Checking in working version of Lanczos.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,54 +0,0 @@
-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

@@ -1,7 +1,3 @@
-# Doxygen stuff
-html/*
-latex/*
-
 # Compiled Object files #
 #########################
 *.slo
@@ -87,36 +83,37 @@ ltmain.sh
 .Trashes
 ehthumbs.db
 Thumbs.db
-.dirstamp
 
 # build directory #
 ###################
 build*/*
-Documentation/_build
 
 # IDE related files #
 #####################
 *.xcodeproj/*
 build.sh
-.vscode
-*.code-workspace
 
 # Eigen source #
 ################
-Grid/Eigen
-Eigen/*
+lib/Eigen/*
+
+# FFTW source #
+################
+lib/fftw/*
 
 # libtool macros #
 ##################
 m4/lt*
 m4/libtool.m4
 
-# github pages #
-################
-gh-pages/
+# Buck files #
+##############
+.buck*
+buck-out
+BUCK
+make-bin-BUCK.sh
 
 # generated sources #
 #####################
-Grid/qcd/spin/gamma-gen/*.h
-Grid/qcd/spin/gamma-gen/*.cc
-Grid/util/Version.h
+lib/qcd/spin/gamma-gen/*.h
+lib/qcd/spin/gamma-gen/*.cc

.travis.yml

@@ -0,0 +1,106 @@
+language: cpp
+
+cache:
+  directories:
+    - clang
+
+matrix:
+  include:
+    - os:        osx
+      osx_image: xcode7.2
+      compiler: clang
+    - compiler: gcc
+      addons:
+        apt:
+          sources:
+            - ubuntu-toolchain-r-test
+          packages:
+            - g++-4.9
+            - libmpfr-dev
+            - libgmp-dev
+            - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
+            - binutils-dev
+      env: VERSION=-4.9
+    - compiler: gcc
+      addons:
+        apt:
+          sources:
+            - ubuntu-toolchain-r-test
+          packages:
+            - g++-5
+            - libmpfr-dev
+            - libgmp-dev
+            - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
+            - binutils-dev
+      env: VERSION=-5
+    - compiler: clang
+      addons:
+        apt:
+          sources:
+            - ubuntu-toolchain-r-test
+          packages:
+            - g++-4.8
+            - libmpfr-dev
+            - libgmp-dev
+            - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
+            - binutils-dev
+      env: CLANG_LINK=http://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
+    - compiler: clang
+      addons:
+        apt:
+          sources:
+            - ubuntu-toolchain-r-test
+          packages:
+            - g++-4.8
+            - libmpfr-dev
+            - libgmp-dev
+            - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
+            - binutils-dev
+      env: CLANG_LINK=http://llvm.org/releases/3.7.0/clang+llvm-3.7.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
+      
+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 [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install openmpi; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]] && [[ "$CC" == "gcc" ]]; then brew install gcc5; fi
+    
+install:
+    - export CC=$CC$VERSION
+    - export CXX=$CXX$VERSION
+    - echo $PATH
+    - which $CC
+    - $CC  --version
+    - which $CXX
+    - $CXX --version
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi
+    
+script:
+    - ./bootstrap.sh
+    - mkdir build
+    - cd build
+    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none
+    - make -j4 
+    - ./benchmarks/Benchmark_dwf --threads 1
+    - echo make clean
+    - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
+    - make -j4
+    - ./benchmarks/Benchmark_dwf --threads 1
+    - echo make clean
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then export CXXFLAGS='-DMPI_UINT32_T=MPI_UNSIGNED -DMPI_UINT64_T=MPI_UNSIGNED_LONG'; fi
+    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto
+    - make -j4
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then mpirun -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+

ChangeLog

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

Grid/DisableWarnings.h

@@ -1,73 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/DisableWarnings.h
-
-Copyright (C) 2016
-
-Author: Guido Cossu <guido.cossu@ed.ac.uk>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#ifndef DISABLE_WARNINGS_H
-#define DISABLE_WARNINGS_H
-
-
-
-#if defined __GNUC__ && __GNUC__>=6
-#pragma GCC diagnostic ignored "-Wignored-attributes"
-#endif
-
- //disables and intel compiler specific warning (in json.hpp)
-#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 extra_semicolon
-#else
- //disables nvcc specific warning in json.hpp
-#pragma diag_suppress unsigned_compare_with_zero
-#pragma diag_suppress cast_to_qualified_type
- //disables nvcc specific warning in many files
-#pragma diag_suppress esa_on_defaulted_function_ignored
-#pragma diag_suppress extra_semicolon
-#endif
-#endif
-
-// Disable vectorisation in Eigen on the Power8/9 and PowerPC
-#ifdef  __ALTIVEC__
-#define  EIGEN_DONT_VECTORIZE
-#endif
-#ifdef  __VSX__
-#define  EIGEN_DONT_VECTORIZE
-#endif
-
-#endif

Grid/GridStd.h

@@ -1,35 +0,0 @@
-#ifndef GRID_STD_H
-#define GRID_STD_H
-
-///////////////////
-// Std C++ dependencies
-///////////////////
-#include <cassert>
-#include <complex>
-#include <memory>
-#include <vector>
-#include <array>
-#include <string>
-#include <iostream>
-#include <iomanip>
-#include <random>
-#include <functional>
-#include <stdio.h>
-#include <stdlib.h>
-#include <strings.h>
-#include <stdio.h>
-#include <signal.h>
-#include <ctime>
-#include <sys/time.h>
-#include <chrono>
-#include <zlib.h>
-
-///////////////////
-// Grid config
-///////////////////
-#include "Config.h"
-
-#ifdef TOFU
-#undef GRID_COMMS_THREADS
-#endif
-#endif /* GRID_STD_H */

Grid/Grid_Eigen_Dense.h

@@ -1,75 +0,0 @@
-#include <Grid/GridCore.h>
-#pragma once
-// Force Eigen to use MKL if Grid has been configured with --enable-mkl
-#ifdef USE_MKL
-#define EIGEN_USE_MKL_ALL
-#endif
-
-
-#if defined __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
-/* NVCC save and restore compile environment*/
-#ifdef __NVCC__
-#pragma push
-#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__
-#define __NVCC__REDEFINE__
-#endif 
-
-/* SYCL save and restore compile environment*/
-#ifdef GRID_SYCL
-#pragma push
-#pragma push_macro("__SYCL_DEVICE_ONLY__")
-#undef __SYCL_DEVICE_ONLY__
-#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>
-
-/* NVCC restore */
-#ifdef __NVCC__REDEFINE__
-#pragma pop_macro("__CUDACC__")
-#pragma pop_macro("__NVCC__")
-#pragma pop_macro("__CUDA_ARCH__")
-#pragma pop
-#endif
-
-/*SYCL restore*/
-#ifdef __SYCL__REDEFINE__
-#pragma pop_macro("__SYCL_DEVICE_ONLY__")
-#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/Grid_Eigen_Tensor.h

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

Grid/Makefile.am

@@ -1,81 +0,0 @@
-extra_sources=
-extra_headers=
-
-if BUILD_COMMS_MPI3
-  extra_sources+=communicator/Communicator_mpi3.cc
-  extra_sources+=communicator/Communicator_base.cc
-  extra_sources+=communicator/SharedMemoryMPI.cc
-  extra_sources+=communicator/SharedMemory.cc
-endif
-
-if BUILD_COMMS_NONE
-  extra_sources+=communicator/Communicator_none.cc
-  extra_sources+=communicator/Communicator_base.cc
-  extra_sources+=communicator/SharedMemoryNone.cc
-  extra_sources+=communicator/SharedMemory.cc
-endif
-
-if BUILD_HDF5
-  extra_sources+=serialisation/Hdf5IO.cc 
-  extra_headers+=serialisation/Hdf5IO.h
-  extra_headers+=serialisation/Hdf5Type.h
-endif
-
-
-all: version-cache Version.h
-
-version-cache:
-	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
-		a="uncommited changes";\
-	else\
-		a="clean";\
-	fi;\
-	echo "`git log -n 1 --format=format:"#define GITHASH \\"%H:%d $$a\\"%n" HEAD`" > vertmp;\
-	if [ -e version-cache ]; then\
-		d=`diff vertmp version-cache`;\
-		if [ "$${d}" != "" ]; then\
-			mv vertmp version-cache;\
-			rm -f Version.h;\
-		fi;\
-	else\
-		mv vertmp version-cache;\
-		rm -f Version.h;\
-	fi;\
-	rm -f vertmp
-
-Version.h: version-cache
-	cp version-cache Version.h
-
-.PHONY: version-cache
-
-#
-# Libraries
-#
-include Make.inc
-include Eigen.inc
-
-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+=$(SP_FERMION_FILES)
-    extra_sources+=$(SP_TWOIND_FERMION_FILES)
-endif
-
-lib_LIBRARIES = libGrid.a
-
-CCFILES += $(extra_sources)
-HFILES  += $(extra_headers) Config.h Version.h
-
-libGrid_a_SOURCES              = $(CCFILES)
-libGrid_adir                   = $(includedir)/Grid
-nobase_dist_pkginclude_HEADERS = $(HFILES) $(eigen_files) $(eigen_unsupp_files)

Grid/Namespace.h

@@ -1,38 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/Namespace.h
-
-Copyright (C) 2016
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-#include <type_traits>
-#include <cassert>
-
-#define NAMESPACE_BEGIN(A) namespace A {
-#define NAMESPACE_END(A)   }
-#define GRID_NAMESPACE_BEGIN NAMESPACE_BEGIN(Grid)
-#define GRID_NAMESPACE_END   NAMESPACE_END(Grid)
-#define NAMESPACE_CHECK(x) struct namespaceTEST##x {};  static_assert(std::is_same<namespaceTEST##x, ::namespaceTEST##x>::value,"Not in :: at"  ); 

Grid/algorithms/FFT.h

@@ -1,296 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Cshift.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef _GRID_FFT_H_
-#define _GRID_FFT_H_
-
-#ifdef HAVE_FFTW
-#if defined(USE_MKL) || defined(GRID_SYCL)
-#include <fftw/fftw3.h>
-#else
-#include <fftw3.h>
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-template<class scalar> struct FFTW { };
-
-#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,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, const 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) {
-    ::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,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, const 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) {
-    ::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;
-  static const int backward=FFTW_BACKWARD;
-    
-  double Flops(void) {return flops;}
-  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)
-  {
-    flops=0;
-    usec =0;
-    Coordinate layout(Nd,1);
-    sgrid = new GridCartesian(dimensions,layout,processors,*grid);
-  };
-    
-  ~FFT ( void)  {
-    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++){
-      if( mask[d] ) {
-	FFT_dim(result,tmp,d,sign);
-	tmp=result;
-      }
-    }
-  }
-
-  template<class vobj>
-  void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
-    Coordinate mask(Nd,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);
-
-    int L = vgrid->_ldimensions[dim];
-    int G = vgrid->_fdimensions[dim];
-      
-    Coordinate layout(Nd,1);
-    Coordinate pencil_gd(vgrid->_fdimensions);
-      
-    pencil_gd[dim] = G*processors[dim];
-      
-    // Pencil global vol LxLxGxLxL per node
-    GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid);
-      
-    // Construct pencils
-    typedef typename vobj::scalar_object sobj;
-    typedef typename sobj::scalar_type   scalar;
-      
-    Lattice<sobj> pgbuf(&pencil_g);
-    autoView(pgbuf_v , pgbuf, CpuWrite);
-
-    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;
-    for(int d=0;d<dim;d++){
-      Nlow*=vgrid->_ldimensions[d];
-    }
-      
-    int rank = 1;  /* 1d transforms */
-    int n[] = {G}; /* 1d transforms of length G */
-    int howmany = Ncomp;
-    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 */
-    int *inembed = n, *onembed = n;
-      
-    scalar div;
-    if ( sign == backward ) div = 1.0/G;
-    else if ( sign == forward ) div = 1.0;
-    else assert(0);
-      
-    FFTW_plan p;
-    {
-      FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
-      FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[0];
-      p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
-					   in,inembed,
-					   istride,idist,
-					   out,onembed,
-					   ostride, odist,
-					   sign,FFTW_ESTIMATE);
-    }
-      
-    // Barrel shift and collect global pencil
-    Coordinate lcoor(Nd), gcoor(Nd);
-    result = source;
-    int pc = processor_coor[dim];
-    for(int p=0;p<processors[dim];p++) {
-      {
-	autoView(r_v,result,CpuRead);
-	autoView(p_v,pgbuf,CpuWrite);
-	thread_for(idx, sgrid->lSites(),{
-          Coordinate cbuf(Nd);
-          sobj s;
-	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,r_v,cbuf);
-	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  pokeLocalSite(s,p_v,cbuf);
-        });
-      }
-      if (p != processors[dim] - 1) {
-	result = Cshift(result,dim,L);
-      }
-    }
-      
-    // 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();
-      
-    // 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
-    {
-      autoView(pgbuf_v,pgbuf,CpuRead);
-      autoView(result_v,result,CpuWrite);
-      thread_for(idx,sgrid->lSites(),{
-	Coordinate clbuf(Nd), cgbuf(Nd);
-	sobj s;
-	sgrid->LocalIndexToLocalCoor(idx,clbuf);
-	cgbuf = clbuf;
-	cgbuf[dim] = clbuf[dim]+L*pc;
-	peekLocalSite(s,pgbuf_v,cgbuf);
-	pokeLocalSite(s,result_v,clbuf);
-      });
-    }
-    result = result*div;
-      
-    // destroying plan
-    FFTW<scalar>::fftw_destroy_plan(p);
-#endif
-  }
-};
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/LinearOperator.h

@@ -1,641 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/LinearOperator.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);
-
-/////////////////////////////////////////////////////////////////////////////////////////////
-// LinearOperators Take a something and return a something.
-/////////////////////////////////////////////////////////////////////////////////////////////
-//
-// Hopefully linearity is satisfied and the AdjOp is indeed the Hermitian Conjugateugate (transpose if real):
-//SBase
-//   i)  F(a x + b y) = aF(x) + b F(y).
-//  ii)  <x|Op|y> = <y|AdjOp|x>^\ast
-//
-// Would be fun to have a test linearity & Herm Conj function!
-/////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class LinearOperatorBase {
-public:
-  // Support for coarsening to a multigrid
-  virtual void OpDiag (const Field &in, Field &out) = 0; // Abstract base
-  virtual void OpDir  (const Field &in, Field &out,int dir,int disp) = 0; // Abstract base
-  virtual void OpDirAll  (const Field &in, std::vector<Field> &out) = 0; // Abstract base
-
-  virtual void Op     (const Field &in, Field &out) = 0; // Abstract base
-  virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
-  virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
-  virtual void HermOp(const Field &in, Field &out)=0;
-  virtual ~LinearOperatorBase(){};
-};
-
-
-/////////////////////////////////////////////////////////////////////////////////////////////
-// By sharing the class for Sparse Matrix across multiple operator wrappers, we can share code
-// between RB and non-RB variants. Sparse matrix is like the fermion action def, and then
-// the wrappers implement the specialisation of "Op" and "AdjOp" to the cases minimising
-// replication of code.
-//
-// 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
-// do it, but I fear it required multiple inheritance and mixed in abstract base classes
-/////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////
-// Construct herm op from non-herm matrix
-////////////////////////////////////////////////////////////////////
-template<class Matrix,class Field>
-class MdagMLinearOperator : public LinearOperatorBase<Field> {
-  Matrix &_Mat;
-public:
-  MdagMLinearOperator(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.MdagM(in,out);
-    ComplexD dot = innerProduct(in,out);
-    n1=real(dot);
-    n2=norm2(out);
-  }
-  void HermOp(const Field &in, Field &out){
-    _Mat.MdagM(in,out);
-  }
-};
-
-////////////////////////////////////////////////////////////////////
-// Construct herm op and shift it for mgrid smoother
-////////////////////////////////////////////////////////////////////
-template<class Matrix,class Field>
-class ShiftedMdagMLinearOperator : public LinearOperatorBase<Field> {
-  Matrix &_Mat;
-  RealD _shift;
-public:
-  ShiftedMdagMLinearOperator(Matrix &Mat,RealD shift): _Mat(Mat), _shift(shift){};
-  // Support for coarsening to a multigrid
-  void OpDiag (const Field &in, Field &out) {
-    _Mat.Mdiag(in,out);
-    assert(0);
-  }
-  void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    _Mat.Mdir(in,out,dir,disp);
-    assert(0);
-  }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
-  };
-  void Op     (const Field &in, Field &out){
-    _Mat.M(in,out);
-    assert(0);
-  }
-  void AdjOp     (const Field &in, Field &out){
-    _Mat.Mdag(in,out);
-    assert(0);
-  }
-  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.MdagM(in,out);
-    out = out + _shift*in;
-  }
-};
-
-////////////////////////////////////////////////////////////////////
-// Wrap an already herm matrix
-////////////////////////////////////////////////////////////////////
-template<class Matrix,class Field>
-class HermitianLinearOperator : public LinearOperatorBase<Field> {
-  Matrix &_Mat;
-public:
-  HermitianLinearOperator(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.M(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.M(in,out);
-  }
-};
-
-template<class Matrix,class Field>
-class NonHermitianLinearOperator : public LinearOperatorBase<Field> {
-  Matrix &_Mat;
-public:
-  NonHermitianLinearOperator(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){
-    assert(0);
-  }
-  void HermOp(const Field &in, Field &out){
-    assert(0);
-  }
-};
-
-//////////////////////////////////////////////////////////
-// Even Odd Schur decomp operators; there are several
-// ways to introduce the even odd checkerboarding
-//////////////////////////////////////////////////////////
-
-template<class Field>
-class SchurOperatorBase :  public LinearOperatorBase<Field> {
- public:
-  virtual  void Mpc      (const Field &in, Field &out) =0;
-  virtual  void MpcDag   (const Field &in, Field &out) =0;
-  virtual  void MpcDagMpc(const Field &in, Field &out) {
-    Field tmp(in.Grid());
-    tmp.Checkerboard() = in.Checkerboard();
-    Mpc(in,tmp);
-    MpcDag(tmp,out);
-  }
-  virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    out.Checkerboard() = in.Checkerboard();
-    MpcDagMpc(in,out);
-    ComplexD dot= innerProduct(in,out); 
-    n1=real(dot);
-    n2=norm2(out);
-  }
-  virtual void HermOp(const Field &in, Field &out){
-    out.Checkerboard() = in.Checkerboard();
-    MpcDagMpc(in,out);
-  }
-  void Op     (const Field &in, Field &out){
-    Mpc(in,out);
-  }
-  void AdjOp     (const Field &in, Field &out){ 
-    MpcDag(in,out);
-  }
-  // Support for coarsening to a multigrid
-  void OpDiag (const Field &in, Field &out) {
-    assert(0); // must coarsen the unpreconditioned system
-  }
-  void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    assert(0);
-  }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
-  };
-};
-template<class Matrix,class Field>
-  class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
- public:
-    Matrix &_Mat;
-    SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
-    virtual  void Mpc      (const Field &in, Field &out) {
-      Field tmp(in.Grid());
-      tmp.Checkerboard() = !in.Checkerboard();
-      
-      _Mat.Meooe(in,tmp);
-      _Mat.MooeeInv(tmp,out);
-      _Mat.Meooe(out,tmp);
-      _Mat.Mooee(in,out);
-      axpy(out,-1.0,tmp,out);
-    }
-    virtual void MpcDag   (const Field &in, Field &out){
-      Field tmp(in.Grid());
-	
-      _Mat.MeooeDag(in,tmp);
-      _Mat.MooeeInvDag(tmp,out);
-      _Mat.MeooeDag(out,tmp);
-      _Mat.MooeeDag(in,out);
-      axpy(out,-1.0,tmp,out);
-    }
-};
-template<class Matrix,class Field>
-  class SchurDiagOneOperator :  public SchurOperatorBase<Field> {
- protected:
-    Matrix &_Mat;
- public:
-    SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
-    
-    virtual void Mpc      (const Field &in, Field &out) {
-      Field tmp(in.Grid());
-
-      _Mat.Meooe(in,out);
-      _Mat.MooeeInv(out,tmp);
-      _Mat.Meooe(tmp,out);
-      _Mat.MooeeInv(out,tmp);
-      axpy(out,-1.0,tmp,in);
-    }
-    virtual void MpcDag   (const Field &in, Field &out){
-      Field tmp(in.Grid());
-      
-      _Mat.MooeeInvDag(in,out);
-      _Mat.MeooeDag(out,tmp);
-      _Mat.MooeeInvDag(tmp,out);
-      _Mat.MeooeDag(out,tmp);
-      axpy(out,-1.0,tmp,in);
-    }
-};
-template<class Matrix,class Field>
-  class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
- protected:
-    Matrix &_Mat;
- public:
-    SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
-    
-    virtual void Mpc      (const Field &in, Field &out) {
-      Field tmp(in.Grid());
-      
-      _Mat.MooeeInv(in,out);
-      _Mat.Meooe(out,tmp);
-      _Mat.MooeeInv(tmp,out);
-      _Mat.Meooe(out,tmp);
-      
-      axpy(out,-1.0,tmp,in);
-    }
-    virtual  void MpcDag   (const Field &in, Field &out){
-      Field tmp(in.Grid());
-
-      _Mat.MeooeDag(in,out);
-      _Mat.MooeeInvDag(out,tmp);
-      _Mat.MeooeDag(tmp,out);
-      _Mat.MooeeInvDag(out,tmp);
-
-      axpy(out,-1.0,tmp,in);
-    }
-};
-
-template<class Field>
-class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field> 
-{
- public:
-  virtual void  Mpc      (const Field& in, Field& out) = 0;
-  virtual void  MpcDag   (const Field& in, Field& out) = 0;
-  virtual void  MpcDagMpc(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-    tmp.Checkerboard() = in.Checkerboard();
-    Mpc(in,tmp);
-    MpcDag(tmp,out);
-  }
-  virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
-    assert(0);
-  }
-  virtual void HermOp(const Field& in, Field& out) {
-    assert(0);
-  }
-  void Op(const Field& in, Field& out) {
-    Mpc(in, out);
-  }
-  void AdjOp(const Field& in, Field& out) { 
-    MpcDag(in, out);
-  }
-  // Support for coarsening to a multigrid
-  void OpDiag(const Field& in, Field& out) {
-    assert(0); // must coarsen the unpreconditioned system
-  }
-  void OpDir(const Field& in, Field& out, int dir, int disp) {
-    assert(0);
-  }
-  void OpDirAll(const Field& in, std::vector<Field>& out){
-    assert(0);
-  };
-};
-
-template<class Matrix, class Field>
-class NonHermitianSchurDiagMooeeOperator :  public NonHermitianSchurOperatorBase<Field> 
-{
- public:
-  Matrix& _Mat;
- NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
-  virtual void Mpc(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-    tmp.Checkerboard() = !in.Checkerboard();
-    
-    _Mat.Meooe(in, tmp);
-    _Mat.MooeeInv(tmp, out);
-    _Mat.Meooe(out, tmp);
-    
-    _Mat.Mooee(in, out);
-    
-    axpy(out, -1.0, tmp, out);
-  }
-  virtual void MpcDag(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-    
-    _Mat.MeooeDag(in, tmp);
-    _Mat.MooeeInvDag(tmp, out);
-    _Mat.MeooeDag(out, tmp);
-	  
-    _Mat.MooeeDag(in, out);
-    
-    axpy(out, -1.0, tmp, out);
-  }
-};
-    
-template<class Matrix,class Field>
-class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Field> 
-{
- protected:
-  Matrix &_Mat;
-  
- public:
-  NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
-  virtual void Mpc(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-	  
-    _Mat.Meooe(in, out);
-    _Mat.MooeeInv(out, tmp);
-    _Mat.Meooe(tmp, out);
-    _Mat.MooeeInv(out, tmp);
-
-    axpy(out, -1.0, tmp, in);
-  }
-  virtual void MpcDag(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-    
-    _Mat.MooeeInvDag(in, out);
-    _Mat.MeooeDag(out, tmp);
-    _Mat.MooeeInvDag(tmp, out);
-    _Mat.MeooeDag(out, tmp);
-    
-    axpy(out, -1.0, tmp, in);
-  }
-};
-
-template<class Matrix, class Field>
-class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Field> 
-{
- protected:
-  Matrix& _Mat;
-  
- public:
- NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
-
-  virtual void Mpc(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-    
-    _Mat.MooeeInv(in, out);
-    _Mat.Meooe(out, tmp);
-    _Mat.MooeeInv(tmp, out);
-    _Mat.Meooe(out, tmp);
-
-    axpy(out, -1.0, tmp, in);
-  }
-  virtual void MpcDag(const Field& in, Field& out) {
-    Field tmp(in.Grid());
-    
-    _Mat.MeooeDag(in, out);
-    _Mat.MooeeInvDag(out, tmp);
-    _Mat.MeooeDag(tmp, out);
-    _Mat.MooeeInvDag(out, tmp);
-
-    axpy(out, -1.0, tmp, in);
-  }
-};
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
-// Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi
-///////////////////////////////////////////////////////////////////////////////////////////////////
-template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
-template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
-///////////////////////////////////////////////////////////////////////////////////////////////////
-//  Staggered use
-///////////////////////////////////////////////////////////////////////////////////////////////////
-template<class Matrix,class Field>
-class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
- protected:
-  Matrix &_Mat;
-  Field tmp;
-  RealD mass;
- public:
-  SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
-  { 
-    assert( _Mat.isTrivialEE() );
-    mass = _Mat.Mass();
-  }
-  virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    Mpc(in,out);
-    ComplexD dot= innerProduct(in,out);
-    n1 = real(dot);
-    n2 =0.0;
-  }
-  virtual void HermOp(const Field &in, Field &out){
-    Mpc(in,out);
-    //    _Mat.Meooe(in,out);
-    //    _Mat.Meooe(out,tmp);
-    //    axpby(out,-1.0,mass*mass,tmp,in);
-  }
-  virtual  void Mpc      (const Field &in, Field &out) 
-  {
-    Field tmp(in.Grid());
-    Field tmp2(in.Grid());
-	
-    //    _Mat.Mooee(in,out);
-    //    _Mat.Mooee(out,tmp);
-
-    _Mat.Meooe(in,out);
-    _Mat.Meooe(out,tmp);
-    axpby(out,-1.0,mass*mass,tmp,in);
-  }
-  virtual  void MpcDag   (const Field &in, Field &out){
-    Mpc(in,out);
-  }
-  virtual void MpcDagMpc(const Field &in, Field &out) {
-    assert(0);// Never need with staggered
-  }
-};
-template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
-
-/////////////////////////////////////////////////////////////
-// Base classes for functions of operators
-/////////////////////////////////////////////////////////////
-template<class Field> class OperatorFunction {
-public:
-  virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
-  virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
-    assert(in.size()==out.size());
-    for(int k=0;k<in.size();k++){
-      (*this)(Linop,in[k],out[k]);
-    }
-  };
-  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)
-  {
-    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> {
-public:
-  void operator() (const Field &in, Field &out){
-    out = in;
-  };
-};
-
-
-/////////////////////////////////////////////////////////////
-// Base classes for Multishift solvers for operators
-/////////////////////////////////////////////////////////////
-template<class Field> class OperatorMultiFunction {
-public:
-  virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, std::vector<Field> &out) = 0;
-};
-
-// FIXME : To think about
-
-// Chroma functionality list defining LinearOperator
-/*
-  virtual void operator() (T& chi, const T& psi, enum PlusMinus isign) const = 0;
-  virtual void operator() (T& chi, const T& psi, enum PlusMinus isign, Real epsilon) const
-  virtual const Subset& subset() const = 0;
-  virtual unsigned long nFlops() const { return 0; }
-  virtual void deriv(P& ds_u, const T& chi, const T& psi, enum PlusMinus isign) const
-  class UnprecLinearOperator : public DiffLinearOperator<T,P,Q>
-  const Subset& subset() const {return all;}
-  };
-*/
-
-////////////////////////////////////////////////////////////////////////////////////////////
-// Hermitian operator Linear function and operator function
-////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field>
-class HermOpOperatorFunction : public OperatorFunction<Field> {
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-    Linop.HermOp(in,out);
-  };
-};
-
-template<typename Field>
-class PlainHermOp : public LinearFunction<Field> {
-public:
-  using LinearFunction<Field>::operator();
-  LinearOperatorBase<Field> &_Linop;
-      
-  PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
-  {}
-      
-  void operator()(const Field& in, Field& out) {
-    _Linop.HermOp(in,out);
-  }
-};
-
-template<typename Field>
-class FunctionHermOp : public LinearFunction<Field> {
-public:
-  using LinearFunction<Field>::operator(); 
-  OperatorFunction<Field>   & _poly;
-  LinearOperatorBase<Field> &_Linop;
-      
-  FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop) 
-    : _poly(poly), _Linop(linop) {};
-      
-  void operator()(const Field& in, Field& out) {
-    _poly(_Linop,in,out);
-  }
-};
-
-template<class Field>
-class Polynomial : public OperatorFunction<Field> {
-private:
-  std::vector<RealD> Coeffs;
-public:
-  using OperatorFunction<Field>::operator();
-
-  Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
-
-  // Implement the required interface
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-
-    Field AtoN(in.Grid());
-    Field Mtmp(in.Grid());
-    AtoN = in;
-    out = AtoN*Coeffs[0];
-    for(int n=1;n<Coeffs.size();n++){
-      Mtmp = AtoN;
-      Linop.HermOp(Mtmp,AtoN);
-      out=out+AtoN*Coeffs[n];
-    }
-  };
-};
-
-NAMESPACE_END(Grid);

Grid/algorithms/SparseMatrix.h

@@ -1,81 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/SparseMatrix.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef  GRID_ALGORITHM_SPARSE_MATRIX_H
-#define  GRID_ALGORITHM_SPARSE_MATRIX_H
-
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////////////////////////////////////
-// Interface defining what I expect of a general sparse matrix, such as a Fermion action
-/////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class SparseMatrixBase {
-public:
-  virtual GridBase *Grid(void) =0;
-  // Full checkerboar operations
-  virtual void  M    (const Field &in, Field &out)=0;
-  virtual void  Mdag (const Field &in, Field &out)=0;
-  virtual void  MdagM(const Field &in, Field &out) {
-    Field tmp (in.Grid());
-    M(in,tmp);
-    Mdag(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() {};
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////////
-// Interface augmented by a red black sparse matrix, such as a Fermion action
-/////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class CheckerBoardedSparseMatrixBase : public SparseMatrixBase<Field> {
-public:
-  virtual GridBase *RedBlackGrid(void)=0;
-
-  //////////////////////////////////////////////////////////////////////
-  // Query the even even properties to make algorithmic decisions
-  //////////////////////////////////////////////////////////////////////
-  virtual RealD  Mass(void)        { return 0.0; };
-  virtual int    ConstEE(void)     { return 1; }; // Disable assumptions unless overridden
-  virtual int    isTrivialEE(void) { return 0; }; // by a derived class that knows better
-
-  // half checkerboard operaions
-  virtual  void Meooe    (const Field &in, Field &out)=0;
-  virtual  void Mooee    (const Field &in, Field &out)=0;
-  virtual  void MooeeInv (const Field &in, Field &out)=0;
-
-  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);
-
-#endif

Grid/algorithms/approx/Chebyshev.h

@@ -1,395 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/approx/Chebyshev.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Christoph Lehner <clehner@bnl.gov>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    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_CHEBYSHEV_H
-#define GRID_CHEBYSHEV_H
-
-#include <Grid/algorithms/LinearOperator.h>
-
-NAMESPACE_BEGIN(Grid);
-
-struct ChebyParams : Serializable {
-  GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
-				  RealD, alpha,  
-				  RealD, beta,   
-				  int, Npoly);
-};
-
-////////////////////////////////////////////////////////////////////////////////////////////
-// Generic Chebyshev approximations
-////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field>
-class Chebyshev : public OperatorFunction<Field> {
-private:
-  using OperatorFunction<Field>::operator();
-
-  std::vector<RealD> Coeffs;
-  int order;
-  RealD hi;
-  RealD lo;
-
-public:
-  void csv(std::ostream &out){
-    RealD diff = hi-lo;
-    RealD delta = diff*1.0e-9;
-    for (RealD x=lo; x<hi; x+=delta) {
-      delta*=1.1;
-      RealD f = approx(x);
-      out<< x<<" "<<f<<std::endl;
-    }
-    return;
-  }
-
-  // Convenience for plotting the approximation
-  void   PlotApprox(std::ostream &out) {
-    out<<"Polynomial approx ["<<lo<<","<<hi<<"]"<<std::endl;
-    for(RealD x=lo;x<hi;x+=(hi-lo)/50.0){
-      out <<x<<"\t"<<approx(x)<<std::endl;
-    }
-  };
-
-  Chebyshev(){};
-  Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
-  Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
-  Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};
-
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  // c.f. numerical recipes "chebft"/"chebev". This is sec 5.8 "Chebyshev approximation".
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  // CJ: the one we need for Lanczos
-  void Init(RealD _lo,RealD _hi,int _order)
-  {
-    lo=_lo;
-    hi=_hi;
-    order=_order;
-      
-    if(order < 2) exit(-1);
-    Coeffs.resize(order);
-    Coeffs.assign(0.,order);
-    Coeffs[order-1] = 1.;
-  };
-  
-  // PB - more efficient low pass drops high modes above the low as 1/x uses all Chebyshev's.
-  // Similar kick effect below the threshold as Lanczos filter approach
-  void InitLowPass(RealD _lo,RealD _hi,int _order)
-  {
-    lo=_lo;
-    hi=_hi;
-    order=_order;
-      
-    if(order < 2) exit(-1);
-    Coeffs.resize(order);
-    for(int j=0;j<order;j++){
-      RealD k=(order-1.0);
-      RealD s=std::cos( j*M_PI*(k+0.5)/order );
-      Coeffs[j] = s * 2.0/order;
-    }
-    
-  };
-
-  void Init(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD))
-  {
-    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){
-    RealD M=order;
-    RealD alpha = M_PI/(M+2);
-    RealD lmax = std::cos(alpha);
-    RealD sumUsq =0;
-    std::vector<RealD> U(M);
-    std::vector<RealD> a(M);
-    std::vector<RealD> g(M);
-    for(int n=0;n<=M;n++){
-      U[n] = std::sin((n+1)*std::acos(lmax))/std::sin(std::acos(lmax));
-      sumUsq += U[n]*U[n];
-    }      
-    sumUsq = std::sqrt(sumUsq);
-
-    for(int i=1;i<=M;i++){
-      a[i] = U[i]/sumUsq;
-    }
-    g[0] = 1.0;
-    for(int m=1;m<=M;m++){
-      g[m] = 0;
-      for(int i=0;i<=M-m;i++){
-	g[m]+= a[i]*a[m+i];
-      }
-    }
-    for(int m=1;m<=M;m++){
-      Coeffs[m]*=g[m];
-    }
-  }
-  RealD approx(RealD x) // Convenience for plotting the approximation
-  {
-    RealD Tn;
-    RealD Tnm;
-    RealD Tnp;
-      
-    RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
-      
-    RealD T0=1;
-    RealD T1=y;
-      
-    RealD sum;
-    sum = 0.5*Coeffs[0]*T0;
-    sum+= Coeffs[1]*T1;
-      
-    Tn =T1;
-    Tnm=T0;
-    for(int i=2;i<order;i++){
-      Tnp=2*y*Tn-Tnm;
-      Tnm=Tn;
-      Tn =Tnp;
-      sum+= Tn*Coeffs[i];
-    }
-    return sum;
-  };
-
-  RealD approxD(RealD x)
-  {
-    RealD Un;
-    RealD Unm;
-    RealD Unp;
-      
-    RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
-      
-    RealD U0=1;
-    RealD U1=2*y;
-      
-    RealD sum;
-    sum = Coeffs[1]*U0;
-    sum+= Coeffs[2]*U1*2.0;
-      
-    Un =U1;
-    Unm=U0;
-    for(int i=2;i<order-1;i++){
-      Unp=2*y*Un-Unm;
-      Unm=Un;
-      Un =Unp;
-      sum+= Un*Coeffs[i+1]*(i+1.0);
-    }
-    return sum/(0.5*(hi-lo));
-  };
-    
-  RealD approxInv(RealD z, RealD x0, int maxiter, RealD resid) {
-    RealD x = x0;
-    RealD eps;
-      
-    int i;
-    for (i=0;i<maxiter;i++) {
-      eps = approx(x) - z;
-      if (fabs(eps / z) < resid)
-	return x;
-      x = x - eps / approxD(x);
-    }
-      
-    return std::numeric_limits<double>::quiet_NaN();
-  }
-    
-  // Implement the required interface
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-
-    GridBase *grid=in.Grid();
-
-    int vol=grid->gSites();
-    typedef typename Field::vector_type vector_type;
-
-    Field T0(grid); T0 = in;  
-    Field T1(grid); 
-    Field T2(grid);
-    Field y(grid);
-      
-    Field *Tnm = &T0;
-    Field *Tn  = &T1;
-    Field *Tnp = &T2;
-
-    // Tn=T1 = (xscale M + mscale)in
-    RealD xscale = 2.0/(hi-lo);
-    RealD mscale = -(hi+lo)/(hi-lo);
-    Linop.HermOp(T0,y);
-    axpby(T1,xscale,mscale,y,in);
-
-    // sum = .5 c[0] T0 + c[1] T1
-    //    out = ()*T0 + Coeffs[1]*T1;
-    axpby(out,0.5*Coeffs[0],Coeffs[1],T0,T1);
-    for(int n=2;n<order;n++){
-
-      Linop.HermOp(*Tn,y);
-      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);
-      }
-
-      // Cycle pointers to avoid copies
-      Field *swizzle = Tnm;
-      Tnm    =Tn;
-      Tn     =Tnp;
-      Tnp    =swizzle;
-	  
-    }
-  }
-};
-
-
-template<class Field>
-class ChebyshevLanczos : public Chebyshev<Field> {
-private:
-  std::vector<RealD> Coeffs;
-  int order;
-  RealD alpha;
-  RealD beta;
-  RealD mu;
-
-public:
-  ChebyshevLanczos(RealD _alpha,RealD _beta,RealD _mu,int _order) :
-    alpha(_alpha),
-    beta(_beta),
-    mu(_mu)
-  {
-    order=_order;
-    Coeffs.resize(order);
-    for(int i=0;i<_order;i++){
-      Coeffs[i] = 0.0;
-    }
-    Coeffs[order-1]=1.0;
-  };
-
-  void csv(std::ostream &out){
-    for (RealD x=-1.2*alpha; x<1.2*alpha; x+=(2.0*alpha)/10000) {
-      RealD f = approx(x);
-      out<< x<<" "<<f<<std::endl;
-    }
-    return;
-  }
-
-  RealD approx(RealD xx) // Convenience for plotting the approximation
-  {
-    RealD Tn;
-    RealD Tnm;
-    RealD Tnp;
-    Real aa = alpha * alpha;
-    Real bb = beta  *  beta;
-      
-    RealD x = ( 2.0 * (xx-mu)*(xx-mu) - (aa+bb) ) / (aa-bb);
-
-    RealD y= x;
-      
-    RealD T0=1;
-    RealD T1=y;
-      
-    RealD sum;
-    sum = 0.5*Coeffs[0]*T0;
-    sum+= Coeffs[1]*T1;
-      
-    Tn =T1;
-    Tnm=T0;
-    for(int i=2;i<order;i++){
-      Tnp=2*y*Tn-Tnm;
-      Tnm=Tn;
-      Tn =Tnp;
-      sum+= Tn*Coeffs[i];
-    }
-    return sum;
-  };
-
-  // shift_Multiply in Rudy's code
-  void AminusMuSq(LinearOperatorBase<Field> &Linop, const Field &in, Field &out) 
-  {
-    GridBase *grid=in.Grid();
-    Field tmp(grid);
-
-    RealD aa= alpha*alpha;
-    RealD bb= beta * beta;
-
-    Linop.HermOp(in,out);
-    out = out - mu*in;
-
-    Linop.HermOp(out,tmp);
-    tmp = tmp - mu * out;
-
-    out = (2.0/ (aa-bb) ) * tmp -  ((aa+bb)/(aa-bb))*in;
-  };
-  // Implement the required interface
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-
-    GridBase *grid=in.Grid();
-
-    int vol=grid->gSites();
-
-    Field T0(grid); T0 = in;  
-    Field T1(grid); 
-    Field T2(grid);
-    Field  y(grid);
-      
-    Field *Tnm = &T0;
-    Field *Tn  = &T1;
-    Field *Tnp = &T2;
-
-    // Tn=T1 = (xscale M )*in
-    AminusMuSq(Linop,T0,T1);
-
-    // sum = .5 c[0] T0 + c[1] T1
-    out = (0.5*Coeffs[0])*T0 + Coeffs[1]*T1;
-    for(int n=2;n<order;n++){
-	
-      AminusMuSq(Linop,*Tn,y);
-
-      *Tnp=2.0*y-(*Tnm);
-
-      out=out+Coeffs[n]* (*Tnp);
-
-      // Cycle pointers to avoid copies
-      Field *swizzle = Tnm;
-      Tnm    =Tn;
-      Tn     =Tnp;
-      Tnp    =swizzle;
-	  
-    }
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/approx/Forecast.h

@@ -1,152 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/algorithms/approx/Forecast.h
-
-Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: David Murphy <dmurphy@phys.columbia.edu>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-			   /*  END LEGAL */
-
-#ifndef INCLUDED_FORECAST_H
-#define INCLUDED_FORECAST_H
-
-NAMESPACE_BEGIN(Grid);
-
-// Abstract base class.
-// Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
-// and returns a forecasted solution to the system D*psi = phi (psi).
-template<class Matrix, class Field>
-class Forecast
-{
-public:
-  virtual Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
-};
-
-// Implementation of Brower et al.'s chronological inverter (arXiv:hep-lat/9509012),
-// used to forecast solutions across poles of the EOFA heatbath.
-//
-// Modified from CPS (cps_pp/src/util/dirac_op/d_op_base/comsrc/minresext.C)
-template<class Matrix, class Field>
-class ChronoForecast : public Forecast<Matrix,Field>
-{
-public:
-  Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& prev_solns)
-  {
-    int degree = prev_solns.size();
-    Field chi(phi); // forecasted solution
-
-    // Trivial cases
-    if(degree == 0){ chi = Zero(); return chi; }
-    else if(degree == 1){ return prev_solns[0]; }
-
-    //    RealD dot;
-    ComplexD xp;
-    Field r(phi); // residual
-    Field Mv(phi);
-    std::vector<Field> v(prev_solns); // orthonormalized previous solutions
-    std::vector<Field> MdagMv(degree,phi);
-
-    // Array to hold the matrix elements
-    std::vector<std::vector<ComplexD>> G(degree, std::vector<ComplexD>(degree));
-
-    // Solution and source vectors
-    std::vector<ComplexD> a(degree);
-    std::vector<ComplexD> b(degree);
-
-    // Orthonormalize the vector basis
-    for(int i=0; i<degree; i++){
-      v[i] *= 1.0/std::sqrt(norm2(v[i]));
-      for(int j=i+1; j<degree; j++){ v[j] -= innerProduct(v[i],v[j]) * v[i]; }
-    }
-
-    // Perform sparse matrix multiplication and construct rhs
-    for(int i=0; i<degree; i++){
-      b[i] = innerProduct(v[i],phi);
-      Mat.M(v[i],Mv);
-      Mat.Mdag(Mv,MdagMv[i]);
-      G[i][i] = innerProduct(v[i],MdagMv[i]);
-    }
-
-    // Construct the matrix
-    for(int j=0; j<degree; j++){
-      for(int k=j+1; k<degree; k++){
-	G[j][k] = innerProduct(v[j],MdagMv[k]);
-	G[k][j] = conjugate(G[j][k]);
-      }}
-
-    // Gauss-Jordan elimination with partial pivoting
-    for(int i=0; i<degree; i++){
-
-      // Perform partial pivoting
-      int k = i;
-      for(int j=i+1; j<degree; j++){ if(abs(G[j][j]) > abs(G[k][k])){ k = j; } }
-      if(k != i){
-	xp = b[k];
-	b[k] = b[i];
-	b[i] = xp;
-	for(int j=0; j<degree; j++){
-	  xp = G[k][j];
-	  G[k][j] = G[i][j];
-	  G[i][j] = xp;
-	}
-      }
-
-      // Convert matrix to upper triangular form
-      for(int j=i+1; j<degree; j++){
-	xp = G[j][i]/G[i][i];
-	b[j] -= xp * b[i];
-	for(int k=0; k<degree; k++){ G[j][k] -= xp*G[i][k]; }
-      }
-    }
-
-    // Use Gaussian elimination to solve equations and calculate initial guess
-    chi = Zero();
-    r = phi;
-    for(int i=degree-1; i>=0; i--){
-      a[i] = 0.0;
-      for(int j=i+1; j<degree; j++){ a[i] += G[i][j] * a[j]; }
-      a[i] = (b[i]-a[i])/G[i][i];
-      chi += a[i]*v[i];
-      r -= a[i]*MdagMv[i];
-    }
-
-    RealD true_r(0.0);
-    ComplexD tmp;
-    for(int i=0; i<degree; i++){
-      tmp = -b[i];
-      for(int j=0; j<degree; j++){ tmp += G[i][j]*a[j]; }
-      tmp = conjugate(tmp)*tmp;
-      true_r += std::sqrt(tmp.real());
-    }
-
-    RealD error = std::sqrt(norm2(r)/norm2(phi));
-    std::cout << GridLogMessage << "ChronoForecast: |res|/|src| = " << error << std::endl;
-
-    return chi;
-  };
-};
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/approx/JacobiPolynomial.h

@@ -1,129 +0,0 @@
-#ifndef GRID_JACOBIPOLYNOMIAL_H
-#define GRID_JACOBIPOLYNOMIAL_H
-
-#include <Grid/algorithms/LinearOperator.h>
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Field>
-class JacobiPolynomial : public OperatorFunction<Field> {
- private:
-  using OperatorFunction<Field>::operator();
-
-  int order;
-  RealD hi;
-  RealD lo;
-  RealD alpha;
-  RealD beta;
-
- public:
-  void csv(std::ostream &out){
-    csv(out,lo,hi);
-  }
-  void csv(std::ostream &out,RealD llo,RealD hhi){
-    RealD diff = hhi-llo;
-    RealD delta = diff*1.0e-5;
-    for (RealD x=llo-delta; x<=hhi; x+=delta) {
-      RealD f = approx(x);
-      out<< x<<" "<<f <<std::endl;
-    }
-    return;
-  }
-
-  JacobiPolynomial(){};
-  JacobiPolynomial(RealD _lo,RealD _hi,int _order,RealD _alpha, RealD _beta)
-  {
-      lo=_lo;
-      hi=_hi;
-      alpha=_alpha;
-      beta=_beta;
-      order=_order;
-  };
-
-  RealD approx(RealD x) // Convenience for plotting the approximation                                                       
-  {
-    RealD Tn;
-    RealD Tnm;
-    RealD Tnp;
-
-    RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
-
-    RealD T0=1.0;
-    RealD T1=(alpha-beta)*0.5+(alpha+beta+2.0)*0.5*y;
-
-    Tn =T1;
-    Tnm=T0;
-    for(int n=2;n<=order;n++){
-      RealD cnp = 2.0*n*(n+alpha+beta)*(2.0*n-2.0+alpha+beta);
-      RealD cny = (2.0*n-2.0+alpha+beta)*(2.0*n-1.0+alpha+beta)*(2.0*n+alpha+beta);
-      RealD cn1 = (2.0*n+alpha+beta-1.0)*(alpha*alpha-beta*beta);
-      RealD cnm = - 2.0*(n+alpha-1.0)*(n+beta-1.0)*(2.0*n+alpha+beta);
-      Tnp= ( cny * y *Tn + cn1 * Tn + cnm * Tnm )/ cnp;
-      Tnm=Tn;
-      Tn =Tnp;
-    }
-    return Tnp;
-  };
-
-  // Implement the required interface                                                                                       
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-    GridBase *grid=in.Grid();
-
-    int vol=grid->gSites();
-
-    Field T0(grid);
-    Field T1(grid);
-    Field T2(grid);
-    Field y(grid);
-
-
-    Field *Tnm = &T0;
-    Field *Tn  = &T1;
-    Field *Tnp = &T2;
-
-    //    RealD T0=1.0;                                                                                                     
-    T0=in;
-
-    //    RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));                                                                           
-    //           = x * 2/(hi-lo) - (hi+lo)/(hi-lo)                                                                          
-    Linop.HermOp(T0,y);
-    RealD xscale = 2.0/(hi-lo);
-    RealD mscale = -(hi+lo)/(hi-lo);
-    Linop.HermOp(T0,y);
-    y=y*xscale+in*mscale;
-
-    // RealD T1=(alpha-beta)*0.5+(alpha+beta+2.0)*0.5*y;
-    RealD halfAmB  = (alpha-beta)*0.5;
-    RealD halfApBp2= (alpha+beta+2.0)*0.5;
-    T1 = halfAmB * in + halfApBp2*y;
-
-    for(int n=2;n<=order;n++){
-
-      Linop.HermOp(*Tn,y);
-      y=xscale*y+mscale*(*Tn);
-
-      RealD cnp = 2.0*n*(n+alpha+beta)*(2.0*n-2.0+alpha+beta);
-      RealD cny = (2.0*n-2.0+alpha+beta)*(2.0*n-1.0+alpha+beta)*(2.0*n+alpha+beta);
-      RealD cn1 = (2.0*n+alpha+beta-1.0)*(alpha*alpha-beta*beta);
-      RealD cnm = - 2.0*(n+alpha-1.0)*(n+beta-1.0)*(2.0*n+alpha+beta);
-
-      //      Tnp= ( cny * y *Tn + cn1 * Tn + cnm * Tnm )/ cnp;                                                             
-      cny=cny/cnp;
-      cn1=cn1/cnp;
-      cn1=cn1/cnp;
-      cnm=cnm/cnp;
-
-      *Tnp=cny*y + cn1 *(*Tn) + cnm * (*Tnm);
-
-      // Cycle pointers to avoid copies                                                                                     
-      Field *swizzle = Tnm;
-      Tnm    =Tn;
-      Tn     =Tnp;
-      Tnp    =swizzle;
-    }
-    out=*Tnp;
-
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/approx/RemezGeneral.cc

@@ -1,473 +0,0 @@
-#include<math.h>
-#include<stdio.h>
-#include<stdlib.h>
-#include<string>
-#include<iostream>
-#include<iomanip>
-#include<cassert>
-
-#include<Grid/algorithms/approx/RemezGeneral.h>
-
-
-// Constructor
-AlgRemezGeneral::AlgRemezGeneral(double lower, double upper, long precision,
-				 bigfloat (*f)(bigfloat x, void *data), void *data): f(f), 
-										     data(data), 
-										     prec(precision),
-										     apstrt(lower), apend(upper), apwidt(upper - lower),
-										     n(0), d(0), pow_n(0), pow_d(0)
-{
-  bigfloat::setDefaultPrecision(prec);
-
-  std::cout<<"Approximation bounds are ["<<apstrt<<","<<apend<<"]\n";
-  std::cout<<"Precision of arithmetic is "<<precision<<std::endl;
-}
-
-//Determine the properties of the numerator and denominator polynomials
-void AlgRemezGeneral::setupPolyProperties(int num_degree, int den_degree, PolyType num_type_in, PolyType den_type_in){
-  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_d % 2 == 0 && den_type_in == PolyType::Odd) assert(0);
-  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) assert(0);
-
-  num_type = num_type_in;
-  den_type = den_type_in;
-
-  num_pows.resize(pow_n+1);
-  den_pows.resize(pow_d+1);
-
-  int n_in = 0;
-  bool odd = num_type == PolyType::Full || num_type == PolyType::Odd;
-  bool even = num_type == PolyType::Full || num_type == PolyType::Even;
-  for(int i=0;i<=pow_n;i++){
-    num_pows[i] = -1;
-    if(i % 2 == 0 && even) num_pows[i] = n_in++;
-    if(i % 2 == 1 && odd) num_pows[i] = n_in++;
-  }
-
-  std::cout << n_in << " terms in numerator" << std::endl;
-  --n_in; //power is 1 less than the number of terms, eg  pow=1   a x^1  + b x^0
-
-  int d_in = 0;
-  odd = den_type == PolyType::Full || den_type == PolyType::Odd;
-  even = den_type == PolyType::Full || den_type == PolyType::Even;
-  for(int i=0;i<=pow_d;i++){
-    den_pows[i] = -1;
-    if(i % 2 == 0 && even) den_pows[i] = d_in++;
-    if(i % 2 == 1 && odd) den_pows[i] = d_in++;
-  }
-
-  std::cout << d_in << " terms in denominator" << std::endl;
-  --d_in;
-
-  n = n_in;
-  d = d_in;
-}
-
-//Setup algorithm
-void AlgRemezGeneral::reinitializeAlgorithm(){
-  spread = 1.0e37;
-  iter = 0;
-
-  neq = n + d + 1; //not +2 because highest-power term in denominator is fixed to 1
-
-  param.resize(neq);
-  yy.resize(neq+1);
-
-  //Initialize linear equation temporaries
-  A.resize(neq*neq);
-  B.resize(neq);
-  IPS.resize(neq);
-
-  //Initialize maximum and minimum errors
-  xx.resize(neq+2);
-  mm.resize(neq+1);
-  initialGuess();
-
-  //Initialize search steps
-  step.resize(neq+1);
-  stpini();
-}
-
-double AlgRemezGeneral::generateApprox(const int num_degree, const int den_degree, 
-				       const PolyType num_type_in, const PolyType den_type_in, 
-				       const double _tolerance, const int report_freq){
-  //Setup the properties of the polynomial
-  setupPolyProperties(num_degree, den_degree, num_type_in, den_type_in);
-
-  //Setup the algorithm
-  reinitializeAlgorithm();
-
-  bigfloat tolerance = _tolerance;
-
-  //Iterate until convergance
-  while (spread > tolerance) { 
-    if (iter++ % report_freq==0)
-      std::cout<<"Iteration " <<iter-1<<" spread "<<(double)spread<<" delta "<<(double)delta << std::endl; 
-
-    equations();
-    if (delta < tolerance) {
-      std::cout<<"Iteration " << iter-1 << " delta too small (" << delta << "<" << tolerance << "), try increasing precision\n";
-      assert(0);
-    };    
-    assert( delta>= tolerance );
-
-    search();
-  }
-
-  int sign;
-  double error = (double)getErr(mm[0],&sign);
-  std::cout<<"Converged at "<<iter<<" iterations; error = "<<error<<std::endl;
-
-  // Return the maximum error in the approximation
-  return error;
-}
-
-
-// Initial values of maximal and minimal errors
-void AlgRemezGeneral::initialGuess(){
-  // Supply initial guesses for solution points
-  long ncheb = neq;			// Degree of Chebyshev error estimate
-
-  // Find ncheb+1 extrema of Chebyshev polynomial
-  bigfloat a = ncheb;
-  bigfloat r;
-
-  mm[0] = apstrt;
-  for (long i = 1; i < ncheb; i++) {
-    r = 0.5 * (1 - cos((M_PI * i)/(double) a));
-    //r *= sqrt_bf(r);
-    r = (exp((double)r)-1.0)/(exp(1.0)-1.0);
-    mm[i] = apstrt + r * apwidt;
-  }
-  mm[ncheb] = apend;
-
-  a = 2.0 * ncheb;
-  for (long i = 0; i <= ncheb; i++) {
-    r = 0.5 * (1 - cos(M_PI * (2*i+1)/(double) a));
-    //r *= sqrt_bf(r); // Squeeze to low end of interval
-    r = (exp((double)r)-1.0)/(exp(1.0)-1.0);
-    xx[i] = apstrt + r * apwidt;
-  }
-}
-
-// Initialise step sizes
-void AlgRemezGeneral::stpini(){
-  xx[neq+1] = apend;
-  delta = 0.25;
-  step[0] = xx[0] - apstrt;
-  for (int i = 1; i < neq; i++) step[i] = xx[i] - xx[i-1];
-  step[neq] = step[neq-1];
-}
-
-// Search for error maxima and minima
-void AlgRemezGeneral::search(){
-  bigfloat a, q, xm, ym, xn, yn, xx1;
-  int emsign, ensign, steps;
-
-  int meq = neq + 1;
-
-  bigfloat eclose = 1.0e30;
-  bigfloat farther = 0l;
-
-  bigfloat xx0 = apstrt;
-
-  for (int i = 0; i < meq; i++) {
-    steps = 0;
-    xx1 = xx[i]; // Next zero
-    if (i == meq-1) xx1 = apend;
-    xm = mm[i];
-    ym = getErr(xm,&emsign);
-    q = step[i];
-    xn = xm + q;
-    if (xn < xx0 || xn >= xx1) {	// Cannot skip over adjacent boundaries
-      q = -q;
-      xn = xm;
-      yn = ym;
-      ensign = emsign;
-    } else {
-      yn = getErr(xn,&ensign);
-      if (yn < ym) {
-	q = -q;
-	xn = xm;
-	yn = ym;
-	ensign = emsign;
-      }
-    }
-  
-    while(yn >= ym) {		// March until error becomes smaller.
-      if (++steps > 10)
-      	break;
-      
-      ym = yn;
-      xm = xn;
-      emsign = ensign;
-      a = xm + q;
-      if (a == xm || a <= xx0 || a >= xx1)
-	break;// Must not skip over the zeros either side.      
-
-      xn = a;
-      yn = getErr(xn,&ensign);
-    }
-
-    mm[i] = xm;			// Position of maximum
-    yy[i] = ym;			// Value of maximum
-
-    if (eclose > ym) eclose = ym;
-    if (farther < ym) farther = ym;
-
-    xx0 = xx1; // Walk to next zero.
-  } // end of search loop
-
-  q = (farther - eclose);	// Decrease step size if error spread increased
-
-  if (eclose != 0.0) q /= eclose; // Relative error spread
-
-  if (q >= spread)
-    delta *= 0.5; // Spread is increasing; decrease step size
-  
-  spread = q;
-
-  for (int i = 0; i < neq; i++) {
-    q = yy[i+1];
-    if (q != 0.0) q = yy[i] / q  - (bigfloat)1l;
-    else q = 0.0625;
-    if (q > (bigfloat)0.25) q = 0.25;
-    q *= mm[i+1] - mm[i];
-    step[i] = q * delta;
-  }
-  step[neq] = step[neq-1];
-  
-  for (int i = 0; i < neq; i++) {	// Insert new locations for the zeros.
-    xm = xx[i] - step[i];
-
-    if (xm <= apstrt)
-      continue;
-
-    if (xm >= apend)
-      continue;
-
-    if (xm <= mm[i])
-      xm = (bigfloat)0.5 * (mm[i] + xx[i]);    
-
-    if (xm >= mm[i+1])
-      xm = (bigfloat)0.5 * (mm[i+1] + xx[i]);
-    
-    xx[i] = xm;
-  }
-}
-
-// Solve the equations
-void AlgRemezGeneral::equations(){
-  bigfloat x, y, z;
-  bigfloat *aa;
-  
-  for (int i = 0; i < neq; i++) {	// set up the equations for solution by simq()
-    int ip = neq * i;		// offset to 1st element of this row of matrix
-    x = xx[i];			// the guess for this row
-    y = func(x);		// right-hand-side vector
-
-    z = (bigfloat)1l;
-    aa = A.data()+ip;
-    int t = 0;
-    for (int j = 0; j <= pow_n; j++) {
-      if(num_pows[j] != -1){ *aa++ = z; t++; }
-      z *= x;
-    }
-    assert(t == n+1);
-
-    z = (bigfloat)1l;
-    t = 0;
-    for (int j = 0; j < pow_d; j++) {
-      if(den_pows[j] != -1){ *aa++ = -y * z; t++; }
-      z *= x;
-    }
-    assert(t == d);
-
-    B[i] = y * z;		// Right hand side vector
-  }
-
-  // Solve the simultaneous linear equations.
-  if (simq()){
-    std::cout<<"simq failed\n";
-    exit(0);
-  }
-}
-
-
-// Evaluate the rational form P(x)/Q(x) using coefficients
-// from the solution vector param
-bigfloat AlgRemezGeneral::approx(const bigfloat x) const{
-  // Work backwards toward the constant term.
-  int c = n;
-  bigfloat yn = param[c--];		// Highest order numerator coefficient
-  for (int i = pow_n-1; i >= 0; i--) yn = x * yn  +  (num_pows[i] != -1 ? param[c--] : bigfloat(0l));  
-
-  c = n+d;
-  bigfloat yd = 1l; //Highest degree coefficient is 1.0
-  for (int i = pow_d-1; i >= 0; i--) yd = x * yd  +  (den_pows[i] != -1 ? param[c--] : bigfloat(0l)); 
-
-  return(yn/yd);
-}
-
-// Compute size and sign of the approximation error at x
-bigfloat AlgRemezGeneral::getErr(bigfloat x, int *sign) const{
-  bigfloat f = func(x);
-  bigfloat e = approx(x) - f;
-  if (f != 0) e /= f;
-  if (e < (bigfloat)0.0) {
-    *sign = -1;
-    e = -e;
-  }
-  else *sign = 1;
-  
-  return(e);
-}
-
-// Solve the system AX=B
-int AlgRemezGeneral::simq(){
-
-  int ip, ipj, ipk, ipn;
-  int idxpiv;
-  int kp, kp1, kpk, kpn;
-  int nip, nkp;
-  bigfloat em, q, rownrm, big, size, pivot, sum;
-  bigfloat *aa;
-  bigfloat *X = param.data();
-
-  int n = neq;
-  int nm1 = n - 1;
-  // Initialize IPS and X
-  
-  int ij = 0;
-  for (int i = 0; i < n; i++) {
-    IPS[i] = i;
-    rownrm = 0.0;
-    for(int j = 0; j < n; j++) {
-      q = abs_bf(A[ij]);
-      if(rownrm < q) rownrm = q;
-      ++ij;
-    }
-    if (rownrm == (bigfloat)0l) {
-      std::cout<<"simq rownrm=0\n";
-      return(1);
-    }
-    X[i] = (bigfloat)1.0 / rownrm;
-  }
-  
-  for (int k = 0; k < nm1; k++) {
-    big = 0.0;
-    idxpiv = 0;
-    
-    for (int i = k; i < n; i++) {
-      ip = IPS[i];
-      ipk = n*ip + k;
-      size = abs_bf(A[ipk]) * X[ip];
-      if (size > big) {
-	big = size;
-	idxpiv = i;
-      }
-    }
-    
-    if (big == (bigfloat)0l) {
-      std::cout<<"simq big=0\n";
-      return(2);
-    }
-    if (idxpiv != k) {
-      int j = IPS[k];
-      IPS[k] = IPS[idxpiv];
-      IPS[idxpiv] = j;
-    }
-    kp = IPS[k];
-    kpk = n*kp + k;
-    pivot = A[kpk];
-    kp1 = k+1;
-    for (int i = kp1; i < n; i++) {
-      ip = IPS[i];
-      ipk = n*ip + k;
-      em = -A[ipk] / pivot;
-      A[ipk] = -em;
-      nip = n*ip;
-      nkp = n*kp;
-      aa = A.data()+nkp+kp1;
-      for (int j = kp1; j < n; j++) {
-	ipj = nip + j;
-	A[ipj] = A[ipj] + em * *aa++;
-      }
-    }
-  }
-  kpn = n * IPS[n-1] + n - 1;	// last element of IPS[n] th row
-  if (A[kpn] == (bigfloat)0l) {
-    std::cout<<"simq A[kpn]=0\n";
-    return(3);
-  }
-
-  
-  ip = IPS[0];
-  X[0] = B[ip];
-  for (int i = 1; i < n; i++) {
-    ip = IPS[i];
-    ipj = n * ip;
-    sum = 0.0;
-    for (int j = 0; j < i; j++) {
-      sum += A[ipj] * X[j];
-      ++ipj;
-    }
-    X[i] = B[ip] - sum;
-  }
-  
-  ipn = n * IPS[n-1] + n - 1;
-  X[n-1] = X[n-1] / A[ipn];
-  
-  for (int iback = 1; iback < n; iback++) {
-    //i goes (n-1),...,1
-    int i = nm1 - iback;
-    ip = IPS[i];
-    nip = n*ip;
-    sum = 0.0;
-    aa = A.data()+nip+i+1;
-    for (int j= i + 1; j < n; j++) 
-      sum += *aa++ * X[j];
-    X[i] = (X[i] - sum) / A[nip+i];
-  }
-  
-  return(0);
-}
-
-void AlgRemezGeneral::csv(std::ostream & os) const{
-  os << "Numerator" << std::endl;
-  for(int i=0;i<=pow_n;i++){
-    os << getCoeffNum(i) << "*x^" << i;
-    if(i!=pow_n) os << " + ";
-  }
-  os << std::endl;
-
-  os << "Denominator" << std::endl;
-  for(int i=0;i<=pow_d;i++){
-    os << getCoeffDen(i) << "*x^" << i;
-    if(i!=pow_d) os << " + ";
-  }
-  os << std::endl;
-
-  //For a true minimax solution the errors should all be equal and the signs should oscillate +-+-+- etc
-  int sign;
-  os << "Errors at maxima: coordinate, error, (sign)" << std::endl;
-  for(int i=0;i<neq+1;i++){ 
-    os << mm[i] << " " << getErr(mm[i],&sign) << " (" << sign << ")" << std::endl;
-  }
-
-  os << "Scan over range:" << std::endl;
-  int npt = 60;
-  bigfloat dlt = (apend - apstrt)/bigfloat(npt-1);
-
-  for (bigfloat x=apstrt; x<=apend; x = x + dlt) {
-    double f = evaluateFunc(x);
-    double r = evaluateApprox(x);
-    os<< x<<","<<r<<","<<f<<","<<r-f<<std::endl;
-  }
-  return;
-}

Grid/algorithms/approx/RemezGeneral.h

@@ -1,170 +0,0 @@
-/*
-  C.Kelly Jan 2020 based on implementation by M. Clark May 2005
-
-  AlgRemezGeneral is an implementation of the Remez algorithm for approximating an arbitrary function by a rational polynomial 
-  It includes optional restriction to odd/even polynomials for the numerator and/or denominator
-*/
-
-#ifndef INCLUDED_ALG_REMEZ_GENERAL_H
-#define INCLUDED_ALG_REMEZ_GENERAL_H
-
-#include <stddef.h>
-#include <Grid/GridStd.h>
-
-#ifdef HAVE_LIBGMP
-#include "bigfloat.h"
-#else
-#include "bigfloat_double.h"
-#endif
-
-
-class AlgRemezGeneral{
- public:
-  enum PolyType { Even, Odd, Full };
-
- private:
-
-  // In GSL-style, pass the function as a function pointer. Any data required to evaluate the function is passed in as a void pointer
-  bigfloat (*f)(bigfloat x, void *data);
-  void *data;
-
-  // The approximation parameters
-  std::vector<bigfloat> param;
-  bigfloat norm;
-
-  // The number of non-zero terms in the numerator and denominator
-  int n, d;
-  // The numerator and denominator degree (i.e.  the largest power)
-  int pow_n, pow_d;
-  
-  // Specify if the numerator and/or denominator are odd/even polynomials
-  PolyType num_type;
-  PolyType den_type;
-  std::vector<int> num_pows; //contains the mapping, with -1 if not present
-  std::vector<int> den_pows;
-
-  // The bounds of the approximation
-  bigfloat apstrt, apwidt, apend;
-
-  // Variables used to calculate the approximation
-  int nd1, iter;
-  std::vector<bigfloat> xx;
-  std::vector<bigfloat> mm;
-  std::vector<bigfloat> step;
-
-  bigfloat delta, spread;
-  
-  // Variables used in search
-  std::vector<bigfloat> yy;
-
-  // Variables used in solving linear equations
-  std::vector<bigfloat> A;
-  std::vector<bigfloat> B;
-  std::vector<int> IPS;
-
-  // The number of equations we must solve at each iteration (n+d+1)
-  int neq;
-
-  // The precision of the GNU MP library
-  long prec;
-
-  // Initialize member variables associated with the polynomial's properties
-  void setupPolyProperties(int num_degree, int den_degree, PolyType num_type_in, PolyType den_type_in);
-
-  // Initial values of maximal and minmal errors
-  void initialGuess();
-
-  // Initialise step sizes
-  void stpini();
-
-  // Initialize the algorithm
-  void reinitializeAlgorithm();
-
-  // Solve the equations
-  void equations();
-
-  // Search for error maxima and minima
-  void search(); 
-
-  // Calculate function required for the approximation
-  inline bigfloat func(bigfloat x) const{
-    return f(x, data);
-  }
-
-  // Compute size and sign of the approximation error at x
-  bigfloat getErr(bigfloat x, int *sign) const;
-
-  // Solve the system AX=B   where X = param
-  int simq();
-
-  // Evaluate the rational form P(x)/Q(x) using coefficients from the solution vector param
-  bigfloat approx(bigfloat x) const;
-
- public:
-  
-  AlgRemezGeneral(double lower, double upper, long prec,
-		  bigfloat (*f)(bigfloat x, void *data), void *data);
-
-  inline int getDegree(void) const{ 
-    assert(n==d);
-    return n;
-  }
-  // Reset the bounds of the approximation
-  inline void setBounds(double lower, double upper) {
-    apstrt = lower;
-    apend = upper;
-    apwidt = apend - apstrt;
-  }
-
-  // Get the bounds of the approximation
-  inline void getBounds(double &lower, double &upper) const{ 
-    lower=(double)apstrt;
-    upper=(double)apend;
-  }
-
-  // Run the algorithm to generate the rational approximation
-  double generateApprox(int num_degree, int den_degree, 
-			PolyType num_type, PolyType den_type,
-			const double tolerance = 1e-15, const int report_freq = 1000);
-  
-  inline double generateApprox(int num_degree, int den_degree, 
-			       const double tolerance = 1e-15, const int report_freq = 1000){
-    return generateApprox(num_degree, den_degree, Full, Full, tolerance, report_freq);
-  }
-  
-  // Evaluate the rational form P(x)/Q(x) using coefficients from the
-  // solution vector param
-  inline double evaluateApprox(double x) const{
-    return (double)approx((bigfloat)x);
-  }
-
-  // Evaluate the rational form Q(x)/P(x) using coefficients from the solution vector param
-  inline double evaluateInverseApprox(double x) const{
-    return 1.0/(double)approx((bigfloat)x);
-  }  
-
-  // Calculate function required for the approximation
-  inline double evaluateFunc(double x) const{
-    return (double)func((bigfloat)x);
-  }
-
-  // Calculate inverse function required for the approximation
-  inline double evaluateInverseFunc(double x) const{
-    return 1.0/(double)func((bigfloat)x);
-  }
-
-  // Dump csv of function, approx and error
-  void csv(std::ostream &os = std::cout) const;
-
-  // Get the coefficient of the term x^i in the numerator
-  inline double getCoeffNum(const int i) const{    
-    return num_pows[i] == -1 ? 0. : double(param[num_pows[i]]);
-  }
-  // Get the coefficient of the term x^i in the denominator
-  inline double getCoeffDen(const int i) const{ 
-    if(i == pow_d) return 1.0;
-    else return den_pows[i] == -1 ? 0. : double(param[den_pows[i]+n+1]); 
-  }
-};
-
-#endif

Grid/algorithms/approx/ZMobius.cc

@@ -1,183 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/approx/ZMobius.cc
-
-    Copyright (C) 2015
-
-Author: Christopher Kelly <ckelly@phys.columbia.edu>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#include <Grid/algorithms/approx/ZMobius.h>
-#include <Grid/algorithms/approx/RemezGeneral.h>
-
-NAMESPACE_BEGIN(Grid);
-NAMESPACE_BEGIN(Approx);
-
-//Compute the tanh approximation
-inline double epsilonMobius(const double x, const std::vector<ComplexD> &w){
-  int Ls = w.size();
-
-  ComplexD fxp = 1., fmp = 1.;
-  for(int i=0;i<Ls;i++){
-    fxp = fxp * ( w[i] + x );
-    fmp = fmp * ( w[i] - x );
-  }
-  return ((fxp - fmp)/(fxp + fmp)).real();
-}
-inline double epsilonMobius(const double x, const std::vector<RealD> &w){
-  int Ls = w.size();
-
-  double fxp = 1., fmp = 1.;
-  for(int i=0;i<Ls;i++){
-    fxp = fxp * ( w[i] + x );
-    fmp = fmp * ( w[i] - x );
-  }
-  return (fxp - fmp)/(fxp + fmp);
-}
-
-
-
-//Compute the tanh approximation in a form suitable for the Remez
-bigfloat epsilonMobius(bigfloat x, void* data){
-  const std::vector<RealD> &omega = *( (std::vector<RealD> const*)data );
-  bigfloat fxp(1.0);
-  bigfloat fmp(1.0);
-
-  for(int i=0;i<omega.size();i++){
-    fxp = fxp * ( bigfloat(omega[i]) + x);
-    fmp = fmp * ( bigfloat(omega[i]) - x);
-  }
-  return (fxp - fmp)/(fxp + fmp);
-}
-
-//Compute the Zmobius Omega parameters suitable for eigenvalue range   -lambda_bound <= lambda <= lambda_bound
-//Note omega_i = 1/(b_i + c_i)   where b_i and c_i are the Mobius parameters
-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);
-  omega_out.resize(Ls_out);
-
-  //Use the Remez algorithm to generate the appropriate rational polynomial
-  //For odd polynomial, to satisfy Haar condition must take either positive or negative half of range (cf https://arxiv.org/pdf/0803.0439.pdf page 6)  
-  AlgRemezGeneral remez(0, lambda_bound, 64, &epsilonMobius, (void*)&omega_in); 
-  remez.generateApprox(Ls_out-1, Ls_out,AlgRemezGeneral::Odd, AlgRemezGeneral::Even, 1e-15, 100);
-  remez.csv(std::cout);
-
-  //The rational approximation has the form  [ f(x) - f(-x) ] / [ f(x) + f(-x) ]  where  f(x) = \Prod_{i=0}^{L_s-1} ( \omega_i + x )
-  //cf https://academiccommons.columbia.edu/doi/10.7916/D8T72HD7  pg 102
-  //omega_i are therefore the negative of the complex roots of f(x)
-
-  //We can find the roots by recognizing that the eigenvalues of a matrix A are the roots of the characteristic polynomial
-  // \rho(\lambda) = det( A - \lambda I )    where I is the unit matrix
-  //The matrix whose characteristic polynomial is an arbitrary monic polynomial a0 + a1 x + a2 x^2 + ... x^n   is the companion matrix 
-  // A = | 0    1   0    0 0 .... 0 |
-  //     | 0    0   1    0 0 .... 0 |
-  //     | :    :   :    : :      : |
-  //     | 0    0   0    0 0      1
-  //     | -a0 -a1 -a2  ...  ... -an|
-
-
-  //Note the Remez defines the largest power to have unit coefficient
-  std::vector<RealD> coeffs(Ls_out+1);
-  for(int i=0;i<Ls_out+1;i+=2) coeffs[i] = coeffs[i] = remez.getCoeffDen(i); //even powers
-  for(int i=1;i<Ls_out+1;i+=2) coeffs[i] = coeffs[i] = remez.getCoeffNum(i); //odd powers
-
-  std::vector<std::complex<RealD> > roots(Ls_out);
-
-  //Form the companion matrix
-  Eigen::MatrixXd compn(Ls_out,Ls_out);
-  for(int i=0;i<Ls_out-1;i++) compn(i,0) = 0.;
-  compn(Ls_out - 1, 0) = -coeffs[0];
-  
-  for(int j=1;j<Ls_out;j++){
-    for(int i=0;i<Ls_out-1;i++) compn(i,j) = i == j-1 ? 1. : 0.;
-    compn(Ls_out - 1, j) = -coeffs[j];
-  }
-
-  //Eigensolve
-  Eigen::EigenSolver<Eigen::MatrixXd> slv(compn, false);
-
-  const auto & ev = slv.eigenvalues();
-  for(int i=0;i<Ls_out;i++)
-    omega_out[i] = -ev(i);
-
-  //Sort ascending (smallest at start of vector!)
-  std::sort(omega_out.begin(), omega_out.end(), 
-	    [&](const ComplexD &a, const ComplexD &b){ return a.real() < b.real() || (a.real() == b.real() && a.imag() < b.imag()); });
-
-  //McGlynn thesis pg 122 suggest improved iteration counts if magnitude of omega diminishes towards the center of the 5th dimension
-  std::vector<ComplexD> omega_tmp = omega_out;
-  int s_low=0, s_high=Ls_out-1, ss=0;
-  for(int s_from = Ls_out-1; s_from >= 0; s_from--){ //loop from largest omega
-    int s_to;
-    if(ss % 2 == 0){
-      s_to = s_low++;
-    }else{
-      s_to = s_high--;
-    }
-    omega_out[s_to] = omega_tmp[s_from];
-    ++ss;
-  }
-  
-  std::cout << "Resulting omega_i:" << std::endl;  
-  for(int i=0;i<Ls_out;i++)
-    std::cout << omega_out[i] << std::endl;
-
-  std::cout << "Test result matches the approximate polynomial found by the Remez" << std::endl;
-  std::cout << "<x> <remez approx> <poly approx> <diff poly approx remez approx> <exact> <diff poly approx exact>\n";
-  
-  int npt = 60;
-  double dlt = lambda_bound/double(npt-1);
-
-  for (int i =0; i<npt; i++){
-    double x = i*dlt;
-    double r = remez.evaluateApprox(x);
-    double p = epsilonMobius(x, omega_out);
-    double e = epsilonMobius(x, omega_in);
-
-    std::cout << x<< " " << r << " " << p <<" " <<r-p << " " << e << " " << e-p << std::endl;
-  }
-
-}
-  
-//mobius_param = b+c   with b-c=1
-void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound){
-  std::vector<RealD> omega_in(Ls_in, 1./mobius_param);
-  computeZmobiusOmega(omega_out, Ls_out, omega_in, Ls_in, lambda_bound);
-}
-
-//ZMobius class takes  gamma_i = (b+c) omega_i as its input, where b, c are factored out
-void computeZmobiusGamma(std::vector<ComplexD> &gamma_out, 
-			 const RealD mobius_param_out, const int Ls_out, 
-			 const RealD mobius_param_in, const int Ls_in,
-			 const RealD lambda_bound){
-  computeZmobiusOmega(gamma_out, Ls_out, mobius_param_in, Ls_in, lambda_bound);
-  for(int i=0;i<Ls_out;i++) gamma_out[i] = gamma_out[i] * mobius_param_out;
-}
-//Assumes mobius_param_out == mobius_param_in
-void computeZmobiusGamma(std::vector<ComplexD> &gamma_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound){
-  computeZmobiusGamma(gamma_out, mobius_param, Ls_out, mobius_param, Ls_in, lambda_bound);
-}
-
-NAMESPACE_END(Approx);
-NAMESPACE_END(Grid);

Grid/algorithms/approx/ZMobius.h

@@ -1,57 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/approx/ZMobius.h
-
-    Copyright (C) 2015
-
-Author: Christopher Kelly <ckelly@phys.columbia.edu>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_ZMOBIUS_APPROX_H
-#define GRID_ZMOBIUS_APPROX_H
-
-#include <Grid/GridCore.h>
-
-NAMESPACE_BEGIN(Grid);
-NAMESPACE_BEGIN(Approx);
-
-//Compute the Zmobius Omega parameters suitable for eigenvalue range   -lambda_bound <= lambda <= lambda_bound
-//Note omega_i = 1/(b_i + c_i)   where b_i and c_i are the Mobius parameters
-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);
-  
-//mobius_param = b+c   with b-c=1
-void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound);
-
-//ZMobius class takes  gamma_i = (b+c) omega_i as its input, where b, c are factored out
-void computeZmobiusGamma(std::vector<ComplexD> &gamma_out, 
-			 const RealD mobius_param_out, const int Ls_out, 
-			 const RealD mobius_param_in, const int Ls_in,
-			 const RealD lambda_bound);
-
-//Assumes mobius_param_out == mobius_param_in
-void computeZmobiusGamma(std::vector<ComplexD> &gamma_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound);
-
-NAMESPACE_END(Approx);
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/blas/BatchedBlas.cc

@@ -1,34 +0,0 @@
-/*************************************************************************************
-
-    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/BatchedBlas.h

@@ -1,727 +0,0 @@
-/*************************************************************************************
-
-    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 */
-#pragma once
-
-#ifdef GRID_HIP
-#include <hipblas/hipblas.h>
-#endif
-#ifdef GRID_CUDA
-#include <cublas_v2.h>
-#endif
-#ifdef GRID_SYCL
-#include <oneapi/mkl.hpp>
-#endif
-#if 0
-#define GRID_ONE_MKL
-#endif
-#ifdef GRID_ONE_MKL
-#include <oneapi/mkl.hpp>
-#endif
-///////////////////////////////////////////////////////////////////////	  
-// Need to rearrange lattice data to be in the right format for a
-// batched multiply. Might as well make these static, dense packed
-///////////////////////////////////////////////////////////////////////
-NAMESPACE_BEGIN(Grid);
-#ifdef GRID_HIP
-  typedef hipblasHandle_t gridblasHandle_t;
-#endif
-#ifdef GRID_CUDA
-  typedef cublasHandle_t gridblasHandle_t;
-#endif
-#ifdef GRID_SYCL
-  typedef cl::sycl::queue *gridblasHandle_t;
-#endif
-#ifdef GRID_ONE_MKL
-  typedef cl::sycl::queue *gridblasHandle_t;
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
-  typedef int32_t gridblasHandle_t;
-#endif
-
-enum GridBLASOperation_t { GridBLAS_OP_N, GridBLAS_OP_T, GridBLAS_OP_C } ;
-
-class GridBLAS {
-public:
-
-  
-  static gridblasHandle_t gridblasHandle;
-  static int            gridblasInit;
-  
-  static void Init(void)
-  {
-    if ( ! gridblasInit ) {
-#ifdef GRID_CUDA
-      std::cout << "cublasCreate"<<std::endl;
-      cublasCreate(&gridblasHandle);
-      cublasSetPointerMode(gridblasHandle, CUBLAS_POINTER_MODE_DEVICE);
-#endif
-#ifdef GRID_HIP
-      std::cout << "hipblasCreate"<<std::endl;
-      hipblasCreate(&gridblasHandle);
-#endif
-#ifdef GRID_SYCL
-      gridblasHandle = theGridAccelerator;
-#endif
-#ifdef GRID_ONE_MKL
-      cl::sycl::cpu_selector selector;
-      cl::sycl::device selectedDevice { selector };
-      gridblasHandle =new sycl::queue (selectedDevice);
-#endif
-      gridblasInit=1;
-    }
-  }
-  
-  // Force construct once
-  GridBLAS() { Init(); };
-  ~GridBLAS() { };
-  
-  /////////////////////////////////////////////////////////////////////////////////////
-  // BLAS GEMM conventions:
-  /////////////////////////////////////////////////////////////////////////////////////
-  // - C = alpha A * B + beta C
-  // Dimensions:
-  // - C_m.n
-  // - A_m.k
-  // - B_k.n
-  // - Flops = 8 M N K
-  // - Bytes = 2*sizeof(word) * (MN+MK+KN)
-  // M=60, N=12
-  // Flop/Byte = 8 . 60.60.12 / (60.12+60.60+60.12)/16 = 4 so expect about 4 TF/s on a GCD
-  /////////////////////////////////////////////////////////////////////////////////////
-  void synchronise(void)
-  {
-#ifdef GRID_HIP
-    auto err = hipDeviceSynchronize();
-    assert(err==hipSuccess);
-#endif
-#ifdef GRID_CUDA
-    auto err = cudaDeviceSynchronize();
-    assert(err==cudaSuccess);
-#endif
-#ifdef GRID_SYCL
-    accelerator_barrier();
-#endif
-#ifdef GRID_ONE_MKL
-    gridblasHandle->wait();
-#endif
-  }
-  
-  void gemmBatched(int m,int n, int k,
-		   ComplexD alpha,
-		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
-		   deviceVector<ComplexD*> &Bkn,
-		   ComplexD beta,
-		   deviceVector<ComplexD*> &Cmn)
-  {
-    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
-		m,n,k,
-		alpha,
-		Amk,
-		Bkn,
-		beta,
-		Cmn);
-  }
-  void gemmBatched(int m,int n, int k,
-		   ComplexF alpha,
-		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
-		   deviceVector<ComplexF*> &Bkn,
-		   ComplexF beta,
-		   deviceVector<ComplexF*> &Cmn)
-  {
-    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
-		m,n,k,
-		alpha,
-		Amk,
-		Bkn,
-		beta,
-		Cmn);
-  }
-  void gemmBatched(int m,int n, int k,
-		   RealD alpha,
-		   deviceVector<RealD*> &Amk,  // pointer list to matrices
-		   deviceVector<RealD*> &Bkn,
-		   RealD beta,
-		   deviceVector<RealD*> &Cmn)
-  {
-    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
-		m,n,k,
-		alpha,
-		Amk,
-		Bkn,
-		beta,
-		Cmn);
-  }
-  void gemmBatched(int m,int n, int k,
-		   RealF alpha,
-		   deviceVector<RealF*> &Amk,  // pointer list to matrices
-		   deviceVector<RealF*> &Bkn,
-		   RealF beta,
-		   deviceVector<RealF*> &Cmn)
-  {
-    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
-		m,n,k,
-		alpha,
-		Amk,
-		Bkn,
-		beta,
-		Cmn);
-  }
-
-  void gemmBatched(GridBLASOperation_t OpA,
-		   GridBLASOperation_t OpB,
-		   int m,int n, int k,
-		   ComplexD alpha,
-		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
-		   deviceVector<ComplexD*> &Bkn,
-		   ComplexD beta,
-		   deviceVector<ComplexD*> &Cmn)
-  {
-    RealD t2=usecond();
-    int32_t batchCount = Amk.size();
-    assert(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
-
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    if(OpA!=GridBLAS_OP_N)
-      lda = k;
-    if(OpB!=GridBLAS_OP_N)
-      ldb = n;
-    
-    static deviceVector<ComplexD> alpha_p(1);
-    static deviceVector<ComplexD> beta_p(1);
-    // can prestore the 1 and the zero on device
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
-    RealD t0=usecond();
-    //    std::cout << "ZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
-#ifdef GRID_HIP
-    hipblasOperation_t hOpA;
-    hipblasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
-    auto err = hipblasZgemmBatched(gridblasHandle,
-				   hOpA,
-				   hOpB,
-				   m,n,k,
-				   (hipblasDoubleComplex *) &alpha_p[0],
-				   (hipblasDoubleComplex **)&Amk[0], lda,
-				   (hipblasDoubleComplex **)&Bkn[0], ldb,
-				   (hipblasDoubleComplex *) &beta_p[0],
-				   (hipblasDoubleComplex **)&Cmn[0], ldc,
-				   batchCount);
-    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
-    assert(err==HIPBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_CUDA
-    cublasOperation_t hOpA;
-    cublasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
-    auto err = cublasZgemmBatched(gridblasHandle,
-				  hOpA,
-				  hOpB,
-				  m,n,k,
-				  (cuDoubleComplex *) &alpha_p[0],
-				  (cuDoubleComplex **)&Amk[0], lda,
-				  (cuDoubleComplex **)&Bkn[0], ldb,
-				  (cuDoubleComplex *) &beta_p[0],
-				  (cuDoubleComplex **)&Cmn[0], ldc,
-				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_SYCL
-    //MKL’s cblas_<T>gemm_batch & OneAPI
-#warning "oneMKL implementation not built "
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
-    // Need a default/reference implementation
-    int sda = lda*k;
-    int sdb = ldb*k;
-    int sdc = ldc*n;
-    for (int p = 0; p < batchCount; ++p) {
-      for (int mm = 0; mm < m; ++mm) {
-	for (int nn = 0; nn < n; ++nn) {
-	  ComplexD c_mn(0.0);
-	  for (int kk = 0; kk < k; ++kk)
-	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
-	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
-	}
-      }
-    }
-#endif
-    //    synchronise();
-     RealD t1=usecond();
-     RealD flops = 8.0*m*n*k*batchCount;
-     RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount;
-     //     std::cout <<GridLogMessage<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
-     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
-     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
-  }
-
-  void gemmBatched(GridBLASOperation_t OpA,
-		   GridBLASOperation_t OpB,
-		   int m,int n, int k,
-		   ComplexF alpha,
-		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
-		   deviceVector<ComplexF*> &Bkn,
-		   ComplexF beta,
-		   deviceVector<ComplexF*> &Cmn)
-  {
-    RealD t2=usecond();
-    int32_t batchCount = Amk.size();
-
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    if(OpA!=GridBLAS_OP_N)
-      lda = k;
-    if(OpB!=GridBLAS_OP_N)
-      ldb = n;
-    static deviceVector<ComplexF> alpha_p(1);
-    static deviceVector<ComplexF> beta_p(1);
-    // can prestore the 1 and the zero on device
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
-    RealD t0=usecond();
-
-    assert(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
-#ifdef GRID_HIP
-    hipblasOperation_t hOpA;
-    hipblasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
-    auto err = hipblasCgemmBatched(gridblasHandle,
-				   hOpA,
-				   hOpB,
-				   m,n,k,
-				   (hipblasComplex *) &alpha_p[0],
-				   (hipblasComplex **)&Amk[0], lda,
-				   (hipblasComplex **)&Bkn[0], ldb,
-				   (hipblasComplex *) &beta_p[0],
-				   (hipblasComplex **)&Cmn[0], ldc,
-				   batchCount);
-
-    assert(err==HIPBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_CUDA
-    cublasOperation_t hOpA;
-    cublasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
-    auto err = cublasCgemmBatched(gridblasHandle,
-				  hOpA,
-				  hOpB,
-				  m,n,k,
-				  (cuComplex *) &alpha_p[0],
-				  (cuComplex **)&Amk[0], lda,
-				  (cuComplex **)&Bkn[0], ldb,
-				  (cuComplex *) &beta_p[0],
-				  (cuComplex **)&Cmn[0], ldc,
-				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_SYCL
-    //MKL’s cblas_<T>gemm_batch & OneAPI
-#warning "oneMKL implementation not built "
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
-    int sda = lda*k;
-    int sdb = ldb*k;
-    int sdc = ldc*n;
-    ComplexF alphaf(real(alpha),imag(alpha));
-    ComplexF betaf(real(beta),imag(beta));
-    // Need a default/reference implementation
-    for (int p = 0; p < batchCount; ++p) {
-      for (int mm = 0; mm < m; ++mm) {
-	for (int nn = 0; nn < n; ++nn) {
-	  ComplexF c_mn(0.0);
-	  for (int kk = 0; kk < k; ++kk)
-	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
-	  Cmn[p][mm + nn*ldc] =  (alphaf)*c_mn + (betaf)*Cmn[p][mm + nn*ldc ];
-	}
-      }
-    }
-#endif
-     RealD t1=usecond();
-     RealD flops = 8.0*m*n*k*batchCount;
-     RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
-  }
-  
-  ///////////////////////////////////////////////////////////////////////////
-  // Single precision real GEMM
-  ///////////////////////////////////////////////////////////////////////////
-
-  void gemmBatched(GridBLASOperation_t OpA,
-		   GridBLASOperation_t OpB,
-		   int m,int n, int k,
-		   RealF alpha,
-		   deviceVector<RealF*> &Amk,  // pointer list to matrices
-		   deviceVector<RealF*> &Bkn,
-		   RealF beta,
-		   deviceVector<RealF*> &Cmn)
-  {
-    RealD t2=usecond();
-    int32_t batchCount = Amk.size();
-
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    if(OpA!=GridBLAS_OP_N)
-      lda = k;
-    if(OpB!=GridBLAS_OP_N)
-      ldb = n;
-    static deviceVector<RealF> alpha_p(1);
-    static deviceVector<RealF> beta_p(1);
-    // can prestore the 1 and the zero on device
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
-    RealD t0=usecond();
-
-    assert(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
-#ifdef GRID_HIP
-    hipblasOperation_t hOpA;
-    hipblasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
-    auto err = hipblasSgemmBatched(gridblasHandle,
-				   hOpA,
-				   hOpB,
-				   m,n,k,
-				   (float *) &alpha_p[0],
-				   (float **)&Amk[0], lda,
-				   (float **)&Bkn[0], ldb,
-				   (float *) &beta_p[0],
-				   (float **)&Cmn[0], ldc,
-				   batchCount);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_CUDA
-    cublasOperation_t hOpA;
-    cublasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
-    auto err = cublasSgemmBatched(gridblasHandle,
-				  hOpA,
-				  hOpB,
-				  m,n,k,
-				  (float *) &alpha_p[0],
-				  (float **)&Amk[0], lda,
-				  (float **)&Bkn[0], ldb,
-				  (float *) &beta_p[0],
-				  (float **)&Cmn[0], ldc,
-				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_SYCL
-    //MKL’s cblas_<T>gemm_batch & OneAPI
-#warning "oneMKL implementation not built "
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
-    int sda = lda*k;
-    int sdb = ldb*k;
-    int sdc = ldc*n;
-    // Need a default/reference implementation
-    for (int p = 0; p < batchCount; ++p) {
-      for (int mm = 0; mm < m; ++mm) {
-	for (int nn = 0; nn < n; ++nn) {
-	  RealD c_mn(0.0);
-	  for (int kk = 0; kk < k; ++kk)
-	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
-	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
-	}
-      }
-    }
-#endif
-     RealD t1=usecond();
-     RealD flops = 2.0*m*n*k*batchCount;
-     RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
-  }
-  
-  
-  ///////////////////////////////////////////////////////////////////////////
-  // Double precision real GEMM
-  ///////////////////////////////////////////////////////////////////////////
-
-  void gemmBatched(GridBLASOperation_t OpA,
-		   GridBLASOperation_t OpB,
-		   int m,int n, int k,
-		   RealD alpha,
-		   deviceVector<RealD*> &Amk,  // pointer list to matrices
-		   deviceVector<RealD*> &Bkn,
-		   RealD beta,
-		   deviceVector<RealD*> &Cmn)
-  {
-    RealD t2=usecond();
-    int32_t batchCount = Amk.size();
-
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    if(OpA!=GridBLAS_OP_N)
-      lda = k;
-    if(OpB!=GridBLAS_OP_N)
-      ldb = n;
-    
-    static deviceVector<RealD> alpha_p(1);
-    static deviceVector<RealD> beta_p(1);
-    // can prestore the 1 and the zero on device
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
-    RealD t0=usecond();
-
-    assert(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
-#ifdef GRID_HIP
-    hipblasOperation_t hOpA;
-    hipblasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
-    auto err = hipblasDgemmBatched(gridblasHandle,
-				   HIPBLAS_OP_N,
-				   HIPBLAS_OP_N,
-				   m,n,k,
-				   (double *) &alpha_p[0],
-				   (double **)&Amk[0], lda,
-				   (double **)&Bkn[0], ldb,
-				   (double *) &beta_p[0],
-				   (double **)&Cmn[0], ldc,
-				   batchCount);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_CUDA
-    cublasOperation_t hOpA;
-    cublasOperation_t hOpB;
-    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
-    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
-    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
-    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
-    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
-    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
-    auto err = cublasDgemmBatched(gridblasHandle,
-				  hOpA,
-				  hOpB,
-				  m,n,k,
-				  (double *) &alpha_p[0],
-				  (double **)&Amk[0], lda,
-				  (double **)&Bkn[0], ldb,
-				  (double *) &beta_p[0],
-				  (double **)&Cmn[0], ldc,
-				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_SYCL
-    /*
-      int64_t m64=m;
-      int64_t n64=n;
-      int64_t k64=k;
-      int64_t batchCount64=batchCount;
-      oneapi::mkl::blas::column_major::gemm_batch(*theGridAccelerator,
-      onemkl::transpose::N,
-      onemkl::transpose::N,
-      &m64,&n64,&k64,
-      (double *) &alpha_p[0],
-      (double **)&Amk[0], lda,
-      (double **)&Bkn[0], ldb,
-      (double *) &beta_p[0],
-      (double **)&Cmn[0], ldc,
-      1,&batchCount64);
-     */
-    //MKL’s cblas_<T>gemm_batch & OneAPI
-#warning "oneMKL implementation not built "
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
-    int sda = lda*k;
-    int sdb = ldb*k;
-    int sdc = ldc*n;
-    // Need a default/reference implementation
-    for (int p = 0; p < batchCount; ++p) {
-      for (int mm = 0; mm < m; ++mm) {
-	for (int nn = 0; nn < n; ++nn) {
-	  RealD c_mn(0.0);
-	  for (int kk = 0; kk < k; ++kk)
-	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
-	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
-	}
-      }
-    }
-#endif
-     RealD t1=usecond();
-     RealD flops = 2.0*m*n*k*batchCount;
-     RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
-  }
-  
-
-  
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // Strided case used by benchmark, but generally unused in Grid
-  // Keep a code example in double complex, but don't generate the single and real variants for now
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  
-  void gemmStridedBatched(int m,int n, int k,
-			  ComplexD alpha,
-			  ComplexD* Amk,  // pointer list to matrices
-			  ComplexD* Bkn,
-			  ComplexD beta,
-			  ComplexD* Cmn,
-			  int batchCount)
-  {
-    // Use C-row major storage, so transpose calls
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    int sda = m*k;
-    int sdb = k*n;
-    int sdc = m*n;
-    deviceVector<ComplexD> alpha_p(1);
-    deviceVector<ComplexD> beta_p(1);
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
-
-    //    std::cout << "blasZgemmStridedBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
-    //    std::cout << "blasZgemmStridedBatched ld   "<<lda<<","<<ldb<<","<<ldc<<std::endl;
-    //    std::cout << "blasZgemmStridedBatched sd   "<<sda<<","<<sdb<<","<<sdc<<std::endl;
-#ifdef GRID_HIP
-    auto err = hipblasZgemmStridedBatched(gridblasHandle,
-					  HIPBLAS_OP_N,
-					  HIPBLAS_OP_N,
-					  m,n,k,
-					  (hipblasDoubleComplex *) &alpha_p[0],
-					  (hipblasDoubleComplex *) Amk, lda, sda,
-					  (hipblasDoubleComplex *) Bkn, ldb, sdb,
-					  (hipblasDoubleComplex *) &beta_p[0],
-					  (hipblasDoubleComplex *) Cmn, ldc, sdc,
-					  batchCount);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_CUDA
-    cublasZgemmStridedBatched(gridblasHandle,
-			      CUBLAS_OP_N,
-			      CUBLAS_OP_N,
-			      m,n,k,
-			      (cuDoubleComplex *) &alpha_p[0],
-			      (cuDoubleComplex *) Amk, lda, sda,
-			      (cuDoubleComplex *) Bkn, ldb, sdb,
-			      (cuDoubleComplex *) &beta_p[0],
-			      (cuDoubleComplex *) Cmn, ldc, sdc,
-			      batchCount);
-#endif
-#if defined(GRID_SYCL) || defined(GRID_ONE_MKL)
-    oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
-						oneapi::mkl::transpose::N,
-						oneapi::mkl::transpose::N,
-						m,n,k,
-						alpha,
-						(const ComplexD *)Amk,lda,sda,
-						(const ComplexD *)Bkn,ldb,sdb,
-						beta,
-						(ComplexD *)Cmn,ldc,sdc,
-						batchCount);
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
-     // Need a default/reference implementation
-     for (int p = 0; p < batchCount; ++p) {
-       for (int mm = 0; mm < m; ++mm) {
-	 for (int nn = 0; nn < n; ++nn) {
-	   ComplexD c_mn(0.0);
-	   for (int kk = 0; kk < k; ++kk)
-	     c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
-	   Cmn[mm + nn*ldc + p*sdc] =  (alpha)*c_mn + (beta)*Cmn[mm + nn*ldc + p*sdc];
-	 }
-       }
-     }
-#endif
-  }
-
-  double benchmark(int M, int N, int K, int BATCH)
-  {
-    int32_t N_A = M*K*BATCH;
-    int32_t N_B = K*N*BATCH;
-    int32_t N_C = M*N*BATCH;
-    deviceVector<ComplexD> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(ComplexD));
-    deviceVector<ComplexD> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(ComplexD));
-    deviceVector<ComplexD> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(ComplexD));
-    ComplexD alpha(1.0);
-    ComplexD beta (1.0);
-    RealD flops = 8.0*M*N*K*BATCH;
-    int ncall=10;
-    RealD t0 = usecond();
-    for(int i=0;i<ncall;i++){
-      gemmStridedBatched(M,N,K,
-			 alpha,
-			 &A[0], // m x k 
-			 &B[0], // k x n
-			 beta, 
-			 &C[0], // m x n
-			 BATCH);
-    }
-    synchronise();
-    RealD t1 = usecond();
-    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K)*BATCH;
-    flops = 8.0*M*N*K*BATCH*ncall;
-    flops = flops/(t1-t0)/1.e3;
-    return flops; // Returns gigaflops
-  }
-
-
-
-
-};
-
-NAMESPACE_END(Grid);

Grid/algorithms/iterative/BiCGSTAB.h

@@ -1,234 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/algorithms/iterative/BiCGSTAB.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>
-Author: juettner <juettner@soton.ac.uk>
-Author: David Murphy <djmurphy@mit.edu>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#ifndef GRID_BICGSTAB_H
-#define GRID_BICGSTAB_H
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////
-// Base classes for iterative processes based on operators
-// single input vec, single output vec.
-/////////////////////////////////////////////////////////////
-
-template <class Field>
-class BiCGSTAB : 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
-  
-    BiCGSTAB(RealD tol, Integer maxit, bool err_on_no_conv = true) : 
-      Tolerance(tol), MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv){};
-
-    void operator()(LinearOperatorBase<Field>& Linop, const Field& src, Field& psi) 
-    {
-      psi.Checkerboard() = src.Checkerboard();
-      conformable(psi, src);
-
-      RealD cp(0), rho(1), rho_prev(0), alpha(1), beta(0), omega(1);
-      RealD a(0), bo(0), b(0), ssq(0);
-
-      Field p(src);
-      Field r(src);
-      Field rhat(src);
-      Field v(src);
-      Field s(src);
-      Field t(src);
-      Field h(src);
-
-      v = Zero();
-      p = Zero();
-
-      // Initial residual computation & set up
-      RealD guess = norm2(psi);
-      assert(std::isnan(guess) == 0);
-    
-      Linop.Op(psi, v);
-      b = norm2(v);
-
-      r = src - v;
-      rhat = r;
-      a = norm2(r);
-      ssq = norm2(src);
-
-      std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: guess " << guess << std::endl;
-      std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB:   src " << ssq << std::endl;
-      std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB:    mp " << b << std::endl;
-      std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB:     r " << a << std::endl;
-
-      RealD rsq = Tolerance * Tolerance * ssq;
-
-      // Check if guess is really REALLY good :)
-      if(a <= rsq){ return; }
-
-      std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: k=0 residual " << a << " target " << rsq << std::endl;
-
-      GridStopWatch LinalgTimer;
-      GridStopWatch InnerTimer;
-      GridStopWatch AxpyNormTimer;
-      GridStopWatch LinearCombTimer;
-      GridStopWatch MatrixTimer;
-      GridStopWatch SolverTimer;
-
-      SolverTimer.Start();
-      int k;
-      for (k = 1; k <= MaxIterations; k++) 
-      {
-        rho_prev = rho;
-
-        LinalgTimer.Start();
-        InnerTimer.Start();
-        ComplexD Crho  = innerProduct(rhat,r);
-        InnerTimer.Stop();
-        rho = Crho.real();
-
-        beta = (rho / rho_prev) * (alpha / omega);
-
-        LinearCombTimer.Start();
-        bo = beta * omega;
-	{
-	  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();
-
-        MatrixTimer.Start();
-        Linop.Op(p,v);
-        MatrixTimer.Stop();
-
-        LinalgTimer.Start();
-        InnerTimer.Start();
-        ComplexD Calpha = innerProduct(rhat,v);
-        InnerTimer.Stop();
-        alpha = rho / Calpha.real();
-
-        LinearCombTimer.Start();
-	{
-	  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();
-
-        MatrixTimer.Start();
-        Linop.Op(s,t);
-        MatrixTimer.Stop();
-
-        LinalgTimer.Start();
-        InnerTimer.Start();
-        ComplexD Comega = innerProduct(t,s);
-        InnerTimer.Stop();
-        omega = Comega.real() / norm2(t);
-
-        LinearCombTimer.Start();
-	{
-	  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();
-
-        std::cout << GridLogIterative << "BiCGSTAB: Iteration " << k << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
-
-        // Stopping condition
-        if(cp <= rsq) 
-        {
-          SolverTimer.Stop();
-          Linop.Op(psi, v);
-          p = v - src;
-
-          RealD srcnorm = sqrt(norm2(src));
-          RealD resnorm = sqrt(norm2(p));
-          RealD true_residual = resnorm / srcnorm;
-
-          std::cout << GridLogMessage << "BiCGSTAB Converged on iteration " << k << std::endl;
-          std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp/ssq) << std::endl;
-          std::cout << GridLogMessage << "\tTrue residual " << true_residual << std::endl;
-          std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
-
-          std::cout << GridLogMessage << "Time breakdown " << std::endl;
-          std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() << std::endl;
-          std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() << std::endl;
-          std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() << std::endl;
-          std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() << std::endl;
-          std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() << std::endl;
-          std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() << std::endl;
-
-          if(ErrorOnNoConverge){ assert(true_residual / Tolerance < 10000.0); }
-
-          IterationsToComplete = k;	
-
-          return;
-        }
-      }
-      
-      std::cout << GridLogMessage << "BiCGSTAB did NOT converge" << std::endl;
-
-      if(ErrorOnNoConverge){ assert(0); }
-      IterationsToComplete = k;
-    }
-};
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/iterative/BiCGSTABMixedPrec.h

@@ -1,159 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid 
-
-Source file: ./lib/algorithms/iterative/BiCGSTABMixedPrec.h
-
-Copyright (C) 2015
-
-Author: Christopher Kelly <ckelly@phys.columbia.edu>
-Author: David Murphy <djmurphy@mit.edu>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#ifndef GRID_BICGSTAB_MIXED_PREC_H
-#define GRID_BICGSTAB_MIXED_PREC_H
-
-NAMESPACE_BEGIN(Grid);
-
-// Mixed precision restarted defect correction BiCGSTAB
-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:
-    using LinearFunction<FieldD>::operator();
-    RealD   Tolerance;
-    RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
-    Integer MaxInnerIterations;
-    Integer MaxOuterIterations;
-    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;
-
-    Integer TotalInnerIterations; //Number of inner CG iterations
-    Integer TotalOuterIterations; //Number of restarts
-    Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
-
-    //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
-    LinearFunction<FieldF> *guesser;
-    
-    MixedPrecisionBiCGSTAB(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, 
-        LinearOperatorBase<FieldF>& _Linop_f, LinearOperatorBase<FieldD>& _Linop_d) : 
-      Linop_f(_Linop_f), Linop_d(_Linop_d), Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), 
-      MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid), OuterLoopNormMult(100.), guesser(NULL) {};
-
-    void useGuesser(LinearFunction<FieldF>& g){
-      guesser = &g;
-    }
-  
-    void operator() (const FieldD& src_d_in, FieldD& sol_d)
-    {
-      TotalInnerIterations = 0;
-    
-      GridStopWatch TotalTimer;
-      TotalTimer.Start();
-      
-      int cb = src_d_in.Checkerboard();
-      sol_d.Checkerboard() = cb;
-      
-      RealD src_norm = norm2(src_d_in);
-      RealD stop = src_norm * Tolerance*Tolerance;
-
-      GridBase* DoublePrecGrid = src_d_in.Grid();
-      FieldD tmp_d(DoublePrecGrid);
-      tmp_d.Checkerboard() = cb;
-      
-      FieldD tmp2_d(DoublePrecGrid);
-      tmp2_d.Checkerboard() = cb;
-      
-      FieldD src_d(DoublePrecGrid);
-      src_d = src_d_in; //source for next inner iteration, computed from residual during operation
-      
-      RealD inner_tol = InnerTolerance;
-      
-      FieldF src_f(SinglePrecGrid);
-      src_f.Checkerboard() = cb;
-      
-      FieldF sol_f(SinglePrecGrid);
-      sol_f.Checkerboard() = cb;
-      
-      BiCGSTAB<FieldF> CG_f(inner_tol, MaxInnerIterations);
-      CG_f.ErrorOnNoConverge = false;
-
-      GridStopWatch InnerCGtimer;
-
-      GridStopWatch PrecChangeTimer;
-      
-      Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
-        
-      for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++)
-      {
-        // Compute double precision rsd and also new RHS vector.
-        Linop_d.Op(sol_d, tmp_d);
-        RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
-        
-        std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Outer iteration " << outer_iter << " residual " << norm << " target " << stop << std::endl;
-
-        if(norm < OuterLoopNormMult * stop){
-          std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: 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) ??
-
-        PrecChangeTimer.Start();
-        precisionChange(src_f, src_d);
-        PrecChangeTimer.Stop();
-        
-        sol_f = Zero();
-
-        //Optionally improve inner solver guess (eg using known eigenvectors)
-        if(guesser != NULL){ (*guesser)(src_f, sol_f); }
-
-        //Inner CG
-        CG_f.Tolerance = inner_tol;
-        InnerCGtimer.Start();
-        CG_f(Linop_f, src_f, sol_f);
-        InnerCGtimer.Stop();
-        TotalInnerIterations += CG_f.IterationsToComplete;
-        
-        //Convert sol back to double and add to double prec solution
-        PrecChangeTimer.Start();
-        precisionChange(tmp_d, sol_f);
-        PrecChangeTimer.Stop();
-        
-        axpy(sol_d, 1.0, tmp_d, sol_d);
-      }
-      
-      //Final trial CG
-      std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Starting final patch-up double-precision solve" << std::endl;
-      
-      BiCGSTAB<FieldD> CG_d(Tolerance, MaxInnerIterations);
-      CG_d(Linop_d, src_d_in, sol_d);
-      TotalFinalStepIterations = CG_d.IterationsToComplete;
-
-      TotalTimer.Stop();
-      std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
-      std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
-  }
-};
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -1,698 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/algorithms/iterative/BlockConjugateGradient.h
-
-Copyright (C) 2017
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
-
-//////////////////////////////////////////////////////////////////////////
-// Block conjugate gradient. Dimension zero should be the block direction
-//////////////////////////////////////////////////////////////////////////
-template <class Field>
-class BlockConjugateGradient : public OperatorFunction<Field> {
- public:
-
-  typedef typename Field::scalar_type scomplex;
-
-  int blockDim ;
-  int Nblock;
-
-  BlockCGtype CGtype;
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
-                           // Defaults true.
-  RealD Tolerance;
-  Integer MaxIterations;
-  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
-  Integer PrintInterval; //GridLogMessages or Iterative
-  RealD TrueResidual;
-  
-  BlockConjugateGradient(BlockCGtype cgtype,int _Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : Tolerance(tol), CGtype(cgtype),   blockDim(_Orthog),  MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv),PrintInterval(100)
-  {};
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Thin QR factorisation (google it)
-////////////////////////////////////////////////////////////////////////////////////////////////////
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  //Dimensions
-  // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
-  //
-  // Rdag R = m_rr = Herm = L L^dag        <-- Cholesky decomposition (LLT routine in Eigen)
-  //
-  //   Q  C = R => Q = R C^{-1}
-  //
-  // Want  Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock} 
-  //
-  // Set C = L^{dag}, and then Q^dag Q = ident 
-  //
-  // Checks:
-  // Cdag C = Rdag R ; passes.
-  // QdagQ  = 1      ; passes
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-void ThinQRfact (Eigen::MatrixXcd &m_rr,
-		 Eigen::MatrixXcd &C,
-		 Eigen::MatrixXcd &Cinv,
-		 Field & Q,
-		 const Field & R)
-{
-  int Orthog = blockDim; // First dimension is block dim; this is an assumption
-  sliceInnerProductMatrix(m_rr,R,R,Orthog);
-
-  // Force manifest hermitian to avoid rounding related
-  m_rr = 0.5*(m_rr+m_rr.adjoint());
-
-  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
-
-  C    = L.adjoint();
-  Cinv = C.inverse();
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Q = R C^{-1}
-  //
-  // Q_j  = R_i Cinv(i,j) 
-  //
-  // NB maddMatrix conventions are Right multiplication X[j] a[j,i] already
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  sliceMulMatrix(Q,Cinv,R,Orthog);
-}
-// see comments above
-void ThinQRfact (Eigen::MatrixXcd &m_rr,
-		 Eigen::MatrixXcd &C,
-		 Eigen::MatrixXcd &Cinv,
-		 std::vector<Field> & Q,
-		 const std::vector<Field> & R)
-{
-  InnerProductMatrix(m_rr,R,R);
-
-  m_rr = 0.5*(m_rr+m_rr.adjoint());
-
-  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
-
-  C    = L.adjoint();
-  Cinv = C.inverse();
-
-  MulMatrix(Q,Cinv,R);
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Call one of several implementations
-////////////////////////////////////////////////////////////////////////////////////////////////////
-void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
-{
-  if ( CGtype == BlockCGrQ ) {
-    BlockCGrQsolve(Linop,Src,Psi);
-  } else if (CGtype == CGmultiRHS ) {
-    CGmultiRHSsolve(Linop,Src,Psi);
-  } else {
-    assert(0);
-  }
-}
-virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Field> &Src, std::vector<Field> &Psi) 
-{
-  if ( CGtype == BlockCGrQVec ) {
-    BlockCGrQsolveVec(Linop,Src,Psi);
-  } else {
-    assert(0);
-  }
-}
-
-////////////////////////////////////////////////////////////////////////////
-// BlockCGrQ implementation:
-//--------------------------
-// X is guess/Solution
-// B is RHS
-// Solve A X_i = B_i    ;        i refers to Nblock index
-////////////////////////////////////////////////////////////////////////////
-void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X) 
-{
-  int Orthog = blockDim; // First dimension is block dim; this is an assumption
-  Nblock = B.Grid()->_fdimensions[Orthog];
-/* FAKE */
-  Nblock=8;
-  std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
-
-  X.Checkerboard() = B.Checkerboard();
-  conformable(X, B);
-
-  Field tmp(B);
-  Field Q(B);
-  Field D(B);
-  Field Z(B);
-  Field AD(B);
-
-  Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-
-  // Initial residual computation & set up
-  std::vector<RealD> residuals(Nblock);
-  std::vector<RealD> ssq(Nblock);
-
-  sliceNorm(ssq,B,Orthog);
-  RealD sssum=0;
-  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
-
-  sliceNorm(residuals,B,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  sliceNorm(residuals,X,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  /************************************************************************
-   * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
-   ************************************************************************
-   * Dimensions:
-   *
-   *   X,B==(Nferm x Nblock)
-   *   A==(Nferm x Nferm)
-   *  
-   * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
-   * 
-   * QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
-   * for k: 
-   *   Z  = AD
-   *   M  = [D^dag Z]^{-1}
-   *   X  = X + D MC
-   *   QS = Q - ZM
-   *   D  = Q + D S^dag
-   *   C  = S C
-   */
-  ///////////////////////////////////////
-  // Initial block: initial search dir is guess
-  ///////////////////////////////////////
-  std::cout << GridLogMessage<<"BlockCGrQ algorithm initialisation " <<std::endl;
-
-  //1.  QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
-  Linop.HermOp(X, AD);
-  tmp = B - AD;  
-
-  ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
-  D=Q;
-
-  std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
-
-  ///////////////////////////////////////
-  // Timers
-  ///////////////////////////////////////
-  GridStopWatch sliceInnerTimer;
-  GridStopWatch sliceMaddTimer;
-  GridStopWatch QRTimer;
-  GridStopWatch MatrixTimer;
-  GridStopWatch SolverTimer;
-  SolverTimer.Start();
-
-  int k;
-  for (k = 1; k <= MaxIterations; k++) {
-
-    //3. Z  = AD
-    MatrixTimer.Start();
-    Linop.HermOp(D, Z);      
-    MatrixTimer.Stop();
-
-    //4. M  = [D^dag Z]^{-1}
-    sliceInnerTimer.Start();
-    sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
-    sliceInnerTimer.Stop();
-    m_M       = m_DZ.inverse();
-    
-    //5. X  = X + D MC
-    m_tmp     = m_M * m_C;
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(X,m_tmp, D,X,Orthog);     
-    sliceMaddTimer.Stop();
-
-    //6. QS = Q - ZM
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(tmp,m_M,Z,Q,Orthog,-1.0);
-    sliceMaddTimer.Stop();
-    QRTimer.Start();
-    ThinQRfact (m_rr, m_S, m_Sinv, Q, tmp);
-    QRTimer.Stop();
-    
-    //7. D  = Q + D S^dag
-    m_tmp = m_S.adjoint();
-
-    sliceMaddTimer.Start();
-    sliceMaddMatrix(D,m_tmp,D,Q,Orthog);
-    sliceMaddTimer.Stop();
-
-    //8. C  = S C
-    m_C = m_S*m_C;
-    
-    /*********************
-     * convergence monitor
-     *********************
-     */
-    m_rr = m_C.adjoint() * m_C;
-
-    RealD max_resid=0;
-    RealD rrsum=0;
-    RealD rr;
-
-    for(int b=0;b<Nblock;b++) {
-      rrsum+=real(m_rr(b,b));
-      rr = real(m_rr(b,b))/ssq[b];
-      if ( rr > max_resid ) max_resid = rr;
-    }
-
-    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
-	      <<" ave "<<std::sqrt(rrsum/sssum) << " max "<< max_resid <<std::endl;
-
-    if ( max_resid < Tolerance*Tolerance ) { 
-
-      SolverTimer.Stop();
-
-      std::cout << GridLogMessage<<"BlockCGrQ converged in "<<k<<" iterations"<<std::endl;
-
-      for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "
-		  << std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
-      }
-      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
-
-      Linop.HermOp(X, AD);
-      AD = AD-B;
-      TrueResidual = std::sqrt(norm2(AD)/norm2(B));
-      std::cout << GridLogMessage <<"\tTrue residual is " << TrueResidual <<std::endl;
-
-      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
-      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
-      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
-      std::cout << GridLogMessage << "\tThinQRfact " << QRTimer.Elapsed()  <<std::endl;
-	    
-      IterationsToComplete = k;
-      return;
-    }
-
-  }
-  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
-
-  if (ErrorOnNoConverge) assert(0);
-  IterationsToComplete = k;
-}
-//////////////////////////////////////////////////////////////////////////
-// multiRHS conjugate gradient. Dimension zero should be the block direction
-// Use this for spread out across nodes
-//////////////////////////////////////////////////////////////////////////
-void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
-{
-  int Orthog = blockDim; // First dimension is block dim
-  Nblock = Src.Grid()->_fdimensions[Orthog];
-
-  std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
-
-  Psi.Checkerboard() = Src.Checkerboard();
-  conformable(Psi, Src);
-
-  Field P(Src);
-  Field AP(Src);
-  Field R(Src);
-  
-  std::vector<ComplexD> v_pAp(Nblock);
-  std::vector<RealD> v_rr (Nblock);
-  std::vector<RealD> v_rr_inv(Nblock);
-  std::vector<RealD> v_alpha(Nblock);
-  std::vector<RealD> v_beta(Nblock);
-
-  // Initial residual computation & set up
-  std::vector<RealD> residuals(Nblock);
-  std::vector<RealD> ssq(Nblock);
-
-  sliceNorm(ssq,Src,Orthog);
-  RealD sssum=0;
-  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
-
-  sliceNorm(residuals,Src,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  sliceNorm(residuals,Psi,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  // Initial search dir is guess
-  Linop.HermOp(Psi, AP);
-
-  R = Src - AP;  
-  P = R;
-  sliceNorm(v_rr,R,Orthog);
-
-  GridStopWatch sliceInnerTimer;
-  GridStopWatch sliceMaddTimer;
-  GridStopWatch sliceNormTimer;
-  GridStopWatch MatrixTimer;
-  GridStopWatch SolverTimer;
-
-  SolverTimer.Start();
-  int k;
-  for (k = 1; k <= MaxIterations; k++) {
-
-    RealD rrsum=0;
-    for(int b=0;b<Nblock;b++) rrsum+=real(v_rr[b]);
-
-    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
-	      <<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
-
-    MatrixTimer.Start();
-    Linop.HermOp(P, AP);
-    MatrixTimer.Stop();
-
-    // Alpha
-    sliceInnerTimer.Start();
-    sliceInnerProductVector(v_pAp,P,AP,Orthog);
-    sliceInnerTimer.Stop();
-    for(int b=0;b<Nblock;b++){
-      v_alpha[b] = v_rr[b]/real(v_pAp[b]);
-    }
-
-    // Psi, R update
-    sliceMaddTimer.Start();
-    sliceMaddVector(Psi,v_alpha, P,Psi,Orthog);     // add alpha *  P to psi
-    sliceMaddVector(R  ,v_alpha,AP,  R,Orthog,-1.0);// sub alpha * AP to resid
-    sliceMaddTimer.Stop();
-
-    // Beta
-    for(int b=0;b<Nblock;b++){
-      v_rr_inv[b] = 1.0/v_rr[b];
-    }
-    sliceNormTimer.Start();
-    sliceNorm(v_rr,R,Orthog);
-    sliceNormTimer.Stop();
-    for(int b=0;b<Nblock;b++){
-      v_beta[b] = v_rr_inv[b] *v_rr[b];
-    }
-
-    // Search update
-    sliceMaddTimer.Start();
-    sliceMaddVector(P,v_beta,P,R,Orthog);
-    sliceMaddTimer.Stop();
-
-    /*********************
-     * convergence monitor
-     *********************
-     */
-    RealD max_resid=0;
-    for(int b=0;b<Nblock;b++){
-      RealD rr = v_rr[b]/ssq[b];
-      if ( rr > max_resid ) max_resid = rr;
-    }
-    
-    if ( max_resid < Tolerance*Tolerance ) { 
-
-      SolverTimer.Stop();
-
-      std::cout << GridLogMessage<<"MultiRHS solver converged in " <<k<<" iterations"<<std::endl;
-      for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tBlock "<<b<<" computed resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
-      }
-      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
-
-      Linop.HermOp(Psi, AP);
-      AP = AP-Src;
-      TrueResidual = std::sqrt(norm2(AP)/norm2(Src));
-      std::cout <<GridLogMessage << "\tTrue residual is " << TrueResidual <<std::endl;
-
-      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
-      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
-      std::cout << GridLogMessage << "\tNorm       " << sliceNormTimer.Elapsed() <<std::endl;
-      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
-
-
-      IterationsToComplete = k;
-      return;
-    }
-
-  }
-  std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
-
-  if (ErrorOnNoConverge) assert(0);
-  IterationsToComplete = k;
-}
-
-void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
-  for(int b=0;b<Nblock;b++){
-  for(int bp=0;bp<Nblock;bp++) {
-    m(b,bp) = innerProduct(X[b],Y[bp]);  
-  }}
-}
-void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
-  // Should make this cache friendly with site outermost, parallel_for
-  // Deal with case AP aliases with either Y or X
-  std::vector<Field> tmp(Nblock,X[0]);
-  for(int b=0;b<Nblock;b++){
-    tmp[b]   = Y[b];
-    for(int bp=0;bp<Nblock;bp++) {
-      tmp[b] = tmp[b] + scomplex(scale*m(bp,b))*X[bp]; 
-    }
-  }
-  for(int b=0;b<Nblock;b++){
-    AP[b] = tmp[b];
-  }
-}
-void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
-  // Should make this cache friendly with site outermost, parallel_for
-  for(int b=0;b<Nblock;b++){
-    AP[b] = Zero();
-    for(int bp=0;bp<Nblock;bp++) {
-      AP[b] += scomplex(m(bp,b))*X[bp]; 
-    }
-  }
-}
-double normv(const std::vector<Field> &P){
-  double nn = 0.0;
-  for(int b=0;b<Nblock;b++) {
-    nn+=norm2(P[b]);
-  }
-  return nn;
-}
-
-////////////////////////////////////////////////////////////////////////////
-// BlockCGrQvec implementation:
-//--------------------------
-// X is guess/Solution
-// B is RHS
-// Solve A X_i = B_i    ;        i refers to Nblock index
-////////////////////////////////////////////////////////////////////////////
-void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field> &B, std::vector<Field> &X) 
-{
-  Nblock = B.size();
-  assert(Nblock == X.size());
-
-  std::cout<<GridLogMessage<<" Block Conjugate Gradient Vec rQ : Nblock "<<Nblock<<std::endl;
-
-  for(int b=0;b<Nblock;b++){ 
-    X[b].Checkerboard() = B[b].Checkerboard();
-    conformable(X[b], B[b]);
-    conformable(X[b], X[0]); 
-  }
-
-  Field Fake(B[0]);
-
-  std::vector<Field> tmp(Nblock,Fake);
-  std::vector<Field>   Q(Nblock,Fake);
-  std::vector<Field>   D(Nblock,Fake);
-  std::vector<Field>   Z(Nblock,Fake);
-  std::vector<Field>  AD(Nblock,Fake);
-
-  Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-  Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-  Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(Nblock,Nblock);
-
-  // Initial residual computation & set up
-  std::vector<RealD> residuals(Nblock);
-  std::vector<RealD> ssq(Nblock);
-
-  RealD sssum=0;
-  for(int b=0;b<Nblock;b++){ ssq[b] = norm2(B[b]);}
-  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
-
-  for(int b=0;b<Nblock;b++){ residuals[b] = norm2(B[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  for(int b=0;b<Nblock;b++){ residuals[b] = norm2(X[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
-
-  /************************************************************************
-   * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
-   ************************************************************************
-   * Dimensions:
-   *
-   *   X,B==(Nferm x Nblock)
-   *   A==(Nferm x Nferm)
-   *  
-   * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
-   * 
-   * QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
-   * for k: 
-   *   Z  = AD
-   *   M  = [D^dag Z]^{-1}
-   *   X  = X + D MC
-   *   QS = Q - ZM
-   *   D  = Q + D S^dag
-   *   C  = S C
-   */
-  ///////////////////////////////////////
-  // Initial block: initial search dir is guess
-  ///////////////////////////////////////
-  std::cout << GridLogMessage<<"BlockCGrQvec algorithm initialisation " <<std::endl;
-
-  //1.  QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
-  for(int b=0;b<Nblock;b++) {
-    Linop.HermOp(X[b], AD[b]);
-    tmp[b] = B[b] - AD[b];  
-  }
-
-  ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
-
-  for(int b=0;b<Nblock;b++) D[b]=Q[b];
-
-  std::cout << GridLogMessage<<"BlockCGrQ vec computed initial residual and QR fact " <<std::endl;
-
-  ///////////////////////////////////////
-  // Timers
-  ///////////////////////////////////////
-  GridStopWatch sliceInnerTimer;
-  GridStopWatch sliceMaddTimer;
-  GridStopWatch QRTimer;
-  GridStopWatch MatrixTimer;
-  GridStopWatch SolverTimer;
-  SolverTimer.Start();
-
-  int k;
-  for (k = 1; k <= MaxIterations; k++) {
-
-    //3. Z  = AD
-    MatrixTimer.Start();
-    for(int b=0;b<Nblock;b++) Linop.HermOp(D[b], Z[b]);      
-    MatrixTimer.Stop();
-
-    //4. M  = [D^dag Z]^{-1}
-    sliceInnerTimer.Start();
-    InnerProductMatrix(m_DZ,D,Z);
-    sliceInnerTimer.Stop();
-    m_M       = m_DZ.inverse();
-    
-    //5. X  = X + D MC
-    m_tmp     = m_M * m_C;
-    sliceMaddTimer.Start();
-    MaddMatrix(X,m_tmp, D,X);     
-    sliceMaddTimer.Stop();
-
-    //6. QS = Q - ZM
-    sliceMaddTimer.Start();
-    MaddMatrix(tmp,m_M,Z,Q,-1.0);
-    sliceMaddTimer.Stop();
-    QRTimer.Start();
-    ThinQRfact (m_rr, m_S, m_Sinv, Q, tmp);
-    QRTimer.Stop();
-    
-    //7. D  = Q + D S^dag
-    m_tmp = m_S.adjoint();
-    sliceMaddTimer.Start();
-    MaddMatrix(D,m_tmp,D,Q);
-    sliceMaddTimer.Stop();
-
-    //8. C  = S C
-    m_C = m_S*m_C;
-    
-    /*********************
-     * convergence monitor
-     *********************
-     */
-    m_rr = m_C.adjoint() * m_C;
-
-    RealD max_resid=0;
-    RealD rrsum=0;
-    RealD rr;
-
-    for(int b=0;b<Nblock;b++) {
-      rrsum+=real(m_rr(b,b));
-      rr = real(m_rr(b,b))/ssq[b];
-      if ( rr > max_resid ) max_resid = rr;
-    }
-
-    std::cout << GridLogIterative << "\t Block Iteration "<<k<<" ave resid "<< std::sqrt(rrsum/sssum) << " max "<< std::sqrt(max_resid) <<std::endl;
-
-    if ( max_resid < Tolerance*Tolerance ) { 
-
-      SolverTimer.Stop();
-
-      std::cout << GridLogMessage<<"BlockCGrQ converged in "<<k<<" iterations"<<std::endl;
-
-      for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
-      }
-      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
-
-      for(int b=0;b<Nblock;b++) Linop.HermOp(X[b], AD[b]);
-      for(int b=0;b<Nblock;b++) AD[b] = AD[b]-B[b];
-      TrueResidual = std::sqrt(normv(AD)/normv(B));
-      std::cout << GridLogMessage << "\tTrue residual is " << TrueResidual <<std::endl;
-
-      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
-      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
-      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
-      std::cout << GridLogMessage << "\tThinQRfact " << QRTimer.Elapsed()  <<std::endl;
-	    
-      IterationsToComplete = k;
-      return;
-    }
-
-  }
-  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
-
-  if (ErrorOnNoConverge) assert(0);
-  IterationsToComplete = k;
-}
-
-};
-
-NAMESPACE_END(Grid);
-

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

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

Grid/algorithms/iterative/ConjugateGradient.h

@@ -1,218 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/algorithms/iterative/ConjugateGradient.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_H
-#define GRID_CONJUGATE_GRADIENT_H
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////
-// Base classes for iterative processes based on operators
-// 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.
-                           // Defaults true.
-  RealD Tolerance;
-  Integer MaxIterations;
-  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
-  RealD TrueResidual;
-  
-  ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : Tolerance(tol),
-      MaxIterations(maxit),
-      ErrorOnNoConverge(err_on_no_conv){};
-
-  void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
-
-    GRID_TRACE("ConjugateGradient");
-    psi.Checkerboard() = src.Checkerboard();
-
-    conformable(psi, src);
-
-    RealD cp, c, a, d, b, ssq, qq;
-    //RealD b_pred;
-
-    Field p(src);
-    Field mmp(src);
-    Field r(src);
-
-    // 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;
-    ssq = norm2(src);
-
-    // 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;
-
-    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);
-      MatrixTimer.Stop();
-
-      LinalgTimer.Start();
-
-      InnerTimer.Start();
-      ComplexD dc  = innerProduct(p,mmp);
-      InnerTimer.Stop();
-      d = dc.real();
-      a = c / d;
-
-      AxpyNormTimer.Start();
-      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();
-
-      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 << GridLogMessage << "Time breakdown "<<std::endl;
-	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\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);
-
-    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
-
-    if (ErrorOnNoConverge) assert(0);
-    IterationsToComplete = k;
-
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -1,170 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrec.h
-
-    Copyright (C) 2015
-
-Author: Christopher Kelly <ckelly@phys.columbia.edu>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
-#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_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 MixedPrecisionConjugateGradient : 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;
-    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;
-
-    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;
-    
-    MixedPrecisionConjugateGradient(RealD tol, 
-				    Integer maxinnerit, 
-				    Integer maxouterit, 
-				    GridBase* _sp_grid, 
-				    LinearOperatorBase<FieldF> &_Linop_f, 
-				    LinearOperatorBase<FieldD> &_Linop_d) :
-      Linop_f(_Linop_f), Linop_d(_Linop_d),
-      Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
-      OuterLoopNormMult(100.), guesser(NULL){ };
-
-    void useGuesser(LinearFunction<FieldF> &g){
-      guesser = &g;
-    }
-  
-  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;
-    TotalTimer.Start();
-    
-    int cb = src_d_in.Checkerboard();
-    sol_d.Checkerboard() = cb;
-    
-    RealD src_norm = norm2(src_d_in);
-    RealD stop = src_norm * Tolerance*Tolerance;
-
-    GridBase* DoublePrecGrid = src_d_in.Grid();
-    FieldD tmp_d(DoublePrecGrid);
-    tmp_d.Checkerboard() = cb;
-    
-    FieldD tmp2_d(DoublePrecGrid);
-    tmp2_d.Checkerboard() = cb;
-    
-    FieldD src_d(DoublePrecGrid);
-    src_d = src_d_in; //source for next inner iteration, computed from residual during operation
-    
-    RealD inner_tol = InnerTolerance;
-    
-    FieldF src_f(SinglePrecGrid);
-    src_f.Checkerboard() = cb;
-    
-    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;
-
-    GridStopWatch InnerCGtimer;
-
-    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<<"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) ??
-
-      PrecChangeTimer.Start();
-      precisionChange(src_f, src_d, pc_wk_dp_to_sp);
-      PrecChangeTimer.Stop();
-      
-      sol_f = Zero();
-
-      //Optionally improve inner solver guess (eg using known eigenvectors)
-      if(guesser != NULL)
-	(*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);
-      InnerCGtimer.Stop();
-      TotalInnerIterations += CG_f.IterationsToComplete;
-      
-      //Convert sol back to double and add to double prec solution
-      PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
-      PrecChangeTimer.Stop();
-      
-      axpy(sol_d, 1.0, tmp_d, sol_d);
-    }
-    
-    //Final trial CG
-    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
-    
-    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;
-    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
-  }
-};
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -1,213 +0,0 @@
-/*************************************************************************************
-
-    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){
-    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

@@ -1,346 +0,0 @@
-/*************************************************************************************
-
-    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>
-
-    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_MULTI_SHIFT_GRADIENT_H
-#define GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////
-// Base classes for iterative processes based on operators
-// single input vec, single output vec.
-/////////////////////////////////////////////////////////////
-
-template<class Field> 
-class ConjugateGradientMultiShift : public OperatorMultiFunction<Field>,
-				    public OperatorFunction<Field>
-{
-public:                                                
-
-  using OperatorFunction<Field>::operator();
-
-  //  RealD   Tolerance;
-  Integer MaxIterations;
-  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
-  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
-  int verbose;
-  MultiShiftFunction shifts;
-  std::vector<RealD> TrueResidualShift;
-
-  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
-    MaxIterations(maxit),
-    shifts(_shifts)
-  { 
-    verbose=1;
-    IterationsToCompleteShift.resize(_shifts.order);
-    TrueResidualShift.resize(_shifts.order);
-  }
-
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
-  {
-    GridBase *grid = src.Grid();
-    int nshift = shifts.order;
-    std::vector<Field> results(nshift,grid);
-    (*this)(Linop,src,results,psi);
-  }
-  void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &results, Field &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<Field> &Linop, const Field &src, std::vector<Field> &psi)
-  {
-    GRID_TRACE("ConjugateGradientMultiShift");
-  
-    GridBase *grid = src.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);
-    std::vector<Field>   ps(nshift,grid);// Search directions
-
-    assert(psi.size()==nshift);
-    assert(mass.size()==nshift);
-    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];
-  
-    const int       primary =0;
-  
-    //Primary shift fields CG iteration
-    RealD a,b,c,d;
-    RealD cp,bp,qq; //prev
-  
-    // Matrix mult fields
-    Field r(grid);
-    Field p(grid);
-    Field tmp(grid);
-    Field mmp(grid);
-  
-    // Check lightest mass
-    for(int s=0;s<nshift;s++){
-      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);
-
-    // Handle trivial case of zero src.
-    if( cp == 0. ){
-      for(int s=0;s<nshift;s++){
-	psi[s] = Zero();
-	IterationsToCompleteShift[s] = 1;
-	TrueResidualShift[s] = 0.;
-      }
-      return;
-    }
-
-    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;
-      ps[s] = src;
-    }
-    // r and p for primary
-    r=src;
-    p=src;
-  
-    //MdagM+m[0]
-    Linop.HermOpAndNorm(p,mmp,d,qq);
-    axpy(mmp,mass[0],p,mmp);
-    RealD rn = norm2(p);
-    d += rn*mass[0];
-  
-    // have verified that inner product of 
-    // p and mmp is equal to d after this since
-    // the d computation is tricky
-    //  qq = real(innerProduct(p,mmp));
-    //  std::cout<<GridLogMessage << "debug equal ?  qq "<<qq<<" d "<< d<<std::endl;
-  
-    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,b,mmp,r);
-  
-    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
-  ///////////////////////////////////////
-  GridStopWatch AXPYTimer;
-  GridStopWatch ShiftTimer;
-  GridStopWatch QRTimer;
-  GridStopWatch MatrixTimer;
-  GridStopWatch SolverTimer;
-  SolverTimer.Start();
-  
-    // Iteration loop
-    int k;
-  
-    for (k=1;k<=MaxIterations;k++){
-    
-      a = c /cp;
-    AXPYTimer.Start();
-      axpy(p,a,p,r);
-    AXPYTimer.Stop();
-    
-      // Note to self - direction ps is iterated seperately
-      // for each shift. Does not appear to have any scope
-      // for avoiding linear algebra in "single" case.
-      // 
-      // However SAME r is used. Could load "r" and update
-      // ALL ps[s]. 2/3 Bandwidth saving
-      // New Kernel: Load r, vector of coeffs, vector of pointers ps
-    AXPYTimer.Start();
-      for(int s=0;s<nshift;s++){
-	if ( ! converged[s] ) { 
-	  if (s==0){
-	    axpy(ps[s],a,ps[s],r);
-	  } else{
-	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
-	    axpby(ps[s],z[s][iz],as,r,ps[s]);
-	  }
-	}
-      }
-    AXPYTimer.Stop();
-    
-      cp=c;
-    MatrixTimer.Start();  
-    //Linop.HermOpAndNorm(p,mmp,d,qq); // d is used
-    // The below is faster on KNL
-    Linop.HermOp(p,mmp); 
-    d=real(innerProduct(p,mmp));
-    
-    MatrixTimer.Stop();  
-
-    AXPYTimer.Start();
-      axpy(mmp,mass[0],p,mmp);
-    AXPYTimer.Stop();
-      RealD rn = norm2(p);
-      d += rn*mass[0];
-    
-      bp=b;
-      b=-cp/d;
-    
-    AXPYTimer.Start();
-      c=axpy_norm(r,b,mmp,r);
-    AXPYTimer.Stop();
-
-      // Toggle the recurrence history
-      bs[0] = b;
-      iz = 1-iz;
-    ShiftTimer.Start();
-      for(int s=1;s<nshift;s++){
-	if((!converged[s])){
-	  RealD z0 = z[s][1-iz];
-	  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();
-    
-      for(int s=0;s<nshift;s++){
-	int ss = s;
-	// Scope for optimisation here in case of "single".
-	// Could load psi[0] and pull all ps[s] in.
-	//      if ( single ) ss=primary;
-	// Bandwith saving in single case is Ls * 3 -> 2+Ls, so ~ 3x saving
-	// Pipelined CG gain:
-	//
-	// New Kernel: Load r, vector of coeffs, vector of pointers ps
-	// New Kernel: Load psi[0], vector of coeffs, vector of pointers ps
-	// If can predict the coefficient bs then we can fuse these and avoid write reread cyce
-	//  on ps[s].
-	// Before:  3 x npole  + 3 x npole
-	// After :  2 x npole (ps[s])        => 3x speed up of multishift CG.
-      
-	if( (!converged[s]) ) { 
-	  axpy(psi[ss],-bs[s]*alpha[s],ps[s],psi[ss]);
-	}
-      }
-    
-      // 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<rsq[s]){
-	    if ( ! converged[s] )
-	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShift k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
-	    converged[s]=1;
-	  } else {
-	    all_converged=0;
-	  }
-
-	}
-      }
-    
-      if ( all_converged ){
-
-    SolverTimer.Stop();
-
-
-	std::cout<<GridLogMessage<< "CGMultiShift: All shifts have converged iteration "<<k<<std::endl;
-	std::cout<<GridLogMessage<< "CGMultiShift: Checking solutions"<<std::endl;
-      
-	// Check answers 
-	for(int s=0; s < nshift; s++) { 
-	  Linop.HermOpAndNorm(psi[s],mmp,d,qq);
-	  axpy(tmp,mass[s],psi[s],mmp);
-	  axpy(r,-alpha[s],src,tmp);
-	  RealD rn = norm2(r);
-	  RealD cn = norm2(src);
-	  TrueResidualShift[s] = std::sqrt(rn/cn);
-	  std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<< TrueResidualShift[s] <<std::endl;
-	}
-
-      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
-      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
-
-      IterationsToComplete = k;	
-
-	return;
-      }
-
-   
-    }
-    // ugly hack
-    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -1,373 +0,0 @@
-/*************************************************************************************
-
-    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);
-
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
-  
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    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++){
-      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;
-    //    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;
-    assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -1,416 +0,0 @@
-/*************************************************************************************
-
-    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){ assert(0); }
-  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
-  
-  void Op     (const Field &in, Field &out){ assert(0); }
-  void AdjOp  (const Field &in, Field &out){ 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);
-
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
-  
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    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++){
-      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;
-    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;
-    assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -1,277 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradientReliableUpdate.h
-
-    Copyright (C) 2015
-
-Author: Christopher Kelly <ckelly@phys.columbia.edu>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
-#define GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
-
-NAMESPACE_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 ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
-public:
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
-  // Defaults true.
-  RealD Tolerance;
-  Integer MaxIterations;
-  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
-  Integer ReliableUpdatesPerformed;
-
-  bool DoFinalCleanup; //Final DP cleanup, defaults to true
-  Integer IterationsToCleanup; //Final DP cleanup step iterations
-    
-  LinearOperatorBase<FieldF> &Linop_f;
-  LinearOperatorBase<FieldD> &Linop_d;
-  GridBase* SinglePrecGrid;
-  RealD Delta; //reliable update parameter. 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;
-  RealD fallback_transition_tol;
-
-    
-  ConjugateGradientReliableUpdate(RealD tol, Integer maxit, RealD _delta, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d, bool err_on_no_conv = true)
-    : Tolerance(tol),
-      MaxIterations(maxit),
-      Delta(_delta),
-      Linop_f(_Linop_f),
-      Linop_d(_Linop_d),
-      SinglePrecGrid(_sp_grid),
-      ErrorOnNoConverge(err_on_no_conv),
-      DoFinalCleanup(true),
-      Linop_fallback(NULL)
-  {
-    assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
-  };
-
-  void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
-    Linop_fallback = &_Linop_fallback;
-    fallback_transition_tol = _fallback_transition_tol;      
-  }
-    
-  void operator()(const FieldD &src, FieldD &psi) {
-    GRID_TRACE("ConjugateGradientReliableUpdate");
-    LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
-    bool using_fallback = false;
-      
-    psi.Checkerboard() = src.Checkerboard();
-    conformable(psi, src);
-
-    RealD cp, c, a, d, b, ssq, qq, b_pred;
-
-    FieldD p(src);
-    FieldD mmp(src);
-    FieldD r(src);
-
-    // Initial residual computation & set up
-    RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
-    
-    Linop_d.HermOpAndNorm(psi, mmp, d, b);
-    
-    r = src - mmp;
-    p = r;
-
-    a = norm2(p);
-    cp = a;
-    ssq = norm2(src);
-
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: guess " << guess << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:   src " << ssq << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:    mp " << d << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:   mmp " << b << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:  cp,r " << cp << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:     p " << a << std::endl;
-
-    RealD rsq = Tolerance * Tolerance * ssq;
-
-    // Check if guess is really REALLY good :)
-    if (cp <= rsq) {
-      std::cout << GridLogMessage << "ConjugateGradientReliableUpdate guess was REALLY good\n";
-      std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
-      return;
-    }
-
-    //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, pc_wk_dp_to_sp);
-
-    FieldF psi_f(r_f);
-    psi_f = Zero();
-
-    FieldF p_f(r_f);
-    FieldF mmp_f(r_f);
-
-    RealD MaxResidSinceLastRelUp = cp; //initial residual    
-    
-    std::cout << GridLogIterative << std::setprecision(4)
-	      << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
-
-    GridStopWatch LinalgTimer;
-    GridStopWatch MatrixTimer;
-    GridStopWatch SolverTimer;
-    GridStopWatch PrecChangeTimer;
-    
-    SolverTimer.Start();
-    int k = 0;
-    int l = 0;
-    
-    for (k = 1; k <= MaxIterations; k++) {
-      c = cp;
-
-      MatrixTimer.Start();
-      Linop_f_use->HermOpAndNorm(p_f, mmp_f, d, qq);
-      MatrixTimer.Stop();
-
-      LinalgTimer.Start();
-
-      a = c / d;
-      b_pred = a * (a * qq - d) / c;
-
-      cp = axpy_norm(r_f, -a, mmp_f, r_f);
-      b = cp / c;
-
-      // Fuse these loops ; should be really easy
-      psi_f = a * p_f + psi_f;
-      //p_f = p_f * b + r_f;
-
-      LinalgTimer.Stop();
-
-      std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: Iteration " << k
-		<< " residual " << cp << " target " << rsq << std::endl;
-      std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << "  b = "<< b << std::endl;
-      std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << "  c = "<< c << std::endl;
-
-      if(cp > MaxResidSinceLastRelUp){
-	std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: updating MaxResidSinceLastRelUp : " << MaxResidSinceLastRelUp << " -> " << cp << std::endl;
-	MaxResidSinceLastRelUp = cp;
-      }
-	  
-      // Stopping condition
-      if (cp <= rsq) {
-	//Although not written in the paper, I assume that I have to add on the final solution
-	PrecChangeTimer.Start();
-	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
-	PrecChangeTimer.Stop();
-	psi = psi + mmp;
-	
-	
-	SolverTimer.Stop();
-	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
-	p = mmp - src;
-
-	RealD srcnorm = std::sqrt(norm2(src));
-	RealD resnorm = std::sqrt(norm2(p));
-	RealD true_residual = resnorm / srcnorm;
-
-	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate Converged on iteration " << k << " after " << l << " reliable updates" << std::endl;
-	std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
-	std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
-	std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
-
-	std::cout << GridLogMessage << "Time breakdown "<<std::endl;
-	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	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;
-	  
-	if(DoFinalCleanup){
-	  //Do a final CG to cleanup
-	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate performing final cleanup.\n";
-	  ConjugateGradient<FieldD> CG(Tolerance,MaxIterations);
-	  CG.ErrorOnNoConverge = ErrorOnNoConverge;
-	  CG(Linop_d,src,psi);
-	  IterationsToCleanup = CG.IterationsToComplete;
-	}
-	else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
-
-	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
-	return;
-      }
-      else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
-	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
-		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
-	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();
-	PrecChangeTimer.Start();
-	precisionChange(r_f, r, pc_wk_dp_to_sp);
-	PrecChangeTimer.Stop();
-	cp = norm2(r);
-	MaxResidSinceLastRelUp = cp;
-
-	b = cp/c;
-	  
-	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate new residual " << cp << std::endl;
-	  
-	l = l+1;
-      }
-
-      p_f = p_f * b + r_f; //update search vector after reliable update appears to help convergence
-
-      if(!using_fallback && Linop_fallback != NULL && cp < fallback_transition_tol){
-	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate switching to fallback linear operator on iteration " << k << " at residual " << cp << std::endl;
-	Linop_f_use = Linop_fallback;
-	using_fallback = true;
-      }
-
-	
-    }
-    std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
-	      << std::endl;
-      
-    if (ErrorOnNoConverge) assert(0);
-    IterationsToComplete = k;
-    ReliableUpdatesPerformed = l;      
-  }    
-};
-
-
-NAMESPACE_END(Grid);
-
-
-
-#endif

Grid/algorithms/iterative/ConjugateResidual.h

@@ -1,113 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateResidual.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_CONJUGATE_RESIDUAL_H
-#define GRID_CONJUGATE_RESIDUAL_H
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////
-// Base classes for iterative processes based on operators
-// single input vec, single output vec.
-/////////////////////////////////////////////////////////////
-
-template<class Field> 
-class ConjugateResidual : public OperatorFunction<Field> {
-public:                                                
-  using OperatorFunction<Field>::operator();
-
-  RealD   Tolerance;
-  Integer MaxIterations;
-  int verbose;
-
-  ConjugateResidual(RealD tol,Integer maxit) : Tolerance(tol), MaxIterations(maxit) { 
-    verbose=0;
-  };
-
-  void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
-
-    RealD a, b; // c, d;
-    RealD cp, ssq,rsq;
-      
-    RealD rAr, rAAr, rArp;
-    RealD pAp, pAAp;
-
-    GridBase *grid = src.Grid();
-    psi=Zero();
-    Field r(grid),  p(grid), Ap(grid), Ar(grid);
-      
-    r=src;
-    p=src;
-
-    Linop.HermOpAndNorm(p,Ap,pAp,pAAp);
-    Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
-
-    cp =norm2(r);
-    ssq=norm2(src);
-    rsq=Tolerance*Tolerance*ssq;
-
-    if (verbose) std::cout<<GridLogMessage<<"ConjugateResidual: iteration " <<0<<" residual "<<cp<< " target"<< rsq<<std::endl;
-
-    for(int k=1;k<MaxIterations;k++){
-
-      a = rAr/pAAp;
-
-      axpy(psi,a,p,psi);
-
-      cp = axpy_norm(r,-a,Ap,r);
-
-      rArp=rAr;
-
-      Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
-
-      b   =rAr/rArp;
- 
-      axpy(p,b,p,r);
-      pAAp=axpy_norm(Ap,b,Ap,Ar);
-	
-      if(verbose) std::cout<<GridLogMessage<<"ConjugateResidual: iteration " <<k<<" residual "<<cp<< " target"<< rsq<<std::endl;
-
-      if(cp<rsq) {
-	Linop.HermOp(psi,Ap);
-	axpy(r,-1.0,src,Ap);
-	RealD true_resid = norm2(r)/ssq;
-	std::cout<<GridLogMessage<<"ConjugateResidual: Converged on iteration " <<k
-		 << " computed residual "<<std::sqrt(cp/ssq)
-		 << " true residual "<<std::sqrt(true_resid)
-		 << " target "       <<Tolerance <<std::endl;
-	return;
-      }
-
-    }
-
-    std::cout<<GridLogMessage<<"ConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/Deflation.h

@@ -1,157 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#ifndef GRID_DEFLATION_H
-#define GRID_DEFLATION_H
-
-namespace Grid { 
-
-template<class Field>
-class ZeroGuesser: public LinearFunction<Field> {
-public:
-  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; };
-};
-
-////////////////////////////////
-// Fine grid deflation
-////////////////////////////////
-template<class Field>
-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)
-  : 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)
-  {
-    assert(evec.size()==eval.size());
-    assert(N <= evec.size());
-  } 
-
-  virtual void operator()(const Field &src,Field &guess) {
-    guess = Zero();
-    for (int i=0;i<N;i++) {
-      const Field& tmp = evec[i];
-      axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
-    }
-    guess.Checkerboard() = src.Checkerboard();
-  }
-};
-
-template<class FineField, class CoarseField>
-class LocalCoherenceDeflatedGuesser: public LinearFunction<FineField> {
-private:
-  const std::vector<FineField>   &subspace;
-  const std::vector<CoarseField> &evec_coarse;
-  const std::vector<RealD>       &eval_coarse;
-public:
-  
-  using LinearFunction<FineField>::operator();
-  LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
-				const std::vector<CoarseField> &_evec_coarse,
-				const std::vector<RealD>       &_eval_coarse)
-    : subspace(_subspace), 
-      evec_coarse(_evec_coarse), 
-      eval_coarse(_eval_coarse)  
-  {
-  }
-  
-  void operator()(const FineField &src,FineField &guess) { 
-    int N = (int)evec_coarse.size();
-    CoarseField src_coarse(evec_coarse[0].Grid());
-    CoarseField guess_coarse(evec_coarse[0].Grid());    guess_coarse = Zero();
-    blockProject(src_coarse,src,subspace);    
-    for (int i=0;i<N;i++) {
-      const CoarseField & tmp = evec_coarse[i];
-      axpy(guess_coarse,TensorRemove(innerProduct(tmp,src_coarse)) / eval_coarse[i],tmp,guess_coarse);
-    }
-    blockPromote(guess_coarse,guess,subspace);
-    guess.Checkerboard() = src.Checkerboard();
-  };
-
-  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();
-    }
-  };
-
-  };
-
-
-
-}
-#endif

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

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

Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h

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

Grid/algorithms/iterative/GeneralisedMinimalResidual.h

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

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -1,737 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Chulwoo Jung <chulwoo@bnl.gov>
-Author: Christoph Lehner <clehner@bnl.gov>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#ifndef GRID_BIRL_H
-#define GRID_BIRL_H
-
-#include <string.h> //memset
-//#include <zlib.h>
-#include <sys/stat.h>
-
-NAMESPACE_BEGIN(Grid); 
-
-/////////////////////////////////////////////////////////////
-// Implicitly restarted lanczos
-/////////////////////////////////////////////////////////////
-template<class Field> class ImplicitlyRestartedLanczosTester 
-{
- public:
-  virtual int TestConvergence(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
-  virtual int ReconstructEval(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
-};
-
-enum IRLdiagonalisation { 
-  IRLdiagonaliseWithDSTEGR,
-  IRLdiagonaliseWithQR,
-  IRLdiagonaliseWithEigen
-};
-
-template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public ImplicitlyRestartedLanczosTester<Field>
-{
- public:
-
-  LinearFunction<Field>       &_HermOp;
-  ImplicitlyRestartedLanczosHermOpTester(LinearFunction<Field> &HermOp) : _HermOp(HermOp)  {  };
-  int ReconstructEval(int j,RealD resid,Field &B, RealD &eval,RealD evalMaxApprox)
-  {
-    return TestConvergence(j,resid,B,eval,evalMaxApprox);
-  }
-  int TestConvergence(int j,RealD eresid,Field &B, RealD &eval,RealD evalMaxApprox)
-  {
-    Field v(B);
-    RealD eval_poly = eval;
-    // Apply operator
-    _HermOp(B,v);
-
-    RealD vnum = real(innerProduct(B,v)); // HermOp.
-    RealD vden = norm2(B);
-    RealD vv0  = norm2(v);
-    eval   = vnum/vden;
-    v -= eval*B;
-
-    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
-
-    std::cout.precision(13);
-    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
-	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
-	     <<std::endl;
-
-    int conv=0;
-    if( (vv<eresid*eresid) ) conv = 1;
-
-    return conv;
-  }
-};
-
-template<class Field> 
-class ImplicitlyRestartedLanczos {
- private:
-  const RealD small = 1.0e-8;
-  int MaxIter;
-  int MinRestart; // Minimum number of restarts; only check for convergence after
-  int Nstop;   // Number of evecs checked for convergence
-  int Nk;      // Number of converged sought
-  //  int Np;      // Np -- Number of spare vecs in krylov space //  == Nm - Nk
-  int Nm;      // Nm -- total number of vectors
-  IRLdiagonalisation diagonalisation;
-  int orth_period;
-    
-  RealD OrthoTime;
-  RealD eresid, betastp;
-  ////////////////////////////////
-  // Embedded objects
-  ////////////////////////////////
-  LinearFunction<Field>       &_PolyOp;
-  LinearFunction<Field>       &_HermOp;
-  ImplicitlyRestartedLanczosTester<Field> &_Tester;
-  // Default tester provided (we need a ref to something in default case)
-  ImplicitlyRestartedLanczosHermOpTester<Field> SimpleTester;
-  /////////////////////////
-  // Constructor
-  /////////////////////////
-  
-public:       
-
-  //////////////////////////////////////////////////////////////////
-  // PAB:
-  //////////////////////////////////////////////////////////////////
-  // Too many options  & knobs. 
-  // Eliminate:
-  //   orth_period
-  //   betastp
-  //   MinRestart
-  //
-  // Do we really need orth_period
-  // What is the theoretical basis & guarantees of betastp ?
-  // Nstop=Nk viable?
-  // MinRestart avoidable with new convergence test?
-  // Could cut to PolyOp, HermOp, Tester, Nk, Nm, resid, maxiter (+diagonalisation)
-  // HermOp could be eliminated if we dropped the Power method for max eval.
-  // -- also: The eval, eval2, eval2_copy stuff is still unnecessarily unclear
-  //////////////////////////////////////////////////////////////////
- ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
-			    LinearFunction<Field> & HermOp,
-			    ImplicitlyRestartedLanczosTester<Field> & Tester,
-			    int _Nstop, // sought vecs
-			    int _Nk, // sought vecs
-			    int _Nm, // spare vecs
-			    RealD _eresid, // resid in lmdue deficit 
-			    int _MaxIter, // Max iterations
-			    RealD _betastp=0.0, // if beta(k) < betastp: converged
-			    int _MinRestart=0, int _orth_period = 1,
-			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
-    SimpleTester(HermOp), _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(Tester),
-    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
-    eresid(_eresid),      betastp(_betastp),
-    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
-    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
-
-    ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
-			       LinearFunction<Field> & HermOp,
-			       int _Nstop, // sought vecs
-			       int _Nk, // sought vecs
-			       int _Nm, // spare vecs
-			       RealD _eresid, // resid in lmdue deficit 
-			       int _MaxIter, // Max iterations
-			       RealD _betastp=0.0, // if beta(k) < betastp: converged
-			       int _MinRestart=0, int _orth_period = 1,
-			       IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
-    SimpleTester(HermOp),  _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(SimpleTester),
-    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
-    eresid(_eresid),      betastp(_betastp),
-    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
-    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
-
-  ////////////////////////////////
-  // Helpers
-  ////////////////////////////////
-  template<typename T>  static RealD normalise(T& v) 
-  {
-    RealD nn = norm2(v);
-    nn = std::sqrt(nn);
-    v = v * (1.0/nn);
-    return nn;
-  }
-
-  void orthogonalize(Field& w, std::vector<Field>& evec,int k)
-  {
-    OrthoTime-=usecond()/1e6;
-    basisOrthogonalize(evec,w,k);
-    normalise(w);
-    OrthoTime+=usecond()/1e6;
-  }
-
-/* Rudy Arthur's thesis pp.137
-------------------------
-Require: M > K P = M − K †
-Compute the factorization AVM = VM HM + fM eM 
-repeat
-  Q=I
-  for i = 1,...,P do
-    QiRi =HM −θiI Q = QQi
-    H M = Q †i H M Q i
-  end for
-  βK =HM(K+1,K) σK =Q(M,K)
-  r=vK+1βK +rσK
-  VK =VM(1:M)Q(1:M,1:K)
-  HK =HM(1:K,1:K)
-  →AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
-until convergence
-*/
-  void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
-  {
-    GridBase *grid = src.Grid();
-    assert(grid == evec[0].Grid());
-    
-    //    GridLogIRL.TimingMode(1);
-    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogIRL <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
-    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogIRL <<" -- seek   Nk    = " << Nk    <<" vectors"<< std::endl;
-    std::cout << GridLogIRL <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
-    std::cout << GridLogIRL <<" -- total  Nm    = " << Nm    <<" vectors"<< std::endl;
-    std::cout << GridLogIRL <<" -- size of eval = " << eval.size() << std::endl;
-    std::cout << GridLogIRL <<" -- size of evec = " << evec.size() << std::endl;
-    if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
-      std::cout << GridLogIRL << "Diagonalisation is DSTEGR "<<std::endl;
-    } else if ( diagonalisation == IRLdiagonaliseWithQR ) { 
-      std::cout << GridLogIRL << "Diagonalisation is QR "<<std::endl;
-    }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
-      std::cout << GridLogIRL << "Diagonalisation is Eigen "<<std::endl;
-    }
-    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-	
-    assert(Nm <= evec.size() && Nm <= eval.size());
-    
-    // quickly get an idea of the largest eigenvalue to more properly normalize the residuum
-    RealD evalMaxApprox = 0.0;
-    {
-      auto src_n = src;
-      auto tmp = src;
-      std::cout << GridLogIRL << " IRL source norm " << norm2(src) << std::endl;
-      const int _MAX_ITER_IRL_MEVAPP_ = 50;
-      for (int i=0;i<_MAX_ITER_IRL_MEVAPP_;i++) {
-	normalise(src_n);
-	_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 vden = norm2(src_n);
-	RealD na = vnum/vden;
-	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
-	  i=_MAX_ITER_IRL_MEVAPP_;
-	evalMaxApprox = na;
-	std::cout << GridLogIRL << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
-	src_n = tmp;
-      }
-    }
-	
-    std::vector<RealD> lme(Nm);  
-    std::vector<RealD> lme2(Nm);
-    std::vector<RealD> eval2(Nm);
-    std::vector<RealD> eval2_copy(Nm);
-    Eigen::MatrixXd Qt = Eigen::MatrixXd::Zero(Nm,Nm);
-
-    Field f(grid);
-    Field v(grid);
-    int k1 = 1;
-    int k2 = Nk;
-    RealD beta_k;
-
-    Nconv = 0;
-  
-    // Set initial vector
-    evec[0] = src;
-    normalise(evec[0]);
-	
-    // Initial Nk steps
-    OrthoTime=0.;
-    for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
-    std::cout<<GridLogIRL <<"Initial "<< Nk <<"steps done "<<std::endl;
-    std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
-
-    //////////////////////////////////
-    // Restarting loop begins
-    //////////////////////////////////
-    int iter;
-    for(iter = 0; iter<MaxIter; ++iter){
-      
-      OrthoTime=0.;
-
-      std::cout<< GridLogMessage <<" **********************"<< std::endl;
-      std::cout<< GridLogMessage <<" Restart iteration = "<< iter << std::endl;
-      std::cout<< GridLogMessage <<" **********************"<< std::endl;
-
-      std::cout<<GridLogIRL <<" running "<<Nm-Nk <<" steps: "<<std::endl;
-      for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
-      f *= lme[Nm-1];
-
-      std::cout<<GridLogIRL <<" "<<Nm-Nk <<" steps done "<<std::endl;
-      std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
-	  
-      //////////////////////////////////
-      // getting eigenvalues
-      //////////////////////////////////
-      for(int k=0; k<Nm; ++k){
-	eval2[k] = eval[k+k1-1];
-	lme2[k] = lme[k+k1-1];
-      }
-      Qt = Eigen::MatrixXd::Identity(Nm,Nm);
-      diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
-      std::cout<<GridLogIRL <<" diagonalized "<<std::endl;
-
-      //////////////////////////////////
-      // sorting
-      //////////////////////////////////
-      eval2_copy = eval2;
-      std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
-      std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
-      const int chunk=8;
-      for(int io=0; io<k2;io+=chunk){
-	std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
-	for(int ii=0;ii<chunk;ii++){
-	  if ( (io+ii)<k2 )
-	    std::cout<< " "<< std::setw(12)<< eval2[io+ii];
-	}
-	std::cout << std::endl;
-      }
-
-      //////////////////////////////////
-      // Implicitly shifted QR transformations
-      //////////////////////////////////
-      Qt = Eigen::MatrixXd::Identity(Nm,Nm);
-      for(int ip=k2; ip<Nm; ++ip){ 
-	QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
-      }
-      std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
-
-      assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
-
-      basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
-      std::cout<<GridLogIRL <<"basisRotated  by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
-      
-      ////////////////////////////////////////////////////
-      // Compressed vector f and beta(k2)
-      ////////////////////////////////////////////////////
-      f *= Qt(k2-1,Nm-1);
-      f += lme[k2-1] * evec[k2];
-      beta_k = norm2(f);
-      beta_k = std::sqrt(beta_k);
-      std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
-	  
-      RealD betar = 1.0/beta_k;
-      evec[k2] = betar * f;
-      lme[k2-1] = beta_k;
-	  
-      ////////////////////////////////////////////////////
-      // Convergence test
-      ////////////////////////////////////////////////////
-      for(int k=0; k<Nm; ++k){    
-	eval2[k] = eval[k];
-	lme2[k] = lme[k];
-      }
-      Qt = Eigen::MatrixXd::Identity(Nm,Nm);
-      diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
-      std::cout<<GridLogIRL <<" Diagonalized "<<std::endl;
-	  
-      Nconv = 0;
-      if (iter >= MinRestart) {
-
-	std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
-
-	Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
-
-	//  power of two search pattern;  not every evalue in eval2 is assessed.
-	int allconv =1;
-	for(int jj = 1; jj<=Nstop; jj*=2){
-	  int j = Nstop-jj;
-	  RealD e = eval2_copy[j]; // Discard the evalue
-	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
-	  if( !_Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) {
-	    allconv=0;
-	  }
-	}
-	// Do evec[0] for good measure
-	{ 
-	  int j=0;
-	  RealD e = eval2_copy[0]; 
-	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
-	  if( !_Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) allconv=0;
-	}
-	if ( allconv ) Nconv = Nstop;
-
-	// test if we converged, if so, terminate
-	std::cout<<GridLogIRL<<" #modes converged: >= "<<Nconv<<"/"<<Nstop<<std::endl;
-	//	if( Nconv>=Nstop || beta_k < betastp){
-	if( Nconv>=Nstop){
-	  goto converged;
-	}
-	  
-      } else {
-	std::cout << GridLogIRL << "iter < MinRestart: do not yet test for convergence\n";
-      } // end of iter loop
-    }
-
-    std::cout<<GridLogError<<"\n NOT converged.\n";
-    abort();
-	
-  converged:
-    {
-      Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
-      basisRotate(evec,Qt,0,Nk,0,Nk,Nm);	    
-      std::cout << GridLogIRL << " Rotated basis"<<std::endl;
-      Nconv=0;
-      //////////////////////////////////////////////////////////////////////
-      // Full final convergence test; unconditionally applied
-      //////////////////////////////////////////////////////////////////////
-      for(int j = 0; j<=Nk; j++){
-	B=evec[j];
-	if( _Tester.ReconstructEval(j,eresid,B,eval2[j],evalMaxApprox) ) {
-	  Nconv++;
-	}
-      }
-
-      if ( Nconv < Nstop ) {
-	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-	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(Nstop);// was Nconv
-      evec.resize(Nstop,grid);// was Nconv
-      basisSortInPlace(evec,eval,reverse);
-      
-    }
-       
-    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogIRL << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
-    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogIRL << " -- Iterations  = "<< iter   << "\n";
-    std::cout << GridLogIRL << " -- beta(k)     = "<< beta_k << "\n";
-    std::cout << GridLogIRL << " -- Nconv       = "<< Nconv  << "\n";
-    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-  }
-
- private:
-/* Saad PP. 195
-1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
-2. For k = 1,2,...,m Do:
-3. wk:=Avk - b_k v_{k-1}      
-4. ak:=(wk,vk)       // 
-5. wk:=wk-akvk       // wk orthog vk 
-6. bk+1 := ||wk||_2. If b_k+1 = 0 then Stop
-7. vk+1 := wk/b_k+1
-8. EndDo
- */
-  void step(std::vector<RealD>& lmd,
-	    std::vector<RealD>& lme, 
-	    std::vector<Field>& evec,
-	    Field& w,int Nm,int k)
-  {
-    std::cout<<GridLogIRL << "Lanczos step " <<k<<std::endl;
-    const RealD tiny = 1.0e-20;
-    assert( k< Nm );
-
-    GridStopWatch gsw_op,gsw_o;
-
-    Field& evec_k = evec[k];
-
-    _PolyOp(evec_k,w);    std::cout<<GridLogIRL << "PolyOp" <<std::endl;
-
-    if(k>0) w -= lme[k-1] * evec[k-1];
-
-    ComplexD zalph = innerProduct(evec_k,w);
-    RealD     alph = real(zalph);
-
-    w = w - alph * evec_k;
-
-    RealD beta = normalise(w); 
-
-    lmd[k] = alph;
-    lme[k] = beta;
-
-    if ( (k>0) && ( (k % orth_period) == 0 )) {
-      std::cout<<GridLogIRL << "Orthogonalising " <<k<<std::endl;
-      orthogonalize(w,evec,k); // orthonormalise
-      std::cout<<GridLogIRL << "Orthogonalised " <<k<<std::endl;
-    }
-
-    if(k < Nm-1) evec[k+1] = w;
-
-    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
-    if ( beta < tiny ) 
-      std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
-
-    std::cout<<GridLogIRL << "Lanczos step complete " <<k<<std::endl;
-  }
-
-  void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
-			 int Nk, int Nm,  
-			 Eigen::MatrixXd & Qt, // Nm x Nm
-			 GridBase *grid)
-  {
-    Eigen::MatrixXd TriDiag = Eigen::MatrixXd::Zero(Nk,Nk);
-
-    for(int i=0;i<Nk;i++)   TriDiag(i,i)   = lmd[i];
-    for(int i=0;i<Nk-1;i++) TriDiag(i,i+1) = lme[i];
-    for(int i=0;i<Nk-1;i++) TriDiag(i+1,i) = lme[i];
-    
-    Eigen::SelfAdjointEigenSolver<Eigen::MatrixXd> eigensolver(TriDiag);
-
-    for (int i = 0; i < Nk; i++) {
-      lmd[Nk-1-i] = eigensolver.eigenvalues()(i);
-    }
-    for (int i = 0; i < Nk; i++) {
-      for (int j = 0; j < Nk; j++) {
-	Qt(Nk-1-i,j) = eigensolver.eigenvectors()(j,i);
-      }
-    }
-  }
-
-  ///////////////////////////////////////////////////////////////////////////
-  // File could end here if settle on Eigen ??? !!!
-  ///////////////////////////////////////////////////////////////////////////
-  void QR_decomp(std::vector<RealD>& lmd,   // Nm 
-		 std::vector<RealD>& lme,   // Nm 
-		 int Nk, int Nm,            // Nk, Nm
-		 Eigen::MatrixXd& Qt,       // Nm x Nm matrix
-		 RealD Dsh, int kmin, int kmax)
-  {
-    int k = kmin-1;
-    RealD x;
-    
-    RealD Fden = 1.0/hypot(lmd[k]-Dsh,lme[k]);
-    RealD c = ( lmd[k] -Dsh) *Fden;
-    RealD s = -lme[k] *Fden;
-      
-    RealD tmpa1 = lmd[k];
-    RealD tmpa2 = lmd[k+1];
-    RealD tmpb  = lme[k];
-
-    lmd[k]   = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
-    lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
-    lme[k]   = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
-    x        =-s*lme[k+1];
-    lme[k+1] = c*lme[k+1];
-      
-    for(int i=0; i<Nk; ++i){
-      RealD Qtmp1 = Qt(k,i);
-      RealD Qtmp2 = Qt(k+1,i);
-      Qt(k,i)  = c*Qtmp1 - s*Qtmp2;
-      Qt(k+1,i)= s*Qtmp1 + c*Qtmp2; 
-    }
-
-    // Givens transformations
-    for(int k = kmin; k < kmax-1; ++k){
-      
-      RealD Fden = 1.0/hypot(x,lme[k-1]);
-      RealD c = lme[k-1]*Fden;
-      RealD s = - x*Fden;
-	
-      RealD tmpa1 = lmd[k];
-      RealD tmpa2 = lmd[k+1];
-      RealD tmpb  = lme[k];
-
-      lmd[k]   = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
-      lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
-      lme[k]   = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
-      lme[k-1] = c*lme[k-1] -s*x;
-
-      if(k != kmax-2){
-	x = -s*lme[k+1];
-	lme[k+1] = c*lme[k+1];
-      }
-
-      for(int i=0; i<Nk; ++i){
-	RealD Qtmp1 = Qt(k,i);
-	RealD Qtmp2 = Qt(k+1,i);
-	Qt(k,i)     = c*Qtmp1 -s*Qtmp2;
-	Qt(k+1,i)   = s*Qtmp1 +c*Qtmp2;
-      }
-    }
-  }
-
-  void diagonalize(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
-		   int Nk, int Nm,   
-		   Eigen::MatrixXd & Qt,
-		   GridBase *grid)
-  {
-    Qt = Eigen::MatrixXd::Identity(Nm,Nm);
-    if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
-      diagonalize_lapack(lmd,lme,Nk,Nm,Qt,grid);
-    } else if ( diagonalisation == IRLdiagonaliseWithQR ) { 
-      diagonalize_QR(lmd,lme,Nk,Nm,Qt,grid);
-    }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
-      diagonalize_Eigen(lmd,lme,Nk,Nm,Qt,grid);
-    } else { 
-      assert(0);
-    }
-  }
-
-#ifdef USE_LAPACK
-void LAPACK_dstegr(char *jobz, char *range, int *n, double *d, double *e,
-                   double *vl, double *vu, int *il, int *iu, double *abstol,
-                   int *m, double *w, double *z, int *ldz, int *isuppz,
-                   double *work, int *lwork, int *iwork, int *liwork,
-                   int *info);
-#endif
-
-void diagonalize_lapack(std::vector<RealD>& lmd,
-			std::vector<RealD>& lme, 
-			int Nk, int Nm,  
-			Eigen::MatrixXd& Qt,
-			GridBase *grid)
-{
-#ifdef USE_LAPACK
-  const int size = Nm;
-  int NN = Nk;
-  double evals_tmp[NN];
-  double evec_tmp[NN][NN];
-  memset(evec_tmp[0],0,sizeof(double)*NN*NN);
-  double DD[NN];
-  double EE[NN];
-  for (int i = 0; i< NN; i++) {
-    for (int j = i - 1; j <= i + 1; j++) {
-      if ( j < NN && j >= 0 ) {
-	if (i==j) DD[i] = lmd[i];
-	if (i==j) evals_tmp[i] = lmd[i];
-	if (j==(i-1)) EE[j] = lme[j];
-      }
-    }
-  }
-  int evals_found;
-  int lwork = ( (18*NN) > (1+4*NN+NN*NN)? (18*NN):(1+4*NN+NN*NN)) ;
-  int liwork =  3+NN*10 ;
-  int iwork[liwork];
-  double work[lwork];
-  int isuppz[2*NN];
-  char jobz = 'V'; // calculate evals & evecs
-  char range = 'I'; // calculate all evals
-  //    char range = 'A'; // calculate all evals
-  char uplo = 'U'; // refer to upper half of original matrix
-  char compz = 'I'; // Compute eigenvectors of tridiagonal matrix
-  int ifail[NN];
-  int info;
-  int total = grid->_Nprocessors;
-  int node  = grid->_processor;
-  int interval = (NN/total)+1;
-  double vl = 0.0, vu = 0.0;
-  int il = interval*node+1 , iu = interval*(node+1);
-  if (iu > NN)  iu=NN;
-  double tol = 0.0;
-  if (1) {
-    memset(evals_tmp,0,sizeof(double)*NN);
-    if ( il <= NN){
-      LAPACK_dstegr(&jobz, &range, &NN,
-		    (double*)DD, (double*)EE,
-		    &vl, &vu, &il, &iu, // these four are ignored if second parameteris 'A'
-		    &tol, // tolerance
-		    &evals_found, evals_tmp, (double*)evec_tmp, &NN,
-		    isuppz,
-		    work, &lwork, iwork, &liwork,
-		    &info);
-      for (int i = iu-1; i>= il-1; i--){
-	evals_tmp[i] = evals_tmp[i - (il-1)];
-	if (il>1) evals_tmp[i-(il-1)]=0.;
-	for (int j = 0; j< NN; j++){
-	  evec_tmp[i][j] = evec_tmp[i - (il-1)][j];
-	  if (il>1) evec_tmp[i-(il-1)][j]=0.;
-	}
-      }
-    }
-    {
-      grid->GlobalSumVector(evals_tmp,NN);
-      grid->GlobalSumVector((double*)evec_tmp,NN*NN);
-    }
-  } 
-  // Safer to sort instead of just reversing it, 
-  // but the document of the routine says evals are sorted in increasing order. 
-  // qr gives evals in decreasing order.
-  for(int i=0;i<NN;i++){
-    lmd [NN-1-i]=evals_tmp[i];
-    for(int j=0;j<NN;j++){
-      Qt((NN-1-i),j)=evec_tmp[i][j];
-    }
-  }
-#else 
-  assert(0);
-#endif
-}
-
-void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
-		    int Nk, int Nm,   
-		    Eigen::MatrixXd & Qt,
-		    GridBase *grid)
-{
-  int QRiter = 100*Nm;
-  int kmin = 1;
-  int kmax = Nk;
-  
-  // (this should be more sophisticated)
-  for(int iter=0; iter<QRiter; ++iter){
-    
-    // determination of 2x2 leading submatrix
-    RealD dsub = lmd[kmax-1]-lmd[kmax-2];
-    RealD dd = std::sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
-    RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
-    // (Dsh: shift)
-    
-    // transformation
-    QR_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
-    
-    // Convergence criterion (redef of kmin and kamx)
-    for(int j=kmax-1; j>= kmin; --j){
-      RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
-      if(fabs(lme[j-1])+dds > dds){
-	kmax = j+1;
-	goto continued;
-      }
-    }
-    QRiter = iter;
-    return;
-    
-  continued:
-    for(int j=0; j<kmax-1; ++j){
-      RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
-      if(fabs(lme[j])+dds > dds){
-	kmin = j+1;
-	break;
-      }
-    }
-  }
-  std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<QRiter<<"\n";
-  abort();
-}
-};
-
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -1,454 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/LocalCoherenceLanczos.h
-
-    Copyright (C) 2015
-
-Author: Christoph Lehner <clehner@bnl.gov>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LOCAL_COHERENCE_IRL_H
-#define GRID_LOCAL_COHERENCE_IRL_H
-
-NAMESPACE_BEGIN(Grid); 
-
-struct LanczosParams : Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
-				  ChebyParams, Cheby,/*Chebyshev*/
-				  int, Nstop,    /*Vecs in Lanczos must converge Nstop < Nk < Nm*/
-				  int, Nk,       /*Vecs in Lanczos seek converge*/
-				  int, Nm,       /*Total vecs in Lanczos include restart*/
-				  RealD, resid,  /*residual*/
- 				  int, MaxIt, 
-				  RealD, betastp,  /* ? */
-				  int, MinRes);    // Must restart
-};
-
-//This class is the input parameter class for some testing programs
-struct LocalCoherenceLanczosParams : Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
-				  bool, saveEvecs,
-				  bool, doFine,
-				  bool, doFineRead,
-				  bool, doCoarse,
-	       			  bool, doCoarseRead,
-				  LanczosParams, FineParams,
-				  LanczosParams, CoarseParams,
-				  ChebyParams,   Smoother,
-				  RealD        , coarse_relax_tol,
-				  std::vector<int>, blockSize,
-				  std::string, config,
-				  std::vector < ComplexD  >, omega,
-				  RealD, mass,
-				  RealD, M5);
-};
-
-// Duplicate functionality; ProjectedFunctionHermOp could be used with the trivial function
-template<class Fobj,class CComplex,int nbasis>
-class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
-public:
-  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
-  typedef Lattice<Fobj>          FineField;
-
-  LinearOperatorBase<FineField> &_Linop;
-  std::vector<FineField>        &subspace;
-
-  ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) : 
-    _Linop(linop), subspace(_subspace)
-  {  
-    assert(subspace.size() >0);
-  };
-
-  void operator()(const CoarseField& in, CoarseField& out) {
-    GridBase *FineGrid = subspace[0].Grid();    
-    int   checkerboard = subspace[0].Checkerboard();
-
-    FineField fin (FineGrid);     fin.Checkerboard()= checkerboard;
-    FineField fout(FineGrid);   fout.Checkerboard() = checkerboard;
-
-    blockPromote(in,fin,subspace);       std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
-    _Linop.HermOp(fin,fout);                   std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
-    blockProject(out,fout,subspace);     std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
-  }
-};
-
-template<class Fobj,class CComplex,int nbasis>
-class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
-public:
-  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
-  typedef Lattice<Fobj>          FineField;
-
-
-  OperatorFunction<FineField>   & _poly;
-  LinearOperatorBase<FineField> &_Linop;
-  std::vector<FineField>        &subspace;
-
-  ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,
-			  LinearOperatorBase<FineField>& linop, 
-			  std::vector<FineField> & _subspace) :
-    _poly(poly),
-    _Linop(linop),
-    subspace(_subspace)
-  {  };
-
-  void operator()(const CoarseField& in, CoarseField& out) {
-
-    GridBase *FineGrid = subspace[0].Grid();    
-    int   checkerboard = subspace[0].Checkerboard();
-
-    FineField fin (FineGrid); fin.Checkerboard() =checkerboard;
-    FineField fout(FineGrid);fout.Checkerboard() =checkerboard;
-    
-    blockPromote(in,fin,subspace);             std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
-    _poly(_Linop,fin,fout);                    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
-    blockProject(out,fout,subspace);           std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
-  }
-};
-
-template<class Fobj,class CComplex,int nbasis>
-class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanczosTester<Lattice<iVector<CComplex,nbasis > > >
-{
-public:
-  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
-  typedef Lattice<CoarseSiteVector>           CoarseField;
-  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
-  typedef Lattice<Fobj>          FineField;
-
-  LinearFunction<CoarseField> & _Poly;
-  OperatorFunction<FineField>   & _smoother;
-  LinearOperatorBase<FineField> &_Linop;
-  RealD                             _coarse_relax_tol;
-  std::vector<FineField>        &_subspace;
-
-  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,
-					   int largestEvalIdxForReport=-1) 
-    : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
-      _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);
-    RealD eval_poly = eval;
-
-    // Apply operator
-    _Poly(B,v);
-
-    RealD vnum = real(innerProduct(B,v)); // HermOp.
-    RealD vden = norm2(B);
-    RealD vv0  = norm2(v);
-    eval   = vnum/vden;
-    v -= eval*B;
-
-    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
-
-    std::cout.precision(13);
-    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
-	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
-	     <<std::endl;
-
-    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;
-  }
-
-  //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;
-    FineField fv(FineGrid);fv.Checkerboard() =checkerboard;
-
-    blockPromote(B,fv,_subspace);  
-    
-    _smoother(_Linop,fv,fB); 
-
-    RealD eval_poly = eval;
-    _Linop.HermOp(fB,fv);
-
-    RealD vnum = real(innerProduct(fB,fv)); // HermOp.
-    RealD vden = norm2(fB);
-    RealD vv0  = norm2(fv);
-    eval   = vnum/vden;
-    fv -= eval*fB;
-    RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
-    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 << " target " << eresid*eresid
-	     <<std::endl;
-    if( (vv<eresid*eresid) ) return 1;
-    return 0;
-  }
-};
-
-////////////////////////////////////////////
-// Make serializable Lanczos params
-////////////////////////////////////////////
-template<class Fobj,class CComplex,int nbasis>
-class LocalCoherenceLanczos 
-{
-public:
-  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
-  typedef Lattice<CComplex>                   CoarseScalar; // used for inner products on fine field
-  typedef Lattice<CoarseSiteVector>           CoarseField;
-  typedef Lattice<Fobj>                       FineField;
-
-protected:
-  GridBase *_CoarseGrid;
-  GridBase *_FineGrid;
-  int _checkerboard;
-  LinearOperatorBase<FineField>                 & _FineOp;
-  
-  std::vector<RealD>                              &evals_fine;
-  std::vector<RealD>                              &evals_coarse; 
-  std::vector<FineField>                          &subspace;
-  std::vector<CoarseField>                        &evec_coarse;
-
-private:
-  std::vector<RealD>                              _evals_fine;
-  std::vector<RealD>                              _evals_coarse; 
-  std::vector<FineField>                          _subspace;
-  std::vector<CoarseField>                        _evec_coarse;
-
-public:
-
-  LocalCoherenceLanczos(GridBase *FineGrid,
-			GridBase *CoarseGrid,
-			LinearOperatorBase<FineField> &FineOp,
-			int checkerboard) :
-    _CoarseGrid(CoarseGrid),
-    _FineGrid(FineGrid),
-    _FineOp(FineOp),
-    _checkerboard(checkerboard),
-    evals_fine  (_evals_fine),
-    evals_coarse(_evals_coarse),
-    subspace    (_subspace),
-    evec_coarse(_evec_coarse)
-  {
-    evals_fine.resize(0);
-    evals_coarse.resize(0);
-  };
-  //////////////////////////////////////////////////////////////////////////
-  // Alternate constructore, external storage for use by Hadrons module
-  //////////////////////////////////////////////////////////////////////////
-  LocalCoherenceLanczos(GridBase *FineGrid,
-			GridBase *CoarseGrid,
-			LinearOperatorBase<FineField> &FineOp,
-			int checkerboard,
-			std::vector<FineField>   &ext_subspace,
-			std::vector<CoarseField> &ext_coarse,
-			std::vector<RealD>       &ext_eval_fine,
-			std::vector<RealD>       &ext_eval_coarse
-			) :
-    _CoarseGrid(CoarseGrid),
-    _FineGrid(FineGrid),
-    _FineOp(FineOp),
-    _checkerboard(checkerboard),
-    evals_fine  (ext_eval_fine), 
-    evals_coarse(ext_eval_coarse),
-    subspace    (ext_subspace),
-    evec_coarse (ext_coarse)
-  {
-    evals_fine.resize(0);
-    evals_coarse.resize(0);
-  };
-
-  //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;
-    blockOrthogonalise(InnerProd,subspace);
-    std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
-    blockOrthogonalise(InnerProd,subspace);
-  };
-
-  template<typename T>  static RealD normalise(T& v) 
-  {
-    RealD nn = norm2(v);
-    nn = ::sqrt(nn);
-    v = v * (1.0/nn);
-    return nn;
-  }
-  /*
-  void fakeFine(void)
-  {
-    int Nk = nbasis;
-    subspace.resize(Nk,_FineGrid);
-    subspace[0]=1.0;
-    subspace[0].Checkerboard()=_checkerboard;
-    normalise(subspace[0]);
-    PlainHermOp<FineField>    Op(_FineOp);
-    for(int k=1;k<Nk;k++){
-      subspace[k].Checkerboard()=_checkerboard;
-      Op(subspace[k-1],subspace[k]);
-      normalise(subspace[k]);
-    }
-  }
-  */
-
-  void testFine(RealD resid) 
-  {
-    assert(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
-    PlainHermOp<FineField>    Op(_FineOp);
-    ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
-    for(int k=0;k<nbasis;k++){
-      assert(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
-    }
-  }
-
-  //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);
-    //////////////////////////////////////////////////////////////////////////////////////////////////
-    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
-    //////////////////////////////////////////////////////////////////////////////////////////////////
-    Chebyshev<FineField>                          ChebySmooth(cheby_smooth);
-    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,subspace);
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
-
-    for(int k=0;k<evec_coarse.size();k++){
-      if ( k < nbasis ) { 
-	assert(ChebySmoothTester.ReconstructEval(k,resid,evec_coarse[k],evals_coarse[k],1.0)==1);
-      } else { 
-	assert(ChebySmoothTester.ReconstructEval(k,resid*relax,evec_coarse[k],evals_coarse[k],1.0)==1);
-      }
-    }
-  }
-
-  void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
-		RealD MaxIt, RealD betastp, int MinRes)
-  {
-    assert(nbasis<=Nm);
-    Chebyshev<FineField>      Cheby(cheby_parms);
-    FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
-    PlainHermOp<FineField>    Op(_FineOp);
-
-    evals_fine.resize(Nm);
-    subspace.resize(Nm,_FineGrid);
-
-    ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
-
-    FineField src(_FineGrid); 
-    typedef typename FineField::scalar_type Scalar;
-    // src=1.0; 
-    src=Scalar(1.0); 
-    src.Checkerboard() = _checkerboard;
-
-    int Nconv;
-    IRL.calc(evals_fine,subspace,src,Nconv,false);
-    
-    // Shrink down to number saved
-    assert(Nstop>=nbasis);
-    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); //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); //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);
-    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);
-#endif

Grid/algorithms/iterative/MinimalResidual.h

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

Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h

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

Grid/algorithms/iterative/NormalEquations.h

@@ -1,112 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/NormalEquations.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_NORMAL_EQUATIONS_H
-#define GRID_NORMAL_EQUATIONS_H
-
-NAMESPACE_BEGIN(Grid);
-
-///////////////////////////////////////////////////////////////////////////////////////////////////////
-// Take a matrix and form an NE solver calling a Herm solver
-///////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class NormalEquations {
-private:
-  SparseMatrixBase<Field> & _Matrix;
-  OperatorFunction<Field> & _HermitianSolver;
-  LinearFunction<Field>   & _Guess;
-public:
-
-  /////////////////////////////////////////////////////
-  // Wrap the usual normal equations trick
-  /////////////////////////////////////////////////////
- NormalEquations(SparseMatrixBase<Field> &Matrix, OperatorFunction<Field> &HermitianSolver,
-		 LinearFunction<Field> &Guess) 
-   :  _Matrix(Matrix), _HermitianSolver(HermitianSolver), _Guess(Guess) {}; 
-
-  void operator() (const Field &in, Field &out){
- 
-    Field src(in.Grid());
-    Field tmp(in.Grid());
-
-    MdagMLinearOperator<SparseMatrixBase<Field>,Field> MdagMOp(_Matrix);
-    _Matrix.Mdag(in,src);
-    _Guess(src,out);
-    _HermitianSolver(MdagMOp,src,out);  // Mdag M out = Mdag in
-
-  }     
-};
-
-template<class Field> class HPDSolver {
-private:
-  LinearOperatorBase<Field> & _Matrix;
-  OperatorFunction<Field> & _HermitianSolver;
-  LinearFunction<Field>   & _Guess;
-public:
-
-  /////////////////////////////////////////////////////
-  // Wrap the usual normal equations trick
-  /////////////////////////////////////////////////////
- HPDSolver(LinearOperatorBase<Field> &Matrix,
-	   OperatorFunction<Field> &HermitianSolver,
-	   LinearFunction<Field> &Guess) 
-   :  _Matrix(Matrix), _HermitianSolver(HermitianSolver), _Guess(Guess) {}; 
-
-  void operator() (const Field &in, Field &out){
- 
-    _Guess(in,out);
-    _HermitianSolver(_Matrix,in,out);  // Mdag M out = Mdag in
-
-  }     
-};
-
-
-template<class Field> class MdagMSolver {
-private:
-  SparseMatrixBase<Field> & _Matrix;
-  OperatorFunction<Field> & _HermitianSolver;
-  LinearFunction<Field>   & _Guess;
-public:
-
-  /////////////////////////////////////////////////////
-  // Wrap the usual normal equations trick
-  /////////////////////////////////////////////////////
- MdagMSolver(SparseMatrixBase<Field> &Matrix, OperatorFunction<Field> &HermitianSolver,
-	     LinearFunction<Field> &Guess) 
-   :  _Matrix(Matrix), _HermitianSolver(HermitianSolver), _Guess(Guess) {}; 
-
-  void operator() (const Field &in, Field &out){
- 
-    MdagMLinearOperator<SparseMatrixBase<Field>,Field> MdagMOp(_Matrix);
-    _Guess(in,out);
-
-    _HermitianSolver(MdagMOp,in,out);  // Mdag M out = Mdag in
-
-  }     
-};
-
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/PowerMethod.h

@@ -1,47 +0,0 @@
-#pragma once
-namespace Grid {
-template<class Field> class PowerMethod  
-{ 
- public: 
-
-  template<typename T>  static RealD normalise(T& v) 
-  {
-    RealD nn = norm2(v);
-    nn = sqrt(nn);
-    v = v * (1.0/nn);
-    return nn;
-  }
-
-  RealD operator()(LinearOperatorBase<Field> &HermOp, const Field &src) 
-  { 
-    GridBase *grid = src.Grid(); 
-    
-    // quickly get an idea of the largest eigenvalue to more properly normalize the residuum 
-    RealD evalMaxApprox = 0.0; 
-    auto src_n = src; 
-    auto tmp = src; 
-    const int _MAX_ITER_EST_ = 50; 
-
-    for (int i=0;i<_MAX_ITER_EST_;i++) { 
-      
-      normalise(src_n); 
-      HermOp.HermOp(src_n,tmp); 
-      RealD vnum = real(innerProduct(src_n,tmp)); // HermOp. 
-      RealD vden = norm2(src_n); 
-      RealD na = vnum/vden; 
-
-      std::cout << GridLogIterative << "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; 
-      } 
-      evalMaxApprox = na; 
-      src_n = tmp;
-    }
-    assert(0);
-    return 0;
-  }
-};
-}

Grid/algorithms/iterative/PrecConjugateResidual.h

@@ -1,119 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/PrecConjugateResidual.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_PREC_CONJUGATE_RESIDUAL_H
-#define GRID_PREC_CONJUGATE_RESIDUAL_H
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////
-// Base classes for iterative processes based on operators
-// single input vec, single output vec.
-/////////////////////////////////////////////////////////////
-
-template<class Field> 
-class PrecConjugateResidual : public OperatorFunction<Field> {
-public:                                                
-  RealD   Tolerance;
-  Integer MaxIterations;
-  int verbose;
-  LinearFunction<Field> &Preconditioner;
-
-  PrecConjugateResidual(RealD tol,Integer maxit,LinearFunction<Field> &Prec) : Tolerance(tol), MaxIterations(maxit),      Preconditioner(Prec)
-  { 
-    verbose=1;
-  };
-
-  void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
-
-    RealD a, b, c, d;
-    RealD cp, ssq,rsq;
-      
-    RealD rAr, rAAr, rArp;
-    RealD pAp, pAAp;
-
-    GridBase *grid = src.Grid();
-    Field r(grid),  p(grid), Ap(grid), Ar(grid), z(grid);
-      
-    psi=zero;
-    r  = src;
-    Preconditioner(r,p);
-
-      
-
-    Linop.HermOpAndNorm(p,Ap,pAp,pAAp);
-    Ar=Ap;
-    rAr=pAp;
-    rAAr=pAAp;
-
-    cp =norm2(r);
-    ssq=norm2(src);
-    rsq=Tolerance*Tolerance*ssq;
-
-    if (verbose) std::cout<<GridLogMessage<<"PrecConjugateResidual: iteration " <<0<<" residual "<<cp<< " target"<< rsq<<std::endl;
-
-    for(int k=0;k<MaxIterations;k++){
-
-
-      Preconditioner(Ap,z);
-      RealD rq= real(innerProduct(Ap,z)); 
-
-      a = rAr/rq;
-
-      axpy(psi,a,p,psi);
-      cp = axpy_norm(r,-a,z,r);
-
-      rArp=rAr;
-
-      Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
-
-      b   =rAr/rArp;
- 
-      axpy(p,b,p,r);
-      pAAp=axpy_norm(Ap,b,Ap,Ar);
-	
-      if(verbose) std::cout<<GridLogMessage<<"PrecConjugateResidual: iteration " <<k<<" residual "<<cp<< " target"<< rsq<<std::endl;
-
-      if(cp<rsq) {
-	Linop.HermOp(psi,Ap);
-	axpy(r,-1.0,src,Ap);
-	RealD true_resid = norm2(r)/ssq;
-	std::cout<<GridLogMessage<<"PrecConjugateResidual: Converged on iteration " <<k
-		 << " computed residual "<<sqrt(cp/ssq)
-		 << " true residual "<<sqrt(true_resid)
-		 << " target "       <<Tolerance <<std::endl;
-	return;
-      }
-
-    }
-
-    std::cout<<GridLogMessage<<"PrecConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -1,239 +0,0 @@
-/*************************************************************************************
-
-    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_H
-#define GRID_PREC_GCR_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 PrecGeneralisedConjugateResidual : 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; };
-
-  PrecGeneralisedConjugateResidual(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)
-  { 
-    level=1;
-    verbose=1;
-  };
-
-  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.HermOp(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;
-	/*
-	  GCRLogLevel<<"PGCR Time elapsed: Precon "<<   PrecTimer.Elapsed() <<std::endl;
-	  GCRLogLevel<<"PGCR Time elapsed: Matrix "<<    MatTimer.Elapsed() <<std::endl;
-	  GCRLogLevel<<"PGCR Time elapsed: Linalg "<< LinalgTimer.Elapsed() <<std::endl;
-	*/
-	return;
-      }
-
-    }
-    GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    assert(0);
-  }
-
-  RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
-
-    RealD cp;
-    RealD a, b;
-    RealD zAz, zAAz;
-    RealD 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);
-    std::vector<Field> p(mmax,grid);
-    std::vector<RealD> qq(mmax);
-      
-    GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
-
-    //////////////////////////////////
-    // initial guess x0 is taken as nonzero.
-    // r0=src-A x0 = src
-    //////////////////////////////////
-    MatTimer.Start();
-    Linop.HermOpAndNorm(psi,Az,zAz,zAAz); 
-    MatTimer.Stop();
-    
-
-    LinalgTimer.Start();
-    r=src-Az;
-    LinalgTimer.Stop();
-    GCRLogLevel<< "PGCR true residual r = src - A psi   "<<norm2(r) <<std::endl;
-    
-    /////////////////////
-    // p = Prec(r)
-    /////////////////////
-
-    PrecTimer.Start();
-    Preconditioner(r,z);
-    PrecTimer.Stop();
-
-    MatTimer.Start();
-    Linop.HermOpAndNorm(z,Az,zAz,zAAz); 
-    MatTimer.Stop();
-
-    LinalgTimer.Start();
-
-    //p[0],q[0],qq[0] 
-    p[0]= z;
-    q[0]= Az;
-    qq[0]= zAAz;
-    
-    cp =norm2(r);
-    LinalgTimer.Stop();
-
-    for(int k=0;k<nstep;k++){
-
-      steps++;
-
-      int kp     = k+1;
-      int peri_k = k %mmax;
-      int peri_kp= kp%mmax;
-
-      LinalgTimer.Start();
-      rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
-      a = rq/qq[peri_k];
-
-      axpy(psi,a,p[peri_k],psi);         
-
-      cp = axpy_norm(r,-a,q[peri_k],r);
-      LinalgTimer.Stop();
-
-      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
-
-      if((k==nstep-1)||(cp<rsq)){
-	return cp;
-      }
-
-
-      PrecTimer.Start();
-      Preconditioner(r,z);// solve Az = r
-      PrecTimer.Stop();
-
-      MatTimer.Start();
-      Linop.HermOpAndNorm(z,Az,zAz,zAAz);
-      MatTimer.Stop();
-
-      LinalgTimer.Start();
-
-      q[peri_kp]=Az;
-      p[peri_kp]=z;
-
-      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);
-
-	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
-	p[peri_kp]=p[peri_kp]+b*p[peri_back];
-	q[peri_kp]=q[peri_kp]+b*q[peri_back];
-
-      }
-      qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
-      LinalgTimer.Stop();
-    }
-    assert(0); // never reached
-    return cp;
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -1,242 +0,0 @@
-/*************************************************************************************
-
-    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)
-  { 
-    level=1;
-    verbose=1;
-  };
-
-  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;
-    //    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);
-    std::vector<Field> p(mmax,grid);
-    std::vector<RealD> qq(mmax);
-      
-    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;
-    
-    /////////////////////
-    // 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;
-    
-    cp =norm2(r);
-    LinalgTimer.Stop();
-
-    for(int k=0;k<nstep;k++){
-
-      steps++;
-
-      int kp     = k+1;
-      int peri_k = k %mmax;
-      int peri_kp= kp%mmax;
-
-      LinalgTimer.Start();
-      rq= innerProduct(q[peri_k],r); // what if rAr not real?
-      a = rq/qq[peri_k];
-
-      axpy(psi,a,p[peri_k],psi);         
-
-      cp = axpy_norm(r,-a,q[peri_k],r);
-      LinalgTimer.Stop();
-
-      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
-
-      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;
-
-      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);
-
-	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
-	p[peri_kp]=p[peri_kp]+b*p[peri_back];
-	q[peri_kp]=q[peri_kp]+b*q[peri_back];
-
-      }
-      qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
-      LinalgTimer.Stop();
-    }
-    assert(0); // never reached
-    return cp;
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/algorithms/iterative/QuasiMinimalResidual.h

@@ -1,371 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/algorithmsf/iterative/QuasiMinimalResidual.h
-
-Copyright (C) 2019
-
-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);
-
-template<class Field> 
-RealD innerG5ProductReal(Field &l, Field &r)
-{
-  Gamma G5(Gamma::Algebra::Gamma5);
-  Field tmp(l.Grid());
-  //  tmp = G5*r;
-  G5R5(tmp,r);
-  ComplexD ip =innerProduct(l,tmp);
-  std::cout << "innerProductRealG5R5 "<<ip<<std::endl;
-  return ip.real();
-}
-
-template<class Field>
-class QuasiMinimalResidual : public OperatorFunction<Field> {
- public:
-  using OperatorFunction<Field>::operator();
-
-  bool ErrorOnNoConverge; 
-  RealD   Tolerance;
-  Integer MaxIterations;
-  Integer IterationCount;
-
-  QuasiMinimalResidual(RealD   tol,
-		       Integer maxit,
-		       bool    err_on_no_conv = true)
-      : Tolerance(tol)
-      , MaxIterations(maxit)
-      , ErrorOnNoConverge(err_on_no_conv) 
-  {};
-
-#if 1
-  void operator()(LinearOperatorBase<Field> &LinOp, const Field &b, Field &x) 
-  {
-    RealD resid;
-    IterationCount=0;
-
-    RealD  rho, rho_1, xi, gamma, gamma_1, theta, theta_1;
-    RealD  eta, delta, ep, beta; 
-
-    GridBase *Grid = b.Grid();
-    Field r(Grid), d(Grid), s(Grid);
-    Field v(Grid), w(Grid), y(Grid),  z(Grid);
-    Field v_tld(Grid), w_tld(Grid), y_tld(Grid), z_tld(Grid);
-    Field p(Grid), q(Grid), p_tld(Grid);
-
-    Real normb = norm2(b);
-
-    LinOp.Op(x,r); r = b - r;
-
-    assert(normb> 0.0);
-
-    resid = norm2(r)/normb;
-    if (resid <= Tolerance) {
-      return;
-    }
-
-    v_tld = r;
-    y = v_tld;
-    rho = norm2(y);
-
-    // Take Gamma5 conjugate
-    //    Gamma G5(Gamma::Algebra::Gamma5);
-    //    G5R5(w_tld,r);
-    //    w_tld = G5* v_tld;
-    w_tld=v_tld;
-    z = w_tld;
-    xi = norm2(z);
-
-    gamma = 1.0;
-    eta   = -1.0;
-    theta = 0.0;
-
-    for (int i = 1; i <= MaxIterations; i++) {
-
-      // Breakdown tests
-      assert( rho != 0.0);
-      assert( xi  != 0.0);
-
-      v = (1. / rho) * v_tld;
-      y = (1. / rho) * y;
-
-      w = (1. / xi) * w_tld;
-      z = (1. / xi) * z;
-
-      ComplexD Zdelta = innerProduct(z, y); // Complex?
-      std::cout << "Zdelta "<<Zdelta<<std::endl;
-      delta = Zdelta.real();
-
-      y_tld = y; 
-      z_tld = z;
-
-      if (i > 1) {
-	p = y_tld - (xi  * delta / ep) * p;
-	q = z_tld - (rho * delta / ep) * q;
-      } else {
-	p = y_tld;
-	q = z_tld;
-      }
-
-      LinOp.Op(p,p_tld);      //     p_tld = A * p;
-      ComplexD Zep = innerProduct(q, p_tld);
-      ep=Zep.real();
-      std::cout << "Zep "<<Zep <<std::endl;
-      // Complex Audit
-      assert(abs(ep)>0);
-
-      beta = ep / delta;
-      assert(abs(beta)>0);
-
-      v_tld = p_tld - beta * v;
-      y = v_tld;
-
-      rho_1 = rho;
-      rho   = norm2(y);
-      LinOp.AdjOp(q,w_tld);
-      w_tld = w_tld - beta * w;
-      z = w_tld;
-
-      xi = norm2(z);
-
-      gamma_1 = gamma;
-      theta_1 = theta;
-
-      theta   = rho / (gamma_1 * beta);
-      gamma   = 1.0 / sqrt(1.0 + theta * theta);
-      std::cout << "theta "<<theta<<std::endl;
-      std::cout << "gamma "<<gamma<<std::endl;
-
-      assert(abs(gamma)> 0.0);
-
-      eta = -eta * rho_1 * gamma* gamma / (beta * gamma_1 * gamma_1);
-
-      if (i > 1) {
-	d = eta * p + (theta_1 * theta_1 * gamma * gamma) * d;
-	s = eta * p_tld + (theta_1 * theta_1 * gamma * gamma) * s;
-      } else {
-	d = eta * p;
-	s = eta * p_tld;
-      }
-
-      x =x+d;                            // update approximation vector
-      r =r-s;                            // compute residual
-
-      if ((resid = norm2(r) / normb) <= Tolerance) {
-	return;
-      }
-      std::cout << "Iteration "<<i<<" resid " << resid<<std::endl;
-    }
-    assert(0);
-    return;                            // no convergence
-  }
-#else
-  // QMRg5 SMP thesis
-  void operator()(LinearOperatorBase<Field> &LinOp, const Field &b, Field &x) 
-  {
-    // Real scalars
-    GridBase *grid = b.Grid();
-
-    Field    r(grid);
-    Field    p_m(grid), p_m_minus_1(grid), p_m_minus_2(grid);
-    Field    v_m(grid), v_m_minus_1(grid), v_m_plus_1(grid);
-    Field    tmp(grid);
-
-    RealD    w;
-    RealD    z1, z2;
-    RealD    delta_m, delta_m_minus_1;
-    RealD    c_m_plus_1, c_m, c_m_minus_1;
-    RealD    s_m_plus_1, s_m, s_m_minus_1;
-    RealD    alpha, beta, gamma, epsilon;
-    RealD    mu, nu, rho, theta, xi, chi;
-    RealD    mod2r, mod2b;
-    RealD    tau2, target2;
-
-    mod2b=norm2(b);
-
-    /////////////////////////
-    // Initial residual
-    /////////////////////////
-    LinOp.Op(x,tmp);
-    r = b - tmp;
-
-    /////////////////////////
-    // \mu = \rho = |r_0|
-    /////////////////////////
-    mod2r = norm2(r);
-    rho = sqrt( mod2r);
-    mu=rho;
-    
-    std::cout << "QuasiMinimalResidual rho "<< rho<<std::endl;
-    /////////////////////////
-    // Zero negative history
-    /////////////////////////
-    v_m_plus_1  = Zero();
-    v_m_minus_1 = Zero();
-    p_m_minus_1 = Zero();
-    p_m_minus_2 = Zero();
-
-    // v0
-    v_m = (1.0/rho)*r;
-
-    /////////////////////////
-    // Initial coeffs
-    /////////////////////////
-    delta_m_minus_1 = 1.0;
-    c_m_minus_1     = 1.0;
-    c_m             = 1.0;
-    s_m_minus_1     = 0.0;
-    s_m             = 0.0;
-
-    /////////////////////////
-    // Set up convergence check
-    /////////////////////////
-    tau2    = mod2r;
-    target2 = mod2b * Tolerance*Tolerance;
- 
-    for(int iter = 0 ; iter < MaxIterations; iter++){
-
-      /////////////////////////
-      // \delta_m = (v_m, \gamma_5 v_m) 
-      /////////////////////////
-      delta_m = innerG5ProductReal(v_m,v_m);
-      std::cout << "QuasiMinimalResidual delta_m "<< delta_m<<std::endl;
-
-      /////////////////////////
-      // tmp = A v_m
-      /////////////////////////
-      LinOp.Op(v_m,tmp);
-
-      /////////////////////////
-      // \alpha = (v_m, \gamma_5 temp) / \delta_m 
-      /////////////////////////
-      alpha = innerG5ProductReal(v_m,tmp);
-      alpha = alpha/delta_m ;
-      std::cout << "QuasiMinimalResidual alpha "<< alpha<<std::endl;
-
-      /////////////////////////
-      // \beta = \rho \delta_m / \delta_{m-1}
-      /////////////////////////
-      beta = rho * delta_m / delta_m_minus_1;
-      std::cout << "QuasiMinimalResidual beta "<< beta<<std::endl;
-
-      /////////////////////////
-      // \tilde{v}_{m+1} = temp - \alpha v_m - \beta v_{m-1}
-      /////////////////////////
-      v_m_plus_1 = tmp - alpha*v_m - beta*v_m_minus_1;
-
-      ///////////////////////////////
-      // \rho = || \tilde{v}_{m+1} ||
-      ///////////////////////////////
-      rho = sqrt( norm2(v_m_plus_1) );
-      std::cout << "QuasiMinimalResidual rho "<< rho<<std::endl;
-
-      ///////////////////////////////
-      //      v_{m+1} = \tilde{v}_{m+1}
-      ///////////////////////////////
-      v_m_plus_1 = (1.0 / rho) * v_m_plus_1;
-
-      ////////////////////////////////
-      // QMR recurrence coefficients.
-      ////////////////////////////////
-      theta      = s_m_minus_1 * beta;
-      gamma      = c_m_minus_1 * beta;
-      epsilon    =  c_m * gamma + s_m * alpha;
-      xi         = -s_m * gamma + c_m * alpha;
-      nu         = sqrt( xi*xi + rho*rho );
-      c_m_plus_1 = fabs(xi) / nu;
-      if ( xi == 0.0 ) {
-	s_m_plus_1 = 1.0;
-      } else {
-	s_m_plus_1 = c_m_plus_1 * rho / xi;
-      }
-      chi = c_m_plus_1 * xi + s_m_plus_1 * rho;
-
-      std::cout << "QuasiMinimalResidual coeffs "<< theta <<" "<<gamma<<" "<< epsilon<<" "<< xi<<" "<< nu<<std::endl;
-      std::cout << "QuasiMinimalResidual coeffs "<< chi   <<std::endl;
-
-      ////////////////////////////////
-      //p_m=(v_m - \epsilon p_{m-1} - \theta p_{m-2}) / \chi
-      ////////////////////////////////
-      p_m = (1.0/chi) * v_m - (epsilon/chi) * p_m_minus_1 - (theta/chi) * p_m_minus_2;
-
-      ////////////////////////////////////////////////////////////////
-      //      \psi = \psi + c_{m+1} \mu p_m	
-      ////////////////////////////////////////////////////////////////
-      x = x + ( c_m_plus_1 * mu ) * p_m;
-
-      ////////////////////////////////////////
-      //
-      ////////////////////////////////////////
-      mu              = -s_m_plus_1 * mu;
-      delta_m_minus_1 = delta_m;
-      c_m_minus_1     = c_m;
-      c_m             = c_m_plus_1;
-      s_m_minus_1     = s_m;
-      s_m             = s_m_plus_1;
-
-      ////////////////////////////////////
-      // Could use pointer swizzle games.
-      ////////////////////////////////////
-      v_m_minus_1 = v_m;
-      v_m         = v_m_plus_1;
-      p_m_minus_2 = p_m_minus_1;
-      p_m_minus_1 = p_m;
-
-
-      /////////////////////////////////////
-      // Convergence checks
-      /////////////////////////////////////
-      z1 = RealD(iter+1.0);
-      z2 = z1 + 1.0;
-      tau2 = tau2 *( z2 / z1 ) * s_m * s_m;
-      std::cout << " QuasiMinimumResidual iteration "<< iter<<std::endl;
-      std::cout << " QuasiMinimumResidual tau bound "<< tau2<<std::endl;
-
-      // Compute true residual
-      mod2r = tau2;
-      if ( 1 || (tau2 < (100.0 * target2)) ) {
-	LinOp.Op(x,tmp);
-	r = b - tmp;
-	mod2r = norm2(r);
-	std::cout << " QuasiMinimumResidual true residual is "<< mod2r<<std::endl;
-      }
-
-
-      if ( mod2r < target2 ) { 
-
-	std::cout << " QuasiMinimumResidual has converged"<<std::endl;
-	return;
-
-      }
-
-    }
-
-
-  }
-#endif
-};
-
-NAMESPACE_END(Grid);

Grid/algorithms/iterative/SchurRedBlack.h

@@ -1,651 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/SchurRedBlack.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#ifndef GRID_SCHUR_RED_BLACK_H
-#define GRID_SCHUR_RED_BLACK_H
-
-
-  /*
-   * Red black Schur decomposition
-   *
-   *  M = (Mee Meo) =  (1             0 )   (Mee   0               )  (1 Mee^{-1} Meo)
-   *      (Moe Moo)    (Moe Mee^-1    1 )   (0   Moo-Moe Mee^-1 Meo)  (0   1         )
-   *                =         L                     D                     U
-   *
-   * L^-1 = (1              0 )
-   *        (-MoeMee^{-1}   1 )   
-   * L^{dag} = ( 1       Mee^{-dag} Moe^{dag} )
-   *           ( 0       1                    )
-   * L^{-d}  = ( 1      -Mee^{-dag} Moe^{dag} )
-   *           ( 0       1                    )
-   *
-   * U^-1 = (1   -Mee^{-1} Meo)
-   *        (0    1           )
-   * U^{dag} = ( 1                 0)
-   *           (Meo^dag Mee^{-dag} 1)
-   * U^{-dag} = (  1                 0)
-   *            (-Meo^dag Mee^{-dag} 1)
-   ***********************
-   *     M psi = eta
-   ***********************
-   *Odd
-   * i)                 D_oo psi_o =  L^{-1}  eta_o
-   *                        eta_o' = (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
-   *
-   * Wilson:
-   *      (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1}  eta_o
-   * Stag:
-   *      D_oo psi_o = L^{-1}  eta =    (eta_o - Moe Mee^{-1} eta_e)
-   *
-   * L^-1 eta_o= (1              0 ) (e
-   *             (-MoeMee^{-1}   1 )   
-   *
-   *Even
-   * ii)  Mee psi_e + Meo psi_o = src_e
-   *
-   *   => sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-   *
-   * 
-   * TODO: Other options:
-   * 
-   * a) change checkerboards for Schur e<->o
-   *
-   * Left precon by Moo^-1
-   * b) Doo^{dag} M_oo^-dag Moo^-1 Doo psi_0 =  (D_oo)^dag M_oo^-dag Moo^-1 L^{-1}  eta_o
-   *                              eta_o'     = (D_oo)^dag  M_oo^-dag Moo^-1 (eta_o - Moe Mee^{-1} eta_e)
-   *
-   * Right precon by Moo^-1
-   * c) M_oo^-dag Doo^{dag} Doo Moo^-1 phi_0 = M_oo^-dag (D_oo)^dag L^{-1}  eta_o
-   *                              eta_o'     = M_oo^-dag (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
-   *                              psi_o = M_oo^-1 phi_o
-   * TODO: Deflation 
-   */
-namespace Grid {
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Use base class to share code
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Take a matrix and form a Red Black solver calling a Herm solver
-  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class SchurRedBlackBase {
-  protected:
-    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
-    OperatorFunction<Field> & _HermitianRBSolver;
-    int CBfactorise;
-    bool subGuess;
-    bool useSolnAsInitGuess; // if true user-supplied solution vector is used as initial guess for solver
-  public:
-
-    SchurRedBlackBase(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-        const bool _solnAsInitGuess = false)  :
-    _HermitianRBSolver(HermitianRBSolver),
-    useSolnAsInitGuess(_solnAsInitGuess)
-    { 
-      CBfactorise = 0;
-      subtractGuess(initSubGuess);
-    };
-    void subtractGuess(const bool initSubGuess)
-    {
-      subGuess = initSubGuess;
-    }
-    bool isSubtractGuess(void)
-    {
-      return subGuess;
-    }
-
-    /////////////////////////////////////////////////////////////
-    // Shared code
-    /////////////////////////////////////////////////////////////
-    void operator() (Matrix & _Matrix,const Field &in, Field &out){
-      ZeroGuesser<Field> guess;
-      (*this)(_Matrix,in,out,guess);
-    }
-    void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out) 
-    {
-      ZeroGuesser<Field> guess;
-      (*this)(_Matrix,in,out,guess);
-    }
-
-    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) 
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-      int nblock = in.size();
-
-      std::vector<Field> src_o(nblock,grid);
-      std::vector<Field> sol_o(nblock,grid);
-      
-      std::vector<Field> guess_save;
-
-      Field resid(fgrid);
-      Field tmp(grid);
-
-      ////////////////////////////////////////////////
-      // Prepare RedBlack source
-      ////////////////////////////////////////////////
-      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);
-
-      
-      if(useSolnAsInitGuess) {
-        for(int b=0;b<nblock;b++){
-          pickCheckerboard(Odd, sol_o[b], out[b]);
-        }
-      } else {
-        guess(src_o, sol_o); 
-      }
-
-	    if ( subGuess ) { 
-        for(int b=0;b<nblock;b++){
-          guess_save[b] = sol_o[b];
-        }
-      }
-      //////////////////////////////////////////////////////////////
-      // Call the block solver
-      //////////////////////////////////////////////////////////////
-      std::cout<<GridLogMessage << "SchurRedBlackBase calling the solver for "<<nblock<<" RHS" <<std::endl;
-      RedBlackSolve(_Matrix,src_o,sol_o);
-
-      ////////////////////////////////////////////////
-      // A2A boolean behavioural control & reconstruct other checkerboard
-      ////////////////////////////////////////////////
-      for(int b=0;b<nblock;b++) {
-
-	if (subGuess)   sol_o[b] = sol_o[b] - guess_save[b];
-
-	///////// Needs even source //////////////
-	pickCheckerboard(Even,tmp,in[b]);
-	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]);
-
-	/////////////////////////////////////////////////
-	// Check unprec residual if possible
-	/////////////////////////////////////////////////
-	if ( ! subGuess ) {
-	  _Matrix.M(out[b],resid); 
-	  resid = resid-in[b];
-	  RealD ns = norm2(in[b]);
-	  RealD nr = norm2(resid);
-	
-	  std::cout<<GridLogMessage<< "SchurRedBlackBase solver true unprec resid["<<b<<"] "<<std::sqrt(nr/ns) << std::endl;
-	} else {
-	  std::cout<<GridLogMessage<< "SchurRedBlackBase Guess subtracted after solve["<<b<<"] " << std::endl;
-	}
-
-      }
-    }
-    template<class Guesser>
-    void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
-
-      // FIXME CGdiagonalMee not implemented virtual function
-      // FIXME use CBfactorise to control schur decomp
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      Field resid(fgrid);
-      Field src_o(grid);
-      Field src_e(grid);
-      Field sol_o(grid);
-
-      ////////////////////////////////////////////////
-      // RedBlack source
-      ////////////////////////////////////////////////
-      RedBlackSource(_Matrix,in,src_e,src_o);
-
-      ////////////////////////////////
-      // Construct the guess
-      ////////////////////////////////
-      if(useSolnAsInitGuess) {
-        pickCheckerboard(Odd, sol_o, out);
-      } else {
-        guess(src_o,sol_o);
-      }
-
-      Field  guess_save(grid);
-      guess_save = sol_o;
-
-      //////////////////////////////////////////////////////////////
-      // Call the red-black solver
-      //////////////////////////////////////////////////////////////
-      RedBlackSolve(_Matrix,src_o,sol_o);
-
-      ////////////////////////////////////////////////
-      // Fionn A2A boolean behavioural control
-      ////////////////////////////////////////////////
-      if (subGuess)      sol_o= sol_o-guess_save;
-
-      ///////////////////////////////////////////////////
-      // RedBlack solution needs the even source
-      ///////////////////////////////////////////////////
-      RedBlackSolution(_Matrix,sol_o,src_e,out);
-
-      // Verify the unprec residual
-      if ( ! subGuess ) {
-        _Matrix.M(out,resid); 
-        resid = resid-in;
-        RealD ns = norm2(in);
-        RealD nr = norm2(resid);
-
-        std::cout<<GridLogMessage << "SchurRedBlackBase solver true unprec resid "<< std::sqrt(nr/ns) << std::endl;
-      } else {
-        std::cout << GridLogMessage << "SchurRedBlackBase Guess subtracted after solve." << std::endl;
-      }
-    }     
-    
-    /////////////////////////////////////////////////////////////
-    // Override in derived. 
-    /////////////////////////////////////////////////////////////
-    virtual void RedBlackSource  (Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)                =0;
-    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)          =0;
-    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)                           =0;
-    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)=0;
-
-  };
-
-  template<class Field> class SchurRedBlackStaggeredSolve : public SchurRedBlackBase<Field> {
-  public:
-    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
-
-    SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-        const bool _solnAsInitGuess = false) 
-      :    SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) 
-    {
-    }
-
-    //////////////////////////////////////////////////////
-    // Override RedBlack specialisation
-    //////////////////////////////////////////////////////
-    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      Field   tmp(grid);
-      Field  Mtmp(grid);
-
-      pickCheckerboard(Even,src_e,src);
-      pickCheckerboard(Odd ,src_o,src);
-
-      /////////////////////////////////////////////////////
-      // src_o = (source_o - Moe MeeInv source_e)
-      /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
-
-      _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
-    }
-    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e_c,Field &sol)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      Field   tmp(grid);
-      Field   sol_e(grid);
-      Field   src_e(grid);
-
-      src_e = src_e_c; // Const correctness
-
-      ///////////////////////////////////////////////////
-      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-      ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.Checkerboard()   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.Checkerboard() ==Even);
-     
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
-    }
-    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
-    {
-      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
-    };
-    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
-    {
-      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
-    }
-  };
-  template<class Field> using SchurRedBlackStagSolve = SchurRedBlackStaggeredSolve<Field>;
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Site diagonal has Mooee on it.
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class SchurRedBlackDiagMooeeSolve : public SchurRedBlackBase<Field> {
-  public:
-    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
-
-    SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-        const bool _solnAsInitGuess = false)  
-      : SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
-
-
-    //////////////////////////////////////////////////////
-    // Override RedBlack specialisation
-    //////////////////////////////////////////////////////
-    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      Field   tmp(grid);
-      Field  Mtmp(grid);
-
-      pickCheckerboard(Even,src_e,src);
-      pickCheckerboard(Odd ,src_o,src);
-
-      /////////////////////////////////////////////////////
-      // src_o = Mdag * (source_o - Moe MeeInv source_e)
-      /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
-
-      // get the right MpcDag
-      SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
-
-    }
-    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      Field   tmp(grid);
-      Field  sol_e(grid);
-      Field  src_e_i(grid);
-      ///////////////////////////////////////////////////
-      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-      ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.Checkerboard()   ==Even);
-      src_e_i = src_e-tmp;               assert(  src_e_i.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.Checkerboard() ==Even);
-     
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
-    }
-    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
-    {
-      SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
-    };
-    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
-    {
-      SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
-    }
-  };
-
-  template<class Field> class NonHermitianSchurRedBlackDiagMooeeSolve : public SchurRedBlackBase<Field> 
-  {
-    public:
-      typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
-
-      NonHermitianSchurRedBlackDiagMooeeSolve(OperatorFunction<Field>& RBSolver, const bool initSubGuess = false,
-          const bool _solnAsInitGuess = false)  
-      : SchurRedBlackBase<Field>(RBSolver, initSubGuess, _solnAsInitGuess) {};
-
-      //////////////////////////////////////////////////////
-      // Override RedBlack specialisation
-      //////////////////////////////////////////////////////
-      virtual void RedBlackSource(Matrix& _Matrix, const Field& src, Field& src_e, Field& src_o)
-      {
-        GridBase* grid  = _Matrix.RedBlackGrid();
-        GridBase* fgrid = _Matrix.Grid();
-
-        Field  tmp(grid);
-        Field Mtmp(grid);
-
-        pickCheckerboard(Even, src_e, src);
-        pickCheckerboard(Odd , src_o, src);
-
-        /////////////////////////////////////////////////////
-        // src_o = Mdag * (source_o - Moe MeeInv source_e)
-        /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
-      }
-      
-      virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
-      {
-        GridBase* grid  = _Matrix.RedBlackGrid();
-        GridBase* fgrid = _Matrix.Grid();
-
-        Field     tmp(grid);
-        Field   sol_e(grid);
-        Field src_e_i(grid);
-        
-        ///////////////////////////////////////////////////
-        // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-        ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o, tmp);         assert(     tmp.Checkerboard() == Even );
-        src_e_i = src_e - tmp;             assert( src_e_i.Checkerboard() == Even );
-        _Matrix.MooeeInv(src_e_i, sol_e);  assert(   sol_e.Checkerboard() == Even );
-       
-        setCheckerboard(sol, sol_e); assert( sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o); assert( sol_o.Checkerboard() == Odd  );
-      }
-
-      virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
-      {
-        NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  assert(sol_o.Checkerboard() == Odd);
-      }
-
-      virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o, std::vector<Field>& sol_o)
-      {
-        NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o); 
-      }
-  };
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Site diagonal is identity, right preconditioned by Mee^inv
-  // ( 1 - Meo Moo^inv Moe Mee^inv  ) phi =( 1 - Meo Moo^inv Moe Mee^inv  ) Mee psi =  = eta  = eta
-  //=> psi = MeeInv phi
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class SchurRedBlackDiagTwoSolve : public SchurRedBlackBase<Field> {
-  public:
-    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
-
-    /////////////////////////////////////////////////////
-    // Wrap the usual normal equations Schur trick
-    /////////////////////////////////////////////////////
-  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-      const bool _solnAsInitGuess = false)  
-    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
-
-    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
-      
-      Field   tmp(grid);
-      Field  Mtmp(grid);
-
-      pickCheckerboard(Even,src_e,src);
-      pickCheckerboard(Odd ,src_o,src);
-    
-      /////////////////////////////////////////////////////
-      // src_o = Mdag * (source_o - Moe MeeInv source_e)
-      /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
-
-      // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
-    }
-
-    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
-    {
-      GridBase *grid = _Matrix.RedBlackGrid();
-      GridBase *fgrid= _Matrix.Grid();
-
-      Field   sol_o_i(grid);
-      Field   tmp(grid);
-      Field   sol_e(grid);
-
-      ////////////////////////////////////////////////
-      // MooeeInv due to pecond
-      ////////////////////////////////////////////////
-      _Matrix.MooeeInv(sol_o,tmp);
-      sol_o_i = tmp;
-
-      ///////////////////////////////////////////////////
-      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-      ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e);   assert(  sol_e.Checkerboard() ==Even);
-     
-      setCheckerboard(sol,sol_e);    assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.Checkerboard() ==Odd );
-    };
-
-    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
-    {
-      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);
-    };
-    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
-    {
-      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
-    }
-  };
-
-  template<class Field> class NonHermitianSchurRedBlackDiagTwoSolve : public SchurRedBlackBase<Field> 
-  {
-    public:
-      typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
-
-      /////////////////////////////////////////////////////
-      // Wrap the usual normal equations Schur trick
-      /////////////////////////////////////////////////////
-      NonHermitianSchurRedBlackDiagTwoSolve(OperatorFunction<Field>& RBSolver, const bool initSubGuess = false,
-          const bool _solnAsInitGuess = false)  
-      : SchurRedBlackBase<Field>(RBSolver, initSubGuess, _solnAsInitGuess) {};
-
-      virtual void RedBlackSource(Matrix& _Matrix, const Field& src, Field& src_e, Field& src_o)
-      {
-        GridBase* grid  = _Matrix.RedBlackGrid();
-        GridBase* fgrid = _Matrix.Grid();
-
-        Field  tmp(grid);
-        Field Mtmp(grid);
-
-        pickCheckerboard(Even, src_e, src);
-        pickCheckerboard(Odd , src_o, src);
-      
-        /////////////////////////////////////////////////////
-        // src_o = Mdag * (source_o - Moe MeeInv source_e)
-        /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
-      }
-
-      virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
-      {
-        GridBase* grid  = _Matrix.RedBlackGrid();
-        GridBase* fgrid = _Matrix.Grid();
-
-        Field sol_o_i(grid);
-        Field     tmp(grid);
-        Field   sol_e(grid);
-
-        ////////////////////////////////////////////////
-        // MooeeInv due to pecond
-        ////////////////////////////////////////////////
-        _Matrix.MooeeInv(sol_o, tmp);
-        sol_o_i = tmp;
-
-        ///////////////////////////////////////////////////
-        // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
-        ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o_i, tmp);    assert(   tmp.Checkerboard() == Even );
-        tmp = src_e - tmp;              assert( src_e.Checkerboard() == Even );
-        _Matrix.MooeeInv(tmp, sol_e);   assert( sol_e.Checkerboard() == Even );
-       
-        setCheckerboard(sol, sol_e);    assert(   sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o_i);  assert( sol_o_i.Checkerboard() == Odd  );
-      };
-
-      virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
-      {
-        NonHermitianSchurDiagTwoOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o);
-      };
-
-      virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o,  std::vector<Field>& sol_o)
-      {
-        NonHermitianSchurDiagTwoOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o); 
-      }
-  };
-}
-
-#endif

Grid/allocator/AlignedAllocator.h

@@ -1,184 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/AlignedAllocator.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);
-
-template<typename _Tp>
-class alignedAllocator {
-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 alignedAllocator<_Tp1> other; };
-  alignedAllocator() throw() { }
-  alignedAllocator(const alignedAllocator&) throw() { }
-  template<typename _Tp1> alignedAllocator(const alignedAllocator<_Tp1>&) throw() { }
-  ~alignedAllocator() 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::CpuAllocate(bytes);
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-    return ptr;
-  }
-
-  void deallocate(pointer __p, size_type __n) 
-  { 
-    size_type bytes = __n * sizeof(_Tp);
-    profilerFree(bytes);
-    MemoryManager::CpuFree((void *)__p,bytes);
-  }
-
-  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
-  void construct(pointer __p, const _Tp& __val) { assert(0);};
-  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);
-    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);
-    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
-////////////////////////////////////////////////////////////////////////////////
-#ifdef ACCELERATOR_CSHIFT
-// Cshift on device
-template<class T> using cshiftAllocator = devAllocator<T>;
-#else
-// Cshift on host
-template<class T> using cshiftAllocator = std::allocator<T>;
-#endif
-
-template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
-template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,devAllocator<T> >;
-template<class T> using deviceVector  = std::vector<T,devAllocator<T> >;
-template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
-
-NAMESPACE_END(Grid);
-
-

Grid/allocator/Allocator.h

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

Grid/allocator/MemoryManager.cc

@@ -1,324 +0,0 @@
-#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;;
-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
-  
-}
-
-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
-  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
-  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

@@ -1,226 +0,0 @@
-/*************************************************************************************
-
-    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
-  static void NotifyDeletion(void * CpuPtr);
-
- public:
-  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

@@ -1,601 +0,0 @@
-#include <Grid/GridCore.h>
-#ifndef GRID_UVM
-
-#warning "Using explicit device memory copies"
-NAMESPACE_BEGIN(Grid);
-
-#define MAXLINE 512
-static char print_buffer [ MAXLINE ];
-
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-//#define dprintf(...) 
-
-
-////////////////////////////////////////////////////////////
-// 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);  assert((count==0)||(count==1));
-  return count;
-}
-void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
-{
-  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)
-{
-  assert(EntryPresent(CpuPtr));
-  auto AccCacheIterator = AccViewTable.find(CpuPtr);
-  assert(AccCacheIterator!=AccViewTable.end());
-  return AccCacheIterator;
-}
-void MemoryManager::EntryErase(uint64_t CpuPtr)
-{
-  auto AccCache = EntryLookup(CpuPtr);
-  AccViewTable.erase(CpuPtr);
-}
-void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
-{
-  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)
-{
-  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.
-  ///////////////////////////////////////////////////////////
-  assert(AccCache.state!=Empty);
-  
-  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-  assert(AccCache.accLock==0);
-  assert(AccCache.cpuLock==0);
-  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\n",(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.
-  ///////////////////////////////////////////////////////////////////////////
-  assert(AccCache.state!=Empty);
-  
-  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
-	  (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(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
-  }
-  //  uint64_t CpuPtr = AccCache.CpuPtr;
-  DeviceEvictions++;
-  //  EntryErase(CpuPtr);
-}
-void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
-{
-  assert(AccCache.state==AccDirty);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
-  assert(AccCache.AccPtr!=(uint64_t)NULL);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
-  acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
-  DeviceToHostBytes+=AccCache.bytes;
-  DeviceToHostXfer++;
-  AccCache.state=Consistent;
-}
-void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
-{
-  assert(AccCache.state==CpuDirty);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
-  if(AccCache.AccPtr==(uint64_t)NULL){
-    AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
-    DeviceBytes+=AccCache.bytes;
-  }
-  mprintf("MemoryManager: Clone %lx <- %lx\n",(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)
-{
-  assert(AccCache.state!=Empty);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
-  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\n",(uint64_t)Ptr);
-    AcceleratorViewClose((uint64_t)Ptr);
-  } else if( (mode==CpuRead)||(mode==CpuWrite)){
-    CpuViewClose((uint64_t)Ptr);
-  } else { 
-    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\n",(uint64_t)CpuPtr);
-    return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
-  } else if( (mode==CpuRead)||(mode==CpuWrite)){
-    return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
-  } else { 
-    assert(0);
-    return NULL;
-  }
-}
-void  MemoryManager::EvictVictims(uint64_t bytes)
-{
-  assert(bytes<DeviceMaxBytes);
-  while(bytes+DeviceLRUBytes > DeviceMaxBytes){
-    if ( DeviceLRUBytes > 0){
-      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); 
-  } 
-  assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
-
-  assert(AccCache.cpuLock==0);  // Programming error
-
-  if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
-		    (uint64_t)AccCache.CpuPtr,
-		    (uint64_t)CpuPtr,
-		    (uint64_t)AccCache.bytes,
-	            (uint64_t)bytes,
-		    (uint64_t)AccCache.accLock);
-    assert(AccCache.CpuPtr == CpuPtr);
-    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) {
-    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\n",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\n",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\n",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\n",AccCache.accLock);
-  } else {
-    assert(0);
-  }
-
-  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 \n");
-    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;
-
-  assert(AccCache.cpuLock==0);
-  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\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
-    LRUinsert(AccCache);
-  } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
-  }
-}
-void MemoryManager::CpuViewClose(uint64_t CpuPtr)
-{
-  auto AccCacheIterator = EntryLookup(CpuPtr);
-  auto & AccCache = AccCacheIterator->second;
-
-  assert(AccCache.cpuLock>0);
-  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);
-  //  }
-
-  assert((mode==CpuRead)||(mode==CpuWrite));
-  assert(AccCache.accLock==0);  // Programming error
-
-  if(AccCache.state!=Empty) {
-    assert(AccCache.CpuPtr == CpuPtr);
-    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) {
-    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) {
-    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 {
-    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;
-  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;
-    assert(EntryPresent(cpuPtr));
-    auto AccCacheIterator = EntryLookup(cpuPtr);
-    auto & AccCache = AccCacheIterator->second;
-    LruBytes2+=AccCache.bytes;
-    assert(AccCache.LRU_valid==1);
-    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 ) {
-      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;
-    }
-
-    assert( AccCache.cpuLock== 0 ) ;
-    assert( AccCache.accLock== 0 ) ;
-  }
-  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
-  assert(LruBytes1==LruBytes2);
-  assert(LruBytes1==DeviceLRUBytes);
-  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
-  assert(AccBytes==DeviceBytes);
-  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
-  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

@@ -1,31 +0,0 @@
-#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

@@ -1,67 +0,0 @@
-#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);
-  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/MemoryStats.h

@@ -1,95 +0,0 @@
-/*************************************************************************************
-
-    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

@@ -1,291 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/cartesian/Cartesian_base.h
-
-    Copyright (C) 2015
-
-    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Guido Cossu <guido.cossu@ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_CARTESIAN_BASE_H
-#define GRID_CARTESIAN_BASE_H
-
-NAMESPACE_BEGIN(Grid);
-
-//////////////////////////////////////////////////////////////////////
-// Commicator provides information on the processor grid
-//////////////////////////////////////////////////////////////////////
-//    unsigned long _ndimension;
-//    Coordinate _processors; // processor grid
-//    int              _processor;  // linear processor rank
-//    Coordinate _processor_coor;  // linear processor rank
-//////////////////////////////////////////////////////////////////////
-class GridBase : public CartesianCommunicator , public GridThread {
-
-public:
-  int dummy;
-  // Give Lattice access
-  template<class object> friend class Lattice;
-
-  GridBase(const Coordinate & processor_grid) : CartesianCommunicator(processor_grid) { LocallyPeriodic=0;}; 
-
-  GridBase(const Coordinate & processor_grid,
-	   const CartesianCommunicator &parent,
-	   int &split_rank) 
-    : CartesianCommunicator(processor_grid,parent,split_rank) {LocallyPeriodic=0;};
-
-  GridBase(const Coordinate & processor_grid,
-	   const CartesianCommunicator &parent) 
-    : CartesianCommunicator(processor_grid,parent,dummy) {LocallyPeriodic=0;};
-
-  virtual ~GridBase() = default;
-
-  // Physics Grid information.
-  Coordinate _simd_layout;// Which dimensions get relayed out over simd lanes.
-  Coordinate _fdimensions;// (full) Global dimensions of array prior to cb removal
-  Coordinate _gdimensions;// Global dimensions of array after cb removal
-  Coordinate _ldimensions;// local dimensions of array with processor images removed
-  Coordinate _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
-  Coordinate _ostride;    // Outer stride for each dimension
-  Coordinate _istride;    // Inner stride i.e. within simd lane
-  int _osites;                  // _isites*_osites = product(dimensions).
-  int _isites;
-  int _fsites;                  // _isites*_osites = product(dimensions).
-  int _gsites;
-  Coordinate _slice_block;// subslice information
-  Coordinate _slice_stride;
-  Coordinate _slice_nblock;
-
-  Coordinate _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
-  Coordinate _lend  ;     // local end of array in gcoors   _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
-
-  bool _isCheckerBoarded; 
-  int        LocallyPeriodic;
-  Coordinate _checker_dim_mask;
-
-public:
-
-  ////////////////////////////////////////////////////////////////
-  // Checkerboarding interface is virtual and overridden by 
-  // GridCartesian / GridRedBlackCartesian
-  ////////////////////////////////////////////////////////////////
-  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;
-  virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;
-  virtual int CheckerBoardFromOindex (int Oindex)=0;
-  virtual int CheckerBoardFromOindexTable (int Oindex)=0;
-
-  //////////////////////////////////////////////////////////////////////////////////////////////
-  // Local layout calculations
-  //////////////////////////////////////////////////////////////////////////////////////////////
-  // These routines are key. Subdivide the linearised cartesian index into
-  //      "inner" index identifying which simd lane of object<vFcomplex> is associated with coord
-  //      "outer" index identifying which element of _odata in class "Lattice" is associated with coord.
-  //
-  // Compared to, say, Blitz++ we simply need to store BOTH an inner stride and an outer
-  // stride per dimension. The cost of evaluating the indexing information is doubled for an n-dimensional
-  // coordinate. Note, however, for data parallel operations the "inner" indexing cost is not paid and all
-  // lanes are operated upon simultaneously.
-  
-  virtual int oIndex(Coordinate &coor)
-  {
-    int idx=0;
-    // Works with either global or local coordinates
-    for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
-    return idx;
-  }
-  virtual int iIndex(Coordinate &lcoor)
-  {
-    int idx=0;
-    for(int d=0;d<_ndimension;d++) idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
-    return idx;
-  }
-  inline int oIndexReduced(Coordinate &ocoor)
-  {
-    int idx=0; 
-    // ocoor is already reduced so can eliminate the modulo operation
-    // for fast indexing and inline the routine
-    for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*ocoor[d];
-    return idx;
-  }
-  inline void oCoorFromOindex (Coordinate& coor,int Oindex){
-    Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
-  }
-
-  inline void InOutCoorToLocalCoor (Coordinate &ocoor, Coordinate &icoor, Coordinate &lcoor) {
-    lcoor.resize(_ndimension);
-    for (int d = 0; d < _ndimension; d++)
-      lcoor[d] = ocoor[d] + _rdimensions[d] * icoor[d];
-  }
-
-  //////////////////////////////////////////////////////////
-  // SIMD lane addressing
-  //////////////////////////////////////////////////////////
-  inline void iCoorFromIindex(Coordinate &coor,int lane)
-  {
-    Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
-  }
-
-  inline int PermuteDim(int dimension){
-    return _simd_layout[dimension]>1;
-  }
-  inline int PermuteType(int dimension){
-    int permute_type=0;
-    //
-    // Best way to encode this would be to present a mask 
-    // for which simd dimensions are rotated, and the rotation
-    // size. If there is only one simd dimension rotated, this is just 
-    // a permute. 
-    //
-    // Cases: PermuteType == 1,2,4,8
-    // Distance should be either 0,1,2..
-    //
-    if ( _simd_layout[dimension] > 2 ) { 
-      for(int d=0;d<_ndimension;d++){
-	if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
-      }
-      permute_type = RotateBit; // How to specify distance; this is not just direction.
-      return permute_type;
-    }
-
-    for(int d=_ndimension-1;d>dimension;d--){
-      if (_simd_layout[d]>1 ) permute_type++;
-    }
-    return permute_type;
-  }
-  ////////////////////////////////////////////////////////////////
-  // Array sizing queries
-  ////////////////////////////////////////////////////////////////
-
-  inline int iSites(void) const { return _isites; };
-  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 int Nd    (void) const { return _ndimension;};
-
-  inline const Coordinate LocalStarts(void)             { return _lstart;    };
-  inline const Coordinate &FullDimensions(void)         { return _fdimensions;};
-  inline const Coordinate &GlobalDimensions(void)       { return _gdimensions;};
-  inline const Coordinate &LocalDimensions(void)        { return _ldimensions;};
-  inline const Coordinate &VirtualLocalDimensions(void) { return _ldimensions;};
-
-  ////////////////////////////////////////////////////////////////
-  // Utility to print the full decomposition details 
-  ////////////////////////////////////////////////////////////////
-
-  void show_decomposition(){
-    std::cout << GridLogMessage << "\tFull Dimensions    : " << _fdimensions << std::endl;
-    std::cout << GridLogMessage << "\tSIMD layout        : " << _simd_layout << std::endl;
-    std::cout << GridLogMessage << "\tGlobal Dimensions  : " << _gdimensions << std::endl;
-    std::cout << GridLogMessage << "\tLocal Dimensions   : " << _ldimensions << std::endl;
-    std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
-    std::cout << GridLogMessage << "\tOuter strides      : " << _ostride << std::endl;
-    std::cout << GridLogMessage << "\tInner strides      : " << _istride << std::endl;
-    std::cout << GridLogMessage << "\tiSites             : " << _isites << std::endl;
-    std::cout << GridLogMessage << "\toSites             : " << _osites << std::endl;
-    std::cout << GridLogMessage << "\tlSites             : " << lSites() << std::endl;        
-    std::cout << GridLogMessage << "\tgSites             : " << gSites() << std::endl;
-    std::cout << GridLogMessage << "\tNd                 : " << _ndimension << std::endl;             
-  } 
-
-  ////////////////////////////////////////////////////////////////
-  // Global addressing
-  ////////////////////////////////////////////////////////////////
-  void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
-    assert(gidx< gSites());
-    Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
-  }
-  void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
-    assert(lidx<lSites());
-    Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
-  }
-  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
-    gidx=0;
-    int mult=1;
-    for(int mu=0;mu<_ndimension;mu++) {
-      gidx+=mult*gcoor[mu];
-      mult*=_gdimensions[mu];
-    }
-  }
-  void GlobalCoorToProcessorCoorLocalCoor(Coordinate &pcoor,Coordinate &lcoor,const Coordinate &gcoor)
-  {
-    pcoor.resize(_ndimension);
-    lcoor.resize(_ndimension);
-    for(int mu=0;mu<_ndimension;mu++){
-      int _fld  = _fdimensions[mu]/_processors[mu];
-      pcoor[mu] = gcoor[mu]/_fld;
-      lcoor[mu] = gcoor[mu]%_fld;
-    }
-  }
-  void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const Coordinate &gcoor)
-  {
-    Coordinate pcoor;
-    Coordinate lcoor;
-    GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
-    rank = RankFromProcessorCoor(pcoor);
-    /*
-      Coordinate cblcoor(lcoor);
-      for(int d=0;d<cblcoor.size();d++){
-      if( this->CheckerBoarded(d) ) {
-      cblcoor[d] = lcoor[d]/2;
-      }
-      }
-    */
-    i_idx= iIndex(lcoor);
-    o_idx= oIndex(lcoor);
-  }
-
-  void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , Coordinate &gcoor)
-  {
-    gcoor.resize(_ndimension);
-    Coordinate coor(_ndimension);
-
-    ProcessorCoorFromRank(rank,coor);
-    for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = _ldimensions[mu]*coor[mu];
-
-    iCoorFromIindex(coor,i_idx);
-    for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += _rdimensions[mu]*coor[mu];
-
-    oCoorFromOindex (coor,o_idx);
-    for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += coor[mu];
-      
-  }
-  void RankIndexCbToFullGlobalCoor(int rank, int o_idx, int i_idx, int cb,Coordinate &fcoor)
-  {
-    RankIndexToGlobalCoor(rank,o_idx,i_idx ,fcoor);
-    if(CheckerBoarded(0)){
-      fcoor[0] = fcoor[0]*2+cb;
-    }
-  }
-  void ProcessorCoorLocalCoorToGlobalCoor(Coordinate &Pcoor,Coordinate &Lcoor,Coordinate &gcoor)
-  {
-    gcoor.resize(_ndimension);
-    for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = Pcoor[mu]*_ldimensions[mu]+Lcoor[mu];
-  }
-};
-
-NAMESPACE_END(Grid);
-#endif

Grid/cartesian/Cartesian_full.h

@@ -1,178 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/cartesian/Cartesian_full.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_CARTESIAN_FULL_H
-#define GRID_CARTESIAN_FULL_H
-
-NAMESPACE_BEGIN(Grid);
-    
-/////////////////////////////////////////////////////////////////////////////////////////
-// Grid Support.
-/////////////////////////////////////////////////////////////////////////////////////////
-
-class GridCartesian: public GridBase {
-
-public:
-  int dummy;
-  Coordinate _checker_dim_mask;
-  virtual int  CheckerBoardFromOindexTable (int Oindex) {
-    return 0;
-  }
-  virtual int  CheckerBoardFromOindex (int Oindex)
-  {
-    return 0;
-  }
-  virtual int CheckerBoarded(int dim){
-    return 0;
-  }
-  virtual int CheckerBoard(const Coordinate &site){
-    return 0;
-  }
-  virtual int CheckerBoardDestination(int cb,int shift,int dim){
-    return 0;
-  }
-  virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift, int ocb){
-    return shift;
-  }
-  virtual int CheckerBoardShift(int source_cb,int dim,int shift, int osite){
-    return shift;
-  }
-  /////////////////////////////////////////////////////////////////////////
-  // Constructor takes a parent grid and possibly subdivides communicator.
-  /////////////////////////////////////////////////////////////////////////
-  GridCartesian(const Coordinate &dimensions,
-		const Coordinate &simd_layout,
-		const Coordinate &processor_grid,
-		const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
-  {
-    Init(dimensions,simd_layout,processor_grid);
-  }
-  GridCartesian(const Coordinate &dimensions,
-		const Coordinate &simd_layout,
-		const Coordinate &processor_grid,
-		const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
-  {
-    Init(dimensions,simd_layout,processor_grid);
-  }
-  /////////////////////////////////////////////////////////////////////////
-  // Construct from comm world
-  /////////////////////////////////////////////////////////////////////////
-  GridCartesian(const Coordinate &dimensions,
-		const Coordinate &simd_layout,
-		const Coordinate &processor_grid) : GridBase(processor_grid)
-  {
-    Init(dimensions,simd_layout,processor_grid);
-  }
-
-  virtual ~GridCartesian() = default;
-
-  void Init(const Coordinate &dimensions,
-	    const Coordinate &simd_layout,
-	    const Coordinate &processor_grid)
-  {
-    ///////////////////////
-    // Grid information
-    ///////////////////////
-      _isCheckerBoarded = false;
-    _ndimension = dimensions.size();
-
-    _fdimensions.resize(_ndimension);
-    _gdimensions.resize(_ndimension);
-    _ldimensions.resize(_ndimension);
-    _rdimensions.resize(_ndimension);
-    _simd_layout.resize(_ndimension);
-    _checker_dim_mask.resize(_ndimension);;
-    _lstart.resize(_ndimension);
-    _lend.resize(_ndimension);
-
-    _ostride.resize(_ndimension);
-    _istride.resize(_ndimension);
-
-    _fsites = _gsites = _osites = _isites = 1;
-
-    for (int d = 0; d < _ndimension; d++)
-      {
-	_checker_dim_mask[d]=0;
-
-        _fdimensions[d] = dimensions[d];   // Global dimensions
-        _gdimensions[d] = _fdimensions[d]; // Global dimensions
-        _simd_layout[d] = simd_layout[d];
-        _fsites = _fsites * _fdimensions[d];
-        _gsites = _gsites * _gdimensions[d];
-
-        // Use a reduced simd grid
-        _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
-        //std::cout << _ldimensions[d] << "  " << _gdimensions[d] << "  " << _processors[d] << std::endl;
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
-
-        _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
-
-        _lstart[d] = _processor_coor[d] * _ldimensions[d];
-        _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
-        _osites *= _rdimensions[d];
-        _isites *= _simd_layout[d];
-
-        // Addressing support
-        if (d == 0)
-	  {
-	    _ostride[d] = 1;
-	    _istride[d] = 1;
-	  }
-        else
-	  {
-	    _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
-	    _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
-	  }
-      }
-
-    ///////////////////////
-    // subplane information
-    ///////////////////////
-    _slice_block.resize(_ndimension);
-    _slice_stride.resize(_ndimension);
-    _slice_nblock.resize(_ndimension);
-
-    int block = 1;
-    int nblock = 1;
-    for (int d = 0; d < _ndimension; d++)
-      nblock *= _rdimensions[d];
-
-    for (int d = 0; d < _ndimension; d++)
-      {
-        nblock /= _rdimensions[d];
-        _slice_block[d] = block;
-        _slice_stride[d] = _ostride[d] * _rdimensions[d];
-        _slice_nblock[d] = nblock;
-        block = block * _rdimensions[d];
-      }
-  };
-
-};
-
-NAMESPACE_END(Grid);
-#endif

Grid/cartesian/Cartesian_red_black.h

@@ -1,305 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/cartesian/Cartesian_red_black.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_CARTESIAN_RED_BLACK_H
-#define GRID_CARTESIAN_RED_BLACK_H
-
-NAMESPACE_BEGIN(Grid);
-
-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;
-  std::vector<int> _checker_board;
-
-  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);
-    for(int d=0;d<_ndimension;d++){ 
-      if(_checker_dim_mask[d])
-	linear=linear+site[d];
-    }
-    return (linear&0x1);
-  }
-
-  // Depending on the cb of site, we toggle source cb.
-  // for block #b, element #e = (b, e)
-  // we need 
-  virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int ocb){
-    if(dim != _checker_dim) return shift;
-
-    int fulldim =_fdimensions[dim];
-    shift = (shift+fulldim)%fulldim;
-
-    // Probably faster with table lookup;
-    // or by looping over x,y,z and multiply rather than computing checkerboard.
-	  
-    if ( (source_cb+ocb)&1 ) {
-      return (shift)/2;
-    } else {
-      return (shift+1)/2;
-    }
-  }
-  virtual int  CheckerBoardFromOindexTable (int Oindex) {
-    return _checker_board[Oindex];
-  }
-  virtual int  CheckerBoardFromOindex (int Oindex)
-  {
-    Coordinate ocoor;
-    oCoorFromOindex(ocoor,Oindex);
-    return CheckerBoard(ocoor);
-  }
-  virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite){
-
-    if(dim != _checker_dim) return shift;
-
-    int ocb=CheckerBoardFromOindex(osite);
-      
-    return CheckerBoardShiftForCB(source_cb,dim,shift,ocb);
-  }
-    
-  virtual int CheckerBoardDestination(int source_cb,int shift,int dim){
-    if ( _checker_dim_mask[dim]  ) {
-      // If _fdimensions[checker_dim] is odd, then shifting by 1 in other dims
-      // does NOT cause a parity hop.
-      int add=(dim==_checker_dim) ? 0 : _fdimensions[_checker_dim];
-      if ( (shift+add) &0x1) {
-	return 1-source_cb;
-      } else {
-	return source_cb;
-      }
-    } else {
-      return source_cb;
-
-    }
-  };
-
-  ////////////////////////////////////////////////////////////
-  // Create Redblack from original grid; require full grid pointer ?
-  ////////////////////////////////////////////////////////////
-  GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
-  {
-    int dims = base->_ndimension;
-    Coordinate checker_dim_mask(dims,1);
-    int checker_dim = 0;
-    Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
-  };
-
-  ////////////////////////////////////////////////////////////
-  // Create redblack from original grid, with non-trivial checker dim mask
-  ////////////////////////////////////////////////////////////
-  GridRedBlackCartesian(const GridBase *base,
-			const Coordinate &checker_dim_mask,
-			int checker_dim
-			) :  GridBase(base->_processors,*base) 
-  {
-    Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim)  ;
-  }
-
-  virtual ~GridRedBlackCartesian() = default;
-
-  void Init(const Coordinate &dimensions,
-	    const Coordinate &simd_layout,
-	    const Coordinate &processor_grid,
-	    const Coordinate &checker_dim_mask,
-	    int checker_dim)
-  {
-
-      _isCheckerBoarded = true;
-    _checker_dim = checker_dim;
-    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);
-
-    _fdimensions.resize(_ndimension);
-    _gdimensions.resize(_ndimension);
-    _ldimensions.resize(_ndimension);
-    _rdimensions.resize(_ndimension);
-    _simd_layout.resize(_ndimension);
-    _lstart.resize(_ndimension);
-    _lend.resize(_ndimension);
-
-    _ostride.resize(_ndimension);
-    _istride.resize(_ndimension);
-
-    _fsites = _gsites = _osites = _isites = 1;
-
-    _checker_dim_mask = checker_dim_mask;
-
-    for (int d = 0; d < _ndimension; d++)
-      {
-        _fdimensions[d] = dimensions[d];
-        _gdimensions[d] = _fdimensions[d];
-        _fsites = _fsites * _fdimensions[d];
-        _gsites = _gsites * _gdimensions[d];
-
-        if (d == _checker_dim)
-	  {
-	    assert((_gdimensions[d] & 0x1) == 0);
-	    _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
-	    _gsites /= 2;
-	  }
-        _ldimensions[d] = _gdimensions[d] / _processors[d];
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
-        _lstart[d] = _processor_coor[d] * _ldimensions[d];
-        _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
-
-        // Use a reduced simd grid
-        _simd_layout[d] = simd_layout[d];
-        _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
-        assert(_rdimensions[d] > 0);
-
-        // all elements of a simd vector must have same checkerboard.
-        // If Ls vectorised, this must still be the case; e.g. dwf rb5d
-        if (_simd_layout[d] > 1)
-	  {
-	    if (checker_dim_mask[d])
-	      {
-		assert((_rdimensions[d] & 0x1) == 0);
-	      }
-	  }
-
-        _osites *= _rdimensions[d];
-        _isites *= _simd_layout[d];
-
-        // Addressing support
-        if (d == 0)
-	  {
-	    _ostride[d] = 1;
-	    _istride[d] = 1;
-	  }
-        else
-	  {
-	    _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
-	    _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
-	  }
-      }
-
-    ////////////////////////////////////////////////////////////////////////////////////////////
-    // subplane information
-    ////////////////////////////////////////////////////////////////////////////////////////////
-    _slice_block.resize(_ndimension);
-    _slice_stride.resize(_ndimension);
-    _slice_nblock.resize(_ndimension);
-
-    int block = 1;
-    int nblock = 1;
-    for (int d = 0; d < _ndimension; d++)
-      nblock *= _rdimensions[d];
-
-    for (int d = 0; d < _ndimension; d++)
-      {
-        nblock /= _rdimensions[d];
-        _slice_block[d] = block;
-        _slice_stride[d] = _ostride[d] * _rdimensions[d];
-        _slice_nblock[d] = nblock;
-        block = block * _rdimensions[d];
-      }
-
-    ////////////////////////////////////////////////
-    // Create a checkerboard lookup table
-    ////////////////////////////////////////////////
-    int rvol = 1;
-    for (int d = 0; d < _ndimension; d++)
-      {
-        rvol = rvol * _rdimensions[d];
-      }
-    _checker_board.resize(rvol);
-    for (int osite = 0; osite < _osites; osite++)
-      {
-        _checker_board[osite] = CheckerBoardFromOindex(osite);
-      }
-  };
-
-protected:
-  virtual int oIndex(Coordinate &coor)
-  {
-    int idx = 0;
-    for (int d = 0; d < _ndimension; d++)
-      {
-        if (d == _checker_dim)
-	  {
-	    idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
-	  }
-        else
-	  {
-	    idx += _ostride[d] * (coor[d] % _rdimensions[d]);
-	  }
-      }
-    return idx;
-  };
-
-  virtual int iIndex(Coordinate &lcoor)
-  {
-    int idx = 0;
-    for (int d = 0; d < _ndimension; d++)
-      {
-        if (d == _checker_dim)
-	  {
-	    idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
-	  }
-        else
-	  {
-	    idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
-	  }
-      }
-    return idx;
-  }
-};
-NAMESPACE_END(Grid);
-#endif

Grid/communicator/Communicator_base.cc

@@ -1,79 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/Communicator_none.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/GridCore.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <limits.h>
-#include <sys/mman.h>
-
-NAMESPACE_BEGIN(Grid);
-
-bool Stencil_force_mpi = true;
-
-///////////////////////////////////////////////////////////////
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////
-CartesianCommunicator::CommunicatorPolicy_t  
-CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
-int CartesianCommunicator::nCommThreads = -1;
-
-/////////////////////////////////
-// Grid information queries
-/////////////////////////////////
-int                      CartesianCommunicator::Dimensions(void)        { return _ndimension; };
-int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
-int                      CartesianCommunicator::BossRank(void)          { return 0; };
-int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
-const Coordinate & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
-const Coordinate & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
-int                      CartesianCommunicator::ProcessorCount(void)    { return _Nprocessors; };
-
-////////////////////////////////////////////////////////////////////////////////
-// very VERY rarely (Log, serial RNG) we need world without a grid
-////////////////////////////////////////////////////////////////////////////////
-
-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);
-}
-void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
-{
-  GlobalSumVector((double *)c,2*N);
-}
-  
-NAMESPACE_END(Grid);
-
-

Grid/communicator/Communicator_mpi3.cc

@@ -1,495 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/communicator/Communicator_mpi.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/GridCore.h>
-#include <Grid/communicator/SharedMemory.h>
-
-NAMESPACE_BEGIN(Grid);
-
-Grid_MPI_Comm       CartesianCommunicator::communicator_world;
-
-////////////////////////////////////////////
-// First initialise of comms system
-////////////////////////////////////////////
-void CartesianCommunicator::Init(int *argc, char ***argv)
-{
-
-  int flag;
-  int provided;
-
-  MPI_Initialized(&flag); // needed to coexist with other libs apparently
-  if ( !flag ) {
-
-#ifndef GRID_COMMS_THREADS
-    nCommThreads=1;
-    // wrong results here too
-    // For now: comms-overlap leads to wrong results in Benchmark_wilson even on single node MPI runs
-    // other comms schemes are ok
-    MPI_Init_thread(argc,argv,MPI_THREAD_SERIALIZED,&provided);
-#else
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
-#endif
-    //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
-    if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
-      assert(0);
-    }
-
-    if( (nCommThreads > 1) && (provided != MPI_THREAD_MULTIPLE) ) {
-      assert(0);
-    }
-  }
-
-  // Never clean up as done once.
-  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
-
-  Grid_quiesce_nodes();
-  GlobalSharedMemory::Init(communicator_world);
-  GlobalSharedMemory::SharedMemoryAllocate(
-		   GlobalSharedMemory::MAX_MPI_SHM_BYTES,
-		   GlobalSharedMemory::Hugepages);
-  Grid_unquiesce_nodes();
-}
-
-///////////////////////////////////////////////////////////////////////////
-// Use cartesian communicators now even in MPI3
-///////////////////////////////////////////////////////////////////////////
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
-{
-  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(ierr==0);
-}
-int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor)
-{
-  int rank;
-  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  assert(ierr==0);
-  return rank;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
-{
-  coor.resize(_ndimension);
-  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-  assert(ierr==0);
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Initialises from communicator_world
-////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
-{
-  MPI_Comm optimal_comm;
-  ////////////////////////////////////////////////////
-  // Remap using the shared memory optimising routine
-  // The remap creates a comm which must be freed
-  ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
-  InitFromMPICommunicator(processors,optimal_comm);
-  SetCommunicator(optimal_comm);
-  ///////////////////////////////////////////////////
-  // Free the temp communicator
-  ///////////////////////////////////////////////////
-  MPI_Comm_free(&optimal_comm);
-}
-
-//////////////////////////////////
-// Try to subdivide communicator
-//////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
-{
-  _ndimension = processors.size();  assert(_ndimension>=1);
-  int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
-  Coordinate parent_processor_coor(_ndimension,0);
-  Coordinate parent_processors    (_ndimension,1);
-  Coordinate shm_processors       (_ndimension,1);
-  // Can make 5d grid from 4d etc...
-  int pad = _ndimension-parent_ndimension;
-  for(int d=0;d<parent_ndimension;d++){
-    parent_processor_coor[pad+d]=parent._processor_coor[d];
-    parent_processors    [pad+d]=parent._processors[d];
-    shm_processors       [pad+d]=parent._shm_processors[d];
-  }
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // split the communicator
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  //  int Nparent = parent._processors ;
-  int Nparent;
-  MPI_Comm_size(parent.communicator,&Nparent);
-
-  int childsize=1;
-  for(int d=0;d<processors.size();d++) {
-    childsize *= processors[d];
-  }
-  int Nchild = Nparent/childsize;
-  assert (childsize * Nchild == Nparent);
-
-  Coordinate ccoor(_ndimension); // coor within subcommunicator
-  Coordinate scoor(_ndimension); // coor of split within parent
-  Coordinate ssize(_ndimension); // coor of split within parent
-
-  for(int d=0;d<_ndimension;d++){
-    ccoor[d] = parent_processor_coor[d] % processors[d];
-    scoor[d] = parent_processor_coor[d] / processors[d];
-    ssize[d] = parent_processors[d]     / processors[d];
-    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
-  }
-
-  // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
-  int crank;
-  // Mpi uses the reverse Lexico convention to us; so reversed routines called
-  Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
-  Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
-
-  MPI_Comm comm_split;
-  if ( Nchild > 1 ) {
-
-    ////////////////////////////////////////////////////////////////
-    // Split the communicator
-    ////////////////////////////////////////////////////////////////
-    int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
-    assert(ierr==0);
-
-  } else {
-    srank = 0;
-    int ierr = MPI_Comm_dup (parent.communicator,&comm_split);
-    assert(ierr==0);
-  }
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Set up from the new split communicator
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  InitFromMPICommunicator(processors,comm_split);
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Take the right SHM buffers
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  SetCommunicator(comm_split);
-
-  ///////////////////////////////////////////////
-  // Free the temp communicator
-  ///////////////////////////////////////////////
-  MPI_Comm_free(&comm_split);
-
-  if(0){
-    std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
-    for(int d=0;d<processors.size();d++){
-      std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
-    }
-  }
-  for(int d=0;d<processors.size();d++){
-    assert(_processor_coor[d] == ccoor[d] );
-  }
-}
-
-void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors, MPI_Comm communicator_base)
-{
-  ////////////////////////////////////////////////////
-  // Creates communicator, and the communicator_halo
-  ////////////////////////////////////////////////////
-  _ndimension = processors.size();
-  _processor_coor.resize(_ndimension);
-
-  /////////////////////////////////
-  // Count the requested nodes
-  /////////////////////////////////
-  _Nprocessors=1;
-  _processors = processors;
-  for(int i=0;i<_ndimension;i++){
-    _Nprocessors*=_processors[i];
-  }
-
-  Coordinate periodic(_ndimension,1);
-  MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
-  MPI_Comm_rank(communicator,&_processor);
-  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
-
-  if ( 0 && (communicator_base != communicator_world) ) {
-    std::cout << "InitFromMPICommunicator Cartesian communicator created with a non-world communicator"<<std::endl;
-    std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";
-    for(int d=0;d<_processors.size();d++){
-      std::cout << _processor_coor[d]<<" ";
-    }
-    std::cout << std::endl;
-  }
-
-  int Size;
-  MPI_Comm_size(communicator,&Size);
-
-  communicator_halo.resize (2*_ndimension);
-  for(int i=0;i<_ndimension*2;i++){
-    MPI_Comm_dup(communicator,&communicator_halo[i]);
-  }
-  assert(Size==_Nprocessors);
-}
-
-CartesianCommunicator::~CartesianCommunicator()
-{
-  int MPI_is_finalised;
-  MPI_Finalized(&MPI_is_finalised);
-  if (communicator && !MPI_is_finalised) {
-    MPI_Comm_free(&communicator);
-    for(int i=0;i<communicator_halo.size();i++){
-      MPI_Comm_free(&communicator_halo[i]);
-    }
-  }
-}
-void CartesianCommunicator::GlobalSum(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(float &f)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(double *d,int N)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFrom(void *xmit,
-					   int dest,
-					   void *recv,
-					   int from,
-					   int bytes)
-{
-  std::vector<CommsRequest_t> reqs(0);
-  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
-
-  int myrank = _processor;
-  int ierr;
-
-  // Enforce no UVM in comms, device or host OK
-  assert(acceleratorIsCommunicable(xmit));
-  assert(acceleratorIsCommunicable(recv));
-
-  // Give the CPU to MPI immediately; can use threads to overlap optionally
-  //  printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
-  ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		    recv,bytes,MPI_CHAR,from, from,
-		    communicator,MPI_STATUS_IGNORE);
-  assert(ierr==0);
-
-  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
-}
-// Basic Halo comms primitive
-double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest, int dox,
-						     void *recv,
-						     int from, int dor,
-						     int bytes,int dir)
-{
-  std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
-  StencilSendToRecvFromComplete(list,dir);
-  return offbytes;
-}
-
-#undef NVLINK_GET // Define to use get instead of put DMA
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,
-							 int dest,int dox,
-							 void *recv,
-							 int from,int dor,
-							 int xbytes,int rbytes,int dir)
-{
-  int ncomm  =communicator_halo.size();
-  int commdir=dir%ncomm;
-
-  MPI_Request xrq;
-  MPI_Request rrq;
-
-  int ierr;
-  int gdest = ShmRanks[dest];
-  int gfrom = ShmRanks[from];
-  int gme   = ShmRanks[_processor];
-
-  assert(dest != _processor);
-  assert(from != _processor);
-  assert(gme  == ShmRank);
-  double off_node_bytes=0.0;
-  int tag;
-
-  if ( dor ) {
-    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+from*32;
-      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-      assert(ierr==0);
-      list.push_back(rrq);
-      off_node_bytes+=rbytes;
-    }
-#ifdef NVLINK_GET
-      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
-      assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
-#endif
-  }
-  
-  if (dox) {
-    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
-      list.push_back(xrq);
-      off_node_bytes+=xbytes;
-    } else {
-#ifndef NVLINK_GET
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
-#endif
-      
-    }
-  }
-
-  return off_node_bytes;
-}
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
-{
-  int nreq=list.size();
-
-  acceleratorCopySynchronise();
-
-  if (nreq==0) return;
-
-  std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
-  list.resize(0);
-}
-void CartesianCommunicator::StencilBarrier(void)
-{
-  MPI_Barrier  (ShmComm);
-}
-//void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-//{
-//}
-void CartesianCommunicator::Barrier(void)
-{
-  int ierr = MPI_Barrier(communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
-{
-  int ierr=MPI_Bcast(data,
-		     bytes,
-		     MPI_BYTE,
-		     root,
-		     communicator);
-  assert(ierr==0);
-}
-int CartesianCommunicator::RankWorld(void){
-  int r;
-  MPI_Comm_rank(communicator_world,&r);
-  return r;
-}
-void CartesianCommunicator::BarrierWorld(void){
-  int ierr = MPI_Barrier(communicator_world);
-  assert(ierr==0);
-}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-  int ierr= MPI_Bcast(data,
-		      bytes,
-		      MPI_BYTE,
-		      root,
-		      communicator_world);
-  assert(ierr==0);
-}
-
-void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
-{
-  Coordinate row(_ndimension,1);
-  assert(dim>=0 && dim<_ndimension);
-
-  //  Split the communicator
-  row[dim] = _processors[dim];
-
-  int me;
-  CartesianCommunicator Comm(row,*this,me);
-  Comm.AllToAll(in,out,words,bytes);
-}
-void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
-{
-  // MPI is a pain and uses "int" arguments
-  // 64*64*64*128*16 == 500Million elements of data.
-  // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
-  // (Turns up on 32^3 x 64 Gparity too)
-  MPI_Datatype object;
-  int iwords;
-  int ibytes;
-  iwords = words;
-  ibytes = bytes;
-  assert(words == iwords); // safe to cast to int ?
-  assert(bytes == ibytes); // safe to cast to int ?
-  MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
-  MPI_Type_commit(&object);
-  MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
-  MPI_Type_free(&object);
-}
-
-NAMESPACE_END(Grid);

Grid/communicator/SharedMemory.cc

@@ -1,172 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/SharedMemory.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#include <Grid/GridCore.h>
-
-NAMESPACE_BEGIN(Grid); 
-
-// static data
-
-int                 GlobalSharedMemory::HPEhypercube = 1;
-uint64_t            GlobalSharedMemory::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
-int                 GlobalSharedMemory::Hugepages = 0;
-int                 GlobalSharedMemory::_ShmSetup;
-int                 GlobalSharedMemory::_ShmAlloc;
-uint64_t            GlobalSharedMemory::_ShmAllocBytes;
-
-std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
-#ifndef ACCELERATOR_AWARE_MPI
-void * GlobalSharedMemory::HostCommBuf;
-#endif
-
-Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm;
-int                 GlobalSharedMemory::WorldShmRank;
-int                 GlobalSharedMemory::WorldShmSize;
-std::vector<int>    GlobalSharedMemory::WorldShmRanks;
-
-Grid_MPI_Comm       GlobalSharedMemory::WorldComm;
-int                 GlobalSharedMemory::WorldSize;
-int                 GlobalSharedMemory::WorldRank;
-
-int                 GlobalSharedMemory::WorldNodes;
-int                 GlobalSharedMemory::WorldNode;
-
-void GlobalSharedMemory::SharedMemoryFree(void)
-{
-  assert(_ShmAlloc);
-  assert(_ShmAllocBytes>0);
-  for(int r=0;r<WorldShmSize;r++){
-    munmap(WorldShmCommBufs[r],_ShmAllocBytes);
-  }
-  _ShmAlloc = 0;
-  _ShmAllocBytes = 0;
-}
-/////////////////////////////////
-// Alloc, free shmem region
-/////////////////////////////////
-#ifndef ACCELERATOR_AWARE_MPI
-void *SharedMemory::HostBufferMalloc(size_t bytes){
-  void *ptr = (void *)host_heap_top;
-  host_heap_top  += bytes;
-  host_heap_bytes+= bytes;
-  if (host_heap_bytes >= host_heap_size) {
-    std::cout<< " HostBufferMalloc exceeded heap size -- try increasing with --shm <MB> flag" <<std::endl;
-    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
-    std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
-    std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
-    std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
-    assert(host_heap_bytes<host_heap_size);
-  }
-  return ptr;
-}
-void SharedMemory::HostBufferFreeAll(void) { 
-  host_heap_top  =(size_t)HostCommBuf;
-  host_heap_bytes=0;
-}
-#endif
-void *SharedMemory::ShmBufferMalloc(size_t bytes){
-  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
-  void *ptr = (void *)heap_top;
-  heap_top  += bytes;
-  heap_bytes+= bytes;
-  if (heap_bytes >= heap_size) {
-    std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
-    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
-    std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
-    std::cout<< " Current bytes is " << (heap_bytes/(1024*1024)) <<"MB"<<std::endl;
-    std::cout<< " Current heap  is " << (heap_size/(1024*1024)) <<"MB"<<std::endl;
-    assert(heap_bytes<heap_size);
-  }
-  //std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;
-  return ptr;
-}
-void SharedMemory::ShmBufferFreeAll(void) { 
-  heap_top  =(size_t)ShmBufferSelf();
-  heap_bytes=0;
-}
-void *SharedMemory::ShmBufferSelf(void)
-{
-  //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
-  return ShmCommBufs[ShmRank];
-}
-static inline int divides(int a,int b)
-{
-  return ( b == ( (b/a)*a ) );
-}
-void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
-{
-  ////////////////////////////////////////////////////////////////
-  // Allow user to configure through environment variable
-  ////////////////////////////////////////////////////////////////
-  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
-  if ( str ) {
-    std::vector<int> IntShmDims;
-    GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
-    long ShmSize = 1;
-    for (int dim=0;dim<WorldDims.size();dim++) {
-      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
-    }
-    assert(ShmSize == WorldShmSize);
-    return;
-  }
-  
-  ////////////////////////////////////////////////////////////////
-  // Powers of 2,3,5 only in prime decomposition for now
-  ////////////////////////////////////////////////////////////////
-  int ndimension = WorldDims.size();
-  ShmDims=Coordinate(ndimension,1);
-
-  std::vector<int> primes({2,3,5});
-
-  int dim = 0;
-  int last_dim = ndimension - 1;
-  int AutoShmSize = 1;
-  while(AutoShmSize != WorldShmSize) {
-    int p;
-    for(p=0;p<primes.size();p++) {
-      int prime=primes[p];
-      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
-        && divides(prime,WorldShmSize/AutoShmSize)  ) {
-  AutoShmSize*=prime;
-  ShmDims[dim]*=prime;
-  last_dim = dim;
-  break;
-      }
-    }
-    if (p == primes.size() && last_dim == dim) {
-      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
-      exit(EXIT_FAILURE);
-    }
-    dim=(dim+1) %ndimension;
-  }
-}
-
-NAMESPACE_END(Grid); 
-

Grid/communicator/SharedMemory.h

@@ -1,176 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/SharedMemory.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once 
-
-#include <Grid/GridCore.h>
-
-#if defined (GRID_COMMS_MPI3) 
-#include <mpi.h>
-#endif 
-#include <semaphore.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <limits.h>
-#include <sys/types.h>
-#include <sys/ipc.h>
-#include <sys/shm.h>
-#include <sys/mman.h>
-#include <zlib.h>
-
-NAMESPACE_BEGIN(Grid);
-
-#if defined (GRID_COMMS_MPI3) 
-typedef MPI_Comm    Grid_MPI_Comm;
-typedef MPI_Request CommsRequest_t;
-#else 
-typedef int CommsRequest_t;
-typedef int Grid_MPI_Comm;
-#endif
-
-class GlobalSharedMemory {
-private:
-  static const int     MAXLOG2RANKSPERNODE = 16;            
-
-
-  // Init once lock on the buffer allocation
-  static int      _ShmSetup;
-  static int      _ShmAlloc;
-  static uint64_t _ShmAllocBytes;
-
-public:
-  ///////////////////////////////////////
-  // HPE 8600 hypercube optimisation
-  ///////////////////////////////////////
-  static int HPEhypercube;
-
-  static int      ShmSetup(void)      { return _ShmSetup; }
-  static int      ShmAlloc(void)      { return _ShmAlloc; }
-  static uint64_t ShmAllocBytes(void) { return _ShmAllocBytes; }
-  static uint64_t      MAX_MPI_SHM_BYTES;
-  static int           Hugepages;
-
-  static std::vector<void *> WorldShmCommBufs;
-#ifndef ACCELERATOR_AWARE_MPI
-  static void *HostCommBuf;
-#endif
-  static Grid_MPI_Comm WorldComm;
-  static int           WorldRank;
-  static int           WorldSize;
-
-  static Grid_MPI_Comm WorldShmComm;
-  static int           WorldShmRank;
-  static int           WorldShmSize;
-
-  static int           WorldNodes;
-  static int           WorldNode;
-
-  static std::vector<int>  WorldShmRanks;
-
-  //////////////////////////////////////////////////////////////////////////////////////
-  // Create an optimal reordered communicator that makes MPI_Cart_create get it right
-  //////////////////////////////////////////////////////////////////////////////////////
-  static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
-  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
-  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
-  static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
-  ///////////////////////////////////////////////////
-  // Provide shared memory facilities off comm world
-  ///////////////////////////////////////////////////
-  static void SharedMemoryAllocate(uint64_t bytes, int flags);
-  static void SharedMemoryFree(void);
-  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
-  static void SharedMemoryZero(void *dest,size_t bytes);
-
-};
-
-//////////////////////////////
-// one per communicator
-//////////////////////////////
-class SharedMemory 
-{
-private:
-  static const int     MAXLOG2RANKSPERNODE = 16;            
-
-  size_t heap_top;
-  size_t heap_bytes;
-  size_t heap_size;
-
-#ifndef ACCELERATOR_AWARE_MPI
-  size_t host_heap_top;  // set in free all
-  size_t host_heap_bytes;// set in free all
-  void *HostCommBuf;     // set in SetCommunicator
-  size_t host_heap_size; // set in SetCommunicator
-#endif
-  
-protected:
-
-  Grid_MPI_Comm    ShmComm; // for barriers
-  int    ShmRank; 
-  int    ShmSize;
-  std::vector<void *> ShmCommBufs;
-  std::vector<int>    ShmRanks;// Mapping comm ranks to Shm ranks
-
-public:
-  SharedMemory() {};
-  ~SharedMemory();
-  ///////////////////////////////////////////////////////////////////////////////////////
-  // set the buffers & sizes
-  ///////////////////////////////////////////////////////////////////////////////////////
-  void SetCommunicator(Grid_MPI_Comm comm);
-
-  ////////////////////////////////////////////////////////////////////////
-  // For this instance ; disjoint buffer sets between splits if split grid
-  ////////////////////////////////////////////////////////////////////////
-  void ShmBarrier(void); 
-
-  ///////////////////////////////////////////////////
-  // Call on any instance
-  ///////////////////////////////////////////////////
-  void SharedMemoryTest(void);
-  
-  void *ShmBufferSelf(void);
-  void *ShmBuffer    (int rank);
-  void *ShmBufferTranslate(int rank,void * local_p);
-  void *ShmBufferMalloc(size_t bytes);
-  void  ShmBufferFreeAll(void) ;
-#ifndef ACCELERATOR_AWARE_MPI
-  void *HostBufferMalloc(size_t bytes);
-  void HostBufferFreeAll(void);
-#endif  
-  //////////////////////////////////////////////////////////////////////////
-  // Make info on Nodes & ranks and Shared memory available
-  //////////////////////////////////////////////////////////////////////////
-  int NodeCount(void) { return GlobalSharedMemory::WorldNodes;};
-  int RankCount(void) { return GlobalSharedMemory::WorldSize;};
-
-};
-
-NAMESPACE_END(Grid);
-

Grid/communicator/SharedMemoryNone.cc

@@ -1,171 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/SharedMemory.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#include <Grid/GridCore.h>
-
-NAMESPACE_BEGIN(Grid); 
-#define header "SharedMemoryNone: "
-
-/*Construct from an MPI communicator*/
-void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
-{
-  assert(_ShmSetup==0);
-  WorldComm = 0;
-  WorldRank = 0;
-  WorldSize = 1;
-  WorldShmComm = 0 ;
-  WorldShmRank = 0 ;
-  WorldShmSize = 1 ;
-  WorldNodes   = 1 ;
-  WorldNode    = 0 ;
-  WorldShmRanks.resize(WorldSize); WorldShmRanks[0] = 0;
-  WorldShmCommBufs.resize(1);
-  _ShmSetup=1;
-}
-
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
-{
-  optimal_comm = WorldComm;
-  SHM = Coordinate(processors.size(),1);
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////
-// Hugetlbfs mapping intended, use anonymous mmap
-////////////////////////////////////////////////////////////////////////////////////////////
-#if 1
-void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
-{
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl;
-  void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Each MPI rank should allocate our own buffer
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
-
-  if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryNone.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
-    exit(EXIT_FAILURE);  
-  }
-  if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryNone.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
-  }
-  SharedMemoryZero(ShmCommBuf,bytes);
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Loop over ranks/gpu's on our node
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  WorldShmCommBufs[0] = ShmCommBuf;
-
-  _ShmAllocBytes=bytes;
-  _ShmAlloc=1;
-}
-#else
-void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
-{
-  void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
-  int mmap_flag =0;
-#ifdef MAP_ANONYMOUS
-  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
-#endif
-#ifdef MAP_ANON
-  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
-#endif
-#ifdef MAP_HUGETLB
-  if ( flags ) mmap_flag |= MAP_HUGETLB;
-#endif
-  ShmCommBuf =(void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag, -1, 0); 
-  if (ShmCommBuf == (void *)MAP_FAILED) {
-    perror("mmap failed ");
-    exit(EXIT_FAILURE);  
-  }
-#ifdef MADV_HUGEPAGE
-  if (!Hugepages ) madvise(ShmCommBuf,bytes,MADV_HUGEPAGE);
-#endif
-  bzero(ShmCommBuf,bytes);
-  WorldShmCommBufs[0] = ShmCommBuf;
-  _ShmAllocBytes=bytes;
-  _ShmAlloc=1;
-};
-#endif
-void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
-{
-  acceleratorMemSet(dest,0,bytes);
-}
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
-  acceleratorCopyToDevice(src,dest,bytes);
-}
-////////////////////////////////////////////////////////
-// Global shared functionality finished
-// Now move to per communicator functionality
-////////////////////////////////////////////////////////
-void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
-{
-  assert(GlobalSharedMemory::ShmAlloc()==1);
-  ShmRanks.resize(1);
-  ShmCommBufs.resize(1);
-  ShmRanks[0] = 0;
-  ShmRank     = 0;
-  ShmSize     = 1;
-  //////////////////////////////////////////////////////////////////////
-  // Map ShmRank to WorldShmRank and use the right buffer
-  //////////////////////////////////////////////////////////////////////
-  ShmCommBufs[0] = GlobalSharedMemory::WorldShmCommBufs[0];
-  heap_size      = GlobalSharedMemory::ShmAllocBytes();
-  ShmBufferFreeAll();
-  return;
-}
-//////////////////////////////////////////////////////////////////
-// On node barrier
-//////////////////////////////////////////////////////////////////
-void SharedMemory::ShmBarrier(void){ return ; }
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Test the shared memory is working
-//////////////////////////////////////////////////////////////////////////////////////////////////////////
-void SharedMemory::SharedMemoryTest(void) { return; }
-
-void *SharedMemory::ShmBuffer(int rank)
-{
-  return NULL;
-}
-void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
-{
-  return NULL;
-}
-SharedMemory::~SharedMemory()
-{};
-
-NAMESPACE_END(Grid); 
-

Grid/cshift/Cshift_common.h

@@ -1,515 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/cshift/Cshift_common.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
-
-NAMESPACE_BEGIN(Grid);
-
-extern std::vector<std::pair<int,int> > Cshift_table; 
-extern commVector<std::pair<int,int> > Cshift_table_device; 
-
-inline std::pair<int,int> *MapCshiftTable(void)
-{
-  // GPU version
-#ifdef ACCELERATOR_CSHIFT    
-  uint64_t sz=Cshift_table.size();
-  if (Cshift_table_device.size()!=sz )    {
-    Cshift_table_device.resize(sz);
-  }
-  acceleratorCopyToDevice((void *)&Cshift_table[0],
-			  (void *)&Cshift_table_device[0],
-			  sizeof(Cshift_table[0])*sz);
-
-  return &Cshift_table_device[0];
-#else 
-  return &Cshift_table[0];
-#endif
-  // CPU version use identify map
-}
-///////////////////////////////////////////////////////////////////
-// Gather for when there is no need to SIMD split 
-///////////////////////////////////////////////////////////////////
-template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
-{
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-    cbmask = 0x3;
-  }
-  
-  int so=plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs.Grid()->_slice_block[dimension];
-  int ent = 0;
-
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
-  int stride=rhs.Grid()->_slice_stride[dimension];
-
-  if ( cbmask == 0x3 ) { 
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int o  = n*stride;
-	int bo = n*e2;
-	Cshift_table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
-      }
-    }
-  } else { 
-     int bo=0;
-     for(int n=0;n<e1;n++){
-       for(int b=0;b<e2;b++){
-	 int o  = n*stride;
-	 int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
-	 if ( ocb &cbmask ) {
-	   Cshift_table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
-	 }
-       }
-     }
-  }
-  {
-    auto buffer_p = & buffer[0];
-    auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT
-    autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for(i,ent,vobj::Nsimd(),{
-	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
-    });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for(i,ent,{
-      buffer_p[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
-  }
-}
-
-///////////////////////////////////////////////////////////////////
-// Gather for when there *is* need to SIMD split 
-///////////////////////////////////////////////////////////////////
-template<class vobj> void 
-Gather_plane_extract(const Lattice<vobj> &rhs,
-		     ExtractPointerArray<typename vobj::scalar_object> pointers,
-		     int dimension,int plane,int cbmask)
-{
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-    cbmask = 0x3;
-  }
-
-  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-
-  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs.Grid()->_slice_block[dimension];
-  int n1=rhs.Grid()->_slice_stride[dimension];
-
-  if ( cbmask ==0x3){
-#ifdef ACCELERATOR_CSHIFT
-    autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for(nn,e1*e2,1,{
-	int n = nn%e1;
-	int b = nn/e1;
-	int o      =   n*n1;
-	int offset = b+n*e2;
-	
-	vobj temp =rhs_v[so+o+b];
-	extract<vobj>(temp,pointers,offset);
-      });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-	int o      =   n*n1;
-	int offset = b+n*e2;
-	
-	vobj temp =rhs_v[so+o+b];
-	extract<vobj>(temp,pointers,offset);
-      });
-#endif
-  } else { 
-    Coordinate rdim=rhs.Grid()->_rdimensions;
-    Coordinate cdm =rhs.Grid()->_checker_dim_mask;
-    std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
-    accelerator_for(nn,e1*e2,1,{
-	int n = nn%e1;
-	int b = nn/e1;
-
-	Coordinate coor;
-
-	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
-	int offset = b+n*e2;
-
-	if ( ocb & cbmask ) {
-	  vobj temp =rhs_v[so+o+b];
-	  extract<vobj>(temp,pointers,offset);
-	}
-      });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-
-	Coordinate coor;
-
-	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
-	int offset = b+n*e2;
-
-	if ( ocb & cbmask ) {
-	  vobj temp =rhs_v[so+o+b];
-	  extract<vobj>(temp,pointers,offset);
-	}
-      });
-#endif
-  }
-}
-
-//////////////////////////////////////////////////////
-// Scatter for when there is no need to SIMD split
-//////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
-{
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-    cbmask=0x3;
-  }
-
-  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-    
-  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs.Grid()->_slice_block[dimension];
-  int stride=rhs.Grid()->_slice_stride[dimension];
-
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
-  int ent    =0;
-
-  if ( cbmask ==0x3 ) {
-
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int o   =n*rhs.Grid()->_slice_stride[dimension];
-	int bo  =n*rhs.Grid()->_slice_block[dimension];
-	Cshift_table[ent++] = std::pair<int,int>(so+o+b,bo+b);
-      }
-    }
-
-  } else { 
-    int bo=0;
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int o   =n*rhs.Grid()->_slice_stride[dimension];
-	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
-	if ( ocb & cbmask ) {
-	  Cshift_table[ent++]=std::pair<int,int> (so+o+b,bo++);
-	}
-      }
-    }
-  }
-  
-  {
-    auto buffer_p = & buffer[0];
-    auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v, rhs, AcceleratorWrite);
-    accelerator_for(i,ent,vobj::Nsimd(),{
-	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
-    });
-#else
-    autoView( rhs_v, rhs, CpuWrite);
-    thread_for(i,ent,{
-      rhs_v[table[i].first]=buffer_p[table[i].second];
-    });
-#endif
-  }
-}
-
-//////////////////////////////////////////////////////
-// Scatter for when there *is* need to SIMD split
-//////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerArray<typename vobj::scalar_object> pointers,int dimension,int plane,int cbmask)
-{
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-    cbmask=0x3;
-  }
-
-  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-    
-  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs.Grid()->_slice_block[dimension];
-
-  if(cbmask ==0x3 ) {
-    int _slice_stride = rhs.Grid()->_slice_stride[dimension];
-    int _slice_block = rhs.Grid()->_slice_block[dimension];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v , rhs, AcceleratorWrite);
-    accelerator_for(nn,e1*e2,1,{
-	int n = nn%e1;
-	int b = nn/e1;
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
-	merge(rhs_v[so+o+b],pointers,offset);
-      });
-#else
-    autoView( rhs_v , rhs, CpuWrite);
-    thread_for2d(n,e1,b,e2,{
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
-	merge(rhs_v[so+o+b],pointers,offset);
-    });
-#endif
-  } else { 
-
-    // Case of SIMD split AND checker dim cannot currently be hit, except in 
-    // Test_cshift_red_black code.
-    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
-    std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
-    assert(0); // This will fail if hit on GPU
-    autoView( rhs_v, rhs, CpuWrite);
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int o      = n*rhs.Grid()->_slice_stride[dimension];
-	int offset = b+n*rhs.Grid()->_slice_block[dimension];
-	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
-	if ( ocb&cbmask ) {
-	  merge(rhs_v[so+o+b],pointers,offset);
-	}
-      }
-    }
-  }
-}
-
-//////////////////////////////////////////////////////
-// local to node block strided copies
-//////////////////////////////////////////////////////
-
-template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask)
-{
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-    cbmask=0x3;
-  }
-
-  int ro  = rplane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-  int lo  = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-
-  int e1=rhs.Grid()->_slice_nblock[dimension]; // clearly loop invariant for icpc
-  int e2=rhs.Grid()->_slice_block[dimension];
-  int stride = rhs.Grid()->_slice_stride[dimension];
-
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
-  int ent=0;
-
-  if(cbmask == 0x3 ){
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-        int o =n*stride+b;
-	Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
-      }
-    }
-  } else { 
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-        int o =n*stride+b;
-        int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o);
-        if ( ocb&cbmask ) {
-	  Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
-	}
-      }
-    }
-  }
-
-  {
-    auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
-    autoView(rhs_v , rhs, AcceleratorRead);
-    autoView(lhs_v , lhs, AcceleratorWrite);
-    accelerator_for(i,ent,vobj::Nsimd(),{
-      coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
-    });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    autoView(lhs_v , lhs, CpuWrite);
-    thread_for(i,ent,{
-      lhs_v[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
-  }
-}
-
-template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask,int permute_type)
-{
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-    cbmask=0x3;
-  }
-
-  int ro  = rplane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-  int lo  = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-
-  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs.Grid()->_slice_block [dimension];
-  int stride = rhs.Grid()->_slice_stride[dimension];
-
-  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
-
-  int ent=0;
-
-  if ( cbmask == 0x3 ) {
-    for(int n=0;n<e1;n++){
-    for(int b=0;b<e2;b++){
-      int o  =n*stride;
-      Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
-    }}
-  } else {
-    for(int n=0;n<e1;n++){
-    for(int b=0;b<e2;b++){
-      int o  =n*stride;
-      int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o+b);
-      if ( ocb&cbmask ) Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
-    }}
-  }
-
-  {
-    auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v, rhs, AcceleratorRead);
-    autoView( lhs_v, lhs, AcceleratorWrite);
-    accelerator_for(i,ent,1,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#else
-    autoView( rhs_v, rhs, CpuRead);
-    autoView( lhs_v, lhs, CpuWrite);
-    thread_for(i,ent,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#endif
-  }
-}
-
-//////////////////////////////////////////////////////
-// Local to node Cshift
-//////////////////////////////////////////////////////
-template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
-{
-  int sshift[2];
-
-  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
-  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
-
-  if ( sshift[0] == sshift[1] ) {
-    Cshift_local(ret,rhs,dimension,shift,0x3);
-  } else {
-    Cshift_local(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
-    Cshift_local(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
-  }
-}
-
-template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  GridBase *grid = rhs.Grid();
-  int fd = grid->_fdimensions[dimension];
-  int rd = grid->_rdimensions[dimension];
-  int ld = grid->_ldimensions[dimension];
-  int gd = grid->_gdimensions[dimension];
-  int ly = grid->_simd_layout[dimension];
-
-  // Map to always positive shift modulo global full dimension.
-  shift = (shift+fd)%fd;
-
-  // the permute type
-  ret.Checkerboard() = grid->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
-  int permute_dim =grid->PermuteDim(dimension);
-  int permute_type=grid->PermuteType(dimension);
-  int permute_type_dist;
-
-  for(int x=0;x<rd;x++){       
-
-    //    int o   = 0;
-    int bo  = x * grid->_ostride[dimension];
-    int cb= (cbmask==0x2)? Odd : Even;
-
-    int sshift = grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-    int sx     = (x+sshift)%rd;
-    
-    // wrap is whether sshift > rd.
-    //  num is sshift mod rd.
-    // 
-    //  shift 7
-    //
-    //  XoXo YcYc 
-    //  oXoX cYcY
-    //  XoXo YcYc
-    //  oXoX cYcY
-    //
-    //  sshift -- 
-    //
-    //  XX YY ; 3
-    //  XX YY ; 0
-    //  XX YY ; 3
-    //  XX YY ; 0
-    //
-    int permute_slice=0;
-    if(permute_dim){
-      int wrap = sshift/rd; wrap=wrap % ly;
-      int  num = sshift%rd;
-
-      if ( x< rd-num ) permute_slice=wrap;
-      else permute_slice = (wrap+1)%ly;
-
-      if ( (ly>2) && (permute_slice) ) {
-	assert(permute_type & RotateBit);
-	permute_type_dist = permute_type|permute_slice;
-      } else {
-	permute_type_dist = permute_type;
-      }
-    }
-
-    if ( permute_slice ) Copy_plane_permute(ret,rhs,dimension,x,sx,cbmask,permute_type_dist);
-    else                 Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-  
-  }
-}
-NAMESPACE_END(Grid);
-

Grid/cshift/Cshift_mpi.h

@@ -1,546 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/cshift/Cshift_mpi.h
-
-    Copyright (C) 2015
-
-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_CSHIFT_MPI_H_
-#define _GRID_CSHIFT_MPI_H_
-
-
-NAMESPACE_BEGIN(Grid); 
-
-template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
-{
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  Lattice<vobj> ret(rhs.Grid()); 
-  
-  int fd = rhs.Grid()->_fdimensions[dimension];
-  int rd = rhs.Grid()->_rdimensions[dimension];
-
-  // Map to always positive shift modulo global full dimension.
-  shift = (shift+fd)%fd;
-
-  ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
-        
-  // the permute type
-  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
-
-  RealD t1,t0;
-  t0=usecond();
-  if ( !comm_dim ) {
-    //std::cout << "CSHIFT: Cshift_local" <<std::endl;
-    Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
-  } else if ( splice_dim ) {
-    //std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
-    Cshift_comms_simd(ret,rhs,dimension,shift);
-  } else {
-    //std::cout << "CSHIFT: Cshift_comms" <<std::endl;
-    Cshift_comms(ret,rhs,dimension,shift);
-  }
-  t1=usecond();
-  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
-  return ret;
-}
-
-template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
-{
-  int sshift[2];
-
-  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
-  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
-
-  //  std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
-  if ( sshift[0] == sshift[1] ) {
-    //    std::cout << "Single pass Cshift_comms" <<std::endl;
-    Cshift_comms(ret,rhs,dimension,shift,0x3);
-  } else {
-    //    std::cout << "Two pass Cshift_comms" <<std::endl;
-    Cshift_comms(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
-    Cshift_comms(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
-  }
-}
-
-template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
-{
-  int sshift[2];
-
-  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
-  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
-
-  //std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
-  if ( sshift[0] == sshift[1] ) {
-    //std::cout << "Single pass Cshift_comms" <<std::endl;
-    Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
-  } else {
-    //std::cout << "Two pass Cshift_comms" <<std::endl;
-    Cshift_comms_simd(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
-    Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
-  }
-}
-#define ACCELERATOR_CSHIFT_NO_COPY
-#ifdef ACCELERATOR_CSHIFT_NO_COPY
-template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  GridBase *grid=rhs.Grid();
-  Lattice<vobj> temp(rhs.Grid());
-
-  int fd              = rhs.Grid()->_fdimensions[dimension];
-  int rd              = rhs.Grid()->_rdimensions[dimension];
-  int pd              = rhs.Grid()->_processors[dimension];
-  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  assert(simd_layout==1);
-  assert(comm_dim==1);
-  assert(shift>=0);
-  assert(shift<fd);
-  
-  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
-    
-  int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-  for(int x=0;x<rd;x++){       
-
-    int sx        =  (x+sshift)%rd;
-    int comm_proc = ((x+sshift)/rd)%pd;
-    
-    if (comm_proc==0) {
-      tcopy-=usecond();
-      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-      tcopy+=usecond();
-    } else {
-
-      int words = buffer_size;
-      if (cbmask != 0x3) words=words>>1;
-
-      int bytes = words * sizeof(vobj);
-
-      tgather-=usecond();
-      Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
-      tgather+=usecond();
-
-      //      int rank           = grid->_processor;
-      int recv_from_rank;
-      int xmit_to_rank;
-      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-      
-      tcomms-=usecond();
-      //      grid->Barrier();
-
-      grid->SendToRecvFrom((void *)&send_buf[0],
-			   xmit_to_rank,
-			   (void *)&recv_buf[0],
-			   recv_from_rank,
-			   bytes);
-      xbytes+=bytes;
-      //      grid->Barrier();
-      tcomms+=usecond();
-
-      tscatter-=usecond();
-      Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
-      tscatter+=usecond();
-    }
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  */
-}
-
-template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  GridBase *grid=rhs.Grid();
-  const int Nsimd = grid->Nsimd();
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-   
-  int fd = grid->_fdimensions[dimension];
-  int rd = grid->_rdimensions[dimension];
-  int ld = grid->_ldimensions[dimension];
-  int pd = grid->_processors[dimension];
-  int simd_layout     = grid->_simd_layout[dimension];
-  int comm_dim        = grid->_processors[dimension] >1 ;
-
-  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
-  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
-  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
-
-  assert(comm_dim==1);
-  assert(simd_layout==2);
-  assert(shift>=0);
-  assert(shift<fd);
-
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-  
-  int permute_type=grid->PermuteType(dimension);
-
-  ///////////////////////////////////////////////
-  // Simd direction uses an extract/merge pair
-  ///////////////////////////////////////////////
-  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
-  //  int words = sizeof(vobj)/sizeof(vector_type);
-
-  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
-  scalar_object *  recv_buf_extract_mpi;
-  scalar_object *  send_buf_extract_mpi;
- 
-  for(int s=0;s<Nsimd;s++){
-    send_buf_extract[s].resize(buffer_size);
-    recv_buf_extract[s].resize(buffer_size);
-  }
-
-  int bytes = buffer_size*sizeof(scalar_object);
-
-  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
-  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
-
-  ///////////////////////////////////////////
-  // Work out what to send where
-  ///////////////////////////////////////////
-  int cb    = (cbmask==0x2)? Odd : Even;
-  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  // loop over outer coord planes orthog to dim
-  for(int x=0;x<rd;x++){       
-
-    // FIXME call local permute copy if none are offnode.
-    for(int i=0;i<Nsimd;i++){       
-      pointers[i] = &send_buf_extract[i][0];
-    }
-    int sx   = (x+sshift)%rd;
-    tgather-=usecond();
-    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
-    tgather+=usecond();
-
-    for(int i=0;i<Nsimd;i++){
-      
-      int inner_bit = (Nsimd>>(permute_type+1));
-      int ic= (i&inner_bit)? 1:0;
-
-      int my_coor          = rd*ic + x;
-      int nbr_coor         = my_coor+sshift;
-      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-
-      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
-      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
-      int nbr_lane = (i&(~inner_bit));
-
-      int recv_from_rank;
-      int xmit_to_rank;
-
-      if (nbr_ic) nbr_lane|=inner_bit;
-
-      assert (sx == nbr_ox);
-
-      if(nbr_proc){
-	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
-
-	tcomms-=usecond();
-	//	grid->Barrier();
-
-	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
-	recv_buf_extract_mpi = &recv_buf_extract[i][0];
-	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
-			     xmit_to_rank,
-			     (void *)recv_buf_extract_mpi,
-			     recv_from_rank,
-			     bytes);
-
-	xbytes+=bytes;
-	//	grid->Barrier();
-	tcomms+=usecond();
-
-	rpointers[i] = &recv_buf_extract[i][0];
-      } else { 
-	rpointers[i] = &send_buf_extract[nbr_lane][0];
-      }
-
-    }
-    tscatter-=usecond();
-    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-    tscatter+=usecond();
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  */
-}
-#else 
-template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  GridBase *grid=rhs.Grid();
-  Lattice<vobj> temp(rhs.Grid());
-
-  int fd              = rhs.Grid()->_fdimensions[dimension];
-  int rd              = rhs.Grid()->_rdimensions[dimension];
-  int pd              = rhs.Grid()->_processors[dimension];
-  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  assert(simd_layout==1);
-  assert(comm_dim==1);
-  assert(shift>=0);
-  assert(shift<fd);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-  
-  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf_v; recv_buf_v.resize(buffer_size);
-  vobj *send_buf;
-  vobj *recv_buf;
-  {
-    grid->ShmBufferFreeAll();
-    size_t bytes = buffer_size*sizeof(vobj);
-    send_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-    recv_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-  }
-    
-  int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  for(int x=0;x<rd;x++){       
-
-    int sx        =  (x+sshift)%rd;
-    int comm_proc = ((x+sshift)/rd)%pd;
-    
-    if (comm_proc==0) {
-
-      tcopy-=usecond();
-      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-      tcopy+=usecond();
-
-    } else {
-
-      int words = buffer_size;
-      if (cbmask != 0x3) words=words>>1;
-
-      int bytes = words * sizeof(vobj);
-
-      tgather-=usecond();
-      Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
-      tgather+=usecond();
-
-      //      int rank           = grid->_processor;
-      int recv_from_rank;
-      int xmit_to_rank;
-      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-
-
-      tcomms-=usecond();
-      //      grid->Barrier();
-
-      acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
-      grid->SendToRecvFrom((void *)&send_buf[0],
-			   xmit_to_rank,
-			   (void *)&recv_buf[0],
-			   recv_from_rank,
-			   bytes);
-      xbytes+=bytes;
-      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
-
-      //      grid->Barrier();
-      tcomms+=usecond();
-
-      tscatter-=usecond();
-      Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
-      tscatter+=usecond();
-    }
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  */
-}
-
-template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  GridBase *grid=rhs.Grid();
-  const int Nsimd = grid->Nsimd();
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-   
-  int fd = grid->_fdimensions[dimension];
-  int rd = grid->_rdimensions[dimension];
-  int ld = grid->_ldimensions[dimension];
-  int pd = grid->_processors[dimension];
-  int simd_layout     = grid->_simd_layout[dimension];
-  int comm_dim        = grid->_processors[dimension] >1 ;
-
-  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
-  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
-  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
-
-  assert(comm_dim==1);
-  assert(simd_layout==2);
-  assert(shift>=0);
-  assert(shift<fd);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-
-  int permute_type=grid->PermuteType(dimension);
-
-  ///////////////////////////////////////////////
-  // Simd direction uses an extract/merge pair
-  ///////////////////////////////////////////////
-  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
-  //  int words = sizeof(vobj)/sizeof(vector_type);
-
-  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
-  scalar_object *  recv_buf_extract_mpi;
-  scalar_object *  send_buf_extract_mpi;
-  {
-    size_t bytes = sizeof(scalar_object)*buffer_size;
-    grid->ShmBufferFreeAll();
-    send_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-    recv_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-  }
-  for(int s=0;s<Nsimd;s++){
-    send_buf_extract[s].resize(buffer_size);
-    recv_buf_extract[s].resize(buffer_size);
-  }
-
-  int bytes = buffer_size*sizeof(scalar_object);
-
-  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
-  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
-
-  ///////////////////////////////////////////
-  // Work out what to send where
-  ///////////////////////////////////////////
-  int cb    = (cbmask==0x2)? Odd : Even;
-  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  // loop over outer coord planes orthog to dim
-  for(int x=0;x<rd;x++){       
-
-    // FIXME call local permute copy if none are offnode.
-    for(int i=0;i<Nsimd;i++){       
-      pointers[i] = &send_buf_extract[i][0];
-    }
-    tgather-=usecond();
-    int sx   = (x+sshift)%rd;
-    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
-    tgather+=usecond();
-
-    for(int i=0;i<Nsimd;i++){
-      
-      int inner_bit = (Nsimd>>(permute_type+1));
-      int ic= (i&inner_bit)? 1:0;
-
-      int my_coor          = rd*ic + x;
-      int nbr_coor         = my_coor+sshift;
-      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-
-      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
-      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
-      int nbr_lane = (i&(~inner_bit));
-
-      int recv_from_rank;
-      int xmit_to_rank;
-
-      if (nbr_ic) nbr_lane|=inner_bit;
-
-      assert (sx == nbr_ox);
-
-      if(nbr_proc){
-	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
-
-	tcomms-=usecond();
-	//	grid->Barrier();
-
-	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
-	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
-			     xmit_to_rank,
-			     (void *)recv_buf_extract_mpi,
-			     recv_from_rank,
-			     bytes);
-	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
-	xbytes+=bytes;
-
-	//	grid->Barrier();
-	tcomms+=usecond();
-	rpointers[i] = &recv_buf_extract[i][0];
-      } else { 
-	rpointers[i] = &send_buf_extract[nbr_lane][0];
-      }
-
-    }
-    tscatter-=usecond();
-    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-    tscatter+=usecond();
-
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
-  */
-}
-#endif
-NAMESPACE_END(Grid); 
-
-#endif

Grid/cshift/Cshift_table.cc

@@ -1,5 +0,0 @@
-#include <Grid/GridCore.h>       
-NAMESPACE_BEGIN(Grid);
-std::vector<std::pair<int,int> > Cshift_table; 
-commVector<std::pair<int,int> > Cshift_table_device; 
-NAMESPACE_END(Grid);

Grid/lattice/Lattice_ET.h

@@ -1,534 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/lattice/Lattice_ET.h
-
-Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-Author: Christoph Lehner <christoph@lhnr.de
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-			   /*  END LEGAL */
-#ifndef GRID_LATTICE_ET_H
-#define GRID_LATTICE_ET_H
-
-#include <iostream>
-#include <tuple>
-#include <typeinfo>
-#include <vector>
-
-NAMESPACE_BEGIN(Grid);
-
-////////////////////////////////////////////////////
-// Predicated where support
-////////////////////////////////////////////////////
-#ifdef GRID_SIMT
-// drop to scalar in SIMT; cleaner in fact
-template <class iobj, class vobj, class robj>
-accelerator_inline vobj predicatedWhere(const iobj &predicate, 
-					const vobj &iftrue, 
-					const robj &iffalse) 
-{
-  Integer mask = TensorRemove(predicate);
-  typename std::remove_const<vobj>::type ret= iffalse;
-  if (mask) ret=iftrue;
-  return ret;
-}
-#else
-template <class iobj, class vobj, class robj>
-accelerator_inline vobj predicatedWhere(const iobj &predicate, 
-					const vobj &iftrue, 
-					const robj &iffalse) 
-{
-  typename std::remove_const<vobj>::type ret;
-
-  typedef typename vobj::scalar_object scalar_object;
-  //  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
-  const int Nsimd = vobj::vector_type::Nsimd();
-
-  ExtractBuffer<Integer> mask(Nsimd);
-  ExtractBuffer<scalar_object> truevals(Nsimd);
-  ExtractBuffer<scalar_object> falsevals(Nsimd);
-
-  extract(iftrue, truevals);
-  extract(iffalse, falsevals);
-  extract<vInteger, Integer>(TensorRemove(predicate), mask);
-
-  for (int s = 0; s < Nsimd; s++) {
-    if (mask[s]) falsevals[s] = truevals[s];
-  }
-
-  merge(ret, falsevals);
-  return ret;
-}
-#endif
-
-/////////////////////////////////////////////////////
-//Specialization of getVectorType for lattices
-/////////////////////////////////////////////////////
-template<typename T>
-struct getVectorType<Lattice<T> >{
-  typedef typename Lattice<T>::vector_object type;
-};
-
-////////////////////////////////////////////
-//--  recursive evaluation of expressions; --
-// handle leaves of syntax tree
-///////////////////////////////////////////////////
-template<class sobj,
-  typename std::enable_if<!is_lattice<sobj>::value&&!is_lattice_expr<sobj>::value,sobj>::type * = nullptr> 
-accelerator_inline 
-sobj eval(const uint64_t ss, const sobj &arg)
-{
-  return arg;
-}
-template <class lobj> accelerator_inline 
-auto eval(const uint64_t ss, const LatticeView<lobj> &arg) -> decltype(arg(ss))
-{
-  return arg(ss);
-}
-
-////////////////////////////////////////////
-//--  recursive evaluation of expressions; --
-// whole vector return, used only for expression return type inference
-///////////////////////////////////////////////////
-template<class sobj> accelerator_inline 
-sobj vecEval(const uint64_t ss, const sobj &arg)
-{
-  return arg;
-}
-template <class lobj> accelerator_inline 
-const lobj & vecEval(const uint64_t ss, const LatticeView<lobj> &arg) 
-{
-  return arg[ss];
-}
-
-///////////////////////////////////////////////////
-// handle nodes in syntax tree- eval one operand
-// vecEval needed (but never called as all expressions offloaded) to infer the return type
-// in SIMT contexts of closure.
-///////////////////////////////////////////////////
-template <typename Op, typename T1> accelerator_inline 
-auto vecEval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
-  -> decltype(expr.op.func( vecEval(ss, expr.arg1)))
-{
-  return expr.op.func( vecEval(ss, expr.arg1) );
-}
-// vecEval two operands
-template <typename Op, typename T1, typename T2> accelerator_inline
-auto vecEval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
-  -> decltype(expr.op.func( vecEval(ss,expr.arg1),vecEval(ss,expr.arg2)))
-{
-  return expr.op.func( vecEval(ss,expr.arg1), vecEval(ss,expr.arg2) );
-}
-// vecEval three operands
-template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
-auto vecEval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
-  -> decltype(expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3)))
-{
-  return expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3));
-}
-
-///////////////////////////////////////////////////
-// handle nodes in syntax tree- eval one operand coalesced
-///////////////////////////////////////////////////
-template <typename Op, typename T1> accelerator_inline 
-auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
-  -> decltype(expr.op.func( eval(ss, expr.arg1)))
-{
-  return expr.op.func( eval(ss, expr.arg1) );
-}
-// eval two operands
-template <typename Op, typename T1, typename T2> accelerator_inline
-auto eval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
-  -> decltype(expr.op.func( eval(ss,expr.arg1),eval(ss,expr.arg2)))
-{
-  return expr.op.func( eval(ss,expr.arg1), eval(ss,expr.arg2) );
-}
-// eval three operands
-template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
-auto eval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
-  -> decltype(expr.op.func(eval(ss, expr.arg1), 
-			   eval(ss, expr.arg2), 
-			   eval(ss, expr.arg3)))
-{
-#ifdef GRID_SIMT
-  // Handles Nsimd (vInteger) != Nsimd(ComplexD)
-  typedef decltype(vecEval(ss, expr.arg2)) rvobj;
-  typedef typename std::remove_reference<rvobj>::type vobj;
-
-  const int Nsimd = vobj::vector_type::Nsimd();
-
-  auto vpred = vecEval(ss,expr.arg1);
-
-  ExtractBuffer<Integer> mask(Nsimd);
-  extract<vInteger, Integer>(TensorRemove(vpred), mask);
-
-  int s = acceleratorSIMTlane(Nsimd);
-  return expr.op.func(mask[s],
-		      eval(ss, expr.arg2), 
-		      eval(ss, expr.arg3));
-#else
-  return expr.op.func(eval(ss, expr.arg1),
-		      eval(ss, expr.arg2), 
-		      eval(ss, expr.arg3));
-#endif
-}
-
-//////////////////////////////////////////////////////////////////////////
-// Obtain the grid from an expression, ensuring conformable. This must follow a
-// tree recursion; must retain grid pointer in the LatticeView class which sucks
-// Use a different method, and make it void *.
-// Perhaps a conformable method.
-//////////////////////////////////////////////////////////////////////////
-template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-accelerator_inline void GridFromExpression(GridBase *&grid, const T1 &lat)  // Lattice leaf
-{
-  lat.Conformable(grid);
-}
-
-template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-accelerator_inline 
-void GridFromExpression(GridBase *&grid,const T1 &notlat)  // non-lattice leaf
-{}
-
-template <typename Op, typename T1>
-accelerator_inline 
-void GridFromExpression(GridBase *&grid,const LatticeUnaryExpression<Op, T1> &expr) 
-{
-  GridFromExpression(grid, expr.arg1);  // recurse
-}
-
-template <typename Op, typename T1, typename T2>
-accelerator_inline 
-void GridFromExpression(GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) 
-{
-  GridFromExpression(grid, expr.arg1);  // recurse
-  GridFromExpression(grid, expr.arg2);
-}
-template <typename Op, typename T1, typename T2, typename T3>
-accelerator_inline 
-void GridFromExpression(GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
-{
-  GridFromExpression(grid, expr.arg1);  // recurse
-  GridFromExpression(grid, expr.arg2);  // recurse
-  GridFromExpression(grid, expr.arg3);  // recurse
-}
-
-//////////////////////////////////////////////////////////////////////////
-// Obtain the CB from an expression, ensuring conformable. This must follow a
-// tree recursion
-//////////////////////////////////////////////////////////////////////////
-template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
-{
-  if ((cb == Odd) || (cb == Even)) {
-    assert(cb == lat.Checkerboard());
-  }
-  cb = lat.Checkerboard();
-}
-template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-inline void CBFromExpression(int &cb, const T1 &notlat) {} // non-lattice leaf
-template <typename Op, typename T1> inline 
-void CBFromExpression(int &cb,const LatticeUnaryExpression<Op, T1> &expr) 
-{
-  CBFromExpression(cb, expr.arg1);  // recurse AST
-}
-template <typename Op, typename T1, typename T2> inline 
-void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr) 
-{
-  CBFromExpression(cb, expr.arg1);  // recurse AST
-  CBFromExpression(cb, expr.arg2);  // recurse AST
-}
-template <typename Op, typename T1, typename T2, typename T3>
-inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
-{
-  CBFromExpression(cb, expr.arg1);  // recurse AST
-  CBFromExpression(cb, expr.arg2);  // recurse AST
-  CBFromExpression(cb, expr.arg3);  // recurse AST
-}
-
-
-//////////////////////////////////////////////////////////////////////////
-// ViewOpen
-//////////////////////////////////////////////////////////////////////////
-template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-inline void ExpressionViewOpen(T1 &lat)  // Lattice leaf
-{
-  lat.ViewOpen(AcceleratorRead);
-}
-template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-  inline void ExpressionViewOpen(T1 &notlat) {}
-
-template <typename Op, typename T1> inline 
-void ExpressionViewOpen(LatticeUnaryExpression<Op, T1> &expr) 
-{  
-  ExpressionViewOpen(expr.arg1); // recurse AST
-}
-
-template <typename Op, typename T1, typename T2> inline 
-void ExpressionViewOpen(LatticeBinaryExpression<Op, T1, T2> &expr) 
-{
-  ExpressionViewOpen(expr.arg1);  // recurse AST
-  ExpressionViewOpen(expr.arg2);  // rrecurse AST
-}
-template <typename Op, typename T1, typename T2, typename T3>
-inline void ExpressionViewOpen(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
-{
-  ExpressionViewOpen(expr.arg1);  // recurse AST
-  ExpressionViewOpen(expr.arg2);  // recurse AST
-  ExpressionViewOpen(expr.arg3);  // recurse AST
-}
-
-//////////////////////////////////////////////////////////////////////////
-// ViewClose
-//////////////////////////////////////////////////////////////////////////
-template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-inline void ExpressionViewClose( T1 &lat)  // Lattice leaf
-{
-  lat.ViewClose();
-}
-template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-inline void ExpressionViewClose(T1 &notlat) {}
-
-template <typename Op, typename T1> inline 
-void ExpressionViewClose(LatticeUnaryExpression<Op, T1> &expr) 
-{  
-  ExpressionViewClose(expr.arg1); // recurse AST
-}
-template <typename Op, typename T1, typename T2> inline 
-void ExpressionViewClose(LatticeBinaryExpression<Op, T1, T2> &expr) 
-{
-  ExpressionViewClose(expr.arg1);  // recurse AST
-  ExpressionViewClose(expr.arg2);  // recurse AST
-}
-template <typename Op, typename T1, typename T2, typename T3>
-inline void ExpressionViewClose(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
-{
-  ExpressionViewClose(expr.arg1);  // recurse AST
-  ExpressionViewClose(expr.arg2);  // recurse AST
-  ExpressionViewClose(expr.arg3);  // recurse AST
-}
-
-////////////////////////////////////////////
-// Unary operators and funcs
-////////////////////////////////////////////
-#define GridUnopClass(name, ret)					\
-  struct name {								\
-    template<class _arg> static auto accelerator_inline func(const _arg a) -> decltype(ret) { return ret; } \
-  };
-
-GridUnopClass(UnarySub, -a);
-GridUnopClass(UnaryNot, Not(a));
-GridUnopClass(UnaryTrace, trace(a));
-GridUnopClass(UnaryTranspose, transpose(a));
-GridUnopClass(UnaryTa, Ta(a));
-GridUnopClass(UnarySpTa, SpTa(a));
-GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
-GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a));
-GridUnopClass(UnaryTimesI, timesI(a));
-GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
-GridUnopClass(UnaryAbs, abs(a));
-GridUnopClass(UnarySqrt, sqrt(a));
-GridUnopClass(UnarySin, sin(a));
-GridUnopClass(UnaryCos, cos(a));
-GridUnopClass(UnaryAsin, asin(a));
-GridUnopClass(UnaryAcos, acos(a));
-GridUnopClass(UnaryLog, log(a));
-GridUnopClass(UnaryExp, exp(a));
-
-////////////////////////////////////////////
-// Binary operators
-////////////////////////////////////////////
-#define GridBinOpClass(name, combination)			\
-  struct name {							\
-    template <class _left, class _right>			\
-    static auto accelerator_inline				\
-    func(const _left &lhs, const _right &rhs)			\
-      -> decltype(combination) const				\
-    {								\
-      return combination;					\
-    }								\
-  };
-
-GridBinOpClass(BinaryAdd, lhs + rhs);
-GridBinOpClass(BinarySub, lhs - rhs);
-GridBinOpClass(BinaryMul, lhs *rhs);
-GridBinOpClass(BinaryDiv, lhs /rhs);
-GridBinOpClass(BinaryAnd, lhs &rhs);
-GridBinOpClass(BinaryOr, lhs | rhs);
-GridBinOpClass(BinaryAndAnd, lhs &&rhs);
-GridBinOpClass(BinaryOrOr, lhs || rhs);
-
-////////////////////////////////////////////////////
-// Trinary conditional op
-////////////////////////////////////////////////////
-#define GridTrinOpClass(name, combination)				\
-  struct name {								\
-    template <class _predicate,class _left, class _right>		\
-    static auto accelerator_inline					\
-    func(const _predicate &pred, const _left &lhs, const _right &rhs)	\
-      -> decltype(combination) const					\
-    {									\
-      return combination;						\
-    }									\
-  };
-
-GridTrinOpClass(TrinaryWhere,
-		(predicatedWhere<
-		 typename std::remove_reference<_predicate>::type, 
-		 typename std::remove_reference<_left>::type,
-		 typename std::remove_reference<_right>::type>(pred, lhs,rhs)));
-
-////////////////////////////////////////////
-// Operator syntactical glue
-////////////////////////////////////////////
-#define GRID_UNOP(name)   name
-#define GRID_BINOP(name)  name
-#define GRID_TRINOP(name) name
-
-#define GRID_DEF_UNOP(op, name)						\
-  template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
-  inline auto op(const T1 &arg) ->decltype(LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg)) \
-  {									\
-    return     LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg); \
-  }
-
-#define GRID_BINOP_LEFT(op, name)					\
-  template <typename T1, typename T2,					\
-            typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
-  inline auto op(const T1 &lhs, const T2 &rhs)				\
-    ->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs)) \
-  {									\
-    return     LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs);\
-  }
-
-#define GRID_BINOP_RIGHT(op, name)					\
-  template <typename T1, typename T2,					\
-            typename std::enable_if<!is_lattice<T1>::value&&!is_lattice_expr<T1>::value,T1>::type * = nullptr, \
-            typename std::enable_if< is_lattice<T2>::value|| is_lattice_expr<T2>::value,T2>::type * = nullptr> \
-  inline auto op(const T1 &lhs, const T2 &rhs)				\
-    ->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs)) \
-  {									\
-    return     LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs); \
-  }
-
-#define GRID_DEF_BINOP(op, name)		\
-  GRID_BINOP_LEFT(op, name);			\
-  GRID_BINOP_RIGHT(op, name);
-
-#define GRID_DEF_TRINOP(op, name)					\
-  template <typename T1, typename T2, typename T3>			\
-  inline auto op(const T1 &pred, const T2 &lhs, const T3 &rhs)		\
-    ->decltype(LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs)) \
-  {									\
-    return LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs); \
-  }
-
-////////////////////////
-// Operator definitions
-////////////////////////
-GRID_DEF_UNOP(operator-, UnarySub);
-GRID_DEF_UNOP(Not, UnaryNot);
-GRID_DEF_UNOP(operator!, UnaryNot);
-//GRID_DEF_UNOP(adj, UnaryAdj);
-//GRID_DEF_UNOP(conjugate, UnaryConj);
-GRID_DEF_UNOP(trace, UnaryTrace);
-GRID_DEF_UNOP(transpose, UnaryTranspose);
-GRID_DEF_UNOP(Ta, UnaryTa);
-GRID_DEF_UNOP(SpTa, UnarySpTa);
-GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
-GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup);
-GRID_DEF_UNOP(timesI, UnaryTimesI);
-GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
-GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the
-                               // abs-fabs-dabs-labs thing
-GRID_DEF_UNOP(sqrt, UnarySqrt);
-GRID_DEF_UNOP(sin, UnarySin);
-GRID_DEF_UNOP(cos, UnaryCos);
-GRID_DEF_UNOP(asin, UnaryAsin);
-GRID_DEF_UNOP(acos, UnaryAcos);
-GRID_DEF_UNOP(log, UnaryLog);
-GRID_DEF_UNOP(exp, UnaryExp);
-
-GRID_DEF_BINOP(operator+, BinaryAdd);
-GRID_DEF_BINOP(operator-, BinarySub);
-GRID_DEF_BINOP(operator*, BinaryMul);
-GRID_DEF_BINOP(operator/, BinaryDiv);
-
-GRID_DEF_BINOP(operator&, BinaryAnd);
-GRID_DEF_BINOP(operator|, BinaryOr);
-GRID_DEF_BINOP(operator&&, BinaryAndAnd);
-GRID_DEF_BINOP(operator||, BinaryOrOr);
-
-GRID_DEF_TRINOP(where, TrinaryWhere);
-
-/////////////////////////////////////////////////////////////
-// Closure convenience to force expression to evaluate
-/////////////////////////////////////////////////////////////
-template <class Op, class T1>
-auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
-{
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > ret(expr);
-  return ret;
-}
-template <class Op, class T1, class T2>
-auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type >
-{
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
-  return ret;
-}
-template <class Op, class T1, class T2, class T3>
-auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
-				   vecEval(0, expr.arg2),
-				   vecEval(0, expr.arg3)))>::type >
-{
-  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
-				vecEval(0, expr.arg2),
-			        vecEval(0, expr.arg3)))>::type >  ret(expr);
-  return ret;
-}
-#define EXPRESSION_CLOSURE(function)					\
-  template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> \
-    auto function(Expression &expr) -> decltype(function(closure(expr))) \
-  {									\
-    return function(closure(expr));					\
-  }
-
-
-#undef GRID_UNOP
-#undef GRID_BINOP
-#undef GRID_TRINOP
-
-#undef GRID_DEF_UNOP
-#undef GRID_DEF_BINOP
-#undef GRID_DEF_TRINOP
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/lattice/Lattice_arith.h

@@ -1,311 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_arith.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Christoph Lehner <christoph@lhnr.de>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_ARITH_H
-#define GRID_LATTICE_ARITH_H
-
-NAMESPACE_BEGIN(Grid);
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-//  avoid copy back routines for mult, mac, sub, add
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class obj1,class obj2,class obj3> inline
-void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mult");
-  ret.Checkerboard() = lhs.Checkerboard();
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  conformable(ret,rhs);
-  conformable(lhs,rhs);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto lhs_t = lhs_v(ss);
-    auto rhs_t = rhs_v(ss);
-    mult(&tmp,&lhs_t,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class obj1,class obj2,class obj3> inline
-void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mac");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(ret,rhs);
-  conformable(lhs,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    auto lhs_t=lhs_v(ss);
-    auto rhs_t=rhs_v(ss);
-    auto tmp  =ret_v(ss);
-    mac(&tmp,&lhs_t,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class obj1,class obj2,class obj3> inline
-void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("sub");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(ret,rhs);
-  conformable(lhs,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto lhs_t=lhs_v(ss);
-    auto rhs_t=rhs_v(ss);
-    sub(&tmp,&lhs_t,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-template<class obj1,class obj2,class obj3> inline
-void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("add");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(ret,rhs);
-  conformable(lhs,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto lhs_t=lhs_v(ss);
-    auto rhs_t=rhs_v(ss);
-    add(&tmp,&lhs_t,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-//  avoid copy back routines for mult, mac, sub, add
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class obj1,class obj2,class obj3> inline
-void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("mult");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(lhs,ret);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    mult(&tmp,&lhs_v(ss),&rhs);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class obj1,class obj2,class obj3> inline
-void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("mac");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(ret,lhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
-    auto lhs_t=lhs_v(ss);
-    mac(&tmp,&lhs_t,&rhs);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class obj1,class obj2,class obj3> inline
-void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("sub");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(ret,lhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto lhs_t=lhs_v(ss);
-    sub(&tmp,&lhs_t,&rhs);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-template<class obj1,class obj2,class obj3> inline
-void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("add");
-  ret.Checkerboard() = lhs.Checkerboard();
-  conformable(lhs,ret);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto lhs_t=lhs_v(ss);
-    add(&tmp,&lhs_t,&rhs);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-//  avoid copy back routines for mult, mac, sub, add
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class obj1,class obj2,class obj3> inline
-void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mult");
-  ret.Checkerboard() = rhs.Checkerboard();
-  conformable(ret,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto rhs_t=rhs_v(ss);
-    mult(&tmp,&lhs,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class obj1,class obj2,class obj3> inline
-void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mac");
-  ret.Checkerboard() = rhs.Checkerboard();
-  conformable(ret,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    auto tmp  =ret_v(ss);
-    auto rhs_t=rhs_v(ss);
-    mac(&tmp,&lhs,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class obj1,class obj2,class obj3> inline
-void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("sub");
-  ret.Checkerboard() = rhs.Checkerboard();
-  conformable(ret,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto rhs_t=rhs_v(ss);
-    sub(&tmp,&lhs,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-template<class obj1,class obj2,class obj3> inline
-void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("add");
-  ret.Checkerboard() = rhs.Checkerboard();
-  conformable(ret,rhs);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( rhs_v , lhs, AcceleratorRead);
-  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
-    auto rhs_t=rhs_v(ss);
-    add(&tmp,&lhs,&rhs_t);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-  
-template<class sobj,class vobj> inline
-void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-  GRID_TRACE("axpy");
-  ret.Checkerboard() = x.Checkerboard();
-  conformable(ret,x);
-  conformable(x,y);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( x_v , x, AcceleratorRead);
-  autoView( y_v , y, AcceleratorRead);
-  accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-    auto tmp = a*coalescedRead(x_v[ss])+coalescedRead(y_v[ss]);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-template<class sobj,class vobj> inline
-void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
-  GRID_TRACE("axpby");
-  ret.Checkerboard() = x.Checkerboard();
-  conformable(ret,x);
-  conformable(x,y);
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( x_v , x, AcceleratorRead);
-  autoView( y_v , y, AcceleratorRead);
-  accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-    auto tmp = a*x_v(ss)+b*y_v(ss);
-    coalescedWrite(ret_v[ss],tmp);
-  });
-}
-
-template<class sobj,class vobj> inline
-RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
-{
-  GRID_TRACE("axpy_norm");
-    return axpy_norm_fast(ret,a,x,y);
-}
-template<class sobj,class vobj> inline
-RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
-{
-  GRID_TRACE("axpby_norm");
-    return axpby_norm_fast(ret,a,b,x,y);
-}
-
-/// Trace product
-template<class obj> auto traceProduct(const Lattice<obj> &rhs_1,const Lattice<obj> &rhs_2)
-  -> Lattice<decltype(trace(obj()))>
-{
-  typedef decltype(trace(obj())) robj;
-  Lattice<robj> ret_i(rhs_1.Grid());
-  autoView( rhs1 , rhs_1, AcceleratorRead);
-  autoView( rhs2 , rhs_2, AcceleratorRead);
-  autoView( ret , ret_i, AcceleratorWrite);
-  ret.Checkerboard() = rhs_1.Checkerboard();
-  accelerator_for(ss,rhs1.size(),obj::Nsimd(),{
-      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2(ss)));
-  });
-  return ret_i;
-}
-
-template<class obj1,class obj2> auto traceProduct(const Lattice<obj1> &rhs_1,const obj2 &rhs2)
-  -> Lattice<decltype(trace(obj1()))>
-{
-  typedef decltype(trace(obj1())) robj;
-  Lattice<robj> ret_i(rhs_1.Grid());
-  autoView( rhs1 , rhs_1, AcceleratorRead);
-  autoView( ret , ret_i, AcceleratorWrite);
-  ret.Checkerboard() = rhs_1.Checkerboard();
-  accelerator_for(ss,rhs1.size(),obj1::Nsimd(),{
-      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2));
-  });
-  return ret_i;
-}
-template<class obj1,class obj2> auto traceProduct(const obj2 &rhs_2,const Lattice<obj1> &rhs_1)
-  -> Lattice<decltype(trace(obj1()))>
-{
-  return traceProduct(rhs_1,rhs_2);
-}
-
-
-
-NAMESPACE_END(Grid);
-#endif

Grid/lattice/Lattice_base.h

@@ -1,382 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/lattice/Lattice_base.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>
-Author: Christoph Lehner <christoph@lhnr.de>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-			   /*  END LEGAL */
-
-#pragma once 
-
-#define STREAMING_STORES
-
-NAMESPACE_BEGIN(Grid);
-
-extern int GridCshiftPermuteMap[4][16];
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// The real lattice class, with normal copy and assignment semantics.
-// This contains extra (host resident) grid pointer data that may be accessed by host code
-/////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj>
-class Lattice : public LatticeAccelerator<vobj>
-{
-public:
-  GridBase *Grid(void) const { return this->_grid; }
-  ///////////////////////////////////////////////////
-  // Member types
-  ///////////////////////////////////////////////////
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_object scalar_object;
-  typedef vobj vector_object;
-
-private:
-  void dealloc(void)
-  {
-    if( this->_odata_size ) {
-      alignedAllocator<vobj> alloc;
-      alloc.deallocate(this->_odata,this->_odata_size);
-      this->_odata=nullptr;
-      this->_odata_size=0;
-    }
-  }
-  void resize(uint64_t size)
-  {
-    if ( this->_odata_size != size ) {
-      alignedAllocator<vobj> alloc;
-
-      dealloc();
-      
-      this->_odata_size = size;
-      if ( size )
-	this->_odata      = alloc.allocate(this->_odata_size);
-      else 
-	this->_odata      = nullptr;
-    }
-  }
-public:
-
-  /////////////////////////////////////////////////////////////////////////////////
-  // Can use to make accelerator dirty without copy from host ; useful for temporaries "dont care" prev contents
-  /////////////////////////////////////////////////////////////////////////////////
-  void SetViewMode(ViewMode mode) {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
-    accessor.ViewClose();
-  }
-
-  // Helper function to print the state of this object in the AccCache
-  void PrintCacheState(void)
-  {
-    MemoryManager::PrintState(this->_odata);
-  }
-
-  /////////////////////////////////////////////////////////////////////////////////
-  // Return a view object that may be dereferenced in site loops.
-  // The view is trivially copy constructible and may be copied to an accelerator device
-  // in device lambdas
-  /////////////////////////////////////////////////////////////////////////////////
-
-  LatticeView<vobj> View (ViewMode mode) const 
-  {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
-    return accessor;
-  }
-
-  ~Lattice() { 
-    if ( this->_odata_size ) {
-      dealloc();
-    }
-   }
-  ////////////////////////////////////////////////////////////////////////////////
-  // Expression Template closure support
-  ////////////////////////////////////////////////////////////////////////////////
-  template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
-  {
-    GRID_TRACE("ExpressionTemplateEval");
-    GridBase *egrid(nullptr);
-    GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
-    conformable(this->_grid,egrid);
-
-    int cb=-1;
-    CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
-    this->checkerboard=cb;
-    
-    auto exprCopy = expr;
-    ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      auto tmp = eval(ss,exprCopy);
-      coalescedWrite(me[ss],tmp);
-    });
-    me.ViewClose();
-    ExpressionViewClose(exprCopy);
-    return *this;
-  }
-  template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
-  {
-    GRID_TRACE("ExpressionTemplateEval");
-    GridBase *egrid(nullptr);
-    GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
-    conformable(this->_grid,egrid);
-
-    int cb=-1;
-    CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
-    this->checkerboard=cb;
-
-    auto exprCopy = expr;
-    ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      auto tmp = eval(ss,exprCopy);
-      coalescedWrite(me[ss],tmp);
-    });
-    me.ViewClose();
-    ExpressionViewClose(exprCopy);
-    return *this;
-  }
-  template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
-  {
-    GRID_TRACE("ExpressionTemplateEval");
-    GridBase *egrid(nullptr);
-    GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
-    conformable(this->_grid,egrid);
-
-    int cb=-1;
-    CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
-    this->checkerboard=cb;
-    auto exprCopy = expr;
-    ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      auto tmp = eval(ss,exprCopy);
-      coalescedWrite(me[ss],tmp);
-    });
-    me.ViewClose();
-    ExpressionViewClose(exprCopy);
-    return *this;
-  }
-  //GridFromExpression is tricky to do
-  template<class Op,class T1>
-  Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
-    this->_grid = nullptr;
-    GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
-
-    int cb=-1;
-    CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
-    this->checkerboard=cb;
-
-    resize(this->_grid->oSites());
-
-    *this = expr;
-  }
-  template<class Op,class T1, class T2>
-  Lattice(const LatticeBinaryExpression<Op,T1,T2> & expr) {
-    this->_grid = nullptr;
-    GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
-
-    int cb=-1;
-    CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
-    this->checkerboard=cb;
-
-    resize(this->_grid->oSites());
-
-    *this = expr;
-  }
-  template<class Op,class T1, class T2, class T3>
-  Lattice(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr) {
-    this->_grid = nullptr;
-    GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
-
-    int cb=-1;
-    CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
-    this->checkerboard=cb;
-
-    resize(this->_grid->oSites());
-
-    *this = expr;
-  }
-
-  template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
-    auto me  = View(CpuWrite);
-    thread_for(ss,me.size(),{
-	me[ss]= r;
-    });
-    me.ViewClose();
-    return *this;
-  }
-
-  //////////////////////////////////////////////////////////////////
-  // Follow rule of five, with Constructor requires "grid" passed
-  // to user defined constructor
-  ///////////////////////////////////////////
-  // user defined constructor
-  ///////////////////////////////////////////
-  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
-    this->_grid = grid;
-    resize(this->_grid->oSites());
-    assert((((uint64_t)&this->_odata[0])&0xF) ==0);
-    this->checkerboard=0;
-    SetViewMode(mode);
-  }
-  
-  //  virtual ~Lattice(void) = default;
-    
-  void reset(GridBase* grid) {
-    if (this->_grid != grid) {
-      this->_grid = grid;
-      this->resize(grid->oSites());
-      this->checkerboard = 0;
-    }
-  }
-  ///////////////////////////////////////////
-  // copy constructor
-  ///////////////////////////////////////////
-  Lattice(const Lattice& r){ 
-    this->_grid = r.Grid();
-    resize(this->_grid->oSites());
-    *this = r;
-  }
-  ///////////////////////////////////////////
-  // move constructor
-  ///////////////////////////////////////////
-  Lattice(Lattice && r){ 
-    this->_grid = r.Grid();
-    this->_odata      = r._odata;
-    this->_odata_size = r._odata_size;
-    this->checkerboard= r.Checkerboard();
-    r._odata      = nullptr;
-    r._odata_size = 0;
-  }
-  ///////////////////////////////////////////
-  // assignment template
-  ///////////////////////////////////////////
-  template<class robj> inline Lattice<vobj> & operator = (const Lattice<robj> & r){
-    typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
-    conformable(*this,r);
-    this->checkerboard = r.Checkerboard();
-    auto him= r.View(AcceleratorRead);
-    auto me =   View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      coalescedWrite(me[ss],him(ss));
-    });
-    me.ViewClose();    him.ViewClose();
-    return *this;
-  }
-
-  ///////////////////////////////////////////
-  // Copy assignment 
-  ///////////////////////////////////////////
-  inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
-    this->checkerboard = r.Checkerboard();
-    conformable(*this,r);
-    auto him= r.View(AcceleratorRead);
-    auto me =   View(AcceleratorWriteDiscard);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      coalescedWrite(me[ss],him(ss));
-    });
-    me.ViewClose();    him.ViewClose();
-    return *this;
-  }
-  ///////////////////////////////////////////
-  // Move assignment possible if same type
-  ///////////////////////////////////////////
-  inline Lattice<vobj> & operator = (Lattice<vobj> && r){
-
-    resize(0); // deletes if appropriate
-    this->_grid       = r.Grid();
-    this->_odata      = r._odata;
-    this->_odata_size = r._odata_size;
-    this->checkerboard= r.Checkerboard();
-
-    r._odata      = nullptr;
-    r._odata_size = 0;
-    
-    return *this;
-  }
-
-  /////////////////////////////////////////////////////////////////////////////
-  // *=,+=,-= operators inherit behvour from correspond */+/- operation
-  /////////////////////////////////////////////////////////////////////////////
-  template<class T> inline Lattice<vobj> &operator *=(const T &r) {
-    *this = (*this)*r;
-    return *this;
-  }
-  
-  template<class T> inline Lattice<vobj> &operator -=(const T &r) {
-    *this = (*this)-r;
-    return *this;
-  }
-  template<class T> inline Lattice<vobj> &operator +=(const T &r) {
-    *this = (*this)+r;
-    return *this;
-  }
-
-  friend inline void swap(Lattice &l, Lattice &r) { 
-    conformable(l,r);
-    LatticeAccelerator<vobj> tmp;
-    LatticeAccelerator<vobj> *lp = (LatticeAccelerator<vobj> *)&l;
-    LatticeAccelerator<vobj> *rp = (LatticeAccelerator<vobj> *)&r;
-    tmp = *lp;    *lp=*rp;    *rp=tmp;
-  }
-
-}; // class Lattice
-
-template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
-  typedef typename vobj::scalar_object sobj;
-  for(int g=0;g<o.Grid()->_gsites;g++){
-
-    Coordinate gcoor;
-    o.Grid()->GlobalIndexToGlobalCoor(g,gcoor);
-
-    sobj ss;
-    peekSite(ss,o,gcoor);
-    stream<<"[";
-    for(int d=0;d<gcoor.size();d++){
-      stream<<gcoor[d];
-      if(d!=gcoor.size()-1) stream<<",";
-    }
-    stream<<"]\t";
-    stream<<ss<<std::endl;
-  }
-  return stream;
-}
-  
-NAMESPACE_END(Grid);
-

Grid/lattice/Lattice_basis.h

@@ -1,248 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/lattice/Lattice_basis.h
-
-Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Christoph Lehner <christoph@lhnr.de>
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-			   /*  END LEGAL */
-
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-template<class Field>
-void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k) 
-{
-  // If assume basis[j] are already orthonormal,
-  // can take all inner products in parallel saving 2x bandwidth
-  // Save 3x bandwidth on the second line of loop.
-  // perhaps 2.5x speed up.
-  // 2x overall in Multigrid Lanczos  
-  for(int j=0; j<k; ++j){
-    auto ip = innerProduct(basis[j],w);
-    w = w - ip*basis[j];
-  }
-}
-
-template<class VField, class Matrix>
-void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm) 
-{
-  typedef decltype(basis[0]) Field;
-  typedef decltype(basis[0].View(AcceleratorRead)) View;
-
-  Vector<View> basis_v; basis_v.reserve(basis.size());
-  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
-  typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
-  GridBase* grid = basis[0].Grid();
-      
-  for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorWrite));
-  }
-
-#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
-  int max_threads = thread_max();
-  Vector < vobj > Bt(Nm * max_threads);
-  thread_region
-    {
-      vobj* B = &Bt[Nm * thread_num()];
-      thread_for_in_region(ss, grid->oSites(),{
-	  for(int j=j0; j<j1; ++j) B[j]=0.;
-      
-	  for(int j=j0; j<j1; ++j){
-	    for(int k=k0; k<k1; ++k){
-	      B[j] +=Qt(j,k) * basis_v[k][ss];
-	    }
-	  }
-	  for(int j=j0; j<j1; ++j){
-	    basis_v[j][ss] = B[j];
-	  }
-	});
-    }
-#else
-  View *basis_vp = &basis_v[0];
-
-  int nrot = j1-j0;
-  if (!nrot) // edge case not handled gracefully by Cuda
-    return;
-
-  uint64_t oSites   =grid->oSites();
-  uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
-
-  Vector <vobj> Bt(siteBlock * nrot); 
-  auto Bp=&Bt[0];
-
-  // GPU readable copy of matrix
-  Vector<Coeff_t> Qt_jv(Nm*Nm);
-  Coeff_t *Qt_p = & Qt_jv[0];
-  thread_for(i,Nm*Nm,{
-      int j = i/Nm;
-      int k = i%Nm;
-      Qt_p[i]=Qt(j,k);
-  });
-
-  // Block the loop to keep storage footprint down
-  for(uint64_t s=0;s<oSites;s+=siteBlock){
-
-    // remaining work in this block
-    int ssites=MIN(siteBlock,oSites-s);
-
-    // zero out the accumulators
-    accelerator_for(ss,siteBlock*nrot,vobj::Nsimd(),{
-	decltype(coalescedRead(Bp[ss])) z;
-	z=Zero();
-	coalescedWrite(Bp[ss],z);
-      });
-
-    accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
-	
-	int j =sj%nrot;
-	int jj  =j0+j;
-	int ss =sj/nrot;
-	int sss=ss+s;
-
-	for(int k=k0; k<k1; ++k){
-	  auto tmp = coalescedRead(Bp[ss*nrot+j]);
-	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_vp[k][sss]));
-	}
-      });
-
-    accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
-	int j =sj%nrot;
-	int jj  =j0+j;
-	int ss =sj/nrot;
-	int sss=ss+s;
-	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
-      });
-  }
-#endif
-
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
-}
-
-// Extract a single rotated vector
-template<class Field>
-void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
-{
-  typedef decltype(basis[0].View(AcceleratorRead)) View;
-  typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0].Grid();
-
-  result.Checkerboard() = basis[0].Checkerboard();
-
-  Vector<View> basis_v; basis_v.reserve(basis.size());
-  for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorRead));
-  }
-  vobj zz=Zero();
-  Vector<double> Qt_jv(Nm);
-  double * Qt_j = & Qt_jv[0];
-  for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
-
-  auto basis_vp=& basis_v[0];
-  autoView(result_v,result,AcceleratorWrite);
-  accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    vobj zzz=Zero();
-    auto B=coalescedRead(zzz);
-    for(int k=k0; k<k1; ++k){
-      B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]);
-    }
-    coalescedWrite(result_v[ss], B);
-  });
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
-}
-
-template<class Field>
-void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx) 
-{
-  int vlen = idx.size();
-
-  assert(vlen>=1);
-  assert(vlen<=sort_vals.size());
-  assert(vlen<=_v.size());
-
-  for (size_t i=0;i<vlen;i++) {
-
-    if (idx[i] != i) {
-
-      //////////////////////////////////////
-      // idx[i] is a table of desired sources giving a permutation.
-      // Swap v[i] with v[idx[i]].
-      // Find  j>i for which _vnew[j] = _vold[i],
-      // track the move idx[j] => idx[i]
-      // track the move idx[i] => i
-      //////////////////////////////////////
-      size_t j;
-      for (j=i;j<idx.size();j++)
-	if (idx[j]==i)
-	  break;
-
-      assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
-
-      swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
-      std::swap(sort_vals[i],sort_vals[idx[i]]);
-
-      idx[j] = idx[i];
-      idx[i] = i;
-    }
-  }
-}
-
-inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals) 
-{
-  std::vector<int> idx(sort_vals.size());
-  std::iota(idx.begin(), idx.end(), 0);
-
-  // sort indexes based on comparing values in v
-  std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
-    return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
-  });
-  return idx;
-}
-
-template<class Field>
-void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse) 
-{
-  std::vector<int> idx = basisSortGetIndex(sort_vals);
-  if (reverse)
-    std::reverse(idx.begin(), idx.end());
-  
-  basisReorderInPlace(_v,sort_vals,idx);
-}
-
-// PAB: faster to compute the inner products first then fuse loops.
-// If performance critical can improve.
-template<class Field>
-void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
-  result = Zero();
-  assert(_v.size()==eval.size());
-  int N = (int)_v.size();
-  for (int i=0;i<N;i++) {
-    Field& tmp = _v[i];
-    axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
-  }
-}
-
-NAMESPACE_END(Grid);

Grid/lattice/Lattice_comparison.h

@@ -1,179 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_comparison.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_LATTICE_COMPARISON_H
-#define GRID_LATTICE_COMPARISON_H
-
-NAMESPACE_BEGIN(Grid);
-
-//////////////////////////////////////////////////////////////////////////
-// relational operators
-// 
-// Support <,>,<=,>=,==,!=
-//
-//Query supporting bitwise &, |, ^, !
-//Query supporting logical &&, ||, 
-//////////////////////////////////////////////////////////////////////////
-
-typedef iScalar<vInteger> vPredicate ;
-
-//////////////////////////////////////////////////////////////////////////
-// compare lattice to lattice
-//////////////////////////////////////////////////////////////////////////
-
-template<class vfunctor,class lobj,class robj>  
-inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
-{
-  Lattice<vPredicate> ret(rhs.Grid());
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( rhs_v, rhs, CpuRead);
-  autoView( ret_v, ret, CpuWrite);
-  thread_for( ss, rhs_v.size(), {
-      ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
-  });
-  return ret;
-}
-//////////////////////////////////////////////////////////////////////////
-// compare lattice to scalar
-//////////////////////////////////////////////////////////////////////////
-template<class vfunctor,class lobj,class robj> 
-inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
-{
-  Lattice<vPredicate> ret(lhs.Grid());
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( ret_v, ret, CpuWrite);
-  thread_for( ss, lhs_v.size(), {
-    ret_v[ss]=op(lhs_v[ss],rhs);
-  });
-  return ret;
-}
-//////////////////////////////////////////////////////////////////////////
-// compare scalar to lattice
-//////////////////////////////////////////////////////////////////////////
-template<class vfunctor,class lobj,class robj> 
-inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
-{
-  Lattice<vPredicate> ret(rhs.Grid());
-  autoView( rhs_v, rhs, CpuRead);
-  autoView( ret_v, ret, CpuWrite);
-  thread_for( ss, rhs_v.size(), {
-    ret_v[ss]=op(lhs,rhs_v[ss]);
-  });
-  return ret;
-}
-  
-//////////////////////////////////////////////////////////////////////////
-// Map to functors
-//////////////////////////////////////////////////////////////////////////
-// Less than
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-  return LLComparison(vlt<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
-  return LSComparison(vlt<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
-  return SLComparison(vlt<lobj,robj>(),lhs,rhs);
-}
-  
-// Less than equal
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-  return LLComparison(vle<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
-  return LSComparison(vle<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
-  return SLComparison(vle<lobj,robj>(),lhs,rhs);
-}
-  
-// Greater than 
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-  return LLComparison(vgt<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
-  return LSComparison(vgt<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
-  return SLComparison(vgt<lobj,robj>(),lhs,rhs);
-}
-  
-  
-// Greater than equal
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-  return LLComparison(vge<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
-  return LSComparison(vge<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
-  return SLComparison(vge<lobj,robj>(),lhs,rhs);
-}
-   
-// equal
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-  return LLComparison(veq<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
-  return LSComparison(veq<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
-  return SLComparison(veq<lobj,robj>(),lhs,rhs);
-}
-   
-   
-// not equal
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-  return LLComparison(vne<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
-  return LSComparison(vne<lobj,robj>(),lhs,rhs);
-}
-template<class lobj,class robj>
-inline Lattice<vPredicate> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
-  return SLComparison(vne<lobj,robj>(),lhs,rhs);
-}
-NAMESPACE_END(Grid);
-#endif

Grid/lattice/Lattice_coordinate.h

@@ -1,55 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_coordinate.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 
-
-NAMESPACE_BEGIN(Grid);
-
-template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
-{
-  typedef typename iobj::scalar_type scalar_type;
-  typedef typename iobj::vector_type vector_type;
-
-  GridBase *grid = l.Grid();
-  int Nsimd = grid->iSites();
-
-  autoView(l_v, l, CpuWrite);
-  thread_for( o, grid->oSites(), {
-    vector_type vI;
-    Coordinate gcoor;
-    ExtractBuffer<scalar_type> mergebuf(Nsimd);
-    for(int i=0;i<grid->iSites();i++){
-      grid->RankIndexToGlobalCoor(grid->ThisRank(),o,i,gcoor);
-      mergebuf[i]=(Integer)gcoor[mu];
-    }
-    merge<vector_type,scalar_type>(vI,mergebuf);
-    l_v[o]=vI;
-  });
-};
-
-NAMESPACE_END(Grid);
-

Grid/lattice/Lattice_crc.h

@@ -1,55 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_crc.h
-
-    Copyright (C) 2021
-
-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);
-
-template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
-{
-  auto ff = localNorm2(f);
-  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
-  typedef typename vobj::tensor_reduced normtype;
-  typedef typename normtype::scalar_object scalar;
-  std::vector<scalar> sff;
-  sliceSum(ff,sff,mu);
-  for(int t=0;t<sff.size();t++){
-    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
-  }
-}
-
-template<class vobj> uint32_t crc(const Lattice<vobj> & buf)
-{
-  autoView( buf_v , buf, CpuRead);
-  return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
-}
-
-#define CRC(U) std::cerr << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
-
-NAMESPACE_END(Grid);
-
-

Grid/lattice/Lattice_local.h

@@ -1,87 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_local.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_LOCALREDUCTION_H
-#define GRID_LATTICE_LOCALREDUCTION_H
-
-///////////////////////////////////////////////
-// localInner, localNorm, outerProduct
-///////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////
-// Non site, reduced locally reduced routines
-/////////////////////////////////////////////////////
-  
-// localNorm2,
-template<class vobj>
-inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
-{
-  Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  autoView( rhs_v , rhs, AcceleratorRead);
-  autoView( ret_v , ret, AcceleratorWrite);
-  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
-    coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
-  });
-  return ret;
-}
-  
-// localInnerProduct
-template<class vobj>
-inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
-{
-  Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  autoView( ret_v , ret, AcceleratorWrite);
-  accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
-    coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
-  });
-  return ret;
-}
-  
-// outerProduct Scalar x Scalar -> Scalar
-//              Vector x Vector -> Matrix
-template<class ll,class rr>
-inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(ll(),rr()))>
-{
-  typedef decltype(coalescedRead(ll())) sll;
-  typedef decltype(coalescedRead(rr())) srr;
-  Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
-  autoView( lhs_v , lhs, AcceleratorRead);
-  autoView( rhs_v , rhs, AcceleratorRead);
-  autoView( ret_v , ret, AcceleratorWrite);
-  accelerator_for(ss,rhs_v.size(),1,{
-    // FIXME had issues with scalar version of outer 
-    // Use vector [] operator and don't read coalesce this loop
-    ret_v[ss]=outerProduct(lhs_v[ss],rhs_v[ss]);
-  });
-  return ret;
-}
-NAMESPACE_END(Grid);
-#endif

Grid/lattice/Lattice_matrix_reduction.h

@@ -1,199 +0,0 @@
-/*************************************************************************************
-    Grid physics library, www.github.com/paboyle/Grid 
-    Source file: ./lib/lattice/Lattice_reduction.h
-    Copyright (C) 2015
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once 
-#include <Grid/Grid_Eigen_Dense.h>
-
-#ifdef GRID_WARN_SUBOPTIMAL
-#warning "Optimisation alert all these reduction loops are NOT threaded "
-#endif     
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj>
-static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
-{    
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-
-  GridBase *FullGrid  = X.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-
-  //  Lattice<vobj> Xslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-  autoView( X_v , X, CpuRead);
-  autoView( Y_v , Y, CpuRead);
-  autoView( R_v , R, CpuWrite);
-  thread_region
-  {
-    std::vector<vobj> s_x(Nblock);
-
-    thread_loop_collapse2( (int n=0;n<nblock;n++),{
-      for(int b=0;b<block;b++){
-	int o  = n*stride + b;
-
-	for(int i=0;i<Nblock;i++){
-	  s_x[i] = X_v[o+i*ostride];
-	}
-
-	vobj dot;
-	for(int i=0;i<Nblock;i++){
-	  dot = Y_v[o+i*ostride];
-	  for(int j=0;j<Nblock;j++){
-	    dot = dot + s_x[j]*(scale*aa(j,i));
-	  }
-	  R_v[o+i*ostride]=dot;
-	}
-      }});
-  }
-};
-
-template<class vobj>
-static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
-{    
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-
-  GridBase *FullGrid  = X.Grid();
-  assert( FullGrid->_simd_layout[Orthog]==1);
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-
-  autoView( X_v , X, CpuRead);
-  autoView( R_v , R, CpuWrite);
-
-  thread_region
-  {
-    std::vector<vobj> s_x(Nblock);
-    
-    thread_loop_collapse2( (int n=0;n<nblock;n++),{
-      for(int b=0;b<block;b++){
-	int o  = n*stride + b;
-
-	for(int i=0;i<Nblock;i++){
-	  s_x[i] = X_v[o+i*ostride];
-	}
-
-	vobj dot;
-	for(int i=0;i<Nblock;i++){
-	  dot = s_x[0]*(scale*aa(0,i));
-	  for(int j=1;j<Nblock;j++){
-	    dot = dot + s_x[j]*(scale*aa(j,i));
-	  }
-	  R_v[o+i*ostride]=dot;
-	}
-    }});
-  }
-
-};
-
-
-template<class vobj>
-static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
-{
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-  
-  GridBase *FullGrid  = lhs.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-  
-  int Nblock = FullGrid->GlobalDimensions()[Orthog];
-  
-  //  Lattice<vobj> Lslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-  
-  mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl = nh-1;
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-
-  typedef typename vobj::vector_typeD vector_typeD;
-  autoView( lhs_v , lhs, CpuRead);
-  autoView( rhs_v , rhs, CpuRead);
-  thread_region {
-    std::vector<vobj> Left(Nblock);
-    std::vector<vobj> Right(Nblock);
-    Eigen::MatrixXcd  mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-    thread_loop_collapse2((int n=0;n<nblock;n++),{
-      for(int b=0;b<block;b++){
-
-	int o  = n*stride + b;
-
-	for(int i=0;i<Nblock;i++){
-	  Left [i] = lhs_v[o+i*ostride];
-	  Right[i] = rhs_v[o+i*ostride];
-	}
-
-	for(int i=0;i<Nblock;i++){
-	  for(int j=0;j<Nblock;j++){
-	    auto tmp = innerProduct(Left[i],Right[j]);
-	    auto rtmp = TensorRemove(tmp);
-	    ComplexD z = Reduce(rtmp);
-	    mat_thread(i,j) += std::complex<double>(real(z),imag(z));
-	  }}
-    }});
-    thread_critical {
-      mat += mat_thread;
-    }  
-  }
-
-  for(int i=0;i<Nblock;i++){
-    for(int j=0;j<Nblock;j++){
-      ComplexD sum = mat(i,j);
-      FullGrid->GlobalSum(sum);
-      mat(i,j)=sum;
-    }}
-
-  return;
-}
-
-NAMESPACE_END(Grid);
-
-
-

Grid/lattice/Lattice_peekpoke.h

@@ -1,231 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_peekpoke.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>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_PEEK_H
-#define GRID_LATTICE_PEEK_H
-
-///////////////////////////////////////////////
-// Peeking and poking around
-///////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-
-// FIXME accelerator_loop and accelerator_inline these
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Peek internal indices of a Lattice object
-////////////////////////////////////////////////////////////////////////////////////////////////////
-template<int Index,class vobj> 
-auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Index>(vobj(),i))>
-{
-  Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
-  ret.Checkerboard()=lhs.Checkerboard();
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
-  });
-  return ret;
-};
-template<int Index,class vobj> 
-auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(vobj(),i,j))>
-{
-  Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
-  ret.Checkerboard()=lhs.Checkerboard();
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
-  });
-  return ret;
-};
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Poke internal indices of a Lattice object
-////////////////////////////////////////////////////////////////////////////////////////////////////
-template<int Index,class vobj>  
-void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
-{
-  autoView( rhs_v, rhs, AcceleratorRead);
-  autoView( lhs_v, lhs, AcceleratorWrite);
-  accelerator_for( ss, lhs_v.size(), 1, {
-    pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
-  });
-}
-template<int Index,class vobj> 
-void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
-{
-  autoView( rhs_v, rhs, AcceleratorRead);
-  autoView( lhs_v, lhs, AcceleratorWrite);
-  accelerator_for( ss, lhs_v.size(), 1, {
-    pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
-  });
-}
-
-//////////////////////////////////////////////////////
-// Poke a scalar object into the SIMD array
-//////////////////////////////////////////////////////
-template<class vobj,class sobj> 
-void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
-
-  GridBase *grid=l.Grid();
-
-  int Nsimd = grid->Nsimd();
-
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
-
-  int rank,odx,idx;
-  // Optional to broadcast from node 0.
-  grid->GlobalCoorToRankIndex(rank,odx,idx,site);
-  grid->Broadcast(grid->BossRank(),s);
-
-  // extract-modify-merge cycle is easiest way and this is not perf critical
-  ExtractBuffer<sobj> buf(Nsimd);
-  autoView( l_v , l, CpuWrite);
-  if ( rank == grid->ThisRank() ) {
-    extract(l_v[odx],buf);
-    buf[idx] = s;
-    merge(l_v[odx],buf);
-  }
-
-  return;
-};
-
-
-//////////////////////////////////////////////////////////
-// Peek a scalar object from the SIMD array
-//////////////////////////////////////////////////////////
-template<class vobj>
-typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
-  typename vobj::scalar_object s;
-  peekSite(s,l,site);
-  return s;
-}        
-template<class vobj,class sobj>
-void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
-        
-  GridBase *grid=l.Grid();
-
-  int Nsimd = grid->Nsimd();
-
-  assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
-
-  int rank,odx,idx;
-  grid->GlobalCoorToRankIndex(rank,odx,idx,site);
-
-  ExtractBuffer<sobj> buf(Nsimd);
-  autoView( l_v , l, CpuWrite);
-  extract(l_v[odx],buf);
-
-  s = buf[idx];
-
-  grid->Broadcast(rank,s);
-
-  return;
-};
-
-//////////////////////////////////////////////////////////
-// Peek a scalar object from the SIMD array
-//////////////////////////////////////////////////////////
-// Must be CPU read view
-template<class vobj,class sobj>
-inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
-{
-  GridBase *grid = l.getGrid();
-  assert(l.mode==CpuRead);
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nsimd = grid->Nsimd();
-
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
-
-  static const int words=sizeof(vobj)/sizeof(vector_type);
-  int odx,idx;
-  idx= grid->iIndex(site);
-  odx= grid->oIndex(site);
-  
-  const vector_type *vp = (const vector_type *) &l[odx];
-  scalar_type * pt = (scalar_type *)&s;
-      
-  for(int w=0;w<words;w++){
-    pt[w] = getlane(vp[w],idx);
-  }
-      
-  return;
-};
-template<class vobj,class sobj>
-inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
-{
-  autoView(lv,l,CpuRead);
-  peekLocalSite(s,lv,site);
-  return;
-};
-
-// Must be CPU write view
-template<class vobj,class sobj>
-inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
-{
-  GridBase *grid=l.getGrid();
-  assert(l.mode==CpuWrite);
-
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nsimd = grid->Nsimd();
-
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
-
-  static const int words=sizeof(vobj)/sizeof(vector_type);
-  int odx,idx;
-  idx= grid->iIndex(site);
-  odx= grid->oIndex(site);
-
-  vector_type * vp = (vector_type *)&l[odx];
-  scalar_type * pt = (scalar_type *)&s;
-  for(int w=0;w<words;w++){
-    putlane(vp[w],pt[w],idx);
-  }
-  return;
-};
-
-template<class vobj,class sobj>
-inline void pokeLocalSite(const sobj &s, Lattice<vobj> &l,Coordinate &site)
-{
-  autoView(lv,l,CpuWrite);
-  pokeLocalSite(s,lv,site);
-  return;
-};
-
-NAMESPACE_END(Grid);
-#endif
-

Grid/lattice/Lattice_real_imag.h

@@ -1,79 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_reality.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_REAL_IMAG_H
-#define GRID_LATTICE_REAL_IMAG_H
-
-
-// FIXME .. this is the sector of the code 
-// I am most worried about the directions
-// The choice of burying complex in the SIMD
-// is making the use of "real" and "imag" very cumbersome
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj> inline Lattice<vobj> real(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard()=lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] =real(lhs_v[ss]);
-  });
-  return ret;
-};
-template<class vobj> inline Lattice<vobj> imag(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard()=lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] =imag(lhs_v[ss]);
-  });
-  return ret;
-};
-
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-  auto real(const Expression &expr) -> decltype(real(closure(expr)))		
-{									
-  return real(closure(expr));					
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-  auto imag(const Expression &expr) -> decltype(imag(closure(expr)))		
-{									
-  return imag(closure(expr));					
-}
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/lattice/Lattice_reality.h

@@ -1,116 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_reality.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_REALITY_H
-#define GRID_LATTICE_REALITY_H
-
-
-// FIXME .. this is the sector of the code 
-// I am most worried about the directions
-// The choice of burying complex in the SIMD
-// is making the use of "real" and "imag" very cumbersome
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard()=lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-     ret_v[ss] = adj(lhs_v[ss]);
-  });
-  return ret;
-};
-
-template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard() = lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
-    coalescedWrite( ret_v[ss] , conjugate(lhs_v(ss)));
-  });
-  return ret;
-};
-
-template<class vobj> inline Lattice<typename vobj::Complexified> toComplex(const Lattice<vobj> &lhs){
-  Lattice<typename vobj::Complexified> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard() = lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] = toComplex(lhs_v[ss]);
-  });
-  return ret;
-};
-template<class vobj> inline Lattice<typename vobj::Realified> toReal(const Lattice<vobj> &lhs){
-  Lattice<typename vobj::Realified> ret(lhs.Grid());
-
-  autoView( lhs_v, lhs, AcceleratorRead);
-  autoView( ret_v, ret, AcceleratorWrite);
-
-  ret.Checkerboard() = lhs.Checkerboard();
-  accelerator_for( ss, lhs_v.size(), 1, {
-    ret_v[ss] = toReal(lhs_v[ss]);
-  });
-  return ret;
-};
-
-
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto toComplex(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return toComplex(closure(expr));
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto toReal(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return toReal(closure(expr));
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto adj(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return adj(closure(expr));
-}
-template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
-auto conjugate(const Expression &expr)  -> decltype(closure(expr)) 
-{
-  return conjugate(closure(expr));
-}
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/lattice/Lattice_reduction.h

@@ -1,871 +0,0 @@
-/*************************************************************************************
-    Grid physics library, www.github.com/paboyle/Grid 
-    Source file: ./lib/lattice/Lattice_reduction.h
-    Copyright (C) 2015
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Christoph Lehner <christoph@lhnr.de>
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#pragma once
-
-#include <Grid/Grid_Eigen_Dense.h>
-
-
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-#include <Grid/lattice/Lattice_reduction_gpu.h>
-#endif
-#if defined(GRID_SYCL)
-#include <Grid/lattice/Lattice_reduction_sycl.h>
-#endif
-#include <Grid/lattice/Lattice_slicesum_core.h>
-
-NAMESPACE_BEGIN(Grid);
-
-//////////////////////////////////////////////////////
-// FIXME this should promote to double and accumulate
-//////////////////////////////////////////////////////
-template<class vobj>
-inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
-{
-  typedef typename vobj::scalar_object  sobj;
-
-  //  const int Nsimd = vobj::Nsimd();
-  const int nthread = GridThread::GetThreads();
-
-  Vector<sobj> sumarray(nthread);
-  for(int i=0;i<nthread;i++){
-    sumarray[i]=Zero();
-  }
-  
-  thread_for(thr,nthread, {
-    int nwork, mywork, myoff;
-    nwork = osites;
-    GridThread::GetWork(nwork,thr,mywork,myoff);
-    vobj vvsum=Zero();
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      vvsum = vvsum + arg[ss];
-    }
-    sumarray[thr]=Reduce(vvsum);
-  });
-  
-  sobj ssum=Zero();  // sum across threads
-  for(int i=0;i<nthread;i++){
-    ssum = ssum+sumarray[i];
-  } 
-  return ssum;
-}
-template<class vobj>
-inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
-{
-  typedef typename vobj::scalar_objectD  sobj;
-
-  const int nthread = GridThread::GetThreads();
-
-  Vector<sobj> sumarray(nthread);
-  for(int i=0;i<nthread;i++){
-    sumarray[i]=Zero();
-  }
-  
-  thread_for(thr,nthread, {
-    int nwork, mywork, myoff;
-    nwork = osites;
-    GridThread::GetWork(nwork,thr,mywork,myoff);
-    vobj vvsum=Zero();
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      vvsum = vvsum + arg[ss];
-    }
-    sumarray[thr]=Reduce(vvsum);
-  });
-  
-  sobj ssum=Zero();  // sum across threads
-  for(int i=0;i<nthread;i++){
-    ssum = ssum+sumarray[i];
-  } 
-  return ssum;
-}
-/*
-Threaded max, don't use for now
-template<class Double>
-inline Double max(const Double *arg, Integer osites)
-{
-  //  const int Nsimd = vobj::Nsimd();
-  const int nthread = GridThread::GetThreads();
-
-  std::vector<Double> maxarray(nthread);
-  
-  thread_for(thr,nthread, {
-    int nwork, mywork, myoff;
-    nwork = osites;
-    GridThread::GetWork(nwork,thr,mywork,myoff);
-    Double max=arg[0];
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      if( arg[ss] > max ) max = arg[ss];
-    }
-    maxarray[thr]=max;
-  });
-  
-  Double tmax=maxarray[0];
-  for(int i=0;i<nthread;i++){
-    if (maxarray[i]>tmax) tmax = maxarray[i];
-  } 
-  return tmax;
-}
-*/
-template<class vobj>
-inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  return sum_gpu(arg,osites);
-#else
-  return sum_cpu(arg,osites);
-#endif  
-}
-template<class vobj>
-inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  return sumD_gpu(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
-template<class vobj>
-inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  return sumD_gpu_large(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
-
-template<class vobj>
-inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
-{
-  Integer osites = arg.Grid()->oSites();
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  autoView( arg_v, arg, AcceleratorRead);
-  return sum_gpu(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  return sum_cpu(&arg_v[0],osites);
-#endif  
-}
-
-template<class vobj>
-inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
-{
-  auto ssum = rankSum(arg);
-  arg.Grid()->GlobalSum(ssum);
-  return ssum;
-}
-
-template<class vobj>
-inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  autoView( arg_v, arg, AcceleratorRead);
-  Integer osites = arg.Grid()->oSites();
-  return sum_gpu_large(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
-  return sum_cpu(&arg_v[0],osites);
-#endif
-}
-
-template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
-{
-  auto ssum = rankSumLarge(arg);
-  arg.Grid()->GlobalSum(ssum);
-  return ssum;
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Deterministic Reduction operations
-////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
-  ComplexD nrm = innerProduct(arg,arg);
-  return real(nrm); 
-}
-
-//The global maximum of the site norm2
-template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
-{
-  typedef typename vobj::tensor_reduced vscalar;  //iScalar<iScalar<.... <vPODtype> > >
-  typedef typename vscalar::scalar_object  scalar;   //iScalar<iScalar<.... <PODtype> > >
-
-  Lattice<vscalar> inner = localNorm2(arg);
-
-  auto grid = arg.Grid();
-
-  RealD max;
-  for(int l=0;l<grid->lSites();l++){
-    Coordinate coor;
-    scalar val;
-    RealD r;
-    grid->LocalIndexToLocalCoor(l,coor);
-    peekLocalSite(val,inner,coor);
-    r=real(TensorRemove(val));
-    if( (l==0) || (r>max)){
-      max=r;
-    }
-  }
-  grid->GlobalMax(max);
-  return max;
-}
-
-// Double inner product
-template<class vobj>
-inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
-{
-  typedef typename vobj::vector_typeD vector_type;
-  ComplexD  nrm;
-  
-  GridBase *grid = left.Grid();
-
-  const uint64_t nsimd = grid->Nsimd();
-  const uint64_t sites = grid->oSites();
-  
-  // Might make all code paths go this way.
-#if 0
-  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  {
-    autoView( left_v , left, AcceleratorRead);
-    autoView( right_v,right, AcceleratorRead);
-    // This code could read coalesce
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
-    });
-  }
-#else
-  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-    
-  {
-    autoView( left_v , left, AcceleratorRead);
-    autoView( right_v,right, AcceleratorRead);
-
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
-    });
-  }
-#endif
-  // This is in single precision and fails some tests
-  auto anrm = sumD(inner_tmp_v,sites);  
-  nrm = anrm;
-  return nrm;
-}
-
-
-template<class vobj>
-inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
-  GridBase *grid = left.Grid();
-
-#ifdef GRID_SYCL
-  uint64_t csum=0;
-  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
-  {
-    // Hack
-    // Fast integer xor checksum. Can also be used in comms now.
-    autoView(l_v,left,AcceleratorRead);
-    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
-    uint64_t *base= (uint64_t *)&l_v[0];
-    csum=svm_xor(base,words);
-  }
-  FlightRecorder::CsumLog(csum);
-#endif
-  ComplexD nrm = rankInnerProduct(left,right);
-  RealD local = real(nrm);
-  FlightRecorder::NormLog(real(nrm)); 
-  grid->GlobalSum(nrm);
-  FlightRecorder::ReductionLog(local,real(nrm)); 
-  return nrm;
-}
-
-
-/////////////////////////
-// Fast axpby_norm
-// z = a x + b y
-// return norm z
-/////////////////////////
-template<class sobj,class vobj> strong_inline RealD 
-axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y) 
-{
-  sobj one(1.0);
-  return axpby_norm_fast(z,a,one,x,y);
-}
-
-template<class sobj,class vobj> strong_inline RealD 
-axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) 
-{
-  z.Checkerboard() = x.Checkerboard();
-  conformable(z,x);
-  conformable(x,y);
-
-  //  typedef typename vobj::vector_typeD vector_type;
-  RealD  nrm;
-  
-  GridBase *grid = x.Grid();
-
-  const uint64_t nsimd = grid->Nsimd();
-  const uint64_t sites = grid->oSites();
-  
-  // GPU
-  autoView( x_v, x, AcceleratorRead);
-  autoView( y_v, y, AcceleratorRead);
-  autoView( z_v, z, AcceleratorWrite);
-#if 0
-  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#else
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-#endif
-  grid->GlobalSum(nrm);
-  return nrm; 
-}
- 
-template<class vobj> strong_inline void
-innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Lattice<vobj> &right)
-{
-  conformable(left,right);
-
-  typedef typename vobj::vector_typeD vector_type;
-  Vector<ComplexD> tmp(2);
-
-  GridBase *grid = left.Grid();
-
-  const uint64_t nsimd = grid->Nsimd();
-  const uint64_t sites = grid->oSites();
-
-  // GPU
-  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
-  typedef decltype(innerProductD(vobj(),vobj())) norm_t;
-  Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t>  norm_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  auto norm_tmp_v = &norm_tmp[0];
-  {
-    autoView(left_v,left, AcceleratorRead);
-    autoView(right_v,right,AcceleratorRead);
-    accelerator_for( ss, sites, 1,{
-	auto left_tmp = left_v[ss];
-	inner_tmp_v[ss]=innerProductD(left_tmp,right_v[ss]);
-        norm_tmp_v [ss]=innerProductD(left_tmp,left_tmp);
-      });
-  }
-
-  tmp[0] = TensorRemove(sum(inner_tmp_v,sites));
-  tmp[1] = TensorRemove(sum(norm_tmp_v,sites));
-
-  grid->GlobalSumVector(&tmp[0],2); // keep norm Complex -> can use GlobalSumVector
-  ip = tmp[0];
-  nrm = real(tmp[1]);
-}
-
-template<class Op,class T1>
-inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
-  ->typename decltype(expr.op.func(eval(0,expr.arg1)))::scalar_object
-{
-  return sum(closure(expr));
-}
-
-template<class Op,class T1,class T2>
-inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
-      ->typename decltype(expr.op.func(eval(0,expr.arg1),eval(0,expr.arg2)))::scalar_object
-{
-  return sum(closure(expr));
-}
-
-
-template<class Op,class T1,class T2,class T3>
-inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
-  ->typename decltype(expr.op.func(eval(0,expr.arg1),
-				      eval(0,expr.arg2),
-				      eval(0,expr.arg3)
-				      ))::scalar_object
-{
-  return sum(closure(expr));
-}
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////
-// sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
-{
-  ///////////////////////////////////////////////////////
-  // FIXME precision promoted summation
-  // may be important for correlation functions
-  // But easily avoided by using double precision fields
-  ///////////////////////////////////////////////////////
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_object::scalar_type scalar_type;
-  GridBase  *grid = Data.Grid();
-  assert(grid!=NULL);
-
-  const int    Nd = grid->_ndimension;
-  const int Nsimd = grid->Nsimd();
-
-  assert(orthogdim >= 0);
-  assert(orthogdim < Nd);
-
-  int fd=grid->_fdimensions[orthogdim];
-  int ld=grid->_ldimensions[orthogdim];
-  int rd=grid->_rdimensions[orthogdim];
-
-  Vector<vobj> lvSum(rd); // will locally sum vectors first
-  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
-  ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
-
-  result.resize(fd); // And then global sum to return the same vector to every node 
-  for(int r=0;r<rd;r++){
-    lvSum[r]=Zero();
-  }
-
-  int e1=    grid->_slice_nblock[orthogdim];
-  int e2=    grid->_slice_block [orthogdim];
-  int stride=grid->_slice_stride[orthogdim];
-  int ostride=grid->_ostride[orthogdim];
-  
-  //Reduce Data down to lvSum
-  sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
-
-  // Sum across simd lanes in the plane, breaking out orthog dir.
-  Coordinate icoor(Nd);
-
-  for(int rt=0;rt<rd;rt++){
-
-    extract(lvSum[rt],extracted);
-
-    for(int idx=0;idx<Nsimd;idx++){
-
-      grid->iCoorFromIindex(icoor,idx);
-
-      int ldx =rt+icoor[orthogdim]*rd;
-
-      lsSum[ldx]=lsSum[ldx]+extracted[idx];
-
-    }
-  }
-  
-  // sum over nodes.
-  for(int t=0;t<fd;t++){
-    int pt = t/ld; // processor plane
-    int lt = t%ld;
-    if ( pt == grid->_processor_coor[orthogdim] ) {
-      result[t]=lsSum[lt];
-    } else {
-      result[t]=Zero();
-    }
-
-  }
-  scalar_type * ptr = (scalar_type *) &result[0];
-  int words = fd*sizeof(sobj)/sizeof(scalar_type);
-  grid->GlobalSumVector(ptr, words);
-}
-template<class vobj> inline
-std::vector<typename vobj::scalar_object> 
-sliceSum(const Lattice<vobj> &Data,int orthogdim)
-{
-  std::vector<typename vobj::scalar_object> result;
-  sliceSum(Data,result,orthogdim);
-  return result;
-}
-
-
-template<class vobj>
-static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
-{
-  typedef typename vobj::vector_type   vector_type;
-  typedef typename vobj::scalar_type   scalar_type;
-  GridBase  *grid = lhs.Grid();
-  assert(grid!=NULL);
-  conformable(grid,rhs.Grid());
-
-  const int    Nd = grid->_ndimension;
-  const int Nsimd = grid->Nsimd();
-
-  assert(orthogdim >= 0);
-  assert(orthogdim < Nd);
-
-  int fd=grid->_fdimensions[orthogdim];
-  int ld=grid->_ldimensions[orthogdim];
-  int rd=grid->_rdimensions[orthogdim];
-
-  Vector<vector_type> lvSum(rd); // will locally sum vectors first
-  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
-  ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
-
-  result.resize(fd); // And then global sum to return the same vector to every node for IO to file
-  for(int r=0;r<rd;r++){
-    lvSum[r]=Zero();
-  }
-
-  int e1=    grid->_slice_nblock[orthogdim];
-  int e2=    grid->_slice_block [orthogdim];
-  int stride=grid->_slice_stride[orthogdim];
-
-  autoView( lhv, lhs, CpuRead);
-  autoView( rhv, rhs, CpuRead);
-  thread_for( r,rd,{
-
-    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int ss= so+n*stride+b;
-	vector_type vv = TensorRemove(innerProduct(lhv[ss],rhv[ss]));
-	lvSum[r]=lvSum[r]+vv;
-      }
-    }
-  });
-
-  // Sum across simd lanes in the plane, breaking out orthog dir.
-  Coordinate icoor(Nd);
-  for(int rt=0;rt<rd;rt++){
-
-    iScalar<vector_type> temp; 
-    temp._internal = lvSum[rt];
-    extract(temp,extracted);
-
-    for(int idx=0;idx<Nsimd;idx++){
-
-      grid->iCoorFromIindex(icoor,idx);
-
-      int ldx =rt+icoor[orthogdim]*rd;
-
-      lsSum[ldx]=lsSum[ldx]+extracted[idx]._internal;
-
-    }
-  }
-  
-  // sum over nodes.
-  scalar_type gsum;
-  for(int t=0;t<fd;t++){
-    int pt = t/ld; // processor plane
-    int lt = t%ld;
-    if ( pt == grid->_processor_coor[orthogdim] ) {
-      gsum=lsSum[lt];
-    } else {
-      gsum=scalar_type(0.0);
-    }
-
-    grid->GlobalSum(gsum);
-
-    result[t]=gsum;
-  }
-}
-template<class vobj>
-static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Orthog) 
-{
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-  
-  int Nblock = rhs.Grid()->GlobalDimensions()[Orthog];
-  std::vector<ComplexD> ip(Nblock);
-  sn.resize(Nblock);
-  
-  sliceInnerProductVector(ip,rhs,rhs,Orthog);
-  for(int ss=0;ss<Nblock;ss++){
-    sn[ss] = real(ip[ss]);
-  }
-};
-
-
-template<class vobj>
-static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
-			    int orthogdim,RealD scale=1.0) 
-{
-  // perhaps easier to just promote A to a field and use regular madd
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::tensor_reduced tensor_reduced;
-  
-  scalar_type zscale(scale);
-
-  GridBase *grid  = X.Grid();
-
-  int Nsimd  =grid->Nsimd();
-  int Nblock =grid->GlobalDimensions()[orthogdim];
-
-  int fd     =grid->_fdimensions[orthogdim];
-  int ld     =grid->_ldimensions[orthogdim];
-  int rd     =grid->_rdimensions[orthogdim];
-
-  int e1     =grid->_slice_nblock[orthogdim];
-  int e2     =grid->_slice_block [orthogdim];
-  int stride =grid->_slice_stride[orthogdim];
-
-  Coordinate icoor;
-  for(int r=0;r<rd;r++){
-
-    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-
-    vector_type    av;
-
-    for(int l=0;l<Nsimd;l++){
-      grid->iCoorFromIindex(icoor,l);
-      int ldx =r+icoor[orthogdim]*rd;
-      av.putlane(scalar_type(a[ldx])*zscale,l);
-    }
-
-    tensor_reduced at; at=av;
-
-    autoView( Rv, R, CpuWrite);
-    autoView( Xv, X, CpuRead);
-    autoView( Yv, Y, CpuRead);
-    thread_for2d( n, e1, b,e2, {
-	int ss= so+n*stride+b;
-	Rv[ss] = at*Xv[ss]+Yv[ss];
-    });
-  }
-};
-
-/*
-inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
-{
-  int NN    = BlockSolverGrid->_ndimension;
-  int nsimd = BlockSolverGrid->Nsimd();
-  
-  std::vector<int> latt_phys(0);
-  std::vector<int> simd_phys(0);
-  std::vector<int>  mpi_phys(0);
-  
-  for(int d=0;d<NN;d++){
-    if( d!=Orthog ) { 
-      latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
-      simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
-      mpi_phys.push_back(BlockSolverGrid->_processors[d]);
-    }
-  }
-  return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
-}
-*/
-
-template<class vobj>
-static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
-{    
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-
-  GridBase *FullGrid  = X.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-
-  //  Lattice<vobj> Xslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl = nh-1;
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-
-  autoView( X_v, X, CpuRead);
-  autoView( Y_v, Y, CpuRead);
-  autoView( R_v, R, CpuWrite);
-  thread_region
-  {
-    Vector<vobj> s_x(Nblock);
-
-    thread_for_collapse_in_region(2, n,nblock, {
-     for(int b=0;b<block;b++){
-      int o  = n*stride + b;
-
-      for(int i=0;i<Nblock;i++){
-	s_x[i] = X_v[o+i*ostride];
-      }
-
-      vobj dot;
-      for(int i=0;i<Nblock;i++){
-	dot = Y_v[o+i*ostride];
-	for(int j=0;j<Nblock;j++){
-	  dot = dot + s_x[j]*(scale*aa(j,i));
-	}
-	R_v[o+i*ostride]=dot;
-      }
-    }});
-  }
-};
-
-template<class vobj>
-static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
-{    
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-
-  GridBase *FullGrid  = X.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-  //  Lattice<vobj> Xslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl=1;
-
-  //FIXME package in a convenient iterator
-  // thread_for2d_in_region
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-  autoView( R_v, R, CpuWrite);
-  autoView( X_v, X, CpuRead);
-  thread_region
-  {
-    std::vector<vobj> s_x(Nblock);
-
-
-    thread_for_collapse_in_region( 2 ,n,nblock,{
-    for(int b=0;b<block;b++){
-      int o  = n*stride + b;
-
-      for(int i=0;i<Nblock;i++){
-	s_x[i] = X_v[o+i*ostride];
-      }
-
-      vobj dot;
-      for(int i=0;i<Nblock;i++){
-	dot = s_x[0]*(scale*aa(0,i));
-	for(int j=1;j<Nblock;j++){
-	  dot = dot + s_x[j]*(scale*aa(j,i));
-	}
-	R_v[o+i*ostride]=dot;
-      }
-    }});
-  }
-};
-
-
-template<class vobj>
-static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
-{
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-  
-  GridBase *FullGrid  = lhs.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-  
-  int Nblock = FullGrid->GlobalDimensions()[Orthog];
-  
-  //  Lattice<vobj> Lslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-  
-  mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl = nh-1;
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-
-  typedef typename vobj::vector_typeD vector_typeD;
-
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( rhs_v, rhs, CpuRead);
-  thread_region
-  {
-    std::vector<vobj> Left(Nblock);
-    std::vector<vobj> Right(Nblock);
-    Eigen::MatrixXcd  mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-    thread_for_collapse_in_region( 2, n,nblock,{
-    for(int b=0;b<block;b++){
-
-      int o  = n*stride + b;
-
-      for(int i=0;i<Nblock;i++){
-	Left [i] = lhs_v[o+i*ostride];
-	Right[i] = rhs_v[o+i*ostride];
-      }
-
-      for(int i=0;i<Nblock;i++){
-      for(int j=0;j<Nblock;j++){
-	auto tmp = innerProduct(Left[i],Right[j]);
-	auto rtmp = TensorRemove(tmp);
-	auto red  =  Reduce(rtmp);
-	mat_thread(i,j) += std::complex<double>(real(red),imag(red));
-      }}
-    }});
-    thread_critical
-    {
-      mat += mat_thread;
-    }  
-  }
-
-  for(int i=0;i<Nblock;i++){
-  for(int j=0;j<Nblock;j++){
-    ComplexD sum = mat(i,j);
-    FullGrid->GlobalSum(sum);
-    mat(i,j)=sum;
-  }}
-
-  return;
-}
-
-NAMESPACE_END(Grid);
-
-
-
-

Grid/lattice/Lattice_reduction_gpu.h

@@ -1,302 +0,0 @@
-NAMESPACE_BEGIN(Grid);
-
-#ifdef GRID_HIP
-extern hipDeviceProp_t *gpu_props;
-#define WARP_SIZE 64
-#endif
-#ifdef GRID_CUDA
-extern cudaDeviceProp *gpu_props;
-#define WARP_SIZE 32
-#endif
-
-__device__ unsigned int retirementCount = 0;
-
-template <class Iterator>
-unsigned int nextPow2(Iterator x) {
-  --x;
-  x |= x >> 1;
-  x |= x >> 2;
-  x |= x >> 4;
-  x |= x >> 8;
-  x |= x >> 16;
-  return ++x;
-}
-
-template <class Iterator>
-int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
-  
-  int device;
-#ifdef GRID_CUDA
-  cudaGetDevice(&device);
-#endif
-#ifdef GRID_HIP
-  auto r=hipGetDevice(&device);
-#endif
-  
-  Iterator warpSize            = gpu_props[device].warpSize;
-  Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
-  Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
-  Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  /*  
-  std::cout << GridLogDebug << "GPU has:" << std::endl;
-  std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
-  std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
-  std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
-  std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  */  
-  if (warpSize != WARP_SIZE) {
-    std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
-    exit(EXIT_FAILURE);
-  }
-  
-  // let the number of threads in a block be a multiple of 2, starting from warpSize
-  threads = warpSize;
-  if ( threads*sizeofsobj > sharedMemPerBlock ) {
-    std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
-    return 0;
-  }
-  while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
-  // keep all the streaming multiprocessors busy
-  blocks = nextPow2(multiProcessorCount);
-  return 1;
-}
-
-template <class sobj, class Iterator>
-__device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid) {
-  
-  Iterator blockSize = blockDim.x;
-  
-  // cannot use overloaded operators for sobj as they are not volatile-qualified
-  memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
-  acceleratorSynchronise();
-  
-  const Iterator VEC = WARP_SIZE;
-  const Iterator vid = tid & (VEC-1);
-  
-  sobj beta, temp;
-  memcpy((void *)&beta, (void *)&mySum, sizeof(sobj));
-  
-  for (int i = VEC/2; i > 0; i>>=1) {
-    if (vid < i) {
-      memcpy((void *)&temp, (void *)&sdata[tid+i], sizeof(sobj));
-      beta += temp;
-      memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
-    }
-    acceleratorSynchronise();
-  }
-  acceleratorSynchroniseAll();
-  
-  if (threadIdx.x == 0) {
-    beta  = Zero();
-    for (Iterator i = 0; i < blockSize; i += VEC) {
-      memcpy((void *)&temp, (void *)&sdata[i], sizeof(sobj));
-      beta  += temp;
-    }
-    memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
-  }
-  acceleratorSynchroniseAll();
-}
-
-
-template <class vobj, class sobj, class Iterator>
-__device__ void reduceBlocks(const vobj *g_idata, sobj *g_odata, Iterator n) 
-{
-  constexpr Iterator nsimd = vobj::Nsimd();
-  
-  Iterator blockSize = blockDim.x;
-  
-  // force shared memory alignment
-  extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
-  // it's not possible to have two extern __shared__ arrays with same name
-  // but different types in different scopes -- need to cast each time
-  sobj *sdata = (sobj *)shmem_pointer;
-  
-  // first level of reduction,
-  // each thread writes result in mySum
-  Iterator tid = threadIdx.x;
-  Iterator i = blockIdx.x*(blockSize*2) + threadIdx.x;
-  Iterator gridSize = blockSize*2*gridDim.x;
-  sobj mySum = Zero();
-  
-  while (i < n) {
-    Iterator lane = i % nsimd;
-    Iterator ss   = i / nsimd;
-    auto tmp = extractLane(lane,g_idata[ss]);
-    sobj tmpD;
-    tmpD=tmp;
-    mySum   +=tmpD;
-    
-    if (i + blockSize < n) {
-      lane = (i+blockSize) % nsimd;
-      ss   = (i+blockSize) / nsimd;
-      tmp = extractLane(lane,g_idata[ss]);
-      tmpD = tmp;
-      mySum += tmpD;
-    }
-    i += gridSize;
-  }
-  
-  // copy mySum to shared memory and perform
-  // reduction for all threads in this block
-  reduceBlock(sdata, mySum, tid);
-  if (tid == 0) g_odata[blockIdx.x] = sdata[0];
-}
-
-template <class vobj, class sobj,class Iterator>
-__global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
-  
-  Iterator blockSize = blockDim.x;
-  
-  // perform reduction for this block and
-  // write result to global memory buffer
-  reduceBlocks(lat, buffer, n);
-  
-  if (gridDim.x > 1) {
-    
-    const Iterator tid = threadIdx.x;
-    __shared__ bool amLast;
-    // force shared memory alignment
-    extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
-    // it's not possible to have two extern __shared__ arrays with same name
-    // but different types in different scopes -- need to cast each time
-    sobj *smem = (sobj *)shmem_pointer;
-    
-    // wait until all outstanding memory instructions in this thread are finished
-    acceleratorFence();
-    
-    if (tid==0) {
-      unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
-      // true if this block is the last block to be done
-      amLast = (ticket == gridDim.x-1);
-    }
-    
-    // each thread must read the correct value of amLast
-    acceleratorSynchroniseAll();
-
-    if (amLast) {
-      // reduce buffer[0], ..., buffer[gridDim.x-1]
-      Iterator i = tid;
-      sobj mySum = Zero();
-      
-      while (i < gridDim.x) {
-        mySum += buffer[i];
-        i += blockSize;
-      }
-      
-      reduceBlock(smem, mySum, tid);
-      
-      if (tid==0) {
-        buffer[0] = smem[0];
-        // reset count variable
-        retirementCount = 0;
-      }
-    }
-  }
-}
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Possibly promote to double and sum
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
-{
-  typedef typename vobj::scalar_objectD sobj;
-  typedef decltype(lat) Iterator;
-  
-  Integer nsimd= vobj::Nsimd();
-  Integer size = osites*nsimd;
-
-  Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
-
-  Integer smemSize = numThreads * sizeof(sobj);
-  // Move out of UVM
-  // Turns out I had messed up the synchronise after move to compute stream
-  // as running this on the default stream fools the synchronise
-#undef UVM_BLOCK_BUFFER  
-#ifndef UVM_BLOCK_BUFFER  
-  commVector<sobj> buffer(numBlocks);
-  sobj *buffer_v = &buffer[0];
-  sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
-  accelerator_barrier();
-  acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
-#else
-  Vector<sobj> buffer(numBlocks);
-  sobj *buffer_v = &buffer[0];
-  sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
-  accelerator_barrier();
-  result = *buffer_v;
-#endif
-  return result;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<vector> buffer(osites);
-  vector *dat = (vector *)lat;
-  vector *buf = &buffer[0];
-  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,1,{
-	buf[ss] = dat[ss*words+w];
-      });
-      
-    ret_p[w] = sumD_gpu_small(tbuf,osites);
-  }
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  
-  Integer nsimd= vobj::Nsimd();
-  Integer size = osites*nsimd;
-  Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  
-  if ( ok ) {
-    ret = sumD_gpu_small(lat,osites);
-  } else {
-    ret = sumD_gpu_large(lat,osites);
-  }
-  return ret;
-}
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Return as same precision as input performing reduction in double precision though
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu(lat,osites);
-  return result;
-}
-
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
-
-NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_sycl.h

@@ -1,126 +0,0 @@
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Possibly promote to double and sum
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
-{
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_objectD sobjD;
-  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
-  sobj identity; zeroit(identity);
-  sobj ret ; 
-
-  Integer nsimd= vobj::Nsimd();
-  
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
-     cgh.parallel_for(cl::sycl::range<1>{osites},
-		      Reduction,
-		      [=] (cl::sycl::id<1> item, auto &sum) {
-      auto osite   = item[0];
-      sum +=Reduce(lat[osite]);
-     });
-   });
-  theGridAccelerator->wait();
-  ret = mysum[0];
-  free(mysum,*theGridAccelerator);
-  sobjD dret; convertType(dret,ret);
-  return dret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  return sumD_gpu_tensor(lat,osites);
-}
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
-{
-  return sumD_gpu_large(lat,osites);
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  return sumD_gpu_large(lat,osites);
-}
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Return as same precision as input performing reduction in double precision though
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu(lat,osites);
-  return result;
-}
-
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
-
-
-template<class Word> Word svm_xor(Word *vec,uint64_t L)
-{
-  Word xorResult; xorResult = 0;
-  Word *d_sum =(Word *)cl::sycl::malloc_shared(sizeof(Word),*theGridAccelerator);
-  Word identity;  identity=0;
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(d_sum,identity,std::bit_xor<>());
-     cgh.parallel_for(cl::sycl::range<1>{L},
-		      Reduction,
-		      [=] (cl::sycl::id<1> index, auto &sum) {
-	 sum ^=vec[index];
-     });
-   });
-  theGridAccelerator->wait();
-  Word ret = d_sum[0];
-  free(d_sum,*theGridAccelerator);
-  return ret;
-}
-
-NAMESPACE_END(Grid);
-
-/*
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_type  scalar;
-
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobjD;
-
-  sobjD ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int nsimd = vobj::Nsimd();
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<scalar> buffer(osites*nsimd);
-  scalar *buf = &buffer[0];
-  vector *dat = (vector *)lat;
-
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,nsimd,{
-	int lane = acceleratorSIMTlane(nsimd);
-	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
-    });
-    //Precision change at this point is to late to gain precision
-    ret_p[w] = svm_reduce(buf,nsimd*osites);
-  }
-  return ret;
-}
-*/

Grid/lattice/Lattice_rng.h

@@ -1,536 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_rng.h
-
-    Copyright (C) 2015
-
-    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: Guido Cossu <guido.cossu@ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_RNG_H
-#define GRID_LATTICE_RNG_H
-
-#include <random>
-
-#ifdef RNG_SITMO
-#include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
-#endif 
-
-#if defined(RNG_SITMO)
-#define RNG_FAST_DISCARD
-#else 
-#undef  RNG_FAST_DISCARD
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-//////////////////////////////////////////////////////////////
-// Allow the RNG state to be less dense than the fine grid
-//////////////////////////////////////////////////////////////
-inline int RNGfillable(GridBase *coarse,GridBase *fine)
-{
-
-  int rngdims = coarse->_ndimension;
-
-  // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
-  int lowerdims   = fine->_ndimension - coarse->_ndimension;
-  assert(lowerdims >= 0);
-  for(int d=0;d<lowerdims;d++){
-    assert(fine->_simd_layout[d]==1);
-    assert(fine->_processors[d]==1);
-  }
-
-  int multiplicity=1;
-  for(int d=0;d<lowerdims;d++){
-    multiplicity=multiplicity*fine->_rdimensions[d];
-  }
-  // local and global volumes subdivide cleanly after SIMDization
-  for(int d=0;d<rngdims;d++){
-    int fd= d+lowerdims;
-    assert(coarse->_processors[d]  == fine->_processors[fd]);
-    assert(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
-    assert(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]); 
-
-    multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d]; 
-  }
-  return multiplicity;
-}
-
-  
-// merge of April 11 2017
-// this function is necessary for the LS vectorised field
-inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
-{
-  int rngdims = coarse->_ndimension;
-    
-  // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
-  int lowerdims   = fine->_ndimension - coarse->_ndimension;  assert(lowerdims >= 0);
-  // assumes that the higher dimensions are not using more processors
-  // all further divisions are local
-  for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
-  for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
-
-  // then divide the number of local sites
-  // check that the total number of sims agree, meanse the iSites are the same
-  assert(fine->Nsimd() == coarse->Nsimd());
-
-  // check that the two grids divide cleanly
-  assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
-
-  return fine->lSites() / coarse->lSites();
-}
-  
-// real scalars are one component
-template<class scalar,class distribution,class generator> 
-void fillScalar(scalar &s,distribution &dist,generator & gen)
-{
-  s=dist(gen);
-}
-template<class distribution,class generator> 
-void fillScalar(ComplexF &s,distribution &dist, generator &gen)
-{
-  //  s=ComplexF(dist(gen),dist(gen));
-  s.real(dist(gen));
-  s.imag(dist(gen));
-}
-template<class distribution,class generator> 
-void fillScalar(ComplexD &s,distribution &dist,generator &gen)
-{
-  //  s=ComplexD(dist(gen),dist(gen));
-  s.real(dist(gen));
-  s.imag(dist(gen));
-}
-  
-class GridRNGbase {
-public:
-  // One generator per site.
-  // Uniform and Gaussian distributions from these generators.
-#ifdef RNG_RANLUX
-  typedef std::ranlux48 RngEngine;
-  typedef uint64_t      RngStateType;
-  static const int RngStateCount = 15;
-#endif 
-#ifdef RNG_MT19937 
-  typedef std::mt19937 RngEngine;
-  typedef uint32_t     RngStateType;
-  static const int     RngStateCount = std::mt19937::state_size;
-#endif
-#ifdef RNG_SITMO
-  typedef sitmo::prng_engine 	RngEngine;
-  typedef uint64_t    	RngStateType;
-  static const int    	RngStateCount = 13;
-#endif
-
-  std::vector<RngEngine>                             _generators;
-  std::vector<std::uniform_real_distribution<RealD> > _uniform;
-  std::vector<std::normal_distribution<RealD> >       _gaussian;
-  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
-  std::vector<std::uniform_int_distribution<uint32_t> > _uid;
-
-  ///////////////////////
-  // support for parallel init
-  ///////////////////////
-#ifdef RNG_FAST_DISCARD
-  static void Skip(RngEngine &eng,uint64_t site)
-  {
-#if 0
-    /////////////////////////////////////////////////////////////////////////////////////
-    // Skip by 2^40 elements between successive lattice sites
-    // This goes by 10^12.
-    // Consider quenched updating; likely never exceeding rate of 1000 sweeps
-    // per second on any machine. This gives us of order 10^9 seconds, or 100 years
-    // skip ahead.
-    // For HMC unlikely to go at faster than a solve per second, and 
-    // tens of seconds per trajectory so this is clean in all reasonable cases,
-    // and margin of safety is orders of magnitude.
-    // We could hack Sitmo to skip in the higher order words of state if necessary
-    //
-    // Replace with 2^30 ; avoid problem on large volumes
-    //
-    /////////////////////////////////////////////////////////////////////////////////////
-    //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
-    const int shift = 30;
-
-    ////////////////////////////////////////////////////////////////////
-    // Weird compiler bug in Intel 2018.1 under O3 was generating 32bit and not 64 bit left shift.
-    ////////////////////////////////////////////////////////////////////
-    volatile uint64_t skip = site;
-
-    skip = skip<<shift;
-
-    assert((skip >> shift)==site); // check for overflow
-
-    eng.discard(skip);
-#else
-    eng.discardhi(site);
-#endif
-    //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
-  } 
-#endif
-  static RngEngine Reseed(RngEngine &eng)
-  {
-    std::vector<uint32_t> newseed;
-    std::uniform_int_distribution<uint32_t> uid;
-    return Reseed(eng,newseed,uid);
-  }
-  static RngEngine Reseed(RngEngine &eng,std::vector<uint32_t> & newseed,
-			  std::uniform_int_distribution<uint32_t> &uid)
-  {
-    const int reseeds=4;
-      
-    newseed.resize(reseeds);
-    for(int i=0;i<reseeds;i++){
-      newseed[i] = uid(eng);
-    }
-    std::seed_seq sseq(newseed.begin(),newseed.end());
-    return RngEngine(sseq);
-  }    
-
-  void GetState(std::vector<RngStateType> & saved,RngEngine &eng) {
-    saved.resize(RngStateCount);
-    std::stringstream ss;
-    ss<<eng;
-    ss.seekg(0,ss.beg);
-    for(int i=0;i<RngStateCount;i++){
-      ss>>saved[i];
-    }
-  }
-  void GetState(std::vector<RngStateType> & saved,int gen) {
-    GetState(saved,_generators[gen]);
-  }
-  void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
-    assert(saved.size()==RngStateCount);
-    std::stringstream ss;
-    for(int i=0;i<RngStateCount;i++){
-      ss<< saved[i]<<" ";
-    }
-    ss.seekg(0,ss.beg);
-    ss>>eng;
-  }
-  void SetState(std::vector<RngStateType> & saved,int gen){
-    SetState(saved,_generators[gen]);
-  }
-  void SetEngine(RngEngine &Eng, int gen){
-    _generators[gen]=Eng;
-  }
-  void GetEngine(RngEngine &Eng, int gen){
-    Eng=_generators[gen];
-  }
-  template<class source> void Seed(source &src, int gen)
-  {
-    _generators[gen] = RngEngine(src);
-  }    
-};
-
-class GridSerialRNG : public GridRNGbase {
-public:
-
-  GridSerialRNG() : GridRNGbase() {
-    _generators.resize(1);
-    _uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
-    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
-    _uid.resize(1,std::uniform_int_distribution<uint32_t>() );
-  }
-
-  template <class sobj,class distribution> inline void fill(sobj &l,std::vector<distribution> &dist){
-
-    typedef typename sobj::scalar_type scalar_type;
- 
-    int words = sizeof(sobj)/sizeof(scalar_type);
-
-    scalar_type *buf = (scalar_type *) & l;
-
-    dist[0].reset();
-    for(int idx=0;idx<words;idx++){
-      fillScalar(buf[idx],dist[0],_generators[0]);
-    }
-
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-
-  }
-
-  template <class distribution>  inline void fill(ComplexF &l,std::vector<distribution> &dist){
-    dist[0].reset();
-    fillScalar(l,dist[0],_generators[0]);
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  template <class distribution>  inline void fill(ComplexD &l,std::vector<distribution> &dist){
-    dist[0].reset();
-    fillScalar(l,dist[0],_generators[0]);
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  template <class distribution>  inline void fill(RealF &l,std::vector<distribution> &dist){
-    dist[0].reset();
-    fillScalar(l,dist[0],_generators[0]);
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  template <class distribution>  inline void fill(RealD &l,std::vector<distribution> &dist){
-    dist[0].reset();
-    fillScalar(l,dist[0],_generators[0]);
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  // vector fill
-  template <class distribution>  inline void fill(vComplexF &l,std::vector<distribution> &dist){
-    RealF *pointer=(RealF *)&l;
-    dist[0].reset();
-    for(int i=0;i<2*vComplexF::Nsimd();i++){
-      fillScalar(pointer[i],dist[0],_generators[0]);
-    }
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  template <class distribution>  inline void fill(vComplexD &l,std::vector<distribution> &dist){
-    RealD *pointer=(RealD *)&l;
-    dist[0].reset();
-    for(int i=0;i<2*vComplexD::Nsimd();i++){
-      fillScalar(pointer[i],dist[0],_generators[0]);
-    }
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  template <class distribution>  inline void fill(vRealF &l,std::vector<distribution> &dist){
-    RealF *pointer=(RealF *)&l;
-    dist[0].reset();
-    for(int i=0;i<vRealF::Nsimd();i++){
-      fillScalar(pointer[i],dist[0],_generators[0]);
-    }
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-  template <class distribution>  inline void fill(vRealD &l,std::vector<distribution> &dist){
-    RealD *pointer=(RealD *)&l;
-    dist[0].reset();
-    for(int i=0;i<vRealD::Nsimd();i++){
-      fillScalar(pointer[i],dist[0],_generators[0]);
-    }
-    CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
-  }
-    
-  void SeedFixedIntegers(const std::vector<int> &seeds){
-    CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
-    std::seed_seq src(seeds.begin(),seeds.end());
-    Seed(src,0);
-  }
-
-    void SeedUniqueString(const std::string &s){
-      std::vector<int> seeds;
-      std::stringstream sha;
-      seeds = GridChecksum::sha256_seeds(s);
-      for(int i=0;i<seeds.size();i++) { 
-        sha << std::hex << seeds[i];
-      }
-      std::cout << GridLogMessage << "Intialising serial RNG with unique string '" 
-                << s << "'" << std::endl;
-      std::cout << GridLogMessage << "Seed SHA256: " << sha.str() << std::endl;
-      SeedFixedIntegers(seeds);
-    }
-};
-
-class GridParallelRNG : public GridRNGbase {
-private:
-  double _time_counter;
-  GridBase *_grid;
-  unsigned int _vol;
-
-public:
-  GridBase *Grid(void) const { return _grid; }
-  int generator_idx(int os,int is) {
-    return is*_grid->oSites()+os;
-  }
-
-  GridParallelRNG(GridBase *grid) : GridRNGbase() {
-    _grid = grid;
-    _vol  =_grid->iSites()*_grid->oSites();
-
-    _generators.resize(_vol);
-    _uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
-    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
-    _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
-  }
-
-  template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
-
-    typedef typename vobj::scalar_object scalar_object;
-    typedef typename vobj::scalar_type scalar_type;
-    typedef typename vobj::vector_type vector_type;
-
-    double inner_time_counter = usecond();
-
-    int multiplicity = RNGfillable_general(_grid, l.Grid()); // l has finer or same grid
-    int Nsimd  = _grid->Nsimd();  // guaranteed to be the same for l.Grid() too
-    int osites = _grid->oSites();  // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
-    int words  = sizeof(scalar_object) / sizeof(scalar_type);
-
-    autoView(l_v, l, CpuWrite);
-    thread_for( ss, osites, {
-      ExtractBuffer<scalar_object> buf(Nsimd);
-      for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
-
-	int sm = multiplicity * ss + m;  // Maps the generator site to the fine site
-
-	for (int si = 0; si < Nsimd; si++) {
-            
-	  int gdx = generator_idx(ss, si);  // index of generator state
-	  scalar_type *pointer = (scalar_type *)&buf[si];
-	  dist[gdx].reset();
-	  for (int idx = 0; idx < words; idx++) 
-	    fillScalar(pointer[idx], dist[gdx], _generators[gdx]);
-	}
-	// merge into SIMD lanes, FIXME suboptimal implementation
-	merge(l_v[sm], buf);
-      }
-      });
-    //    });
-
-    _time_counter += usecond()- inner_time_counter;
-  }
-
-    void SeedUniqueString(const std::string &s){
-      std::vector<int> seeds;
-      seeds = GridChecksum::sha256_seeds(s);
-      std::cout << GridLogMessage << "Intialising parallel RNG with unique string '" 
-                << s << "'" << std::endl;
-      std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
-      SeedFixedIntegers(seeds);
-    }
-  void SeedFixedIntegers(const std::vector<int> &seeds, int britney=0){
-
-    // Everyone generates the same seed_seq based on input seeds
-    CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
-
-    std::seed_seq source(seeds.begin(),seeds.end());
-
-    RngEngine master_engine(source);
-
-#ifdef RNG_FAST_DISCARD
-    ////////////////////////////////////////////////
-    // Skip ahead through a single stream.
-    // Applicable to SITMO and other has based/crypto RNGs
-    // Should be applicable to Mersenne Twister, but the C++11
-    // MT implementation does not implement fast discard even though
-    // in principle this is possible
-    ////////////////////////////////////////////////
-    thread_for( lidx, _grid->lSites(), {
-
-	int gidx;
-	int o_idx;
-	int i_idx;
-	int rank;
-	Coordinate pcoor;
-	Coordinate lcoor;
-	Coordinate gcoor;
-	_grid->LocalIndexToLocalCoor(lidx,lcoor);
-	pcoor=_grid->ThisProcessorCoor();
-	_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
-	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
-
-	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-
-	assert(rank == _grid->ThisRank() );
-	
-	int l_idx=generator_idx(o_idx,i_idx);
-	_generators[l_idx] = master_engine;
-	if ( britney ) { 
-	  Skip(_generators[l_idx],l_idx); // Skip to next RNG sequence
-	} else { 	
-	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
-	}
-    });
-#else 
-    ////////////////////////////////////////////////////////////////
-    // Machine and thread decomposition dependent seeding is efficient
-    // and maximally parallel; but NOT reproducible from machine to machine. 
-    // Not ideal, but fastest way to reseed all nodes.
-    ////////////////////////////////////////////////////////////////
-    {
-      // Obtain one Reseed per processor
-      int Nproc = _grid->ProcessorCount();
-      std::vector<RngEngine> seeders(Nproc);
-      int me= _grid->ThisRank();
-      for(int p=0;p<Nproc;p++){
-	seeders[p] = Reseed(master_engine);
-      }
-      master_engine = seeders[me];
-    }
-
-    {
-      // Obtain one reseeded generator per thread      
-      int Nthread = 32; // Hardwire a good level or parallelism
-      std::vector<RngEngine> seeders(Nthread);
-      for(int t=0;t<Nthread;t++){
-	seeders[t] = Reseed(master_engine);
-      }
-
-      thread_for( t, Nthread, {
-	// set up one per local site in threaded fashion
-	std::vector<uint32_t> newseeds;
-	std::uniform_int_distribution<uint32_t> uid;	
-	for(int l=0;l<_grid->lSites();l++) {
-	  if ( (l%Nthread)==t ) {
-	    _generators[l] = Reseed(seeders[t],newseeds,uid);
-	  }
-	}
-      });
-    }
-#endif
-  }
-
-  void Report(){
-    std::cout << GridLogMessage << "Time spent in the fill() routine by GridParallelRNG: "<< _time_counter/1e3 << " ms" << std::endl;
-  }
-
-
-  ////////////////////////////////////////////////////////////////////////
-  // Support for rigorous test of RNG's
-  // Return uniform random uint32_t from requested site generator
-  ////////////////////////////////////////////////////////////////////////
-  uint32_t GlobalU01(int gsite){
-
-    uint32_t the_number;
-    // who
-    int rank,o_idx,i_idx;
-    Coordinate gcoor;
-    _grid->GlobalIndexToGlobalCoor(gsite,gcoor);
-    _grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-
-    // draw
-    int l_idx=generator_idx(o_idx,i_idx);
-    if( rank == _grid->ThisRank() ){
-      the_number = _uid[l_idx](_generators[l_idx]);
-    }
-      
-    // share & return
-    _grid->Broadcast(rank,(void *)&the_number,sizeof(the_number));
-    return the_number;
-  }
-
-};
-
-template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l)   { rng.fill(l,rng._uniform);  }
-template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
-template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
-
-template <class sobj> inline void random(GridSerialRNG &rng,sobj &l)   { rng.fill(l,rng._uniform  ); }
-template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
-template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
-
-NAMESPACE_END(Grid);
-#endif

Grid/lattice/Lattice_slicesum_core.h

@@ -1,224 +0,0 @@
-#pragma once
-
-#if defined(GRID_CUDA)
-
-#include <cub/cub.cuh>
-#define gpucub cub
-#define gpuError_t cudaError_t
-#define gpuSuccess cudaSuccess
-
-#elif defined(GRID_HIP)
-
-#include <hipcub/hipcub.hpp>
-#define gpucub hipcub
-#define gpuError_t hipError_t
-#define gpuSuccess hipSuccess
-
-#endif
-
-
-NAMESPACE_BEGIN(Grid);
-
-
-#if defined(GRID_CUDA) || defined(GRID_HIP)
-template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
-  size_t subvol_size = e1*e2;
-  commVector<vobj> reduction_buffer(rd*subvol_size);
-  auto rb_p = &reduction_buffer[0];
-  vobj zero_init;
-  zeroit(zero_init);
-
-  
-  void *temp_storage_array = NULL;
-  size_t temp_storage_bytes = 0;
-  vobj *d_out;
-  int* d_offsets;
-
-  std::vector<int> offsets(rd+1,0);
-
-  for (int i = 0; i < offsets.size(); i++) {
-    offsets[i] = i*subvol_size;
-  }
-  
-  //Allocate memory for output and offset arrays on device
-  d_out = static_cast<vobj*>(acceleratorAllocDevice(rd*sizeof(vobj)));
-  
-  d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
-  
-  //copy offsets to device
-  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
-  
-  
-  gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
-  if (gpuErr!=gpuSuccess) {
-    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl;
-    exit(EXIT_FAILURE);
-  }
-
-  //allocate memory for temp_storage_array  
-  temp_storage_array = acceleratorAllocDevice(temp_storage_bytes);
-  
-  //prepare buffer for reduction
-  //use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream)
-  //use 2d accelerator_for to avoid launch latencies found when serially looping over rd 
-  accelerator_for2dNB( s,subvol_size, r,rd, Nsimd,{ 
-  
-    int n = s / e2;
-    int b = s % e2;
-    int so=r*ostride; // base offset for start of plane 
-    int ss= so+n*stride+b;
-
-    coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
-
-  });
-  
-  //issue segmented reductions in computeStream
-  gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), zero_init, computeStream);
-  if (gpuErr!=gpuSuccess) {
-    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl;
-    exit(EXIT_FAILURE);
-  }
-  
-  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
-  
-  //sync after copy
-  accelerator_barrier();
- 
-  acceleratorFreeDevice(temp_storage_array);
-  acceleratorFreeDevice(d_out);
-  acceleratorFreeDevice(d_offsets);
-  
-
-}
-#endif 
-
-
-#if defined(GRID_SYCL)
-template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
-{
-  size_t subvol_size = e1*e2;
-
-  vobj *mysum = (vobj *) malloc_shared(rd*sizeof(vobj),*theGridAccelerator);
-  vobj vobj_zero;
-  zeroit(vobj_zero);
-  for (int r = 0; r<rd; r++) { 
-    mysum[r] = vobj_zero; 
-  }
-
-  commVector<vobj> reduction_buffer(rd*subvol_size);    
-
-  auto rb_p = &reduction_buffer[0];
-
-  // autoView(Data_v, Data, AcceleratorRead);
-
-  //prepare reduction buffer 
-  accelerator_for2d( s,subvol_size, r,rd, (size_t)Nsimd,{ 
-  
-      int n = s / e2;
-      int b = s % e2;
-      int so=r*ostride; // base offset for start of plane 
-      int ss= so+n*stride+b;
-
-      coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
-
-  });
-
-  for (int r = 0; r < rd; r++) {
-      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-          auto Reduction = cl::sycl::reduction(&mysum[r],std::plus<>());
-          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
-          Reduction,
-          [=](cl::sycl::id<1> item, auto &sum) {
-              auto s = item[0];
-              sum += rb_p[r*subvol_size+s];
-          });
-      });
-      
-     
-  }
-  theGridAccelerator->wait();
-  for (int r = 0; r < rd; r++) {
-    lvSum[r] = mysum[r];
-  }
-  free(mysum,*theGridAccelerator);
-}
-#endif
-
-template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
-  typedef typename vobj::vector_type vector;
-  const int words = sizeof(vobj)/sizeof(vector);
-  const int osites = rd*e1*e2;
-  commVector<vector>buffer(osites);
-  vector *dat = (vector *)Data;
-  vector *buf = &buffer[0];
-  Vector<vector> lvSum_small(rd);
-  vector *lvSum_ptr = (vector *)&lvSum[0];
-
-  for (int w = 0; w < words; w++) {
-    accelerator_for(ss,osites,1,{
-	    buf[ss] = dat[ss*words+w];
-    });
-
-    #if defined(GRID_CUDA) || defined(GRID_HIP)
-      sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
-    #elif defined(GRID_SYCL)
-      sliceSumReduction_sycl_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
-    #endif
-
-    for (int r = 0; r < rd; r++) {
-      lvSum_ptr[w+words*r]=lvSum_small[r];
-    }
-
-  }
-
-  
-}
-
-template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
-{
-  autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
-    if constexpr (sizeof(vobj) <= 256) { 
-
-      #if defined(GRID_CUDA) || defined(GRID_HIP)
-        sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
-      #elif defined (GRID_SYCL)
-        sliceSumReduction_sycl_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
-      #endif
-
-    }
-    else {
-      sliceSumReduction_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
-    }
-}
-
-
-template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
-{
-  // sum over reduced dimension planes, breaking out orthog dir
-  // Parallel over orthog direction
-  autoView( Data_v, Data, CpuRead);
-  thread_for( r,rd, {
-    int so=r*ostride; // base offset for start of plane 
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-        int ss= so+n*stride+b;
-        lvSum[r]=lvSum[r]+Data_v[ss];
-      }
-    }
-  });
-}
-
-template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) 
-{
-  #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  
-  sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-  
-  #else
-  sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-
-  #endif
-}
-
-
-NAMESPACE_END(Grid);

Grid/lattice/Lattice_trace.h

@@ -1,130 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_trace.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#ifndef GRID_LATTICE_TRACE_H
-#define GRID_LATTICE_TRACE_H
-
-///////////////////////////////////////////////
-// Tracing, transposing, peeking, poking
-///////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Trace
-////////////////////////////////////////////////////////////////////////////////////////////////////
-/*
-template<class vobj>
-inline auto trace(const Lattice<vobj> &lhs)  -> Lattice<decltype(trace(vobj()))>
-{
-  Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
-  autoView(ret_v , ret, AcceleratorWrite);
-  autoView(lhs_v , lhs, AcceleratorRead);
-  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
-    coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
-  });
-  return ret;
-};
-*/
-    
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Trace Index level dependent operation
-////////////////////////////////////////////////////////////////////////////////////////////////////
-template<int Index,class vobj>
-inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
-{
-  Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
-  autoView( ret_v , ret, AcceleratorWrite);
-  autoView( lhs_v , lhs, AcceleratorRead);
-  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
-    coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
-  });
-  return ret;
-};
-
-template<int N, class Vec>
-Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
-{
-  GridBase *grid=Umu.Grid();
-  auto lvol = grid->lSites();
-  Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
-  typedef typename Vec::scalar_type scalar;
-  autoView(Umu_v,Umu,CpuRead);
-  autoView(ret_v,ret,CpuWrite);
-  thread_for(site,lvol,{
-    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
-    Coordinate lcoor;
-    grid->LocalIndexToLocalCoor(site, lcoor);
-    iScalar<iScalar<iMatrix<scalar, N> > > Us;
-    peekLocalSite(Us, Umu_v, lcoor);
-    for(int i=0;i<N;i++){
-      for(int j=0;j<N;j++){
-	scalar tmp= Us()()(i,j);
-	ComplexD ztmp(real(tmp),imag(tmp));
-	EigenU(i,j)=ztmp;
-      }}
-    ComplexD detD  = EigenU.determinant();
-    typename Vec::scalar_type det(detD.real(),detD.imag());
-    pokeLocalSite(det,ret_v,lcoor);
-  });
-  return ret;
-}
-
-template<int N>
-Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
-{
-  GridBase *grid=Umu.Grid();
-  auto lvol = grid->lSites();
-  Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
-  
-  autoView(Umu_v,Umu,CpuRead);
-  autoView(ret_v,ret,CpuWrite);
-  thread_for(site,lvol,{
-    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
-    Coordinate lcoor;
-    grid->LocalIndexToLocalCoor(site, lcoor);
-    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
-    iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
-    peekLocalSite(Us, Umu_v, lcoor);
-    for(int i=0;i<N;i++){
-      for(int j=0;j<N;j++){
-	EigenU(i,j) = Us()()(i,j);
-      }}
-    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
-    for(int i=0;i<N;i++){
-      for(int j=0;j<N;j++){
-	Ui()()(i,j) = EigenUinv(i,j);
-      }}
-    pokeLocalSite(Ui,ret_v,lcoor);
-  });
-  return ret;
-}
-
-
-NAMESPACE_END(Grid);
-#endif
-

Grid/lattice/Lattice_transpose.h

@@ -1,70 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_transpose.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_LATTICE_TRANSPOSE_H
-#define GRID_LATTICE_TRANSPOSE_H
-
-///////////////////////////////////////////////
-// Transpose
-///////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Transpose
-////////////////////////////////////////////////////////////////////////////////////////////////////
-/*
-template<class vobj>
-inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
-  Lattice<vobj> ret(lhs.Grid());
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
-    coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
-  });
-  return ret;
-};
-*/    
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Index level dependent transpose
-////////////////////////////////////////////////////////////////////////////////////////////////////
-template<int Index,class vobj>
-inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
-{
-  Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
-  autoView( ret_v, ret, AcceleratorWrite);
-  autoView( lhs_v, lhs, AcceleratorRead);
-  accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
-    coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
-  });
-  return ret;
-};
-
-NAMESPACE_END(Grid);
-#endif

Grid/lattice/Lattice_unary.h

@@ -1,80 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_unary.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-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_LATTICE_UNARY_H
-#define GRID_LATTICE_UNARY_H
-
-NAMESPACE_BEGIN(Grid);
-
-template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
-  Lattice<obj> ret_i(rhs_i.Grid());
-  autoView( rhs, rhs_i, AcceleratorRead);
-  autoView( ret, ret_i, AcceleratorWrite);
-  ret.Checkerboard() = rhs.Checkerboard();
-  accelerator_for(ss,rhs.size(),1,{
-      ret[ss]=pow(rhs[ss],y);
-  });
-  return ret_i;
-}
-template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
-  Lattice<obj> ret_i(rhs_i.Grid());
-  autoView( rhs , rhs_i, AcceleratorRead);
-  autoView( ret , ret_i, AcceleratorWrite);
-  ret.Checkerboard() = rhs.Checkerboard();
-  accelerator_for(ss,rhs.size(),obj::Nsimd(),{
-    coalescedWrite(ret[ss],mod(rhs(ss),y));
-  });
-  return ret_i;
-}
-
-template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
-  Lattice<obj> ret_i(rhs_i.Grid());
-  autoView( ret , ret_i, AcceleratorWrite);
-  autoView( rhs , rhs_i, AcceleratorRead);
-  ret.Checkerboard() = rhs_i.Checkerboard();
-  accelerator_for(ss,rhs.size(),obj::Nsimd(),{
-    coalescedWrite(ret[ss],div(rhs(ss),y));
-  });
-  return ret_i;
-}
-
-template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
-  Lattice<obj> ret_i(rhs_i.Grid());
-  autoView( rhs , rhs_i, AcceleratorRead);
-  autoView( ret , ret_i, AcceleratorWrite);
-  ret.Checkerboard() = rhs.Checkerboard();
-  accelerator_for(ss,rhs.size(),obj::Nsimd(),{
-    coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));
-  });
-  return ret_i;
-}
-
-NAMESPACE_END(Grid);
-#endif

Grid/lattice/Lattice_view.h

@@ -1,174 +0,0 @@
-#pragma once
-NAMESPACE_BEGIN(Grid);
-///////////////////////////////////////////////////////////////////
-// Base class which can be used by traits to pick up behaviour
-///////////////////////////////////////////////////////////////////
-class LatticeBase {};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Conformable checks; same instance of Grid required
-/////////////////////////////////////////////////////////////////////////////////////////
-void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
-{
-  assert(lhs == rhs);
-}
-
-////////////////////////////////////////////////////////////////////////////
-// Minimal base class containing only data valid to access from accelerator
-// _odata will be a managed pointer in CUDA
-////////////////////////////////////////////////////////////////////////////
-// Force access to lattice through a view object.
-// prevents writing of code that will not offload to GPU, but perhaps annoyingly
-// strict since host could could in principle direct access through the lattice object
-// Need to decide programming model.
-#define LATTICE_VIEW_STRICT
-template<class vobj> class LatticeAccelerator : public LatticeBase
-{
-protected:
-  //public:
-  GridBase *_grid;
-  int checkerboard;
-  vobj     *_odata;    // A managed pointer
-  uint64_t _odata_size;    
-  ViewAdvise advise;
-public:
-  accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr), advise(AdviseDefault) { }; 
-  accelerator_inline uint64_t oSites(void) const { return _odata_size; };
-  accelerator_inline int  Checkerboard(void) const { return checkerboard; };
-  accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
-  accelerator_inline ViewAdvise Advise(void) const { return advise; };
-  accelerator_inline ViewAdvise &Advise(void) { return this->advise; }; // can assign advise on a container, not a view
-  accelerator_inline void Conformable(GridBase * &grid) const
-  { 
-    if (grid) conformable(grid, _grid);
-    else      grid = _grid;
-  };
-  // Host only
-  GridBase * getGrid(void) const { return _grid; };
-  vobj* getHostPointer(void) const { return _odata; };
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// A View class which provides accessor to the data.
-// This will be safe to call from accelerator_for and is trivially copy constructible
-// The copy constructor for this will need to be used by device lambda functions
-/////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> 
-class LatticeView : public LatticeAccelerator<vobj>
-{
-public:
-  // Rvalue
-  ViewMode mode;
-  void * cpu_ptr;
-#ifdef GRID_SIMT
-  accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { 
-    return coalescedRead(this->_odata[i]); 
-  }
-#else 
-  accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
-#endif
-
-#if 1
-  //  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i) const { return this->_odata[i]; };
-#else
-  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
-#endif
-  
-  accelerator_inline uint64_t begin(void) const { return 0;};
-  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
-  accelerator_inline uint64_t size(void)  const { return this->_odata_size; };
-
-  LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me){}
-  LatticeView(const LatticeView<vobj> &refer_to_me) = default; // Trivially copyable
-  LatticeView(const LatticeAccelerator<vobj> &refer_to_me,ViewMode mode) : LatticeAccelerator<vobj> (refer_to_me)
-  {
-    this->ViewOpen(mode);
-  }
-
-  // Host functions
-  void ViewOpen(ViewMode mode)
-  { // Translate the pointer, could save a copy. Could use a "Handle" and not save _odata originally in base
-    //    std::cout << "View Open"<<std::hex<<this->_odata<<std::dec <<std::endl;
-    this->cpu_ptr = (void *)this->_odata;
-    this->mode    = mode;
-    this->_odata  =(vobj *)
-      MemoryManager::ViewOpen(this->cpu_ptr,
-				this->_odata_size*sizeof(vobj),
-				mode,
-				this->advise);    
-  }
-  void ViewClose(void)
-  { // Inform the manager
-    //    std::cout << "View Close"<<std::hex<<this->cpu_ptr<<std::dec <<std::endl;
-    MemoryManager::ViewClose(this->cpu_ptr,this->mode);    
-  }
-
-};
-// Little autoscope assister
-template<class View> 
-class ViewCloser
-{
-  View v;  // Take a copy of view and call view close when I go out of scope automatically
- public:
-  ViewCloser(View &_v) : v(_v) {};
-  ~ViewCloser() { v.ViewClose(); }
-};
-
-#define autoView(l_v,l,mode)				\
-	  auto l_v = l.View(mode);			\
-	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v);
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Lattice expression types used by ET to assemble the AST
-// 
-// Need to be able to detect code paths according to the whether a lattice object or not
-// so introduce some trait type things
-/////////////////////////////////////////////////////////////////////////////////////////
-
-class LatticeExpressionBase {};
-
-template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
-template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
-
-template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
-template<class T>                 struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
-template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
-
-template <typename Op, typename _T1>                           
-class LatticeUnaryExpression : public  LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  Op op;
-  T1 arg1;
-  LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
-};
-
-template <typename Op, typename _T1, typename _T2>              
-class LatticeBinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
-};
-
-template <typename Op, typename _T1, typename _T2, typename _T3> 
-class LatticeTrinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  typedef typename ViewMap<_T3>::Type T3;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  T3 arg3;
-  LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
-};
-NAMESPACE_END(Grid);

Grid/lattice/PaddedCell.h

@@ -1,174 +0,0 @@
-/*************************************************************************************
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/PaddedCell.h
-
-    Copyright (C) 2019
-
-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/cshift/Cshift.h>
-
-NAMESPACE_BEGIN(Grid);
-
-//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
-template<typename vobj>
-struct CshiftImplBase{
-  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
-  virtual ~CshiftImplBase(){}
-};
-template<typename vobj>
-struct CshiftImplDefault: public CshiftImplBase<vobj>{
-  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
-};
-template<typename Gimpl>
-struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
-  typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
-};  
-
-class PaddedCell {
-public:
-  GridCartesian * unpadded_grid;
-  int dims;
-  int depth;
-  std::vector<GridCartesian *> grids;
-
-  ~PaddedCell()
-  {
-    DeleteGrids();
-  }
-  PaddedCell(int _depth,GridCartesian *_grid)
-  {
-    unpadded_grid = _grid;
-    depth=_depth;
-    dims=_grid->Nd();
-    AllocateGrids();
-    Coordinate local     =unpadded_grid->LocalDimensions();
-    for(int d=0;d<dims;d++){
-      assert(local[d]>=depth);
-    }
-  }
-  void DeleteGrids(void)
-  {
-    for(int d=0;d<grids.size();d++){
-      delete grids[d];
-    }
-    grids.resize(0);
-  };
-  void AllocateGrids(void)
-  {
-    Coordinate local     =unpadded_grid->LocalDimensions();
-    Coordinate simd      =unpadded_grid->_simd_layout;
-    Coordinate processors=unpadded_grid->_processors;
-    Coordinate plocal    =unpadded_grid->LocalDimensions();
-    Coordinate global(dims);
-
-    // expand up one dim at a time
-    for(int d=0;d<dims;d++){
-
-      plocal[d] += 2*depth; 
-
-      for(int d=0;d<dims;d++){
-	global[d] = plocal[d]*processors[d];
-      }
-
-      grids.push_back(new GridCartesian(global,simd,processors));
-    }
-  };
-  template<class vobj>
-  inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
-  {
-    Lattice<vobj> out(unpadded_grid);
-
-    Coordinate local     =unpadded_grid->LocalDimensions();
-    Coordinate fll(dims,depth); // depends on the MPI spread
-    Coordinate tll(dims,0); // depends on the MPI spread
-    localCopyRegion(in,out,fll,tll,local);
-    return out;
-  }
-  template<class vobj>
-  inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
-  {
-    GridBase *old_grid = in.Grid();
-    int dims = old_grid->Nd();
-    Lattice<vobj> tmp = in;
-    for(int d=0;d<dims;d++){
-      tmp = Expand(d,tmp,cshift); // rvalue && assignment
-    }
-    return tmp;
-  }
-  // expand up one dim at a time
-  template<class vobj>
-  inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
-  {
-    GridBase *old_grid = in.Grid();
-    GridCartesian *new_grid = grids[dim];//These are new grids
-    Lattice<vobj>  padded(new_grid);
-    Lattice<vobj> shifted(old_grid);    
-    Coordinate local     =old_grid->LocalDimensions();
-    Coordinate plocal    =new_grid->LocalDimensions();
-    if(dim==0) conformable(old_grid,unpadded_grid);
-    else       conformable(old_grid,grids[dim-1]);
-
-    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
-
-    double tins=0, tshift=0;
-    
-    // Middle bit
-    double t = usecond();
-    for(int x=0;x<local[dim];x++){
-      InsertSliceLocal(in,padded,x,depth+x,dim);
-    }
-    tins += usecond() - t;
-    
-    // High bit
-    t = usecond();
-    shifted = cshift.Cshift(in,dim,depth);
-    tshift += usecond() - t;
-
-    t=usecond();
-    for(int x=0;x<depth;x++){
-      InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
-    }
-    tins += usecond() - t;
-    
-    // Low bit
-    t = usecond();
-    shifted = cshift.Cshift(in,dim,-depth);
-    tshift += usecond() - t;
-    
-    t = usecond();
-    for(int x=0;x<depth;x++){
-      InsertSliceLocal(shifted,padded,x,x,dim);
-    }
-    tins += usecond() - t;
-
-    std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
-    
-    return padded;
-  }
-
-};
- 
-
-NAMESPACE_END(Grid);
-

Grid/log/Log.h

@@ -1,262 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Log.h
-
-    Copyright (C) 2015
-
-    Author: Antonin Portelli <antonin.portelli@me.com>
-    Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-
-#include <map>
-
-#ifndef GRID_LOG_H
-#define GRID_LOG_H
-
-#ifdef HAVE_EXECINFO_H
-#include <execinfo.h>
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-//////////////////////////////////////////////////////////////////////////////////////////////////
-// Dress the output; use std::chrono for time stamping via the StopWatch class
-//////////////////////////////////////////////////////////////////////////////////////////////////
-
-class Colours{
-protected:
-  bool is_active;
-public:
-  std::map<std::string, std::string> colour;
-
-  Colours(bool activate=false){
-    Active(activate);
-  };
-
-  void Active(bool activate){
-    is_active=activate;
-    if (is_active){
-      colour["BLACK"]  ="\033[30m";
-      colour["RED"]    ="\033[31m";
-      colour["GREEN"]  ="\033[32m";
-      colour["YELLOW"] ="\033[33m";
-      colour["BLUE"]   ="\033[34m";
-      colour["PURPLE"] ="\033[35m";
-      colour["CYAN"]   ="\033[36m";
-      colour["WHITE"]  ="\033[37m";
-      colour["NORMAL"] ="\033[0;39m";
-    } else {
-      colour["BLACK"] ="";
-      colour["RED"]   ="";
-      colour["GREEN"] ="";
-      colour["YELLOW"]="";
-      colour["BLUE"]  ="";
-      colour["PURPLE"]="";
-      colour["CYAN"]  ="";
-      colour["WHITE"] ="";
-      colour["NORMAL"]="";
-    }
-  };
-};
-
-
-class Logger {
-protected:
-  Colours &Painter;
-  int active;
-  int timing_mode;
-  int topWidth{-1}, chanWidth{-1};
-  static int timestamp;
-  std::string name, topName;
-  std::string COLOUR;
-
-public:
-  static GridStopWatch GlobalStopWatch;
-  GridStopWatch         LocalStopWatch;
-  GridStopWatch *StopWatch;
-  static std::ostream devnull;
-
-  std::string background() {return Painter.colour["NORMAL"];}
-  std::string evidence() {return Painter.colour["YELLOW"];}
-  std::string colour() {return Painter.colour[COLOUR];}
-
-  Logger(std::string topNm, int on, std::string nm, Colours& col_class, std::string col)  : active(on),
-											    name(nm),
-											    topName(topNm),
-											    Painter(col_class),
-											    timing_mode(0),
-											    COLOUR(col) 
-  {
-    StopWatch = & GlobalStopWatch;
-  };
-  
-  void Active(int on) {active = on;};
-  int  isActive(void) {return active;};
-  static void Timestamp(int on) {timestamp = on;};
-  void Reset(void) { 
-    StopWatch->Reset(); 
-    StopWatch->Start(); 
-  }
-  void TimingMode(int on) { 
-    timing_mode = on; 
-    if(on) { 
-      StopWatch = &LocalStopWatch;
-      Reset(); 
-    }
-  }
-  void setTopWidth(const int w) {topWidth = w;}
-  void setChanWidth(const int w) {chanWidth = w;}
-
-  friend std::ostream& operator<< (std::ostream& stream, Logger& log){
-
-    if ( log.active ) {
-      std::ios_base::fmtflags f(stream.flags());
-
-      stream << log.background()<<  std::left;
-      if (log.topWidth > 0)
-      {
-        stream << std::setw(log.topWidth);
-      }
-      stream << log.topName << log.background()<< " : ";
-      //      stream << log.colour() <<  std::left;
-      stream <<  std::left;
-      if (log.chanWidth > 0)
-      {
-        stream << std::setw(log.chanWidth);
-      }
-      stream << log.name << log.background() << " : ";
-      if ( log.timestamp ) {
-	log.StopWatch->Stop();
-	GridTime now = log.StopWatch->Elapsed();
-	
-	if ( log.timing_mode==1 ) log.StopWatch->Reset();
-	log.StopWatch->Start();
-	stream << log.evidence()
-	       << now	       << log.background() << " : " ;
-      }
-      //      stream << log.colour();
-      stream <<  std::right;
-      stream.flags(f);
-      return stream;
-    } else { 
-      return devnull;
-    }
-  }
-
-};
-
-class GridLogger: public Logger {
-public:
-  GridLogger(int on, std::string nm, Colours&col_class, std::string col_key = "NORMAL"):
-    Logger("Grid", on, nm, col_class, col_key){};
-};
-
-void GridLogConfigure(std::vector<std::string> &logstreams);
-
-extern GridLogger GridLogMG;
-extern GridLogger GridLogIRL;
-extern GridLogger GridLogSolver;
-extern GridLogger GridLogError;
-extern GridLogger GridLogWarning;
-extern GridLogger GridLogMessage;
-extern GridLogger GridLogDebug;
-extern GridLogger GridLogPerformance;
-extern GridLogger GridLogDslash;
-extern GridLogger GridLogIterative;
-extern GridLogger GridLogIntegrator;
-extern GridLogger GridLogHMC;
-extern GridLogger GridLogMemory;
-extern GridLogger GridLogTracing;
-extern Colours    GridLogColours;
-
-std::string demangle(const char* name) ;
-
-template<typename... Args>
-inline std::string sjoin(Args&&... args) noexcept {
-    std::ostringstream msg;
-    (msg << ... << args);
-    return msg.str();
-}
-
-/*!  @brief make log messages work like python print */
-template <typename... Args>
-inline void Grid_log(Args&&... args) {
-    std::string msg = sjoin(std::forward<Args>(args)...);
-    std::cout << GridLogMessage << msg << std::endl;
-}
-
-/*!  @brief make warning messages work like python print */
-template <typename... Args>
-inline void Grid_warn(Args&&... args) {
-    std::string msg = sjoin(std::forward<Args>(args)...);
-    std::cout << "\033[33m" << GridLogWarning << msg << "\033[0m" << std::endl;
-}
-
-/*!  @brief make error messages work like python print */
-template <typename... Args>
-inline void Grid_error(Args&&... args) {
-    std::string msg = sjoin(std::forward<Args>(args)...);
-    std::cout << "\033[31m" << GridLogError << msg << "\033[0m" << std::endl;
-}
-
-/*!  @brief make pass messages work like python print */
-template <typename... Args>
-inline void Grid_pass(Args&&... args) {
-    std::string msg = sjoin(std::forward<Args>(args)...);
-    std::cout << "\033[32m" << GridLogMessage << msg << "\033[0m" << std::endl;
-}
-
-#define _NBACKTRACE (256)
-extern void * Grid_backtrace_buffer[_NBACKTRACE];
-
-#define BACKTRACEFILE() {						\
-    char string[20];							\
-    std::sprintf(string,"backtrace.%d",CartesianCommunicator::RankWorld()); \
-    std::FILE * fp = std::fopen(string,"w");				\
-    BACKTRACEFP(fp)							\
-      std::fclose(fp);							\
-  }
-
-
-#ifdef HAVE_EXECINFO_H
-#define BACKTRACEFP(fp) {						\
-    int symbols    = backtrace        (Grid_backtrace_buffer,_NBACKTRACE); \
-    char **strings = backtrace_symbols(Grid_backtrace_buffer,symbols);	\
-    for (int i = 0; i < symbols; i++){					\
-      std::fprintf (fp,"BackTrace Strings: %d %s\n",i, demangle(strings[i]).c_str()); std::fflush(fp); \
-    }									\
-  }
-#else 
-#define BACKTRACEFP(fp) {						\
-    std::fprintf (fp,"BT %d %lx\n",0, __builtin_return_address(0)); std::fflush(fp); \
-    std::fprintf (fp,"BT %d %lx\n",1, __builtin_return_address(1)); std::fflush(fp); \
-    std::fprintf (fp,"BT %d %lx\n",2, __builtin_return_address(2)); std::fflush(fp); \
-    std::fprintf (fp,"BT %d %lx\n",3, __builtin_return_address(3)); std::fflush(fp); \
-  }
-#endif
-
-#define BACKTRACE() BACKTRACEFP(stdout) 
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/parallelIO/BinaryIO.cc

@@ -1,4 +0,0 @@
-#include <Grid/GridCore.h>
-
-int                    Grid::BinaryIO::latticeWriteMaxRetry = -1;
-Grid::BinaryIO::IoPerf Grid::BinaryIO::lastPerf;

Grid/parallelIO/BinaryIO.h

@@ -1,773 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/parallelIO/BinaryIO.h
-
-    Copyright (C) 2015
-
-    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: Guido Cossu<guido.cossu@ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#pragma once
-
-#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT) 
-#define USE_MPI_IO
-#else
-#undef  USE_MPI_IO
-#endif
-
-#ifdef HAVE_ENDIAN_H
-#include <endian.h>
-#endif
-
-#include <arpa/inet.h>
-#include <algorithm>
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////////////////////////
-// Byte reversal garbage
-/////////////////////////////////////////////////////////////////////////////////
-inline uint32_t byte_reverse32(uint32_t f) { 
-      f = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
-      return f;
-}
-inline uint64_t byte_reverse64(uint64_t f) { 
-  uint64_t g;
-  g = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
-  g = g << 32;
-  f = f >> 32;
-  g|= ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
-  return g;
-}
-
-#if BYTE_ORDER == BIG_ENDIAN 
-inline uint64_t Grid_ntohll(uint64_t A) { return A; }
-#else
-inline uint64_t Grid_ntohll(uint64_t A) { 
-  return byte_reverse64(A);
-}
-#endif
-
-// A little helper
-inline void removeWhitespace(std::string &key)
-{
-  key.erase(std::remove_if(key.begin(), key.end(), ::isspace),key.end());
-}
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Static class holding the parallel IO code
-// Could just use a namespace
-///////////////////////////////////////////////////////////////////////////////////////////////////
-class BinaryIO {
- public:
-  struct IoPerf
-  {
-    uint64_t size{0},time{0};
-    double   mbytesPerSecond{0.};
-  };
-
-  static IoPerf lastPerf;
-  static int latticeWriteMaxRetry;
-
-  /////////////////////////////////////////////////////////////////////////////
-  // more byte manipulation helpers
-  /////////////////////////////////////////////////////////////////////////////
-
-  template<class vobj> static inline void Uint32Checksum(Lattice<vobj> &lat,uint32_t &nersc_csum)
-  {
-    typedef typename vobj::scalar_object sobj;
-
-    GridBase *grid = lat.Grid();
-    uint64_t lsites = grid->lSites();
-
-    std::vector<sobj> scalardata(lsites); 
-    unvectorizeToLexOrdArray(scalardata,lat);    
-
-    NerscChecksum(grid,scalardata,nersc_csum);
-  }
-
-  template <class fobj>
-  static inline void NerscChecksum(GridBase *grid, std::vector<fobj> &fbuf, uint32_t &nersc_csum)
-  {
-    const uint64_t size32 = sizeof(fobj) / sizeof(uint32_t);
-
-    uint64_t lsites = grid->lSites();
-    if (fbuf.size() == 1)
-    {
-      lsites = 1;
-    }
-
-    thread_region
-    {
-      uint32_t nersc_csum_thr = 0;
-
-      thread_for_in_region( local_site, lsites, 
-      {
-        uint32_t *site_buf = (uint32_t *)&fbuf[local_site];
-        for (uint64_t j = 0; j < size32; j++)
-        {
-          nersc_csum_thr = nersc_csum_thr + site_buf[j];
-        }
-      });
-
-      thread_critical
-      {
-        nersc_csum += nersc_csum_thr;
-      }
-    }
-  }
-
-  template<class fobj> static inline void ScidacChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &scidac_csuma,uint32_t &scidac_csumb)
-  {
-    int nd = grid->_ndimension;
-
-    uint64_t lsites              =grid->lSites();
-    if (fbuf.size()==1) {
-      lsites=1;
-    }
-    Coordinate local_vol   =grid->LocalDimensions();
-    Coordinate local_start =grid->LocalStarts();
-    Coordinate global_vol  =grid->FullDimensions();
-
-    thread_region
-    { 
-      Coordinate coor(nd);
-      uint32_t scidac_csuma_thr=0;
-      uint32_t scidac_csumb_thr=0;
-      uint32_t site_crc=0;
-
-      thread_for_in_region( local_site, lsites, 
-      {
-
-	uint32_t * site_buf = (uint32_t *)&fbuf[local_site];
-
-	/* 
-	 * Scidac csum  is rather more heavyweight
-	 * FIXME -- 128^3 x 256 x 16 will overflow.
-	 */
-	
-	int global_site;
-
-	Lexicographic::CoorFromIndex(coor,local_site,local_vol);
-
-	for(int d=0;d<nd;d++) {
-	  coor[d] = coor[d]+local_start[d];
-	}
-
-	Lexicographic::IndexFromCoor(coor,global_site,global_vol);
-
-	uint32_t gsite29   = global_site%29;
-	uint32_t gsite31   = global_site%31;
-	
-	site_crc = crc32(0,(unsigned char *)site_buf,sizeof(fobj));
-	//	std::cout << "Site "<<local_site << " crc "<<std::hex<<site_crc<<std::dec<<std::endl;
-	//	std::cout << "Site "<<local_site << std::hex<<site_buf[0] <<site_buf[1]<<std::dec <<std::endl;
-	scidac_csuma_thr ^= site_crc<<gsite29 | site_crc>>(32-gsite29);
-	scidac_csumb_thr ^= site_crc<<gsite31 | site_crc>>(32-gsite31);
-      });
-
-      thread_critical
-      {
-	scidac_csuma^= scidac_csuma_thr;
-	scidac_csumb^= scidac_csumb_thr;
-      }
-    }
-  }
-
-  // Network is big endian
-  static inline void htobe32_v(void *file_object,uint32_t bytes){ be32toh_v(file_object,bytes);} 
-  static inline void htobe64_v(void *file_object,uint32_t bytes){ be64toh_v(file_object,bytes);} 
-  static inline void htole32_v(void *file_object,uint32_t bytes){ le32toh_v(file_object,bytes);} 
-  static inline void htole64_v(void *file_object,uint32_t bytes){ le64toh_v(file_object,bytes);} 
-
-  static inline void be32toh_v(void *file_object,uint64_t bytes)
-  {
-    uint32_t * f = (uint32_t *)file_object;
-    uint64_t count = bytes/sizeof(uint32_t);
-    thread_for( i, count, {  
-      f[i] = ntohl(f[i]);
-    });
-  }
-  // LE must Swap and switch to host
-  static inline void le32toh_v(void *file_object,uint64_t bytes)
-  {
-    uint32_t *fp = (uint32_t *)file_object;
-
-    uint64_t count = bytes/sizeof(uint32_t);
-    thread_for(i,count,{
-      uint32_t f;
-      f = fp[i];
-      // got network order and the network to host
-      f = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
-      fp[i] = ntohl(f);
-    });
-  }
-
-  // BE is same as network
-  static inline void be64toh_v(void *file_object,uint64_t bytes)
-  {
-    uint64_t * f = (uint64_t *)file_object;
-    uint64_t count = bytes/sizeof(uint64_t);
-    thread_for( i, count, {
-      f[i] = Grid_ntohll(f[i]);
-    });
-  }
-  
-  // LE must swap and switch;
-  static inline void le64toh_v(void *file_object,uint64_t bytes)
-  {
-    uint64_t *fp = (uint64_t *)file_object;
-    uint64_t count = bytes/sizeof(uint64_t);
-    thread_for( i, count, {
-      uint64_t f,g;
-      f = fp[i];
-      // got network order and the network to host
-      g = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
-      g = g << 32;
-      f = f >> 32;
-      g|= ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ; 
-      fp[i] = Grid_ntohll(g);
-    });
-  }
-  /////////////////////////////////////////////////////////////////////////////
-  // Real action:
-  // Read or Write distributed lexico array of ANY object to a specific location in file 
-  //////////////////////////////////////////////////////////////////////////////////////
-
-  static const int BINARYIO_MASTER_APPEND = 0x10;
-  static const int BINARYIO_UNORDERED     = 0x08;
-  static const int BINARYIO_LEXICOGRAPHIC = 0x04;
-  static const int BINARYIO_READ          = 0x02;
-  static const int BINARYIO_WRITE         = 0x01;
-
-  template<class word,class fobj>
-  static inline void IOobject(word w,
-			      GridBase *grid,
-			      std::vector<fobj> &iodata,
-			      std::string file,
-			      uint64_t& offset,
-			      const std::string &format, int control,
-			      uint32_t &nersc_csum,
-			      uint32_t &scidac_csuma,
-			      uint32_t &scidac_csumb)
-  {
-    grid->Barrier();
-    GridStopWatch timer; 
-    GridStopWatch bstimer;
-    
-    nersc_csum=0;
-    scidac_csuma=0;
-    scidac_csumb=0;
-
-    int ndim                 = grid->Dimensions();
-    int nrank                = grid->ProcessorCount();
-    int myrank               = grid->ThisRank();
-
-    Coordinate  psizes = grid->ProcessorGrid(); 
-    Coordinate  pcoor  = grid->ThisProcessorCoor();
-    Coordinate gLattice= grid->GlobalDimensions();
-    Coordinate lLattice= grid->LocalDimensions();
-
-    Coordinate lStart(ndim);
-    Coordinate gStart(ndim);
-
-    // Flatten the file
-    uint64_t lsites = grid->lSites();
-    if ( control & BINARYIO_MASTER_APPEND )  {
-      assert(iodata.size()==1);
-    } else {
-      assert(lsites==iodata.size());
-    }
-    for(int d=0;d<ndim;d++){
-      gStart[d] = lLattice[d]*pcoor[d];
-      lStart[d] = 0;
-    }
-
-#ifdef USE_MPI_IO
-    std::vector<int> distribs(ndim,MPI_DISTRIBUTE_BLOCK);
-    std::vector<int> dargs   (ndim,MPI_DISTRIBUTE_DFLT_DARG);
-    MPI_Datatype mpiObject;
-    MPI_Datatype fileArray;
-    MPI_Datatype localArray;
-    MPI_Datatype mpiword;
-    MPI_Offset disp = offset;
-    MPI_File fh ;
-    MPI_Status status;
-    int numword;
-
-    if ( sizeof( word ) == sizeof(float ) ) {
-      numword = sizeof(fobj)/sizeof(float);
-      mpiword = MPI_FLOAT;
-    } else {
-      numword = sizeof(fobj)/sizeof(double);
-      mpiword = MPI_DOUBLE;
-    }
-
-    //////////////////////////////////////////////////////////////////////////////
-    // Sobj in MPI phrasing
-    //////////////////////////////////////////////////////////////////////////////
-    int ierr;
-    ierr = MPI_Type_contiguous(numword,mpiword,&mpiObject);    assert(ierr==0);
-    ierr = MPI_Type_commit(&mpiObject);
-
-    //////////////////////////////////////////////////////////////////////////////
-    // File global array data type
-    //////////////////////////////////////////////////////////////////////////////
-    ierr=MPI_Type_create_subarray(ndim,&gLattice[0],&lLattice[0],&gStart[0],MPI_ORDER_FORTRAN, mpiObject,&fileArray);    assert(ierr==0);
-    ierr=MPI_Type_commit(&fileArray);    assert(ierr==0);
-
-    //////////////////////////////////////////////////////////////////////////////
-    // local lattice array
-    //////////////////////////////////////////////////////////////////////////////
-    ierr=MPI_Type_create_subarray(ndim,&lLattice[0],&lLattice[0],&lStart[0],MPI_ORDER_FORTRAN, mpiObject,&localArray);    assert(ierr==0);
-    ierr=MPI_Type_commit(&localArray);    assert(ierr==0);
-#endif
-
-    //////////////////////////////////////////////////////////////////////////////
-    // Byte order
-    //////////////////////////////////////////////////////////////////////////////
-    int ieee32big = (format == std::string("IEEE32BIG"));
-    int ieee32    = (format == std::string("IEEE32"));
-    int ieee64big = (format == std::string("IEEE64BIG"));
-    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
-    assert(ieee64||ieee32|ieee64big||ieee32big);
-    assert((ieee64+ieee32+ieee64big+ieee32big)==1);
-    //////////////////////////////////////////////////////////////////////////////
-    // Do the I/O
-    //////////////////////////////////////////////////////////////////////////////
-    if ( control & BINARYIO_READ ) { 
-
-      timer.Start();
-
-      if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
-#ifdef USE_MPI_IO
-	std::cout<< GridLogMessage<<"IOobject: MPI read I/O "<< file<< std::endl;
-	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);    assert(ierr==0);
-	ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);    assert(ierr==0);
-	ierr=MPI_File_read_all(fh, &iodata[0], 1, localArray, &status);    assert(ierr==0);
-	MPI_File_close(&fh);
-	MPI_Type_free(&fileArray);
-	MPI_Type_free(&localArray);
-#else 
-	assert(0);
-#endif
-      } else {
-	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
-        std::ifstream fin;
-	fin.open(file, std::ios::binary | std::ios::in);
-        if (control & BINARYIO_MASTER_APPEND)
-        {
-          fin.seekg(-sizeof(fobj), fin.end);
-        }
-        else
-        {
-          fin.seekg(offset + myrank * lsites * sizeof(fobj));
-        }
-        fin.read((char *)&iodata[0], iodata.size() * sizeof(fobj));
-        assert(fin.fail() == 0);
-        fin.close();
-      }
-      timer.Stop();
-
-      grid->Barrier();
-
-      bstimer.Start();
-      ScidacChecksum(grid,iodata,scidac_csuma,scidac_csumb);
-      if (ieee32big) be32toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      if (ieee32)    le32toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      if (ieee64big) be64toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      if (ieee64)    le64toh_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      NerscChecksum(grid,iodata,nersc_csum);
-      bstimer.Stop();
-    }
-    
-    if ( control & BINARYIO_WRITE ) { 
-
-      bstimer.Start();
-      NerscChecksum(grid,iodata,nersc_csum);
-      if (ieee32big) htobe32_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      if (ieee32)    htole32_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      if (ieee64big) htobe64_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      if (ieee64)    htole64_v((void *)&iodata[0], sizeof(fobj)*iodata.size());
-      ScidacChecksum(grid,iodata,scidac_csuma,scidac_csumb);
-      bstimer.Stop();
-
-      grid->Barrier();
-
-      timer.Start();
-      if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
-#ifdef USE_MPI_IO
-        std::cout << GridLogMessage <<"IOobject: MPI write I/O " << file << std::endl;
-        ierr = MPI_File_open(grid->communicator, (char *)file.c_str(), MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh);
-	//        std::cout << GridLogMessage << "Checking for errors" << std::endl;
-        if (ierr != MPI_SUCCESS)
-        {
-          char error_string[BUFSIZ];
-          int length_of_error_string, error_class;
-
-          MPI_Error_class(ierr, &error_class);
-          MPI_Error_string(error_class, error_string, &length_of_error_string);
-          fprintf(stderr, "%3d: %s\n", myrank, error_string);
-          MPI_Error_string(ierr, error_string, &length_of_error_string);
-          fprintf(stderr, "%3d: %s\n", myrank, error_string);
-          MPI_Abort(MPI_COMM_WORLD, 1); //assert(ierr == 0);
-        }
-
-        std::cout << GridLogDebug << "MPI write I/O set view " << file << std::endl;
-        ierr = MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);
-        assert(ierr == 0);
-
-        std::cout << GridLogDebug << "MPI write I/O write all " << file << std::endl;
-        ierr = MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);
-        assert(ierr == 0);
-
-        MPI_Offset os;
-        MPI_File_get_position(fh, &os);
-        MPI_File_get_byte_offset(fh, os, &disp);
-        offset = disp;
-
-
-        MPI_File_close(&fh);
-        MPI_Type_free(&fileArray);
-        MPI_Type_free(&localArray);
-#else 
-	assert(0);
-#endif
-      } else { 
-
-        std::cout << GridLogMessage << "IOobject: C++ write I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
-        
-	std::ofstream fout; 
-	fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
-	try {
-	  if (offset) { // Must already exist and contain data
-	    fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
-	  } else {     // Allow create
-	    fout.open(file,std::ios::binary|std::ios::out);
-	  }
-	} catch (const std::fstream::failure& exc) {
-	  std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
-	  std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
-	  //	  std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
-#ifdef USE_MPI_IO
-	  MPI_Abort(MPI_COMM_WORLD,1);
-#else
-	  exit(1);
-#endif
-	}
-	
-	if ( control & BINARYIO_MASTER_APPEND )  {
-	  try {
-	    fout.seekp(0,fout.end);
-	  } catch (const std::fstream::failure& exc) {
-	    std::cout << "Exception in seeking file end " << file << std::endl;
-	  }
-	} else {
-	  try { 
-	    fout.seekp(offset+myrank*lsites*sizeof(fobj));
-	  } catch (const std::fstream::failure& exc) {
-	    std::cout << "Exception in seeking file " << file <<" offset "<< offset << std::endl;
-	  }
-	}
-
-	try {
-	  fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));//assert( fout.fail()==0);
-	}
-	catch (const std::fstream::failure& exc) {
-	  std::cout << "Exception in writing file " << file << std::endl;
-	  std::cout << GridLogError << "Exception description: "<< exc.what() << std::endl;
-#ifdef USE_MPI_IO
-	  MPI_Abort(MPI_COMM_WORLD,1);
-#else
-	  exit(1);
-#endif
-	}
-  offset  = fout.tellp();
-	fout.close();
-      }
-      timer.Stop();
-    }
-    
-    lastPerf.size            = sizeof(fobj)*iodata.size()*nrank;
-    lastPerf.time            = timer.useconds();
-    lastPerf.mbytesPerSecond = lastPerf.size/1024./1024./(lastPerf.time/1.0e6);
-    std::cout<<GridLogMessage<<"IOobject: ";
-    if ( control & BINARYIO_READ) std::cout << " read  ";
-    else                          std::cout << " write ";
-    uint64_t bytes = sizeof(fobj)*iodata.size()*nrank;
-    std::cout<< lastPerf.size <<" bytes in "<< timer.Elapsed() <<" "
-	     << lastPerf.mbytesPerSecond <<" MB/s "<<std::endl;
-
-    std::cout<<GridLogMessage<<"IOobject: endian and checksum overhead "<<bstimer.Elapsed()  <<std::endl;
-
-    //////////////////////////////////////////////////////////////////////////////
-    // Safety check
-    //////////////////////////////////////////////////////////////////////////////
-    // if the data size is 1 we do not want to sum over the MPI ranks
-    if (iodata.size() != 1){
-      grid->Barrier();
-      grid->GlobalSum(nersc_csum);
-      grid->GlobalXOR(scidac_csuma);
-      grid->GlobalXOR(scidac_csumb);
-      grid->Barrier();
-    }
-  }
-
-  /////////////////////////////////////////////////////////////////////////////
-  // Read a Lattice of object
-  //////////////////////////////////////////////////////////////////////////////////////
-  template<class vobj,class fobj,class munger>
-  static inline void readLatticeObject(Lattice<vobj> &Umu,
-				       std::string file,
-				       munger munge,
-				       uint64_t offset,
-				       const std::string &format,
-				       uint32_t &nersc_csum,
-				       uint32_t &scidac_csuma,
-				       uint32_t &scidac_csumb)
-  {
-    typedef typename vobj::scalar_object sobj;
-    typedef typename vobj::Realified::scalar_type word;    word w=0;
-
-    GridBase *grid = Umu.Grid();
-    uint64_t lsites = grid->lSites();
-
-    std::vector<sobj> scalardata(lsites); 
-    std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
-    
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
-	     nersc_csum,scidac_csuma,scidac_csumb);
-
-    GridStopWatch timer; 
-    timer.Start();
-
-    thread_for(x,lsites, { munge(iodata[x], scalardata[x]); });
-
-    vectorizeFromLexOrdArray(scalardata,Umu);    
-    grid->Barrier();
-
-    timer.Stop();
-    std::cout<<GridLogMessage<<"readLatticeObject: vectorize overhead "<<timer.Elapsed()  <<std::endl;
-  }
-
-  /////////////////////////////////////////////////////////////////////////////
-  // Write a Lattice of object
-  //////////////////////////////////////////////////////////////////////////////////////
-  template<class vobj,class fobj,class munger>
-    static inline void writeLatticeObject(Lattice<vobj> &Umu,
-					  std::string file,
-					  munger munge,
-					  uint64_t offset,
-					  const std::string &format,
-					  uint32_t &nersc_csum,
-					  uint32_t &scidac_csuma,
-					  uint32_t &scidac_csumb)
-  {
-    typedef typename vobj::scalar_object sobj;
-    typedef typename vobj::Realified::scalar_type word;    word w=0;
-    GridBase *grid = Umu.Grid();
-    uint64_t lsites = grid->lSites(), offsetCopy = offset;
-    int attemptsLeft = std::max(0, BinaryIO::latticeWriteMaxRetry);
-    bool checkWrite = (BinaryIO::latticeWriteMaxRetry >= 0);
-
-    std::vector<sobj> scalardata(lsites); 
-    std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
-
-    //////////////////////////////////////////////////////////////////////////////
-    // Munge [ .e.g 3rd row recon ]
-    //////////////////////////////////////////////////////////////////////////////
-    GridStopWatch timer; timer.Start();
-    unvectorizeToLexOrdArray(scalardata,Umu);    
-
-    thread_for(x, lsites, { munge(scalardata[x],iodata[x]); });
-
-    grid->Barrier();
-    timer.Stop();
-    while (attemptsLeft >= 0)
-    {
-      grid->Barrier();
-      IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
-	             nersc_csum,scidac_csuma,scidac_csumb);
-      if (checkWrite)
-      {
-        std::vector<fobj> ckiodata(lsites);
-        uint32_t          cknersc_csum, ckscidac_csuma, ckscidac_csumb;
-        uint64_t          ckoffset = offsetCopy;
-
-        std::cout << GridLogMessage << "writeLatticeObject: read back object" << std::endl;
-        grid->Barrier();
-        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
-	               cknersc_csum,ckscidac_csuma,ckscidac_csumb);
-        if ((cknersc_csum != nersc_csum) or (ckscidac_csuma != scidac_csuma) or (ckscidac_csumb != scidac_csumb))
-        {
-          std::cout << GridLogMessage << "writeLatticeObject: read test checksum failure, re-writing (" << attemptsLeft << " attempt(s) remaining)" << std::endl;
-          offset = offsetCopy;
-          thread_for(x,lsites, { munge(scalardata[x],iodata[x]); });
-        }
-        else
-        {
-          std::cout << GridLogMessage << "writeLatticeObject: read test checksum correct" << std::endl;
-          break;
-        }
-      }
-      attemptsLeft--;
-    }
-    
-
-    std::cout<<GridLogMessage<<"writeLatticeObject: unvectorize overhead "<<timer.Elapsed()  <<std::endl;
-  }
-  
-  /////////////////////////////////////////////////////////////////////////////
-  // Read a RNG;  use IOobject and lexico map to an array of state 
-  //////////////////////////////////////////////////////////////////////////////////////
-  static inline void readRNG(GridSerialRNG &serial_rng,
-			     GridParallelRNG &parallel_rng,
-			     std::string file,
-			     uint64_t offset,
-			     uint32_t &nersc_csum,
-			     uint32_t &scidac_csuma,
-			     uint32_t &scidac_csumb)
-  {
-    typedef typename GridSerialRNG::RngStateType RngStateType;
-    const int RngStateCount = GridSerialRNG::RngStateCount;
-    typedef std::array<RngStateType,RngStateCount> RNGstate;
-    typedef RngStateType word;    word w=0;
-
-    std::string format = "IEEE32BIG";
-
-    GridBase *grid = parallel_rng.Grid();
-    uint64_t gsites = grid->gSites();
-    uint64_t lsites = grid->lSites();
-
-    uint32_t nersc_csum_tmp   = 0;
-    uint32_t scidac_csuma_tmp = 0;
-    uint32_t scidac_csumb_tmp = 0;
-
-    GridStopWatch timer;
-
-    std::cout << GridLogMessage << "RNG read I/O on file " << file << std::endl;
-
-    std::vector<RNGstate> iodata(lsites);
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
-	     nersc_csum,scidac_csuma,scidac_csumb);
-
-    timer.Start();
-    thread_for(lidx,lsites,{  // FIX ME, suboptimal implementation
-      std::vector<RngStateType> tmp(RngStateCount);
-      std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
-      Coordinate lcoor;
-      grid->LocalIndexToLocalCoor(lidx, lcoor);
-      int o_idx=grid->oIndex(lcoor);
-      int i_idx=grid->iIndex(lcoor);
-      int gidx=parallel_rng.generator_idx(o_idx,i_idx);
-      parallel_rng.SetState(tmp,gidx);
-      });
-    timer.Stop();
-
-    iodata.resize(1);
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_MASTER_APPEND,
-	     nersc_csum_tmp,scidac_csuma_tmp,scidac_csumb_tmp);
-
-    {
-      std::vector<RngStateType> tmp(RngStateCount);
-      std::copy(iodata[0].begin(),iodata[0].end(),tmp.begin());
-      serial_rng.SetState(tmp,0);
-    }
-
-    nersc_csum   = nersc_csum   + nersc_csum_tmp;
-    scidac_csuma = scidac_csuma ^ scidac_csuma_tmp;
-    scidac_csumb = scidac_csumb ^ scidac_csumb_tmp;
-
-    std::cout << GridLogMessage << "RNG file nersc_checksum   " << std::hex << nersc_csum << std::dec << std::endl;
-    std::cout << GridLogMessage << "RNG file scidac_checksuma " << std::hex << scidac_csuma << std::dec << std::endl;
-    std::cout << GridLogMessage << "RNG file scidac_checksumb " << std::hex << scidac_csumb << std::dec << std::endl;
-
-    std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
-  }
-  /////////////////////////////////////////////////////////////////////////////
-  // Write a RNG; lexico map to an array of state and use IOobject
-  //////////////////////////////////////////////////////////////////////////////////////
-  static inline void writeRNG(GridSerialRNG &serial_rng,
-			      GridParallelRNG &parallel_rng,
-			      std::string file,
-			      uint64_t offset,
-			      uint32_t &nersc_csum,
-			      uint32_t &scidac_csuma,
-			      uint32_t &scidac_csumb)
-  {
-    typedef typename GridSerialRNG::RngStateType RngStateType;
-    typedef RngStateType word; word w=0;
-    const int RngStateCount = GridSerialRNG::RngStateCount;
-    typedef std::array<RngStateType,RngStateCount> RNGstate;
-
-    GridBase *grid = parallel_rng.Grid();
-    uint64_t gsites = grid->gSites();
-    uint64_t lsites = grid->lSites();
-
-    uint32_t nersc_csum_tmp;
-    uint32_t scidac_csuma_tmp;
-    uint32_t scidac_csumb_tmp;
-
-    GridStopWatch timer;
-    std::string format = "IEEE32BIG";
-
-    std::cout << GridLogMessage << "RNG write I/O on file " << file << std::endl;
-
-    timer.Start();
-    std::vector<RNGstate> iodata(lsites);
-    thread_for(lidx,lsites,{
-      std::vector<RngStateType> tmp(RngStateCount);
-      Coordinate lcoor;
-      grid->LocalIndexToLocalCoor(lidx, lcoor);
-      int o_idx=grid->oIndex(lcoor);
-      int i_idx=grid->iIndex(lcoor);
-      int gidx=parallel_rng.generator_idx(o_idx,i_idx);
-      parallel_rng.GetState(tmp,gidx);
-      std::copy(tmp.begin(),tmp.end(),iodata[lidx].begin());
-    });
-    timer.Stop();
-
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
-	     nersc_csum,scidac_csuma,scidac_csumb);
-    iodata.resize(1);
-    {
-      std::vector<RngStateType> tmp(RngStateCount);
-      serial_rng.GetState(tmp,0);
-      std::copy(tmp.begin(),tmp.end(),iodata[0].begin());
-    }
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_MASTER_APPEND,
-	     nersc_csum_tmp,scidac_csuma_tmp,scidac_csumb_tmp);
-
-    nersc_csum   = nersc_csum   + nersc_csum_tmp;
-    scidac_csuma = scidac_csuma ^ scidac_csuma_tmp;
-    scidac_csumb = scidac_csumb ^ scidac_csumb_tmp;
-    
-    std::cout << GridLogMessage << "RNG file checksum " << std::hex << nersc_csum    << std::dec << std::endl;
-    std::cout << GridLogMessage << "RNG file checksuma " << std::hex << scidac_csuma << std::dec << std::endl;
-    std::cout << GridLogMessage << "RNG file checksumb " << std::hex << scidac_csumb << std::dec << std::endl;
-    std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
-  }
-};
-
-NAMESPACE_END(Grid);

Grid/parallelIO/IldgIO.h

@@ -1,946 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/parallelIO/IldgIO.h
-
-Copyright (C) 2015
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-#ifdef HAVE_LIME
-#include <algorithm>
-#include <fstream>
-#include <iomanip>
-#include <iostream>
-#include <string>
-#include <map>
-
-#include <pwd.h>
-#include <sys/utsname.h>
-#include <unistd.h>
-
-//C-Lime is a must have for this functionality
-extern "C" {  
-#include "lime.h"
-}
-
-NAMESPACE_BEGIN(Grid);
-
-#define GRID_FIELD_NORM "FieldNormMetaData"
-#define GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) \
-0.5*fabs(FieldNormMetaData_.norm2 - n2ck)/(FieldNormMetaData_.norm2 + n2ck)
-#define GRID_FIELD_NORM_CHECK(FieldNormMetaData_, n2ck) \
-assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
-
-  /////////////////////////////////
-  // Encode word types as strings
-  /////////////////////////////////
- template<class word> inline std::string ScidacWordMnemonic(void){ return std::string("unknown"); }
- template<> inline std::string ScidacWordMnemonic<double>  (void){ return std::string("D"); }
- template<> inline std::string ScidacWordMnemonic<float>   (void){ return std::string("F"); }
- template<> inline std::string ScidacWordMnemonic< int32_t>(void){ return std::string("I32_t"); }
- template<> inline std::string ScidacWordMnemonic<uint32_t>(void){ return std::string("U32_t"); }
- template<> inline std::string ScidacWordMnemonic< int64_t>(void){ return std::string("I64_t"); }
- template<> inline std::string ScidacWordMnemonic<uint64_t>(void){ return std::string("U64_t"); }
-
-  /////////////////////////////////////////
-  // Encode a generic tensor as a string
-  /////////////////////////////////////////
- template<class vobj> std::string ScidacRecordTypeString(int &colors, int &spins, int & typesize,int &datacount) { 
-
-   typedef typename getPrecision<vobj>::real_scalar_type stype;
-
-   int _ColourN       = indexRank<ColourIndex,vobj>();
-   int _ColourScalar  =  isScalar<ColourIndex,vobj>();
-   int _ColourVector  =  isVector<ColourIndex,vobj>();
-   int _ColourMatrix  =  isMatrix<ColourIndex,vobj>();
-
-   int _SpinN       = indexRank<SpinIndex,vobj>();
-   int _SpinScalar  =  isScalar<SpinIndex,vobj>();
-   int _SpinVector  =  isVector<SpinIndex,vobj>();
-   int _SpinMatrix  =  isMatrix<SpinIndex,vobj>();
-
-   int _LorentzN       = indexRank<LorentzIndex,vobj>();
-   int _LorentzScalar  =  isScalar<LorentzIndex,vobj>();
-   int _LorentzVector  =  isVector<LorentzIndex,vobj>();
-   int _LorentzMatrix  =  isMatrix<LorentzIndex,vobj>();
-
-   std::stringstream stream;
-
-   stream << "GRID_";
-   stream << ScidacWordMnemonic<stype>();
-
-   if ( _LorentzVector )   stream << "_LorentzVector"<<_LorentzN;
-   if ( _LorentzMatrix )   stream << "_LorentzMatrix"<<_LorentzN;
-
-   if ( _SpinVector )   stream << "_SpinVector"<<_SpinN;
-   if ( _SpinMatrix )   stream << "_SpinMatrix"<<_SpinN;
-
-   if ( _ColourVector )   stream << "_ColourVector"<<_ColourN;
-   if ( _ColourMatrix )   stream << "_ColourMatrix"<<_ColourN;
-
-   if ( _ColourScalar && _LorentzScalar && _SpinScalar )   stream << "_Complex";
-
-
-   typesize = sizeof(typename vobj::scalar_type);
-
-   if ( _ColourMatrix ) typesize*= _ColourN*_ColourN;
-   else                 typesize*= _ColourN;
-
-   if ( _SpinMatrix )   typesize*= _SpinN*_SpinN;
-   else                 typesize*= _SpinN;
-
-   colors    = _ColourN;
-   spins     = _SpinN;
-   datacount = _LorentzN;
-
-   return stream.str();
- }
- 
- template<class vobj> std::string ScidacRecordTypeString(Lattice<vobj> & lat,int &colors, int &spins, int & typesize,int &datacount) { 
-   return ScidacRecordTypeString<vobj>(colors,spins,typesize,datacount);
- };
-
-
- ////////////////////////////////////////////////////////////
- // Helper to fill out metadata
- ////////////////////////////////////////////////////////////
-template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
-					  FieldMetaData &header,
-					  scidacRecord & _scidacRecord,
-					  scidacFile   & _scidacFile) 
- {
-   typedef typename getPrecision<vobj>::real_scalar_type stype;
-
-   /////////////////////////////////////
-   // Pull Grid's metadata
-   /////////////////////////////////////
-   PrepareMetaData(field,header);
-
-   /////////////////////////////////////
-   // Scidac Private File structure
-   /////////////////////////////////////
-   _scidacFile              = scidacFile(field.Grid());
-
-   /////////////////////////////////////
-   // Scidac Private Record structure
-   /////////////////////////////////////
-   scidacRecord sr;
-   sr.datatype   = ScidacRecordTypeString(field,sr.colors,sr.spins,sr.typesize,sr.datacount);
-   sr.date       = header.creation_date;
-   sr.precision  = ScidacWordMnemonic<stype>();
-   sr.recordtype = GRID_IO_FIELD;
-
-   _scidacRecord = sr;
-
-   //   std::cout << GridLogMessage << "Build SciDAC datatype " <<sr.datatype<<std::endl;
- }
- 
- ///////////////////////////////////////////////////////
- // Scidac checksum
- ///////////////////////////////////////////////////////
- static int scidacChecksumVerify(scidacChecksum &scidacChecksum_,uint32_t scidac_csuma,uint32_t scidac_csumb)
- {
-   uint32_t scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
-   uint32_t scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
-   if ( scidac_csuma !=scidac_checksuma) return 0;
-   if ( scidac_csumb !=scidac_checksumb) return 0;
-   return 1;
- }
-
-////////////////////////////////////////////////////////////////////////////////////
-// Lime, ILDG and Scidac I/O classes
-////////////////////////////////////////////////////////////////////////////////////
-class GridLimeReader : public BinaryIO {
- public:
-   ///////////////////////////////////////////////////
-   // FIXME: format for RNG? Now just binary out instead
-   ///////////////////////////////////////////////////
-
-   FILE       *File;
-   LimeReader *LimeR;
-   std::string filename;
-
-   /////////////////////////////////////////////
-   // Open the file
-   /////////////////////////////////////////////
-   void open(const std::string &_filename) 
-   {
-     filename= _filename;
-     File = fopen(filename.c_str(), "r");
-     if (File == nullptr)
-     {
-       std::cerr << "cannot open file '" << filename << "'" << std::endl;
-       abort();
-     }
-     LimeR = limeCreateReader(File);
-   }
-   /////////////////////////////////////////////
-   // Close the file
-   /////////////////////////////////////////////
-   void close(void){
-     fclose(File);
-     //     limeDestroyReader(LimeR);
-   }
-
-  ////////////////////////////////////////////
-  // Read a generic lattice field and verify checksum
-  ////////////////////////////////////////////
-  template<class vobj>
-  void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
-  {
-    typedef typename vobj::scalar_object sobj;
-    scidacChecksum scidacChecksum_;
-    FieldNormMetaData  FieldNormMetaData_;
-    uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-
-    std::string format = getFormatString<vobj>();
-
-    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
-
-      uint64_t file_bytes =limeReaderBytes(LimeR);
-
-      //      std::cout << GridLogMessage << limeReaderType(LimeR) << " "<< file_bytes <<" bytes "<<std::endl;
-      //      std::cout << GridLogMessage<< " readLimeObject seeking "<<  record_name <<" found record :" <<limeReaderType(LimeR) <<std::endl;
-
-      if ( !strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
-
-	//	std::cout << GridLogMessage<< " readLimeLatticeBinaryObject matches ! " <<std::endl;
-
-	uint64_t PayloadSize = sizeof(sobj) * field.Grid()->_gsites;
-
-	//	std::cout << "R sizeof(sobj)= " <<sizeof(sobj)<<std::endl;
-	//	std::cout << "R Gsites " <<field.Grid()->_gsites<<std::endl;
-	//	std::cout << "R Payload expected " <<PayloadSize<<std::endl;
-	//	std::cout << "R file size " <<file_bytes <<std::endl;
-
-	assert(PayloadSize == file_bytes);// Must match or user error
-
-	uint64_t offset= ftello(File);
-	//	std::cout << " ReadLatticeObject from offset "<<offset << std::endl;
-	BinarySimpleMunger<sobj,sobj> munge;
-	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
-	std::cout << GridLogMessage << "SciDAC checksum A " << std::hex << scidac_csuma << std::dec << std::endl;
-	std::cout << GridLogMessage << "SciDAC checksum B " << std::hex << scidac_csumb << std::dec << std::endl;
-	/////////////////////////////////////////////
-	// Insist checksum is next record
-	/////////////////////////////////////////////
-	readScidacChecksum(scidacChecksum_,FieldNormMetaData_);
-	/////////////////////////////////////////////
-	// Verify checksums
-	/////////////////////////////////////////////
-	if(FieldNormMetaData_.norm2 != 0.0){ 
-	  RealD n2ck = norm2(field);
-	  std::cout << GridLogMessage << "Field norm: metadata= " << FieldNormMetaData_.norm2 
-              << " / field= " << n2ck << " / rdiff= " << GRID_FIELD_NORM_CALC(FieldNormMetaData_,n2ck) << std::endl;
-	  GRID_FIELD_NORM_CHECK(FieldNormMetaData_,n2ck);
-	}
-	assert(scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb)==1);
-
-	// find out if next field is a GridFieldNorm
-	return;
-      }
-    }
-  }
-  void readScidacChecksum(scidacChecksum     &scidacChecksum_,
-			  FieldNormMetaData  &FieldNormMetaData_)
-  {
-    FieldNormMetaData_.norm2 =0.0;
-    std::string scidac_str(SCIDAC_CHECKSUM);
-    std::string field_norm_str(GRID_FIELD_NORM);
-    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
-      uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
-      std::vector<char> xmlc(nbytes+1,'\0');
-      limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
-      std::string xmlstring = std::string(&xmlc[0]);
-      XmlReader RD(xmlstring, true, "");
-      if ( !strncmp(limeReaderType(LimeR), field_norm_str.c_str(),strlen(field_norm_str.c_str()) )  ) {
-	//	std::cout << "FieldNormMetaData "<<xmlstring<<std::endl;
-	read(RD,field_norm_str,FieldNormMetaData_);
-      }
-      if ( !strncmp(limeReaderType(LimeR), scidac_str.c_str(),strlen(scidac_str.c_str()) )  ) {
-	//	std::cout << SCIDAC_CHECKSUM << " " <<xmlstring<<std::endl;
-	read(RD,std::string("scidacChecksum"),scidacChecksum_);
-	return;
-      }      
-    }
-    assert(0);
-  }
-  ////////////////////////////////////////////
-  // Read a generic serialisable object
-  ////////////////////////////////////////////
-  void readLimeObject(std::string &xmlstring,std::string record_name)
-  {
-    // should this be a do while; can we miss a first record??
-    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
-
-      //      std::cout << GridLogMessage<< " readLimeObject seeking "<< record_name <<" found record :" <<limeReaderType(LimeR) <<std::endl;
-      uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
-
-      if ( !strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
-
-	//	std::cout << GridLogMessage<< " readLimeObject matches ! " << record_name <<std::endl;
-	std::vector<char> xmlc(nbytes+1,'\0');
-	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
-	//	std::cout << GridLogMessage<< " readLimeObject matches XML " << &xmlc[0] <<std::endl;
-
-	xmlstring = std::string(&xmlc[0]);
-	return;
-      }
-
-    }  
-    assert(0);
-  }
-
-  template<class serialisable_object>
-  void readLimeObject(serialisable_object &object,std::string object_name,std::string record_name)
-  {
-    std::string xmlstring;
-
-    readLimeObject(xmlstring, record_name);
-    XmlReader RD(xmlstring, true, "");
-    read(RD,object_name,object);
-  }
-};
-
-class GridLimeWriter : public BinaryIO 
-{
- public:
-
-   ///////////////////////////////////////////////////
-   // FIXME: format for RNG? Now just binary out instead
-   // FIXME: collective calls or not ?
-   //      : must know if I am the I/O boss
-   ///////////////////////////////////////////////////
-   FILE       *File;
-   LimeWriter *LimeW;
-   std::string filename;
-   bool        boss_node;
-   GridLimeWriter( bool isboss = true) {
-     boss_node = isboss;
-   }
-   void open(const std::string &_filename) { 
-     filename= _filename;
-     if ( boss_node ) {
-       File = fopen(filename.c_str(), "w");
-       LimeW = limeCreateWriter(File); assert(LimeW != NULL );
-     }
-   }
-   /////////////////////////////////////////////
-   // Close the file
-   /////////////////////////////////////////////
-   void close(void) {
-     if ( boss_node ) {
-       fclose(File);
-     }
-     //  limeDestroyWriter(LimeW);
-   }
-  ///////////////////////////////////////////////////////
-  // Lime utility functions
-  ///////////////////////////////////////////////////////
-  int createLimeRecordHeader(std::string message, int MB, int ME, size_t PayloadSize)
-  {
-    if ( boss_node ) {
-      LimeRecordHeader *h;
-      h = limeCreateHeader(MB, ME, const_cast<char *>(message.c_str()), PayloadSize);
-      assert(limeWriteRecordHeader(h, LimeW) >= 0);
-      limeDestroyHeader(h);
-    }
-    return LIME_SUCCESS;
-  }
-  ////////////////////////////////////////////
-  // Write a generic serialisable object
-  ////////////////////////////////////////////
-  void writeLimeObject(int MB,int ME,XmlWriter &writer,std::string object_name,std::string record_name)
-  {
-    if ( boss_node ) {
-      std::string xmlstring = writer.docString();
-
-      //    std::cout << "WriteLimeObject" << record_name <<std::endl;
-      uint64_t nbytes = xmlstring.size();
-      //    std::cout << " xmlstring "<< nbytes<< " " << xmlstring <<std::endl;
-      int err;
-      LimeRecordHeader *h = limeCreateHeader(MB, ME,const_cast<char *>(record_name.c_str()), nbytes); 
-      assert(h!= NULL);
-      
-      err=limeWriteRecordHeader(h, LimeW);                    assert(err>=0);
-      err=limeWriteRecordData(&xmlstring[0], &nbytes, LimeW); assert(err>=0);
-      err=limeWriterCloseRecord(LimeW);                       assert(err>=0);
-      limeDestroyHeader(h);
-    }
-  }
-
-  template<class serialisable_object>
-  void writeLimeObject(int MB,int ME,serialisable_object &object,std::string object_name,std::string record_name, const unsigned int scientificPrec = 0)
-  {
-    XmlWriter WR("","");
-
-    if (scientificPrec)
-    {
-      WR.scientificFormat(true);
-      WR.setPrecision(scientificPrec);
-    }
-    write(WR,object_name,object);
-    writeLimeObject(MB, ME, WR, object_name, record_name);
-  }
-  ////////////////////////////////////////////////////
-  // Write a generic lattice field and csum
-  // This routine is Collectively called by all nodes
-  // in communicator used by the field.Grid()
-  ////////////////////////////////////////////////////
-  template<class vobj>
-  void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
-  {
-    ////////////////////////////////////////////////////////////////////
-    // NB: FILE and iostream are jointly writing disjoint sequences in the
-    // the same file through different file handles (integer units).
-    // 
-    // These are both buffered, so why I think this code is right is as follows.
-    //
-    // i)  write record header to FILE *File, telegraphing the size; flush
-    // ii) ftello reads the offset from FILE *File . 
-    // iii) iostream / MPI Open independently seek this offset. Write sequence direct to disk.
-    //      Closes iostream and flushes.
-    // iv) fseek on FILE * to end of this disjoint section.
-    //  v) Continue writing scidac record.
-    ////////////////////////////////////////////////////////////////////
-    
-    GridBase *grid = field.Grid();
-    assert(boss_node == field.Grid()->IsBoss() );
-
-    FieldNormMetaData FNMD; FNMD.norm2 = norm2(field);
-
-    ////////////////////////////////////////////
-    // Create record header
-    ////////////////////////////////////////////
-    typedef typename vobj::scalar_object sobj;
-    int err;
-    uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    uint64_t PayloadSize = sizeof(sobj) * grid->_gsites;
-    if ( boss_node ) {
-      createLimeRecordHeader(record_name, 0, 0, PayloadSize);
-      fflush(File);
-    }
-    
-    //    std::cout << "W sizeof(sobj)"      <<sizeof(sobj)<<std::endl;
-    //    std::cout << "W Gsites "           <<field.Grid()->_gsites<<std::endl;
-    //    std::cout << "W Payload expected " <<PayloadSize<<std::endl;
-
-    ////////////////////////////////////////////////
-    // Check all nodes agree on file position
-    ////////////////////////////////////////////////
-    uint64_t offset1;
-    if ( boss_node ) {
-      offset1 = ftello(File);    
-    }
-    grid->Broadcast(0,(void *)&offset1,sizeof(offset1));
-
-    ///////////////////////////////////////////
-    // The above is collective. Write by other means into the binary record
-    ///////////////////////////////////////////
-    std::string format = getFormatString<vobj>();
-    BinarySimpleMunger<sobj,sobj> munge;
-    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb);
-
-    ///////////////////////////////////////////
-    // Wind forward and close the record
-    ///////////////////////////////////////////
-    if ( boss_node ) {
-      fseek(File,0,SEEK_END);             
-      uint64_t offset2 = ftello(File);     //    std::cout << " now at offset "<<offset2 << std::endl;
-      assert( (offset2-offset1) == PayloadSize);
-    }
-
-    /////////////////////////////////////////////////////////////
-    // Check MPI-2 I/O did what we expect to file
-    /////////////////////////////////////////////////////////////
-
-    if ( boss_node ) { 
-      err=limeWriterCloseRecord(LimeW);  assert(err>=0);
-    }
-    ////////////////////////////////////////
-    // Write checksum element, propagaing forward from the BinaryIO
-    // Always pair a checksum with a binary object, and close message
-    ////////////////////////////////////////
-    scidacChecksum checksum;
-    std::stringstream streama; streama << std::hex << scidac_csuma;
-    std::stringstream streamb; streamb << std::hex << scidac_csumb;
-    checksum.suma= streama.str();
-    checksum.sumb= streamb.str();
-    if ( boss_node ) { 
-      writeLimeObject(0,0,FNMD,std::string(GRID_FIELD_NORM),std::string(GRID_FIELD_NORM));
-      writeLimeObject(0,1,checksum,std::string("scidacChecksum"),std::string(SCIDAC_CHECKSUM));
-    }
-  }
-};
-
-class ScidacWriter : public GridLimeWriter {
- public:
-
-  ScidacWriter(bool isboss =true ) : GridLimeWriter(isboss)  { };
-
-  template<class SerialisableUserFile>
-  void writeScidacFileRecord(GridBase *grid,SerialisableUserFile &_userFile)
-  {
-    scidacFile    _scidacFile(grid);
-    if ( this->boss_node ) {
-      writeLimeObject(1,0,_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
-      writeLimeObject(0,1,_userFile,_userFile.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
-    }
-  }
-  ////////////////////////////////////////////////
-  // Write generic lattice field in scidac format
-  ////////////////////////////////////////////////
-  template <class vobj, class userRecord>
-  void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
-                              const unsigned int recordScientificPrec = 0) 
-  {
-    GridBase * grid = field.Grid();
-
-    ////////////////////////////////////////
-    // fill the Grid header
-    ////////////////////////////////////////
-    FieldMetaData header;
-    scidacRecord  _scidacRecord;
-    scidacFile    _scidacFile;
-
-    ScidacMetaData(field,header,_scidacRecord,_scidacFile);
-
-    //////////////////////////////////////////////
-    // Fill the Lime file record by record
-    //////////////////////////////////////////////
-    if ( this->boss_node ) {
-      writeLimeObject(1,0,header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
-      writeLimeObject(0,0,_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML), recordScientificPrec);
-      writeLimeObject(0,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
-    }
-    // Collective call
-    writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
-  }
-};
-
-
-class ScidacReader : public GridLimeReader {
- public:
-
-   template<class SerialisableUserFile>
-   void readScidacFileRecord(GridBase *grid,SerialisableUserFile &_userFile)
-   {
-     scidacFile    _scidacFile(grid);
-     readLimeObject(_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
-     readLimeObject(_userFile,_userFile.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
-   }
-  ////////////////////////////////////////////////
-  // Write generic lattice field in scidac format
-  ////////////////////////////////////////////////
-  template <class vobj, class userRecord>
-  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord) 
-  {
-    typedef typename vobj::scalar_object sobj;
-    GridBase * grid = field.Grid();
-
-    ////////////////////////////////////////
-    // fill the Grid header
-    ////////////////////////////////////////
-    FieldMetaData header;
-    scidacRecord  _scidacRecord;
-    scidacFile    _scidacFile;
-
-    //////////////////////////////////////////////
-    // Fill the Lime file record by record
-    //////////////////////////////////////////////
-    readLimeObject(header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
-    readLimeObject(_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
-    readLimeObject(_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
-    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));
-  }
-  void skipPastBinaryRecord(void) {
-    std::string rec_name(ILDG_BINARY_DATA);
-    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
-      if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
-  // in principle should do the line below, but that breaks backard compatibility with old data
-  // skipPastObjectRecord(std::string(GRID_FIELD_NORM));
-	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
-	return;
-      }
-    }    
-  }
-  void skipPastObjectRecord(std::string rec_name) {
-    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
-      if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
-	return;
-      }
-    }
-  }
-  void skipScidacFieldRecord() {
-    skipPastObjectRecord(std::string(GRID_FORMAT));
-    skipPastObjectRecord(std::string(SCIDAC_RECORD_XML));
-    skipPastObjectRecord(std::string(SCIDAC_PRIVATE_RECORD_XML));
-    skipPastBinaryRecord();
-  }
-};
-
-
-class IldgWriter : public ScidacWriter {
- public:
-  
-  IldgWriter(bool isboss) : ScidacWriter(isboss) {};
-
-  ///////////////////////////////////
-  // A little helper
-  ///////////////////////////////////
-  void writeLimeIldgLFN(std::string &LFN)
-  {
-    uint64_t PayloadSize = LFN.size();
-    int err;
-    createLimeRecordHeader(ILDG_DATA_LFN, 0 , 0, PayloadSize);
-    err=limeWriteRecordData(const_cast<char*>(LFN.c_str()), &PayloadSize,LimeW); assert(err>=0);
-    err=limeWriterCloseRecord(LimeW); assert(err>=0);
-  }
-
-  ////////////////////////////////////////////////////////////////
-  // Special ILDG operations ; gauge configs only.
-  // Don't require scidac records EXCEPT checksum
-  // Use Grid MetaData object if present.
-  ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
-  {
-    GridBase * grid = Umu.Grid();
-    typedef Lattice<vLorentzColourMatrixD> GaugeField;
-    typedef vLorentzColourMatrixD vobj;
-    typedef typename vobj::scalar_object sobj;
-
-    ////////////////////////////////////////
-    // fill the Grid header
-    ////////////////////////////////////////
-    FieldMetaData header;
-    scidacRecord  _scidacRecord;
-    scidacFile    _scidacFile;
-
-    ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
-
-    stats Stats;
-    Stats(Umu,header);
-    
-    std::string format = header.floating_point;
-    header.ensemble_id    = description;
-    header.ensemble_label = description;
-    header.sequence_number = sequence;
-    header.ildg_lfn = LFN;
-
-    assert ( (format == std::string("IEEE32BIG"))  
-           ||(format == std::string("IEEE64BIG")) );
-
-    //////////////////////////////////////////////////////
-    // Fill ILDG header data struct
-    //////////////////////////////////////////////////////
-    ildgFormat ildgfmt ;
-    const std::string stNC = std::to_string( Nc ) ;
-    ildgfmt.field          = std::string("su"+stNC+"gauge");
-
-    if ( format == std::string("IEEE32BIG") ) { 
-      ildgfmt.precision = 32;
-    } else { 
-      ildgfmt.precision = 64;
-    }
-    ildgfmt.version = 1.0;
-    ildgfmt.lx = header.dimension[0];
-    ildgfmt.ly = header.dimension[1];
-    ildgfmt.lz = header.dimension[2];
-    ildgfmt.lt = header.dimension[3];
-    assert(header.nd==4);
-    assert(header.nd==header.dimension.size());
-
-    //////////////////////////////////////////////////////////////////////////////
-    // Field norm tests
-    //////////////////////////////////////////////////////////////////////////////
-    FieldNormMetaData FieldNormMetaData_;
-    FieldNormMetaData_.norm2 = norm2(Umu);
-
-    //////////////////////////////////////////////////////////////////////////////
-    // Fill the USQCD info field
-    //////////////////////////////////////////////////////////////////////////////
-    usqcdInfo info;
-    info.version=1.0;
-    info.plaq   = header.plaquette;
-    info.linktr = header.link_trace;
-
-    //    std::cout << GridLogMessage << " Writing config; IldgIO n2 "<< FieldNormMetaData_.norm2<<std::endl;
-    //////////////////////////////////////////////
-    // Fill the Lime file record by record
-    //////////////////////////////////////////////
-    writeLimeObject(1,0,header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
-    writeLimeObject(0,0,FieldNormMetaData_,FieldNormMetaData_.SerialisableClassName(),std::string(GRID_FIELD_NORM));
-    writeLimeObject(0,0,_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
-    writeLimeObject(0,1,info,info.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
-    writeLimeObject(1,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
-    writeLimeObject(0,0,info,info.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
-    writeLimeObject(0,0,ildgfmt,std::string("ildgFormat")   ,std::string(ILDG_FORMAT)); // rec
-    writeLimeIldgLFN(header.ildg_lfn);                                                 // rec
-    writeLimeLatticeBinaryObject(Umu,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
-    //    limeDestroyWriter(LimeW);
-  }
-};
-
-class IldgReader : public GridLimeReader {
- public:
-
-  ////////////////////////////////////////////////////////////////
-  // Read either Grid/SciDAC/ILDG configuration
-  // Don't require scidac records EXCEPT checksum
-  // Use Grid MetaData object if present.
-  // Else use ILDG MetaData object if present.
-  // Else use SciDAC MetaData object if present.
-  ////////////////////////////////////////////////////////////////
-  template <class stats = PeriodicGaugeStatistics>
-  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
-
-    typedef Lattice<vLorentzColourMatrixD > GaugeField;
-    typedef typename GaugeField::vector_object  vobj;
-    typedef typename vobj::scalar_object sobj;
-
-    typedef LorentzColourMatrixF fobj;
-    typedef LorentzColourMatrixD dobj;
-
-    GridBase *grid = Umu.Grid();
-
-    Coordinate dims = Umu.Grid()->FullDimensions();
-
-    assert(dims.size()==4);
-
-    // Metadata holders
-    ildgFormat     ildgFormat_    ;
-    std::string    ildgLFN_       ;
-    scidacChecksum scidacChecksum_; 
-    usqcdInfo      usqcdInfo_     ;
-    FieldNormMetaData FieldNormMetaData_;
-
-    // track what we read from file
-    int found_ildgFormat    =0;
-    int found_ildgLFN       =0;
-    int found_scidacChecksum=0;
-    int found_usqcdInfo     =0;
-    int found_ildgBinary =0;
-    int found_FieldMetaData =0;
-    int found_FieldNormMetaData =0;
-    uint32_t nersc_csum;
-    uint32_t scidac_csuma;
-    uint32_t scidac_csumb;
-
-    // Binary format
-    std::string format;
-
-    //////////////////////////////////////////////////////////////////////////
-    // Loop over all records
-    // -- Order is poorly guaranteed except ILDG header preceeds binary section.
-    // -- Run like an event loop.
-    // -- Impose trust hierarchy. Grid takes precedence & look for ILDG, and failing
-    //    that Scidac. 
-    // -- Insist on Scidac checksum record.
-    //////////////////////////////////////////////////////////////////////////
-
-    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
-
-      uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
-      
-      //////////////////////////////////////////////////////////////////
-      // If not BINARY_DATA read a string and parse
-      //////////////////////////////////////////////////////////////////
-      if ( strncmp(limeReaderType(LimeR), ILDG_BINARY_DATA,strlen(ILDG_BINARY_DATA) )  ) {
-	
-	// Copy out the string
-	std::vector<char> xmlc(nbytes+1,'\0');
-	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
-	//	std::cout << GridLogMessage<< "Non binary record :" <<limeReaderType(LimeR) <<std::endl; //<<"\n"<<(&xmlc[0])<<std::endl;
-
-	//////////////////////////////////
-	// ILDG format record
-
-	std::string xmlstring(&xmlc[0]);
-	if ( !strncmp(limeReaderType(LimeR), ILDG_FORMAT,strlen(ILDG_FORMAT)) ) { 
-
-	  XmlReader RD(xmlstring, true, "");
-	  read(RD,"ildgFormat",ildgFormat_);
-
-	  if ( ildgFormat_.precision == 64 ) format = std::string("IEEE64BIG");
-	  if ( ildgFormat_.precision == 32 ) format = std::string("IEEE32BIG");
-
-	  assert( ildgFormat_.lx == dims[0]);
-	  assert( ildgFormat_.ly == dims[1]);
-	  assert( ildgFormat_.lz == dims[2]);
-	  assert( ildgFormat_.lt == dims[3]);
-
-	  found_ildgFormat = 1;
-	}
-
-	if ( !strncmp(limeReaderType(LimeR), ILDG_DATA_LFN,strlen(ILDG_DATA_LFN)) ) {
-	  FieldMetaData_.ildg_lfn = xmlstring;
-	  found_ildgLFN = 1;
-	}
-
-	if ( !strncmp(limeReaderType(LimeR), GRID_FORMAT,strlen(ILDG_FORMAT)) ) { 
-
-	  XmlReader RD(xmlstring, true, "");
-	  read(RD,"FieldMetaData",FieldMetaData_);
-
-	  format = FieldMetaData_.floating_point;
-
-	  assert(FieldMetaData_.dimension[0] == dims[0]);
-	  assert(FieldMetaData_.dimension[1] == dims[1]);
-	  assert(FieldMetaData_.dimension[2] == dims[2]);
-	  assert(FieldMetaData_.dimension[3] == dims[3]);
-
-	  found_FieldMetaData = 1;
-	}
-
-	if ( !strncmp(limeReaderType(LimeR), SCIDAC_RECORD_XML,strlen(SCIDAC_RECORD_XML)) ) { 
-	  // is it a USQCD info field
-	  if ( xmlstring.find(std::string("usqcdInfo")) != std::string::npos ) { 
-	    //	    std::cout << GridLogMessage<<"...found a usqcdInfo field"<<std::endl;
-	    XmlReader RD(xmlstring, true, "");
-	    read(RD,"usqcdInfo",usqcdInfo_);
-	    found_usqcdInfo = 1;
-	  }
-	}
-
-	if ( !strncmp(limeReaderType(LimeR), SCIDAC_CHECKSUM,strlen(SCIDAC_CHECKSUM)) ) { 
-	  XmlReader RD(xmlstring, true, "");
-	  read(RD,"scidacChecksum",scidacChecksum_);
-	  found_scidacChecksum = 1;
-	}
-
-	if ( !strncmp(limeReaderType(LimeR), GRID_FIELD_NORM,strlen(GRID_FIELD_NORM)) ) { 
-	  XmlReader RD(xmlstring, true, "");
-	  read(RD,GRID_FIELD_NORM,FieldNormMetaData_);
-	  found_FieldNormMetaData = 1;
-	}
-
-      } else {  
-	/////////////////////////////////
-	// Binary data
-	/////////////////////////////////
-	//	std::cout << GridLogMessage << "ILDG Binary record found : "  ILDG_BINARY_DATA << std::endl;
-	uint64_t offset= ftello(File);
-	if ( format == std::string("IEEE64BIG") ) {
-	  GaugeSimpleMunger<dobj, sobj> munge;
-	  BinaryIO::readLatticeObject< vobj, dobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
-	} else { 
-	  GaugeSimpleMunger<fobj, sobj> munge;
-	  BinaryIO::readLatticeObject< vobj, fobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
-	}
-
-	found_ildgBinary = 1;
-      }
-
-    }
-
-    //////////////////////////////////////////////////////
-    // Minimally must find binary segment and checksum
-    // Since this is an ILDG reader require ILDG format
-    //////////////////////////////////////////////////////
-    assert(found_ildgLFN);
-    assert(found_ildgBinary);
-    assert(found_ildgFormat);
-    assert(found_scidacChecksum);
-
-    // Must find something with the lattice dimensions
-    assert(found_FieldMetaData||found_ildgFormat);
-
-    if ( found_FieldMetaData ) {
-
-      std::cout << GridLogMessage<<"Grid MetaData was record found: configuration was probably written by Grid ! Yay ! "<<std::endl;
-
-    } else { 
-
-      assert(found_ildgFormat);
-      const std::string stNC = std::to_string( Nc ) ;
-      assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
-
-      ///////////////////////////////////////////////////////////////////////////////////////
-      // Populate our Grid metadata as best we can
-      ///////////////////////////////////////////////////////////////////////////////////////
-
-      std::ostringstream vers; vers << ildgFormat_.version;
-      FieldMetaData_.hdr_version = vers.str();
-      FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
-
-      FieldMetaData_.nd=4;
-      FieldMetaData_.dimension.resize(4);
-
-      FieldMetaData_.dimension[0] = ildgFormat_.lx ;
-      FieldMetaData_.dimension[1] = ildgFormat_.ly ;
-      FieldMetaData_.dimension[2] = ildgFormat_.lz ;
-      FieldMetaData_.dimension[3] = ildgFormat_.lt ;
-
-      if ( found_usqcdInfo ) { 
-	FieldMetaData_.plaquette = usqcdInfo_.plaq;
-	FieldMetaData_.link_trace= usqcdInfo_.linktr;
-	std::cout << GridLogMessage <<"This configuration was probably written by USQCD "<<std::endl;
-	std::cout << GridLogMessage <<"USQCD xml record Plaquette : "<<FieldMetaData_.plaquette<<std::endl;
-	std::cout << GridLogMessage <<"USQCD xml record LinkTrace : "<<FieldMetaData_.link_trace<<std::endl;
-      } else { 
-	FieldMetaData_.plaquette = 0.0;
-	FieldMetaData_.link_trace= 0.0;
-	std::cout << GridLogWarning << "This configuration is unsafe with no plaquette records that can verify it !!! "<<std::endl;
-      }
-    }
-
-    ////////////////////////////////////////////////////////////
-    // Really really want to mandate a scidac checksum
-    ////////////////////////////////////////////////////////////
-    if ( found_FieldNormMetaData ) { 
-      RealD nn = norm2(Umu);
-      GRID_FIELD_NORM_CHECK(FieldNormMetaData_,nn);
-      std::cout << GridLogMessage<<"FieldNormMetaData matches " << std::endl;
-    }  else { 
-      std::cout << GridLogWarning<<"FieldNormMetaData not found. " << std::endl;
-    }
-    if ( found_scidacChecksum ) {
-      FieldMetaData_.scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
-      FieldMetaData_.scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
-      scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb);
-      assert( scidac_csuma ==FieldMetaData_.scidac_checksuma);
-      assert( scidac_csumb ==FieldMetaData_.scidac_checksumb);
-      std::cout << GridLogMessage<<"SciDAC checksums match " << std::endl;
-    } else { 
-      std::cout << GridLogWarning<<"SciDAC checksums not found. This is unsafe. " << std::endl;
-      assert(0); // Can I insist always checksum ?
-    }
-
-    if ( found_FieldMetaData || found_usqcdInfo ) {
-      FieldMetaData checker;
-      stats Stats;
-      Stats(Umu,checker);
-      assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
-      assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
-      std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;
-    }
-  }
- };
-
-NAMESPACE_END(Grid);
-
-
-//HAVE_LIME
-#endif
-

Grid/parallelIO/IldgIOtypes.h

@@ -1,237 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/parallelIO/IldgIO.h
-
-Copyright (C) 2015
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-			   /*  END LEGAL */
-#ifndef GRID_ILDGTYPES_IO_H
-#define GRID_ILDGTYPES_IO_H
-
-#ifdef HAVE_LIME
-extern "C" { // for linkage
-#include "lime.h"
-}
-
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////////////////////////
-// Data representation of records that enter ILDG and SciDac formats
-/////////////////////////////////////////////////////////////////////////////////
-
-#define GRID_FORMAT      "grid-format"
-#define ILDG_FORMAT      "ildg-format"
-#define ILDG_BINARY_DATA "ildg-binary-data"
-#define ILDG_DATA_LFN    "ildg-data-lfn"
-#define SCIDAC_CHECKSUM           "scidac-checksum"
-#define SCIDAC_PRIVATE_FILE_XML   "scidac-private-file-xml"
-#define SCIDAC_FILE_XML           "scidac-file-xml"
-#define SCIDAC_PRIVATE_RECORD_XML "scidac-private-record-xml"
-#define SCIDAC_RECORD_XML         "scidac-record-xml"
-#define SCIDAC_BINARY_DATA        "scidac-binary-data"
-// Unused SCIDAC records names; could move to support this functionality
-#define SCIDAC_SITELIST           "scidac-sitelist"
-
-////////////////////////////////////////////////////////////
-const int GRID_IO_SINGLEFILE = 0; // hardcode lift from QIO compat
-const int GRID_IO_MULTIFILE  = 1; // hardcode lift from QIO compat
-const int GRID_IO_FIELD      = 0; // hardcode lift from QIO compat
-const int GRID_IO_GLOBAL     = 1; // hardcode lift from QIO compat
-////////////////////////////////////////////////////////////
-
-/////////////////////////////////////////////////////////////////////////////////
-// QIO uses mandatory "private" records fixed format
-// Private is in principle "opaque" however it can't be changed now because that would break existing 
-// file compatability, so should be correct to assume the undocumented but defacto file structure.
-/////////////////////////////////////////////////////////////////////////////////
-
-struct emptyUserRecord : Serializable { 
-  GRID_SERIALIZABLE_CLASS_MEMBERS(emptyUserRecord,int,dummy);
-  emptyUserRecord() { dummy=0; };
-};
-
-////////////////////////
-// Scidac private file xml
-// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile>
-////////////////////////
-struct scidacFile : Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(scidacFile,
-                                  double, version,
-                                  int, spacetime,
-				  std::string, dims, // must convert to int
-                                  int, volfmt);
-
-  std::vector<int> getDimensions(void) { 
-    std::stringstream stream(dims);
-    std::vector<int> dimensions;
-    int n;
-    while(stream >> n){
-      dimensions.push_back(n);
-    }
-    return dimensions;
-  }
-
-  void setDimensions(Coordinate dimensions) { 
-    char delimiter = ' ';
-    std::stringstream stream;
-    for(int i=0;i<dimensions.size();i++){ 
-      stream << dimensions[i];
-      if ( i != dimensions.size()-1) { 
-	stream << delimiter <<std::endl;
-      }
-    }
-    dims = stream.str();
-  }
-
-  // Constructor provides Grid
-  scidacFile() =default; // default constructor
-  scidacFile(GridBase * grid){
-    version      = 1.0;
-    spacetime    = grid->_ndimension;
-    setDimensions(grid->FullDimensions()); 
-    volfmt       = GRID_IO_SINGLEFILE;
-  }
-
-};
-
-///////////////////////////////////////////////////////////////////////
-// scidac-private-record-xml : example
-// <scidacRecord>
-// <version>1.1</version><date>Tue Jul 26 21:14:44 2011 UTC</date><recordtype>0</recordtype>
-// <datatype>QDP_D3_ColorMatrix</datatype><precision>D</precision><colors>3</colors><spins>4</spins>
-// <typesize>144</typesize><datacount>4</datacount>
-// </scidacRecord>
-///////////////////////////////////////////////////////////////////////
-
-struct scidacRecord : Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(scidacRecord,
-                                  double, version,
-                                  std::string, date,
-				  int, recordtype,
-				  std::string, datatype,
-				  std::string, precision,
-				  int, colors,
-				  int, spins,
-				  int, typesize,
-				  int, datacount);
-
-  scidacRecord()
-  : version(1.0), recordtype(0), colors(0), spins(0), typesize(0), datacount(0)
-  {}
-};
-
-////////////////////////
-// ILDG format
-////////////////////////
-struct ildgFormat : Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(ildgFormat,
-				  double, version,
-				  std::string, field,
-				  int, precision,
-				  int, lx,
-				  int, ly,
-				  int, lz,
-				  int, lt);
-  ildgFormat() { version=1.0; };
-};
-////////////////////////
-// USQCD info
-////////////////////////
-struct usqcdInfo : Serializable { 
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdInfo,
-				  double, version,
-				  double, plaq,
-				  double, linktr,
-				  std::string, info);
-  usqcdInfo() { 
-    version=1.0; 
-  };
-};
-////////////////////////
-// Scidac Checksum
-////////////////////////
-struct scidacChecksum : Serializable { 
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(scidacChecksum,
-				  double, version,
-				  std::string, suma,
-				  std::string, sumb);
-  scidacChecksum() { 
-    version=1.0; 
-  };
-};
-////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Type:           scidac-file-xml         <title>MILC ILDG archival gauge configuration</title>
-////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Type:           
-////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////
-// Scidac private file xml 
-// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile> 
-////////////////////////                                                                                                                                                                              
-
-#if 0
-////////////////////////////////////////////////////////////////////////////////////////
-// From http://www.physics.utah.edu/~detar/scidac/qio_2p3.pdf
-////////////////////////////////////////////////////////////////////////////////////////
-struct usqcdPropFile : Serializable { 
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropFile,
-				  double, version,
-				  std::string, type,
-				  std::string, info);
-  usqcdPropFile() { 
-    version=1.0; 
-  };
-};
-struct usqcdSourceInfo : Serializable { 
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdSourceInfo,
-				  double, version,
-				  std::string, info);
-  usqcdSourceInfo() { 
-    version=1.0; 
-  };
-};
-struct usqcdPropInfo : Serializable { 
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropInfo,
-				  double, version,
-				  int, spin,
-				  int, color,
-				  std::string, info);
-  usqcdPropInfo() { 
-    version=1.0; 
-  };
-};
-#endif
-
-NAMESPACE_END(Grid);
-#endif
-#endif