid/allocator/AlignedAllocator.cc

@@ -1,154 +0,0 @@
-#include <Grid/GridCore.h>
-#include <fcntl.h>
-
-NAMESPACE_BEGIN(Grid);
-
-MemoryStats *MemoryProfiler::stats = nullptr;
-bool         MemoryProfiler::debug = false;
-
-int PointerCache::NcacheSmall = PointerCache::NcacheSmallMax;
-#ifdef GRID_CUDA
-int PointerCache::Ncache      = 32;
-#else 
-int PointerCache::Ncache      = 8;
-#endif
-int PointerCache::Victim;
-int PointerCache::VictimSmall;
-PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::NcacheMax];
-PointerCache::PointerCacheEntry PointerCache::EntriesSmall[PointerCache::NcacheSmallMax];
-
-void PointerCache::Init(void)
-{
-  char * str;
-
-  str= getenv("GRID_ALLOC_NCACHE_LARGE");
-  if ( str ) Ncache = atoi(str);
-  if ( (Ncache<0) || (Ncache > NcacheMax)) Ncache = NcacheMax;
-
-  str= getenv("GRID_ALLOC_NCACHE_SMALL");
-  if ( str ) NcacheSmall = atoi(str);
-  if ( (NcacheSmall<0) || (NcacheSmall > NcacheSmallMax)) NcacheSmall = NcacheSmallMax;
-
-  //  printf("Aligned alloocator cache: large %d/%d small %d/%d\n",Ncache,NcacheMax,NcacheSmall,NcacheSmallMax);
-}
-void *PointerCache::Insert(void *ptr,size_t bytes) 
-{
-  if (bytes < GRID_ALLOC_SMALL_LIMIT ) 
-    return Insert(ptr,bytes,EntriesSmall,NcacheSmall,VictimSmall);
-  return Insert(ptr,bytes,Entries,Ncache,Victim);  
-}
-void *PointerCache::Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) 
-{
-#ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
-#endif 
-
-  void * ret = NULL;
-  int v = -1;
-
-  for(int e=0;e<ncache;e++) {
-    if ( entries[e].valid==0 ) {
-      v=e; 
-      break;
-    }
-  }
-
-  if ( v==-1 ) {
-    v=victim;
-    victim = (victim+1)%ncache;
-  }
-
-  if ( entries[v].valid ) {
-    ret = entries[v].address;
-    entries[v].valid = 0;
-    entries[v].address = NULL;
-    entries[v].bytes = 0;
-  }
-
-  entries[v].address=ptr;
-  entries[v].bytes  =bytes;
-  entries[v].valid  =1;
-
-  return ret;
-}
-
-void *PointerCache::Lookup(size_t bytes)
-{
-  if (bytes < GRID_ALLOC_SMALL_LIMIT ) 
-    return Lookup(bytes,EntriesSmall,NcacheSmall);
-  return Lookup(bytes,Entries,Ncache);
-}
-void *PointerCache::Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) 
-{
-#ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
-#endif 
-  for(int e=0;e<ncache;e++){
-    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
-      entries[e].valid = 0;
-      return entries[e].address;
-    }
-  }
-  return NULL;
-}
-
-
-void check_huge_pages(void *Buf,uint64_t BYTES)
-{
-#ifdef __linux__
-  int fd = open("/proc/self/pagemap", O_RDONLY);
-  assert(fd >= 0);
-  const int page_size = 4096;
-  uint64_t virt_pfn = (uint64_t)Buf / page_size;
-  off_t offset = sizeof(uint64_t) * virt_pfn;
-  uint64_t npages = (BYTES + page_size-1) / page_size;
-  uint64_t pagedata[npages];
-  uint64_t ret = lseek(fd, offset, SEEK_SET);
-  assert(ret == offset);
-  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
-  assert(ret == sizeof(uint64_t) * npages);
-  int nhugepages = npages / 512;
-  int n4ktotal, nnothuge;
-  n4ktotal = 0;
-  nnothuge = 0;
-  for (int i = 0; i < nhugepages; ++i) {
-    uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
-    for (int j = 0; j < 512; ++j) {
-      uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
-      ++n4ktotal;
-      if (pageaddr != baseaddr + j * page_size)
-	++nnothuge;
-    }
-  }
-  int rank = CartesianCommunicator::RankWorld();
-  printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
-#endif
-}
-
-std::string sizeString(const size_t bytes)
-{
-  constexpr unsigned int bufSize = 256;
-  const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
-  char                   buf[256];
-  size_t                 s     = 0;
-  double                 count = bytes;
-  
-  while (count >= 1024 && s < 7)
-    {
-      s++;
-      count /= 1024;
-    }
-  if (count - floor(count) == 0.0)
-    {
-      snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
-    }
-  else
-    {
-      snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
-    }
-  
-  return std::string(buf);
-}
-
-NAMESPACE_END(Grid);
-

Grid/allocator/AlignedAllocator.h

@@ -26,129 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_ALIGNED_ALLOCATOR_H
-#define GRID_ALIGNED_ALLOCATOR_H
-
-#ifdef HAVE_MALLOC_MALLOC_H
-#include <malloc/malloc.h>
-#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_MM_MALLOC_H
-#include <mm_malloc.h>
-#endif
-
-#define POINTER_CACHE
-#define GRID_ALLOC_ALIGN (2*1024*1024)
-#define GRID_ALLOC_SMALL_LIMIT (4096)
+#pragma once
 
 NAMESPACE_BEGIN(Grid);
 
-// Move control to configure.ac and Config.h?
-
-class PointerCache {
-private:
-/*Pinning pages is costly*/
-/*Could maintain separate large and small allocation caches*/
-/* Could make these configurable, perhaps up to a max size*/
-  static const int NcacheSmallMax=128; 
-  static const int NcacheMax=16;
-  static int NcacheSmall;
-  static int Ncache;
-
-  typedef struct { 
-    void *address;
-    size_t bytes;
-    int valid;
-  } PointerCacheEntry;
-    
-  static PointerCacheEntry Entries[NcacheMax];
-  static int Victim;
-  static PointerCacheEntry EntriesSmall[NcacheSmallMax];
-  static int VictimSmall;
-
-public:
-  static void Init(void);
-  static void *Insert(void *ptr,size_t bytes) ;
-  static void *Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) ;
-  static void *Lookup(size_t bytes) ;
-  static void *Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) ;
-};
-
-std::string sizeString(size_t bytes);
-
-struct MemoryStats
-{
-  size_t totalAllocated{0}, maxAllocated{0}, 
-    currentlyAllocated{0}, totalFreed{0};
-};
-    
-class MemoryProfiler
-{
-public:
-  static MemoryStats *stats;
-  static bool        debug;
-};
-
-#ifdef GRID_NVCC
-#define profilerCudaMeminfo \
-  { size_t f, t ; cudaMemGetInfo ( &f,&t); std::cout << GridLogDebug << "[Memory debug] Cuda free "<<f<<"/"<<t << std::endl;}
-#else
-#define profilerCudaMeminfo
-#endif
-
-#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
-#define profilerDebugPrint						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl; \
-      std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
-		<< std::endl;						\
-      std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
-		<< std::endl;						\
-    }									\
-  profilerCudaMeminfo;
-
-#define profilerAllocate(bytes)						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      s->totalAllocated     += (bytes);					\
-      s->currentlyAllocated += (bytes);					\
-      s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated); \
-    }									\
-  if (MemoryProfiler::debug)						\
-    {									\
-      std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl; \
-      profilerDebugPrint;						\
-    }
-
-#define profilerFree(bytes)						\
-  if (MemoryProfiler::stats)						\
-    {									\
-      auto s = MemoryProfiler::stats;					\
-      s->totalFreed         += (bytes);					\
-      s->currentlyAllocated -= (bytes);					\
-    }									\
-  if (MemoryProfiler::debug)						\
-    {									\
-      std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl; \
-      profilerDebugPrint;						\
-    }
-
-void check_huge_pages(void *Buf,uint64_t BYTES);
-
-////////////////////////////////////////////////////////////////////
-// A lattice of something, but assume the something is SIMDized.
-////////////////////////////////////////////////////////////////////
-
 template<typename _Tp>
 class alignedAllocator {
 public: 
@@ -172,89 +53,170 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
-
-
-#ifdef POINTER_CACHE
-    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
-#else
-    pointer ptr = nullptr;
-#endif
-
-#ifdef GRID_NVCC
-    ////////////////////////////////////
-    // Unified (managed) memory
-    ////////////////////////////////////
-    if ( ptr == (_Tp *) NULL ) {
-      //      printf(" alignedAllocater cache miss %ld bytes ",bytes);      BACKTRACEFP(stdout);
-      auto err = cudaMallocManaged((void **)&ptr,bytes);
-      if( err != cudaSuccess ) {
-	ptr = (_Tp *) NULL;
-	std::cerr << " cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
-	assert(0);
-      }
-    } 
-    assert( ptr != (_Tp *)NULL);
-#else 
-    //////////////////////////////////////////////////////////////////////////////////////////
-    // 2MB align; could make option probably doesn't need configurability
-    //////////////////////////////////////////////////////////////////////////////////////////
-  #ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
-  #else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
-  #endif
-    assert( ptr != (_Tp *)NULL);
-
-    //////////////////////////////////////////////////
-    // First touch optimise in threaded loop 
-    //////////////////////////////////////////////////
-    uint64_t *cp = (uint64_t *)ptr;
-    thread_for(n,bytes/sizeof(uint64_t), { // need only one touch per page
-      cp[n]=0;
-    });
-#endif
+    _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
+    if ( (_Tp*)ptr == (_Tp *) NULL ) {
+      printf("Grid CPU Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
+    }
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
-  void deallocate(pointer __p, size_type __n) { 
+  void deallocate(pointer __p, size_type __n) 
+  { 
     size_type bytes = __n * sizeof(_Tp);
-
     profilerFree(bytes);
-
-#ifdef POINTER_CACHE
-    pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
-#else 
-    pointer __freeme = __p;
-#endif
-
-#ifdef GRID_NVCC
-    if ( __freeme ) cudaFree((void *)__freeme);
-#else 
-  #ifdef HAVE_MM_MALLOC_H
-    if ( __freeme ) _mm_free((void *)__freeme); 
-  #else
-    if ( __freeme ) free((void *)__freeme);
-  #endif
-#endif
+    MemoryManager::CpuFree((void *)__p,bytes);
   }
 
-  // FIXME: hack for the copy constructor, eventually it must be avoided
-  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
-  //void construct(pointer __p, const _Tp& __val) { };
+  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
+  void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
 template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
+//////////////////////////////////////////////////////////////////////////////////////
+// Unified virtual memory
+//////////////////////////////////////////////////////////////////////////////////////
+template<typename _Tp>
+class uvmAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef uvmAllocator<_Tp1> other; };
+  uvmAllocator() throw() { }
+  uvmAllocator(const uvmAllocator&) throw() { }
+  template<typename _Tp1> uvmAllocator(const uvmAllocator<_Tp1>&) throw() { }
+  ~uvmAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+  pointer allocate(size_type __n, const void* _p= 0)
+  { 
+    size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
+    _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
+    if ( (_Tp*)ptr == (_Tp *) NULL ) {
+      printf("Grid Shared Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
+    }
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
+    profilerFree(bytes);
+    MemoryManager::SharedFree((void *)__p,bytes);
+  }
+
+  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
+
+////////////////////////////////////////////////////////////////////////////////
+// Device memory
+////////////////////////////////////////////////////////////////////////////////
+template<typename _Tp>
+class devAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef devAllocator<_Tp1> other; };
+  devAllocator() throw() { }
+  devAllocator(const devAllocator&) throw() { }
+  template<typename _Tp1> devAllocator(const devAllocator<_Tp1>&) throw() { }
+  ~devAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+  pointer allocate(size_type __n, const void* _p= 0)
+  { 
+    size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
+    _Tp *ptr = (_Tp*) MemoryManager::AcceleratorAllocate(bytes);
+    if ( (_Tp*)ptr == (_Tp *) NULL ) {
+      printf("Grid Device Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
+    }
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
+    profilerFree(bytes);
+    MemoryManager::AcceleratorFree((void *)__p,bytes);
+  }
+  void construct(pointer __p, const _Tp& __val) { };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return false; }
+
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using commAllocator = alignedAllocator<T>;
-template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
-template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
+template<class T> using hostVector          = std::vector<T,alignedAllocator<T> >;           // Needs autoview
+template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // Really want to deprecate
+template<class T> using uvmVector           = std::vector<T,uvmAllocator<T> >;               // auto migrating page
+template<class T> using deviceVector        = std::vector<T,devAllocator<T> >;               // device vector
+
+/*
+template<class T> class vecView
+{
+ protected:
+  T * data;
+  uint64_t size;
+  ViewMode mode;
+  void * cpu_ptr;
+ public:
+  // Rvalue accessor
+  accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
+  vecView(Vector<T> &refer_to_me,ViewMode _mode)
+  {
+    cpu_ptr = &refer_to_me[0];
+    size = refer_to_me.size();
+    mode = _mode;
+    data =(T *) MemoryManager::ViewOpen(cpu_ptr,
+					size*sizeof(T),
+					mode,
+					AdviseDefault);
+  }
+  void ViewClose(void)
+  { // Inform the manager
+    MemoryManager::ViewClose(this->cpu_ptr,this->mode);    
+  }
+};
+
+template<class T> vecView<T> VectorView(Vector<T> &vec,ViewMode _mode)
+{
+  vecView<T> ret(vec,_mode); // does the open
+  return ret;                // must be closed
+}
+
+#define autoVecView(v_v,v,mode)					\
+  auto v_v = VectorView(v,mode);				\
+  ViewCloser<decltype(v_v)> _autoView##v_v(v_v);
+*/
 
 NAMESPACE_END(Grid);
 
-#endif
+

Grid/allocator/Allocator.h

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

Grid/allocator/MemoryManager.cc

@@ -0,0 +1,362 @@
+#include <Grid/GridCore.h>
+
+NAMESPACE_BEGIN(Grid);
+
+/*Allocation types, saying which pointer cache should be used*/
+#define Cpu      (0)
+#define CpuHuge  (1)
+#define CpuSmall (2)
+#define Acc      (3)
+#define AccHuge  (4)
+#define AccSmall (5)
+#define Shared   (6)
+#define SharedHuge  (7)
+#define SharedSmall (8)
+#undef GRID_MM_VERBOSE 
+uint64_t total_shared;
+uint64_t total_device;
+uint64_t total_host;;
+
+#if defined(__has_feature)
+#if __has_feature(leak_sanitizer)
+#define ASAN_LEAK_CHECK
+#endif
+#endif
+
+#ifdef ASAN_LEAK_CHECK
+#include <sanitizer/asan_interface.h>
+#include <sanitizer/common_interface_defs.h>
+#include <sanitizer/lsan_interface.h>
+#define LEAK_CHECK(A) { __lsan_do_recoverable_leak_check(); }
+#else
+#define LEAK_CHECK(A) { }
+#endif
+
+void MemoryManager::DisplayMallinfo(void)
+{
+#ifdef __linux__
+  struct mallinfo mi; // really want mallinfo2, but glibc version isn't uniform
+  
+  mi = mallinfo();
+
+  std::cout << "MemoryManager: Total non-mmapped bytes (arena):       "<< (size_t)mi.arena<<std::endl;
+  std::cout << "MemoryManager: # of free chunks (ordblks):            "<< (size_t)mi.ordblks<<std::endl;
+  std::cout << "MemoryManager: # of free fastbin blocks (smblks):     "<< (size_t)mi.smblks<<std::endl;
+  std::cout << "MemoryManager: # of mapped regions (hblks):           "<< (size_t)mi.hblks<<std::endl;
+  std::cout << "MemoryManager: Bytes in mapped regions (hblkhd):      "<< (size_t)mi.hblkhd<<std::endl;
+  std::cout << "MemoryManager: Max. total allocated space (usmblks):  "<< (size_t)mi.usmblks<<std::endl;
+  std::cout << "MemoryManager: Free bytes held in fastbins (fsmblks): "<< (size_t)mi.fsmblks<<std::endl;
+  std::cout << "MemoryManager: Total allocated space (uordblks):      "<< (size_t)mi.uordblks<<std::endl;
+  std::cout << "MemoryManager: Total free space (fordblks):           "<< (size_t)mi.fordblks<<std::endl;
+  std::cout << "MemoryManager: Topmost releasable block (keepcost):   "<< (size_t)mi.keepcost<<std::endl;
+#endif
+  LEAK_CHECK();
+ 
+}
+
+void MemoryManager::PrintBytes(void)
+{
+  std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
+  std::cout << " MemoryManager : PrintBytes "<<std::endl;
+  std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
+  std::cout << " MemoryManager : "<<(total_shared>>20)<<" shared      Mbytes "<<std::endl;
+  std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
+  std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl;
+  uint64_t cacheBytes;
+  cacheBytes = CacheBytes[Cpu];
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
+  cacheBytes = CacheBytes[Acc];
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
+  cacheBytes = CacheBytes[Shared];
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
+  
+#ifdef GRID_CUDA
+  cuda_mem();
+#endif
+  DisplayMallinfo();
+}
+
+uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
+uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
+
+//////////////////////////////////////////////////////////////////////
+// Data tables for recently freed pooiniter caches
+//////////////////////////////////////////////////////////////////////
+MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
+int MemoryManager::Victim[MemoryManager::NallocType];
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
+uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
+//////////////////////////////////////////////////////////////////////
+// Actual allocation and deallocation utils
+//////////////////////////////////////////////////////////////////////
+void *MemoryManager::AcceleratorAllocate(size_t bytes)
+{
+  total_device+=bytes;
+  void *ptr = (void *) Lookup(bytes,Acc);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocDevice(bytes);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"AcceleratorAllocate "<<std::endl;
+  PrintBytes();
+#endif
+  return ptr;
+}
+void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes)
+{
+  total_device-=bytes;
+  void *__freeme = Insert(ptr,bytes,Acc);
+  if ( __freeme ) {
+    acceleratorFreeDevice(__freeme);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"AcceleratorFree "<<std::endl;
+  PrintBytes();
+#endif
+}
+void *MemoryManager::SharedAllocate(size_t bytes)
+{
+  total_shared+=bytes;
+  void *ptr = (void *) Lookup(bytes,Shared);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"SharedAllocate "<<std::endl;
+  PrintBytes();
+#endif
+  return ptr;
+}
+void  MemoryManager::SharedFree    (void *ptr,size_t bytes)
+{
+  total_shared-=bytes;
+  void *__freeme = Insert(ptr,bytes,Shared);
+  if ( __freeme ) {
+    acceleratorFreeShared(__freeme);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"SharedFree "<<std::endl;
+  PrintBytes();
+#endif
+}
+#ifdef GRID_UVM
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  total_host+=bytes;
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuAllocate "<<std::endl;
+  PrintBytes();
+#endif
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  total_host-=bytes;
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeShared(__freeme);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuFree "<<std::endl;
+  PrintBytes();
+#endif
+}
+#else
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  total_host+=bytes;
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocCpu(bytes);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuAllocate "<<std::endl;
+  PrintBytes();
+#endif
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  total_host-=bytes;
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeCpu(__freeme);
+  }
+#ifdef GRID_MM_VERBOSE
+  std::cout <<"CpuFree "<<std::endl;
+  PrintBytes();
+#endif
+}
+#endif
+
+//////////////////////////////////////////
+// call only once
+//////////////////////////////////////////
+void MemoryManager::Init(void)
+{
+
+  char * str;
+  int Nc;
+  
+  str= getenv("GRID_ALLOC_NCACHE_LARGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[Cpu]=Nc;
+      Ncache[Acc]=Nc;
+      Ncache[Shared]=Nc;
+    }
+  }
+
+  str= getenv("GRID_ALLOC_NCACHE_HUGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuHuge]=Nc;
+      Ncache[AccHuge]=Nc;
+      Ncache[SharedHuge]=Nc;
+    }
+  }
+
+  str= getenv("GRID_ALLOC_NCACHE_SMALL");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuSmall]=Nc;
+      Ncache[AccSmall]=Nc;
+      Ncache[SharedSmall]=Nc;
+    }
+  }
+
+}
+
+void MemoryManager::InitMessage(void) {
+
+#ifndef GRID_UVM
+  std::cout << GridLogMessage << "MemoryManager Cache "<< MemoryManager::DeviceMaxBytes <<" bytes "<<std::endl;
+#endif
+  
+  std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
+#ifdef ALLOCATION_CACHE
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
+#endif
+  
+#ifdef GRID_UVM
+  std::cout << GridLogMessage<< "MemoryManager::Init() Unified memory space"<<std::endl;
+#ifdef GRID_CUDA
+  std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMallocManaged"<<std::endl;
+#endif
+#ifdef GRID_HIP
+  std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMallocManaged"<<std::endl;
+#endif
+#ifdef GRID_SYCL
+  std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_shared"<<std::endl;
+#endif
+#else
+  std::cout << GridLogMessage<< "MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory"<<std::endl;
+#ifdef GRID_CUDA
+  std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMalloc"<<std::endl;
+#endif
+#ifdef GRID_HIP
+  std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMalloc"<<std::endl;
+#endif
+#ifdef GRID_SYCL
+  std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_device"<<std::endl;
+#endif
+#endif
+
+}
+
+void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
+{
+#ifdef ALLOCATION_CACHE
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
+  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
+#else
+  return ptr;
+#endif
+}
+
+void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
+{
+#ifdef GRID_OMP
+  GRID_ASSERT(omp_in_parallel()==0);
+#endif 
+
+  if (ncache == 0) return ptr;
+
+  void * ret = NULL;
+  int v = -1;
+
+  for(int e=0;e<ncache;e++) {
+    if ( entries[e].valid==0 ) {
+      v=e; 
+      break;
+    }
+  }
+
+  if ( v==-1 ) {
+    v=victim;
+    victim = (victim+1)%ncache;
+  }
+
+  if ( entries[v].valid ) {
+    ret = entries[v].address;
+    cacheBytes -= entries[v].bytes;
+    entries[v].valid = 0;
+    entries[v].address = NULL;
+    entries[v].bytes = 0;
+  }
+
+  entries[v].address=ptr;
+  entries[v].bytes  =bytes;
+  entries[v].valid  =1;
+  cacheBytes += bytes;
+
+  return ret;
+}
+
+void *MemoryManager::Lookup(size_t bytes,int type)
+{
+#ifdef ALLOCATION_CACHE
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
+  return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
+#else
+  return NULL;
+#endif
+}
+
+void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
+{
+#ifdef GRID_OMP
+  GRID_ASSERT(omp_in_parallel()==0);
+#endif 
+  for(int e=0;e<ncache;e++){
+    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
+      entries[e].valid = 0;
+      cacheBytes -= entries[e].bytes;
+      return entries[e].address;
+    }
+  }
+  return NULL;
+}
+
+
+NAMESPACE_END(Grid);
+

Grid/allocator/MemoryManager.h

@@ -0,0 +1,227 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/MemoryManager.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+#include <list> 
+#include <unordered_map>  
+
+NAMESPACE_BEGIN(Grid);
+
+// Move control to configure.ac and Config.h?
+
+#define GRID_ALLOC_SMALL_LIMIT (4096)
+#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
+
+/*Pinning pages is costly*/
+////////////////////////////////////////////////////////////////////////////
+// Advise the LatticeAccelerator class
+////////////////////////////////////////////////////////////////////////////
+enum ViewAdvise {
+ AdviseDefault       = 0x0,    // Regular data
+ AdviseInfrequentUse = 0x1     // Advise that the data is used infrequently.  This can
+                               // significantly influence performance of bulk storage.
+ 
+ // AdviseTransient      = 0x2,   // Data will mostly be read.  On some architectures
+                               // enables read-only copies of memory to be kept on
+                               // host and device.
+
+ // AdviseAcceleratorWriteDiscard = 0x4  // Field will be written in entirety on device
+
+};
+
+////////////////////////////////////////////////////////////////////////////
+// View Access Mode
+////////////////////////////////////////////////////////////////////////////
+enum ViewMode {
+  AcceleratorRead  = 0x01,
+  AcceleratorWrite = 0x02,
+  AcceleratorWriteDiscard = 0x04,
+  CpuRead  = 0x08,
+  CpuWrite = 0x10,
+  CpuWriteDiscard = 0x10 // same for now
+};
+
+struct MemoryStatus {
+  uint64_t     DeviceBytes;
+  uint64_t     DeviceLRUBytes;
+  uint64_t     DeviceMaxBytes;
+  uint64_t     HostToDeviceBytes;
+  uint64_t     DeviceToHostBytes;
+  uint64_t     HostToDeviceXfer;
+  uint64_t     DeviceToHostXfer;
+  uint64_t     DeviceEvictions;
+  uint64_t     DeviceDestroy;
+  uint64_t     DeviceAllocCacheBytes;
+  uint64_t     HostAllocCacheBytes;
+};
+
+
+class MemoryManager {
+private:
+
+  ////////////////////////////////////////////////////////////
+  // For caching recently freed allocations
+  ////////////////////////////////////////////////////////////
+  typedef struct { 
+    void *address;
+    size_t bytes;
+    int valid;
+  } AllocationCacheEntry;
+
+  static const int NallocCacheMax=128; 
+  static const int NallocType=9;
+  static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
+  static int Victim[NallocType];
+  static int Ncache[NallocType];
+  static uint64_t CacheBytes[NallocType];
+
+  /////////////////////////////////////////////////
+  // Free pool
+  /////////////////////////////////////////////////
+  static void *Insert(void *ptr,size_t bytes,int type) ;
+  static void *Lookup(size_t bytes,int type) ;
+  static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
+  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
+
+ public:
+  static void PrintBytes(void);
+  static void Audit(std::string s);
+  static void Init(void);
+  static void InitMessage(void);
+  static void *AcceleratorAllocate(size_t bytes);
+  static void  AcceleratorFree    (void *ptr,size_t bytes);
+  static void *SharedAllocate(size_t bytes);
+  static void  SharedFree    (void *ptr,size_t bytes);
+  static void *CpuAllocate(size_t bytes);
+  static void  CpuFree    (void *ptr,size_t bytes);
+
+  ////////////////////////////////////////////////////////
+  // Footprint tracking
+  ////////////////////////////////////////////////////////
+  static uint64_t     DeviceBytes;
+  static uint64_t     DeviceLRUBytes;
+  static uint64_t     DeviceMaxBytes;
+  static uint64_t     HostToDeviceBytes;
+  static uint64_t     DeviceToHostBytes;
+  static uint64_t     HostToDeviceXfer;
+  static uint64_t     DeviceToHostXfer;
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
+  
+  static uint64_t     DeviceCacheBytes();
+  static uint64_t     HostCacheBytes();
+
+  static MemoryStatus GetFootprint(void) {
+    MemoryStatus stat;
+    stat.DeviceBytes       = DeviceBytes;
+    stat.DeviceLRUBytes    = DeviceLRUBytes;
+    stat.DeviceMaxBytes    = DeviceMaxBytes;
+    stat.HostToDeviceBytes = HostToDeviceBytes;
+    stat.DeviceToHostBytes = DeviceToHostBytes;
+    stat.HostToDeviceXfer  = HostToDeviceXfer;
+    stat.DeviceToHostXfer  = DeviceToHostXfer;
+    stat.DeviceEvictions   = DeviceEvictions;
+    stat.DeviceDestroy     = DeviceDestroy;
+    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
+    stat.HostAllocCacheBytes   = HostCacheBytes();
+    return stat;
+  };
+  
+ private:
+#ifndef GRID_UVM
+  //////////////////////////////////////////////////////////////////////
+  // Data tables for ViewCache
+  //////////////////////////////////////////////////////////////////////
+  typedef std::list<uint64_t> LRU_t;
+  typedef typename LRU_t::iterator LRUiterator;
+  typedef struct { 
+    int        LRU_valid;
+    LRUiterator LRU_entry;
+    uint64_t CpuPtr;
+    uint64_t AccPtr;
+    size_t   bytes;
+    uint32_t transient;
+    uint32_t state;
+    uint32_t accLock;
+    uint32_t cpuLock;
+  } AcceleratorViewEntry;
+  
+  typedef std::unordered_map<uint64_t,AcceleratorViewEntry> AccViewTable_t;
+  typedef typename AccViewTable_t::iterator AccViewTableIterator ;
+
+  static AccViewTable_t AccViewTable;
+  static LRU_t LRU;
+
+  /////////////////////////////////////////////////
+  // Device motion
+  /////////////////////////////////////////////////
+  static void  Create(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+  static void  EvictVictims(uint64_t bytes); // Frees up <bytes>
+  static void  Evict(AcceleratorViewEntry &AccCache);
+  static void  Flush(AcceleratorViewEntry &AccCache);
+  static void  Clone(AcceleratorViewEntry &AccCache);
+  static void  AccDiscard(AcceleratorViewEntry &AccCache);
+  static void  CpuDiscard(AcceleratorViewEntry &AccCache);
+
+  //  static void  LRUupdate(AcceleratorViewEntry &AccCache);
+  static void  LRUinsert(AcceleratorViewEntry &AccCache);
+  static void  LRUremove(AcceleratorViewEntry &AccCache);
+  
+  // manage entries in the table
+  static int                  EntryPresent(uint64_t CpuPtr);
+  static void                 EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+  static void                 EntryErase (uint64_t CpuPtr);
+  static AccViewTableIterator EntryLookup(uint64_t CpuPtr);
+  static void                 EntrySet   (uint64_t CpuPtr,AcceleratorViewEntry &entry);
+
+  static void     AcceleratorViewClose(uint64_t AccPtr);
+  static uint64_t AcceleratorViewOpen(uint64_t  CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+  static void     CpuViewClose(uint64_t Ptr);
+  static uint64_t CpuViewOpen(uint64_t  CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+#endif
+
+ public:
+  static void DisplayMallinfo(void);
+  static void NotifyDeletion(void * CpuPtr);
+  static void Print(void);
+  static void PrintAll(void);
+  static void PrintState( void* CpuPtr);
+  static int   isOpen   (void* CpuPtr);
+  static void  ViewClose(void* CpuPtr,ViewMode mode);
+  static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+
+};
+
+NAMESPACE_END(Grid);
+
+

Grid/allocator/MemoryManagerCache.cc

@@ -0,0 +1,606 @@
+#include <Grid/GridCore.h>
+#ifndef GRID_UVM
+
+NAMESPACE_BEGIN(Grid);
+
+#define MAXLINE 512
+static char print_buffer [ MAXLINE ];
+
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer << std::endl;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug  << print_buffer << std::endl;
+//#define dprintf(...) 
+//#define mprintf(...) 
+
+////////////////////////////////////////////////////////////
+// For caching copies of data on device
+////////////////////////////////////////////////////////////
+MemoryManager::AccViewTable_t MemoryManager::AccViewTable;
+MemoryManager::LRU_t MemoryManager::LRU;
+  
+////////////////////////////////////////////////////////
+// Footprint tracking
+////////////////////////////////////////////////////////
+uint64_t  MemoryManager::DeviceBytes;
+uint64_t  MemoryManager::DeviceLRUBytes;
+uint64_t  MemoryManager::DeviceMaxBytes = 1024*1024*128;
+uint64_t  MemoryManager::HostToDeviceBytes;
+uint64_t  MemoryManager::DeviceToHostBytes;
+uint64_t  MemoryManager::HostToDeviceXfer;
+uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
+
+////////////////////////////////////
+// Priority ordering for unlocked entries
+//  Empty
+//  CpuDirty 
+//  Consistent
+//  AccDirty
+////////////////////////////////////
+#define Empty         (0x0)  /*Entry unoccupied  */
+#define CpuDirty      (0x1)  /*CPU copy is golden, Acc buffer MAY not be allocated*/
+#define Consistent    (0x2)  /*ACC copy AND CPU copy are valid */
+#define AccDirty      (0x4)  /*ACC copy is golden */
+#define EvictNext     (0x8)  /*Priority for eviction*/
+
+/////////////////////////////////////////////////
+// Mechanics of data table maintenance
+/////////////////////////////////////////////////
+int   MemoryManager::EntryPresent(uint64_t CpuPtr)
+{
+  if(AccViewTable.empty()) return 0;
+
+  auto count = AccViewTable.count(CpuPtr);  GRID_ASSERT((count==0)||(count==1));
+  return count;
+}
+void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
+{
+  GRID_ASSERT(!EntryPresent(CpuPtr));
+  AcceleratorViewEntry AccCache;
+  AccCache.CpuPtr = CpuPtr;
+  AccCache.AccPtr = (uint64_t)NULL;
+  AccCache.bytes  = bytes;
+  AccCache.state  = CpuDirty;
+  AccCache.LRU_valid=0;
+  AccCache.transient=0;
+  AccCache.accLock=0;
+  AccCache.cpuLock=0;
+  AccViewTable[CpuPtr] = AccCache;
+}
+MemoryManager::AccViewTableIterator MemoryManager::EntryLookup(uint64_t CpuPtr)
+{
+  GRID_ASSERT(EntryPresent(CpuPtr));
+  auto AccCacheIterator = AccViewTable.find(CpuPtr);
+  GRID_ASSERT(AccCacheIterator!=AccViewTable.end());
+  return AccCacheIterator;
+}
+void MemoryManager::EntryErase(uint64_t CpuPtr)
+{
+  auto AccCache = EntryLookup(CpuPtr);
+  AccViewTable.erase(CpuPtr);
+}
+void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
+{
+  GRID_ASSERT(AccCache.LRU_valid==0);
+  if (AccCache.transient) { 
+    LRU.push_back(AccCache.CpuPtr);
+    AccCache.LRU_entry = --LRU.end();
+  } else {
+    LRU.push_front(AccCache.CpuPtr);
+    AccCache.LRU_entry = LRU.begin();
+  }
+  AccCache.LRU_valid = 1;
+  DeviceLRUBytes+=AccCache.bytes;
+}
+void  MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
+{
+  GRID_ASSERT(AccCache.LRU_valid==1);
+  LRU.erase(AccCache.LRU_entry);
+  AccCache.LRU_valid = 0;
+  DeviceLRUBytes-=AccCache.bytes;
+}
+/////////////////////////////////////////////////
+// Accelerator cache motion & consistency logic
+/////////////////////////////////////////////////
+void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
+{
+  ///////////////////////////////////////////////////////////
+  // Remove from Accelerator, remove entry, without flush
+  // Cannot be locked. If allocated Must be in LRU pool.
+  ///////////////////////////////////////////////////////////
+  GRID_ASSERT(AccCache.state!=Empty);
+  
+  dprintf("MemoryManager: Discard(%lx) %lx",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr) {
+    AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
+    DeviceBytes   -=AccCache.bytes;
+    LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t) NULL;
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+  }
+  uint64_t CpuPtr = AccCache.CpuPtr;
+  EntryErase(CpuPtr);
+}
+
+void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
+{
+  ///////////////////////////////////////////////////////////////////////////
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
+  ///////////////////////////////////////////////////////////////////////////
+  GRID_ASSERT(AccCache.state!=Empty);
+  
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld",
+	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
+	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  if (AccCache.accLock!=0) return;
+  if (AccCache.cpuLock!=0) return;
+  if(AccCache.state==AccDirty) {
+    Flush(AccCache);
+  }
+  if(AccCache.AccPtr) {
+    AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld ",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+  }
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
+}
+void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
+{
+  GRID_ASSERT(AccCache.state==AccDirty);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.AccPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
+  acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  DeviceToHostBytes+=AccCache.bytes;
+  DeviceToHostXfer++;
+  AccCache.state=Consistent;
+}
+void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
+{
+  GRID_ASSERT(AccCache.state==CpuDirty);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr==(uint64_t)NULL){
+    AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
+    DeviceBytes+=AccCache.bytes;
+  }
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx",
+	  (uint64_t)AccCache.bytes,
+	  (uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
+  HostToDeviceBytes+=AccCache.bytes;
+  HostToDeviceXfer++;
+  AccCache.state=Consistent;
+}
+
+void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
+{
+  GRID_ASSERT(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr==(uint64_t)NULL){
+    AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
+    DeviceBytes+=AccCache.bytes;
+  }
+  AccCache.state=AccDirty;
+}
+
+/////////////////////////////////////////////////////////////////////////////////
+// View management
+/////////////////////////////////////////////////////////////////////////////////
+void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
+{
+  if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewClose %lx",(uint64_t)Ptr);
+    AcceleratorViewClose((uint64_t)Ptr);
+  } else if( (mode==CpuRead)||(mode==CpuWrite)){
+    CpuViewClose((uint64_t)Ptr);
+  } else { 
+    GRID_ASSERT(0);
+  }
+}
+void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
+{
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+  if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewOpen %lx",(uint64_t)CpuPtr);
+    return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
+  } else if( (mode==CpuRead)||(mode==CpuWrite)){
+    return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
+  } else { 
+    GRID_ASSERT(0);
+    return NULL;
+  }
+}
+void  MemoryManager::EvictVictims(uint64_t bytes)
+{
+  if(bytes>=DeviceMaxBytes) {
+    printf("EvictVictims bytes %ld DeviceMaxBytes %ld\n",bytes,DeviceMaxBytes);
+  }
+  GRID_ASSERT(bytes<DeviceMaxBytes);
+  while(bytes+DeviceLRUBytes > DeviceMaxBytes){
+    if ( DeviceLRUBytes > 0){
+      GRID_ASSERT(LRU.size()>0);
+      uint64_t victim = LRU.back(); // From the LRU
+      auto AccCacheIterator = EntryLookup(victim);
+      auto & AccCache = AccCacheIterator->second;
+      Evict(AccCache);
+    } else {
+      return;
+    }
+  }
+}
+uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
+{
+  ////////////////////////////////////////////////////////////////////////////
+  // Find if present, otherwise get or force an empty
+  ////////////////////////////////////////////////////////////////////////////
+  if ( EntryPresent(CpuPtr)==0 ){
+    EntryCreate(CpuPtr,bytes,mode,hint);
+  }
+
+  auto AccCacheIterator = EntryLookup(CpuPtr);
+  auto & AccCache = AccCacheIterator->second;
+  if (!AccCache.AccPtr) {
+    EvictVictims(bytes); 
+  } 
+  GRID_ASSERT((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
+
+  GRID_ASSERT(AccCache.cpuLock==0);  // Programming error
+
+  if(AccCache.state!=Empty) {
+    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld",
+		    (uint64_t)AccCache.CpuPtr,
+		    (uint64_t)CpuPtr,
+		    (uint64_t)AccCache.bytes,
+	            (uint64_t)bytes,
+		    (uint64_t)AccCache.accLock);
+    GRID_ASSERT(AccCache.CpuPtr == CpuPtr);
+    GRID_ASSERT(AccCache.bytes  ==bytes);
+  }
+/*
+ *  State transitions and actions
+ *
+ *  Action  State   StateNext         Flush    Clone
+ *
+ *  AccRead  Empty   Consistent        -        Y
+ *  AccWrite Empty   AccDirty          -        Y
+ *  AccRead  CpuDirty Consistent       -        Y
+ *  AccWrite CpuDirty AccDirty         -        Y
+ *  AccRead  Consistent Consistent     -        - 
+ *  AccWrite Consistent AccDirty       -        - 
+ *  AccRead  AccDirty   AccDirty       -        - 
+ *  AccWrite AccDirty   AccDirty       -        - 
+ */
+  if(AccCache.state==Empty) {
+    GRID_ASSERT(AccCache.LRU_valid==0);
+    AccCache.CpuPtr = CpuPtr;
+    AccCache.AccPtr = (uint64_t)NULL;
+    AccCache.bytes  = bytes;
+    AccCache.state  = CpuDirty;   // Cpu starts primary
+    if(mode==AcceleratorWriteDiscard){
+      CpuDiscard(AccCache);
+      AccCache.state  = AccDirty;   // Empty + AcceleratorWrite=> AccDirty
+    } else if(mode==AcceleratorWrite){
+      Clone(AccCache);
+      AccCache.state  = AccDirty;   // Empty + AcceleratorWrite=> AccDirty
+    } else {
+      Clone(AccCache);
+      AccCache.state  = Consistent; // Empty + AccRead => Consistent
+    }
+    AccCache.accLock= 1;
+    dprintf("Copied Empty entry into device accLock= %d",AccCache.accLock);
+  } else if(AccCache.state==CpuDirty ){
+    if(mode==AcceleratorWriteDiscard) {
+      CpuDiscard(AccCache);
+      AccCache.state  = AccDirty;   // CpuDirty + AcceleratorWrite=> AccDirty
+    } else if(mode==AcceleratorWrite) {
+      Clone(AccCache);
+      AccCache.state  = AccDirty;   // CpuDirty + AcceleratorWrite=> AccDirty
+    } else {
+      Clone(AccCache);
+      AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
+    }
+    AccCache.accLock++;
+    dprintf("CpuDirty entry into device ++accLock= %d",AccCache.accLock);
+  } else if(AccCache.state==Consistent) {
+    if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
+      AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
+    else
+      AccCache.state  = Consistent; // Consistent + AccRead => Consistent
+    AccCache.accLock++;
+    dprintf("Consistent entry into device ++accLock= %d",AccCache.accLock);
+  } else if(AccCache.state==AccDirty) {
+    if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
+      AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
+    else
+      AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
+    AccCache.accLock++;
+    dprintf("AccDirty entry ++accLock= %d",AccCache.accLock);
+  } else {
+    GRID_ASSERT(0);
+  }
+
+  GRID_ASSERT(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
+  if(AccCache.LRU_valid==1){
+    // must possibly remove from LRU as now locked on GPU
+    dprintf("AccCache entry removed from LRU ");
+    LRUremove(AccCache);
+  }
+
+  int transient =hint;
+  AccCache.transient= transient? EvictNext : 0;
+
+  return AccCache.AccPtr;
+}
+////////////////////////////////////
+// look up & decrement lock count
+////////////////////////////////////
+void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
+{
+  auto AccCacheIterator = EntryLookup(CpuPtr);
+  auto & AccCache = AccCacheIterator->second;
+
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock>0);
+
+  AccCache.accLock--;
+  // Move to LRU queue if not locked and close on device
+  if(AccCache.accLock==0) {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    LRUinsert(AccCache);
+  } else {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+  }
+}
+void MemoryManager::CpuViewClose(uint64_t CpuPtr)
+{
+  auto AccCacheIterator = EntryLookup(CpuPtr);
+  auto & AccCache = AccCacheIterator->second;
+
+  GRID_ASSERT(AccCache.cpuLock>0);
+  GRID_ASSERT(AccCache.accLock==0);
+
+  AccCache.cpuLock--;
+}
+/*
+ *  Action  State   StateNext         Flush    Clone
+ *
+ *  CpuRead  Empty   CpuDirty          -        -
+ *  CpuWrite Empty   CpuDirty          -        -
+ *  CpuRead  CpuDirty CpuDirty         -        -
+ *  CpuWrite CpuDirty CpuDirty         -        - 
+ *  CpuRead  Consistent Consistent     -        - 
+ *  CpuWrite Consistent CpuDirty       -        - 
+ *  CpuRead  AccDirty   Consistent     Y        -
+ *  CpuWrite AccDirty   CpuDirty       Y        -
+ */
+uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise transient)
+{
+  ////////////////////////////////////////////////////////////////////////////
+  // Find if present, otherwise get or force an empty
+  ////////////////////////////////////////////////////////////////////////////
+  if ( EntryPresent(CpuPtr)==0 ){
+    EntryCreate(CpuPtr,bytes,mode,transient);
+  }
+
+  auto AccCacheIterator = EntryLookup(CpuPtr);
+  auto & AccCache = AccCacheIterator->second;
+
+  // CPU doesn't need to free space
+  //  if (!AccCache.AccPtr) {
+  //    EvictVictims(bytes);
+  //  }
+
+  GRID_ASSERT((mode==CpuRead)||(mode==CpuWrite));
+  GRID_ASSERT(AccCache.accLock==0);  // Programming error
+
+  if(AccCache.state!=Empty) {
+    GRID_ASSERT(AccCache.CpuPtr == CpuPtr);
+    GRID_ASSERT(AccCache.bytes==bytes);
+  }
+
+  if(AccCache.state==Empty) {
+    AccCache.CpuPtr = CpuPtr;
+    AccCache.AccPtr = (uint64_t)NULL;
+    AccCache.bytes  = bytes;
+    AccCache.state  = CpuDirty; // Empty + CpuRead/CpuWrite => CpuDirty
+    AccCache.accLock= 0;
+    AccCache.cpuLock= 1;
+  } else if(AccCache.state==CpuDirty ){
+    // AccPtr dont care, deferred allocate
+    AccCache.state = CpuDirty; // CpuDirty +CpuRead/CpuWrite => CpuDirty
+    AccCache.cpuLock++;
+  } else if(AccCache.state==Consistent) {
+    GRID_ASSERT(AccCache.AccPtr != (uint64_t)NULL);
+    if(mode==CpuWrite)
+      AccCache.state = CpuDirty;   // Consistent +CpuWrite => CpuDirty
+    else 
+      AccCache.state = Consistent; // Consistent +CpuRead  => Consistent
+    AccCache.cpuLock++;
+  } else if(AccCache.state==AccDirty) {
+    GRID_ASSERT(AccCache.AccPtr != (uint64_t)NULL);
+    Flush(AccCache);
+    if(mode==CpuWrite) AccCache.state = CpuDirty;   // AccDirty +CpuWrite => CpuDirty, Flush
+    else            AccCache.state = Consistent; // AccDirty +CpuRead  => Consistent, Flush
+    AccCache.cpuLock++;
+  } else {
+    GRID_ASSERT(0); // should be unreachable
+  }
+
+  AccCache.transient= transient? EvictNext : 0;
+
+  return AccCache.CpuPtr;
+}
+void  MemoryManager::NotifyDeletion(void *_ptr)
+{
+  // Look up in ViewCache
+  uint64_t ptr = (uint64_t)_ptr;
+  if(EntryPresent(ptr)) {
+    auto e = EntryLookup(ptr);
+    AccDiscard(e->second);
+  }
+}
+void  MemoryManager::Print(void)
+{
+  PrintBytes();
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  acceleratorMem();
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+	      << "\t" << AccCache.cpuLock
+	      << "\t" << AccCache.accLock
+	      << "\t" << AccCache.LRU_valid<<std::endl;
+  }
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+
+};
+int   MemoryManager::isOpen   (void* _CpuPtr) 
+{ 
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    return AccCache.cpuLock+AccCache.accLock;
+  } else { 
+    return 0;
+  }
+}
+void MemoryManager::Audit(std::string s)
+{
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    GRID_ASSERT(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    GRID_ASSERT(AccCache.LRU_valid==1);
+    GRID_ASSERT(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      GRID_ASSERT(AccCache.LRU_valid==0);
+
+      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+		<< "\t cpuLock  " << AccCache.cpuLock
+		<< "\t accLock  " << AccCache.accLock
+		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
+    }
+
+    GRID_ASSERT( AccCache.cpuLock== 0 ) ;
+    GRID_ASSERT( AccCache.accLock== 0 ) ;
+  }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  GRID_ASSERT(LruBytes1==LruBytes2);
+  GRID_ASSERT(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  GRID_ASSERT(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  GRID_ASSERT(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+
+}
+
+void MemoryManager::PrintState(void* _CpuPtr)
+{
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+    if ( AccCache.state==EvictNext) str = std::string("EvictNext");
+
+    std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\t" << AccCache.cpuLock
+    << "\t" << AccCache.accLock
+    << "\t" << AccCache.LRU_valid<<std::endl;
+
+  } else {
+    std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl; 
+  }
+}
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/allocator/MemoryManagerShared.cc

@@ -0,0 +1,31 @@
+#include <Grid/GridCore.h>
+#ifdef GRID_UVM
+
+NAMESPACE_BEGIN(Grid);
+/////////////////////////////////////////////////////////////////////////////////
+// View management is 1:1 address space mapping
+/////////////////////////////////////////////////////////////////////////////////
+uint64_t  MemoryManager::DeviceBytes;
+uint64_t  MemoryManager::DeviceLRUBytes;
+uint64_t  MemoryManager::DeviceMaxBytes = 1024*1024*128;
+uint64_t  MemoryManager::HostToDeviceBytes;
+uint64_t  MemoryManager::DeviceToHostBytes;
+uint64_t  MemoryManager::HostToDeviceXfer;
+uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
+
+void  MemoryManager::Audit(std::string s){};
+void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
+void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
+int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
+void  MemoryManager::PrintState(void* CpuPtr)
+{
+std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
+};
+void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
+void  MemoryManager::NotifyDeletion(void *ptr){};
+
+NAMESPACE_END(Grid);
+#endif

Grid/allocator/MemoryStats.cc

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

Grid/allocator/MemoryStats.h

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

Grid/cartesian/Cartesian_base.h

@@ -70,8 +70,8 @@ public:
   Coordinate _istride;    // Inner stride i.e. within simd lane
   int _osites;                  // _isites*_osites = product(dimensions).
   int _isites;
-  int _fsites;                  // _isites*_osites = product(dimensions).
-  int _gsites;
+  int64_t _fsites;                  // _isites*_osites = product(dimensions).
+  int64_t _gsites;
   Coordinate _slice_block;// subslice information
   Coordinate _slice_stride;
   Coordinate _slice_nblock;
@@ -81,6 +81,8 @@ public:
 
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
+  Coordinate _checker_dim_mask;
+  int              _checker_dim;
 
 public:
 
@@ -88,7 +90,7 @@ public:
   // Checkerboarding interface is virtual and overridden by 
   // GridCartesian / GridRedBlackCartesian
   ////////////////////////////////////////////////////////////////
-  virtual int CheckerBoarded(int dim)=0;
+  virtual int CheckerBoarded(int dim) =0;
   virtual int CheckerBoard(const Coordinate &site)=0;
   virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
   virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
@@ -163,7 +165,7 @@ public:
     //
     if ( _simd_layout[dimension] > 2 ) { 
       for(int d=0;d<_ndimension;d++){
-	if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
+	if ( d != dimension ) GRID_ASSERT ( (_simd_layout[d]==1)  );
       }
       permute_type = RotateBit; // How to specify distance; this is not just direction.
       return permute_type;
@@ -182,10 +184,10 @@ public:
   inline int Nsimd(void)  const { return _isites; };// Synonymous with iSites
   inline int oSites(void) const { return _osites; };
   inline int lSites(void) const { return _isites*_osites; }; 
-  inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
+  inline int64_t gSites(void) const { return (int64_t)_isites*(int64_t)_osites*(int64_t)_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
-  inline const Coordinate LocalStarts(void)             { return _lstart;    };
+  inline const Coordinate &LocalStarts(void)            { return _lstart;    };
   inline const Coordinate &FullDimensions(void)         { return _fdimensions;};
   inline const Coordinate &GlobalDimensions(void)       { return _gdimensions;};
   inline const Coordinate &LocalDimensions(void)        { return _ldimensions;};
@@ -213,15 +215,15 @@ public:
   ////////////////////////////////////////////////////////////////
   // Global addressing
   ////////////////////////////////////////////////////////////////
-  void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
-    assert(gidx< gSites());
+  void GlobalIndexToGlobalCoor(int64_t gidx,Coordinate &gcoor){
+    GRID_ASSERT(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
   void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
-    assert(lidx<lSites());
+    GRID_ASSERT(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
-  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
+  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int64_t & gidx){
     gidx=0;
     int mult=1;
     for(int mu=0;mu<_ndimension;mu++) {

Grid/cartesian/Cartesian_full.h

@@ -38,6 +38,7 @@ class GridCartesian: public GridBase {
 
 public:
   int dummy;
+  //  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -45,7 +46,7 @@ public:
   {
     return 0;
   }
-  virtual int CheckerBoarded(int dim){
+  virtual int CheckerBoarded(int dim) {
     return 0;
   }
   virtual int CheckerBoard(const Coordinate &site){
@@ -104,6 +105,8 @@ public:
     _ldimensions.resize(_ndimension);
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
+    _checker_dim_mask.resize(_ndimension);;
+    _checker_dim = -1;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -114,6 +117,8 @@ public:
 
     for (int d = 0; d < _ndimension; d++)
       {
+	_checker_dim_mask[d]=0;
+
         _fdimensions[d] = dimensions[d];   // Global dimensions
         _gdimensions[d] = _fdimensions[d]; // Global dimensions
         _simd_layout[d] = simd_layout[d];
@@ -123,10 +128,10 @@ public:
         // Use a reduced simd grid
         _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
         //std::cout << _ldimensions[d] << "  " << _gdimensions[d] << "  " << _processors[d] << std::endl;
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        GRID_ASSERT(_ldimensions[d] * _processors[d] == _gdimensions[d]);
 
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        GRID_ASSERT(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
 
         _lstart[d] = _processor_coor[d] * _ldimensions[d];
         _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;

Grid/cartesian/Cartesian_red_black.h

@@ -35,22 +35,39 @@ static const int CbRed  =0;
 static const int CbBlack=1;
 static const int Even   =CbRed;
 static const int Odd    =CbBlack;
+
+accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex,const Coordinate &rdim,const Coordinate &chk_dim_msk)
+{
+  int nd=rdim.size();
+  Coordinate coor(nd);
+
+  Lexicographic::CoorFromIndex(coor,oindex,rdim);
+
+  int linear=0;
+  for(int d=0;d<nd;d++){
+    if(chk_dim_msk[d])
+      linear=linear+coor[d];
+  }
+  return (linear&0x1);
+}
+
     
 // Specialise this for red black grids storing half the data like a chess board.
 class GridRedBlackCartesian : public GridBase
 {
 public:
-  Coordinate _checker_dim_mask;
-  int              _checker_dim;
+  //  Coordinate _checker_dim_mask;
+  //  int              _checker_dim;
   std::vector<int> _checker_board;
 
+  virtual int isCheckerBoarded(void) const { return 1; };
   virtual int CheckerBoarded(int dim){
     if( dim==_checker_dim) return 1;
     else return 0;
   }
   virtual int CheckerBoard(const Coordinate &site){
     int linear=0;
-    assert(site.size()==_ndimension);
+    GRID_ASSERT(site.size()==_ndimension);
     for(int d=0;d<_ndimension;d++){ 
       if(_checker_dim_mask[d])
 	linear=linear+site[d];
@@ -131,7 +148,7 @@ public:
   {
     Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim)  ;
   }
-
+  
   virtual ~GridRedBlackCartesian() = default;
 
   void Init(const Coordinate &dimensions,
@@ -143,11 +160,11 @@ public:
 
       _isCheckerBoarded = true;
     _checker_dim = checker_dim;
-    assert(checker_dim_mask[checker_dim] == 1);
+    GRID_ASSERT(checker_dim_mask[checker_dim] == 1);
     _ndimension = dimensions.size();
-    assert(checker_dim_mask.size() == _ndimension);
-    assert(processor_grid.size() == _ndimension);
-    assert(simd_layout.size() == _ndimension);
+    GRID_ASSERT(checker_dim_mask.size() == _ndimension);
+    GRID_ASSERT(processor_grid.size() == _ndimension);
+    GRID_ASSERT(simd_layout.size() == _ndimension);
 
     _fdimensions.resize(_ndimension);
     _gdimensions.resize(_ndimension);
@@ -173,20 +190,20 @@ public:
 
         if (d == _checker_dim)
 	  {
-	    assert((_gdimensions[d] & 0x1) == 0);
+	    GRID_ASSERT((_gdimensions[d] & 0x1) == 0);
 	    _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
 	    _gsites /= 2;
 	  }
         _ldimensions[d] = _gdimensions[d] / _processors[d];
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        GRID_ASSERT(_ldimensions[d] * _processors[d] == _gdimensions[d]);
         _lstart[d] = _processor_coor[d] * _ldimensions[d];
         _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
 
         // Use a reduced simd grid
         _simd_layout[d] = simd_layout[d];
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
-        assert(_rdimensions[d] > 0);
+        GRID_ASSERT(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        GRID_ASSERT(_rdimensions[d] > 0);
 
         // all elements of a simd vector must have same checkerboard.
         // If Ls vectorised, this must still be the case; e.g. dwf rb5d

Grid/communicator/Communicator_base.cc

@@ -33,6 +33,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+bool Stencil_force_mpi = true;
+
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////
@@ -55,18 +57,29 @@ int                      CartesianCommunicator::ProcessorCount(void)    { return
 // very VERY rarely (Log, serial RNG) we need world without a grid
 ////////////////////////////////////////////////////////////////////////////////
 
+#ifdef USE_GRID_REDUCTION
+void CartesianCommunicator::GlobalSum(ComplexF &c)
+{
+  GlobalSumP2P(c);
+}
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumP2P(c);
+}
+#else
 void CartesianCommunicator::GlobalSum(ComplexF &c)
 {
   GlobalSumVector((float *)&c,2);
 }
-void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
-{
-  GlobalSumVector((float *)c,2*N);
-}
 void CartesianCommunicator::GlobalSum(ComplexD &c)
 {
   GlobalSumVector((double *)&c,2);
 }
+#endif
+void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
+{
+  GlobalSumVector((float *)c,2*N);
+}
 void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);

Grid/communicator/Communicator_base.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -34,8 +33,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
+#define NVLINK_GET
+
 NAMESPACE_BEGIN(Grid);
 
+extern bool Stencil_force_mpi ;
+
 class CartesianCommunicator : public SharedMemory {
 
 public:    
@@ -52,10 +55,11 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
+  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processor_coor;  // linear processor coordinate
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -96,18 +100,22 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                      ProcessorCount(void)    ;
+  int                ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // 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);
   
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
+  void GlobalMax(RealD &);
+  void GlobalMax(RealF &);
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
   void GlobalSum(RealD &);
@@ -121,50 +129,84 @@ public:
   void GlobalSumVector(ComplexD *c,int N);
   void GlobalXOR(uint32_t &);
   void GlobalXOR(uint64_t &);
-  
+
+  template<class obj> void GlobalSumP2P(obj &o)
+  {
+    std::vector<obj> column;
+    obj accum = o;
+    int source,dest;
+    for(int d=0;d<_ndimension;d++){
+      column.resize(_processors[d]);
+      column[0] = accum;
+      std::vector<MpiCommsRequest_t> list;
+      for(int p=1;p<_processors[d];p++){
+	ShiftedRanks(d,p,source,dest);
+	SendToRecvFromBegin(list,
+			    &column[0],
+			    dest,
+			    &column[p],
+			    source,
+			    sizeof(obj),d*100+p);
+
+      }
+      if (!list.empty()) // avoid triggering GRID_ASSERT in comms == none
+	CommsComplete(list);
+      for(int p=1;p<_processors[d];p++){
+	accum = accum + column[p];
+      }
+    }
+    Broadcast(0,accum);
+    o=accum;
+  }
+
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
-    scalar_type * ptr = (scalar_type *)& o;
+    scalar_type * ptr = (scalar_type *)& o; // Safe alias 
     GlobalSumVector(ptr,words);
   }
   
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
+  void CommsComplete(std::vector<MpiCommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
+			   void *xmit,
+			   int dest,
+			   void *recv,
+			   int from,
+			   uint64_t bytes,int dir);
+  
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,
 		      int recv_from_rank,
-		      int bytes);
+		      uint64_t bytes);
   
-  void SendRecvPacket(void *xmit,
-		      void *recv,
-		      int xmit_to_rank,
-		      int recv_from_rank,
-		      int bytes);
-  
-  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-			   void *xmit,
-			   int xmit_to_rank,
-			   void *recv,
-			   int recv_from_rank,
-			   int bytes);
-  
-  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
-
+  int IsOffNode(int rank);
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,
+			       int xmit_to_rank,int do_xmit,
 			       void *recv,
-			       int recv_from_rank,
-			       int bytes,int dir);
+			       int recv_from_rank,int do_recv,
+			       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,
+				      uint64_t xbytes,uint64_t rbytes,int dir);
+
+  // Could do a PollHtoD and have a CommsMerge dependence
+  void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
+  void StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-				    void *xmit,
-				    int xmit_to_rank,
-				    void *recv,
-				    int recv_from_rank,
-				    int bytes,int dir);
+				    void *xmit,void *xmit_comp,
+				    int xmit_to_rank,int do_xmit,
+				    void *recv,void *recv_comp,
+				    int recv_from_rank,int do_recv,
+				    uint64_t xbytes,uint64_t rbytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
@@ -178,20 +220,20 @@ 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
   ////////////////////////////////////////////////////////////
   template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
-    assert(dim>=0);
-    assert(dim<_ndimension);
-    assert(in.size()==out.size());
+    GRID_ASSERT(dim>=0);
+    GRID_ASSERT(dim<_ndimension);
+    GRID_ASSERT(in.size()==out.size());
     int numnode = _processors[dim];
     uint64_t bytes=sizeof(T);
     uint64_t words=in.size()/numnode;
-    assert(numnode * words == in.size());
-    assert(words < (1ULL<<31));
+    GRID_ASSERT(numnode * words == in.size());
+    GRID_ASSERT(words < (1ULL<<31));
     AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
   }
   void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);