Fixed test for very much non-unit det

Revised dslash/dwf kernels for A64FX

.gitignore

@@ -114,3 +114,4 @@ gh-pages/
 #####################
 Grid/qcd/spin/gamma-gen/*.h
 Grid/qcd/spin/gamma-gen/*.cc
+Grid/util/Version.h

.travis.yml

@@ -9,11 +9,6 @@ matrix:
     - os:        osx
       osx_image: xcode8.3
       compiler: clang
-      env: PREC=single
-    - os:        osx
-      osx_image: xcode8.3
-      compiler: clang
-      env: PREC=double
       
 before_install:
     - export GRIDDIR=`pwd`
@@ -55,7 +50,7 @@ script:
     - make -j4
     - make install
     - cd $CWD/build
-    - ../configure --enable-precision=$PREC --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
+    - ../configure --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF}
     - make -j4 
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - make check

ChangeLog

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

Grid/DisableWarnings.h

@@ -30,8 +30,36 @@ directory
 #ifndef DISABLE_WARNINGS_H
 #define DISABLE_WARNINGS_H
 
- //disables and intel compiler specific warning (in json.hpp)
-#pragma warning disable 488  
 
 
+#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"
+#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
+
+//Eigen only
+#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/Grid.h

@@ -42,6 +42,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridQCDcore.h>
 #include <Grid/qcd/action/Action.h>
 #include <Grid/qcd/utils/GaugeFix.h>
+#include <Grid/qcd/utils/CovariantSmearing.h>
 #include <Grid/qcd/smearing/Smearing.h>
 #include <Grid/parallelIO/MetaData.h>
 #include <Grid/qcd/hmc/HMC_aggregate.h>

Grid/GridCore.h

@@ -38,16 +38,19 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_BASE_H
 #define GRID_BASE_H
 
-#include <Grid/GridStd.h>
 
+#include <Grid/DisableWarnings.h>
+#include <Grid/Namespace.h>
+#include <Grid/GridStd.h>
+#include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
 #include <Grid/perfmon/PerfCount.h>
-#include <Grid/log/Log.h>
-#include <Grid/allocator/AlignedAllocator.h>
-#include <Grid/simd/Simd.h>
-#include <Grid/serialisation/Serialisation.h>
-#include <Grid/threads/Threads.h>
 #include <Grid/util/Util.h>
+#include <Grid/log/Log.h>
+#include <Grid/allocator/Allocator.h>
+#include <Grid/simd/Simd.h>
+#include <Grid/threads/ThreadReduction.h>
+#include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 
 #include <Grid/cartesian/Cartesian.h>    
@@ -57,5 +60,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/stencil/Stencil.h>      
 #include <Grid/parallelIO/BinaryIO.h>
 #include <Grid/algorithms/Algorithms.h>   
+NAMESPACE_CHECK(GridCore)
 
 #endif

Grid/GridQCDcore.h

@@ -38,5 +38,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/qcd/spin/Spin.h>
 #include <Grid/qcd/utils/Utils.h>
 #include <Grid/qcd/representations/Representations.h>
+NAMESPACE_CHECK(GridQCDCore);
 
 #endif

Grid/GridStd.h

@@ -6,7 +6,9 @@
 ///////////////////
 #include <cassert>
 #include <complex>
+#include <memory>
 #include <vector>
+#include <array>
 #include <string>
 #include <iostream>
 #include <iomanip>
@@ -26,4 +28,7 @@
 ///////////////////
 #include "Config.h"
 
+#ifdef TOFU
+#undef GRID_COMMS_THREADS
+#endif
 #endif /* GRID_STD_H */

Grid/Grid_Eigen_Dense.h

@@ -1,14 +1,71 @@
+#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
+#pragma diag_suppress code_is_unreachable
+#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

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

Grid/Makefile.am

@@ -21,7 +21,8 @@ if BUILD_HDF5
   extra_headers+=serialisation/Hdf5Type.h
 endif
 
-all: version-cache
+
+all: version-cache Version.h
 
 version-cache:
 	@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
@@ -42,7 +43,7 @@ version-cache:
 	fi;\
 	rm -f vertmp
 
-Version.h:
+Version.h: version-cache
 	cp version-cache Version.h
 
 .PHONY: version-cache
@@ -53,6 +54,17 @@ Version.h:
 include Make.inc
 include Eigen.inc
 
+extra_sources+=$(ZWILS_FERMION_FILES)
+extra_sources+=$(WILS_FERMION_FILES)
+extra_sources+=$(STAG_FERMION_FILES)
+if BUILD_GPARITY
+  extra_sources+=$(GP_FERMION_FILES)
+endif
+if BUILD_FERMION_REPS
+  extra_sources+=$(ADJ_FERMION_FILES)
+  extra_sources+=$(TWOIND_FERMION_FILES)
+endif
+
 lib_LIBRARIES = libGrid.a
 
 CCFILES += $(extra_sources)

Grid/Namespace.h

@@ -0,0 +1,38 @@
+/*************************************************************************************
+
+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/Algorithms.h

@@ -29,33 +29,46 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_H
 #define GRID_ALGORITHMS_H
 
+NAMESPACE_CHECK(algorithms);
 #include <Grid/algorithms/SparseMatrix.h>
 #include <Grid/algorithms/LinearOperator.h>
 #include <Grid/algorithms/Preconditioner.h>
+NAMESPACE_CHECK(SparseMatrix);
 
 #include <Grid/algorithms/approx/Zolotarev.h>
 #include <Grid/algorithms/approx/Chebyshev.h>
+#include <Grid/algorithms/approx/JacobiPolynomial.h>
 #include <Grid/algorithms/approx/Remez.h>
 #include <Grid/algorithms/approx/MultiShiftFunction.h>
 #include <Grid/algorithms/approx/Forecast.h>
-
+#include <Grid/algorithms/approx/RemezGeneral.h>
+#include <Grid/algorithms/approx/ZMobius.h>
+NAMESPACE_CHECK(approx);
 #include <Grid/algorithms/iterative/Deflation.h>
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
+NAMESPACE_CHECK(ConjGrad);
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
+#include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
+#include <Grid/algorithms/iterative/MinimalResidual.h>
+#include <Grid/algorithms/iterative/GeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
+#include <Grid/algorithms/iterative/PowerMethod.h>
+
+NAMESPACE_CHECK(PowerMethod);
 #include <Grid/algorithms/CoarsenedMatrix.h>
+NAMESPACE_CHECK(CoarsendMatrix);
 #include <Grid/algorithms/FFT.h>
 
-// EigCg
-// Pcg
-// Hdcg
-// GCR
-// etc..
-
 #endif

Grid/algorithms/FFT.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -37,8 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif
 #endif
 
-
+NAMESPACE_BEGIN(Grid);
-namespace Grid {
 
 template<class scalar> struct FFTW { };
 
@@ -115,9 +113,9 @@ namespace Grid {
   double flops_call;
   uint64_t usec;
     
-    std::vector<int> dimensions;
+  Coordinate dimensions;
-    std::vector<int> processors;
+  Coordinate processors;
-    std::vector<int> processor_coor;
+  Coordinate processor_coor;
     
 public:
     
@@ -137,7 +135,7 @@ namespace Grid {
   {
     flops=0;
     usec =0;
-      std::vector<int> layout(Nd,1);
+    Coordinate layout(Nd,1);
     sgrid = new GridCartesian(dimensions,layout,processors);
   };
     
@@ -146,10 +144,10 @@ namespace Grid {
   }
     
   template<class vobj>
-    void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,std::vector<int> mask,int sign){
+  void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
 
-      conformable(result._grid,vgrid);
+    conformable(result.Grid(),vgrid);
-      conformable(source._grid,vgrid);
+    conformable(source.Grid(),vgrid);
     Lattice<vobj> tmp(vgrid);
     tmp = source;
     for(int d=0;d<Nd;d++){
@@ -162,7 +160,7 @@ namespace Grid {
 
   template<class vobj>
   void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
-      std::vector<int> mask(Nd,1);
+    Coordinate mask(Nd,1);
     FFT_dim_mask(result,source,mask,sign);
   }
 
@@ -172,14 +170,14 @@ namespace Grid {
 #ifndef HAVE_FFTW
     assert(0);
 #else
-      conformable(result._grid,vgrid);
+    conformable(result.Grid(),vgrid);
-      conformable(source._grid,vgrid);
+    conformable(source.Grid(),vgrid);
 
     int L = vgrid->_ldimensions[dim];
     int G = vgrid->_fdimensions[dim];
       
-      std::vector<int> layout(Nd,1);
+    Coordinate layout(Nd,1);
-      std::vector<int> pencil_gd(vgrid->_fdimensions);
+    Coordinate pencil_gd(vgrid->_fdimensions);
       
     pencil_gd[dim] = G*processors[dim];
       
@@ -191,7 +189,7 @@ namespace Grid {
     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;
@@ -217,8 +215,8 @@ namespace Grid {
       
     FFTW_plan p;
     {
-        FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[0];
+      FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
-        FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[0];
+      FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[0];
       p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
 					   in,inembed,
 					   istride,idist,
@@ -228,26 +226,23 @@ namespace Grid {
     }
       
     // Barrel shift and collect global pencil
-      std::vector<int> lcoor(Nd), gcoor(Nd);
+    Coordinate lcoor(Nd), gcoor(Nd);
     result = source;
     int pc = processor_coor[dim];
     for(int p=0;p<processors[dim];p++) {
-        PARALLEL_REGION
       {
-          std::vector<int> cbuf(Nd);
+	autoView(r_v,result,CpuRead);
+	autoView(p_v,pgbuf,CpuWrite);
+	thread_for(idx, sgrid->lSites(),{
+          Coordinate cbuf(Nd);
           sobj s;
-          
-          PARALLEL_FOR_LOOP_INTERN
-          for(int idx=0;idx<sgrid->lSites();idx++) {
 	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-            peekLocalSite(s,result,cbuf);
+	  peekLocalSite(s,r_v,cbuf);
 	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	    //            cbuf[dim]+=p*L;
+	  pokeLocalSite(s,p_v,cbuf);
-            pokeLocalSite(s,pgbuf,cbuf);
+        });
       }
-        }
+      if (p != processors[dim] - 1) {
-        if (p != processors[dim] - 1)
-        {
 	result = Cshift(result,dim,L);
       }
     }
@@ -256,20 +251,15 @@ namespace Grid {
     int NN=pencil_g.lSites();
     GridStopWatch timer;
     timer.Start();
-      PARALLEL_REGION
+    thread_for( idx,NN,{
-      {
+        Coordinate cbuf(Nd);
-        std::vector<int> cbuf(Nd);
-        
-        PARALLEL_FOR_LOOP_INTERN
-        for(int idx=0;idx<NN;idx++) {
 	pencil_g.LocalIndexToLocalCoor(idx, cbuf);
 	if ( cbuf[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
-            FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[idx];
+	  FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[idx];
-            FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[idx];
+	  FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[idx];
 	  FFTW<scalar>::fftw_execute_dft(p,in,out);
 	}
-        }
+    });
-      }
     timer.Stop();
       
     // performance counting
@@ -280,19 +270,18 @@ namespace Grid {
     flops+= flops_call*NN;
       
     // writing out result
-      PARALLEL_REGION
     {
-        std::vector<int> clbuf(Nd), cgbuf(Nd);
+      autoView(pgbuf_v,pgbuf,CpuRead);
+      autoView(result_v,result,CpuWrite);
+      thread_for(idx,sgrid->lSites(),{
+	Coordinate clbuf(Nd), cgbuf(Nd);
 	sobj s;
-        
-        PARALLEL_FOR_LOOP_INTERN
-        for(int idx=0;idx<sgrid->lSites();idx++) {
 	sgrid->LocalIndexToLocalCoor(idx,clbuf);
 	cgbuf = clbuf;
 	cgbuf[dim] = clbuf[dim]+L*pc;
-          peekLocalSite(s,pgbuf,cgbuf);
+	peekLocalSite(s,pgbuf_v,cgbuf);
-          pokeLocalSite(s,result,clbuf);
+	pokeLocalSite(s,result_v,clbuf);
-        }
+      });
     }
     result = result*div;
       
@@ -301,6 +290,7 @@ namespace Grid {
 #endif
   }
 };
-}
+
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/LinearOperator.h

@@ -26,16 +26,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef  GRID_ALGORITHM_LINEAR_OP_H
+#pragma once 
-#define  GRID_ALGORITHM_LINEAR_OP_H
 
-namespace Grid {
+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):
+// 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
@@ -44,10 +43,10 @@ namespace Grid {
 /////////////////////////////////////////////////////////////////////////////////////////////
 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
@@ -84,6 +83,9 @@ namespace Grid {
   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);
   }
@@ -91,11 +93,13 @@ namespace Grid {
     _Mat.Mdag(in,out);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	_Mat.MdagM(in,out,n1,n2);
+    _Mat.MdagM(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
   }
   void HermOp(const Field &in, Field &out){
-	RealD n1,n2;
+    _Mat.MdagM(in,out);
-	HermOpAndNorm(in,out,n1,n2);
   }
 };
 
@@ -117,6 +121,9 @@ namespace Grid {
     _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);
@@ -126,17 +133,14 @@ namespace Grid {
     assert(0);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	_Mat.MdagM(in,out,n1,n2);
+    HermOp(in,out);
-	out = out + _shift*in;
+    ComplexD dot = innerProduct(in,out);
-
-	ComplexD dot;	
-	dot= innerProduct(in,out);
     n1=real(dot);
     n2=norm2(out);
   }
   void HermOp(const Field &in, Field &out){
-	RealD n1,n2;
+    _Mat.MdagM(in,out);
-	HermOpAndNorm(in,out,n1,n2);
+    out = out + _shift*in;
   }
 };
 
@@ -155,6 +159,9 @@ namespace Grid {
   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);
   }
@@ -162,8 +169,7 @@ namespace Grid {
     _Mat.M(in,out);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	_Mat.M(in,out);
+    HermOp(in,out);
-	
     ComplexD dot= innerProduct(in,out); n1=real(dot);
     n2=norm2(out);
   }
@@ -172,6 +178,35 @@ namespace Grid {
   }
 };
 
+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
@@ -180,21 +215,24 @@ namespace Grid {
 template<class Field>
 class SchurOperatorBase :  public LinearOperatorBase<Field> {
  public:
-      virtual  RealD Mpc      (const Field &in, Field &out) =0;
+  virtual  void Mpc      (const Field &in, Field &out) =0;
-      virtual  RealD MpcDag   (const Field &in, Field &out) =0;
+  virtual  void MpcDag   (const Field &in, Field &out) =0;
-      virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+  virtual  void MpcDagMpc(const Field &in, Field &out) {
-      Field tmp(in._grid);
+    Field tmp(in.Grid());
-      tmp.checkerboard = in.checkerboard;
+    tmp.Checkerboard() = in.Checkerboard();
-	ni=Mpc(in,tmp);
+    Mpc(in,tmp);
-	no=MpcDag(tmp,out);
+    MpcDag(tmp,out);
   }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-      out.checkerboard = in.checkerboard;
+    out.Checkerboard() = in.Checkerboard();
-	MpcDagMpc(in,out,n1,n2);
+    MpcDagMpc(in,out);
+    ComplexD dot= innerProduct(in,out); 
+    n1=real(dot);
+    n2=norm2(out);
   }
   virtual void HermOp(const Field &in, Field &out){
-	RealD n1,n2;
+    out.Checkerboard() = in.Checkerboard();
-	HermOpAndNorm(in,out,n1,n2);
+    MpcDagMpc(in,out);
   }
   void Op     (const Field &in, Field &out){
     Mpc(in,out);
@@ -209,35 +247,33 @@ namespace Grid {
   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> {
-    protected:
-      Matrix &_Mat;
  public:
+    Matrix &_Mat;
     SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
-      virtual  RealD Mpc      (const Field &in, Field &out) {
+    virtual  void Mpc      (const Field &in, Field &out) {
-      Field tmp(in._grid);
+      Field tmp(in.Grid());
-      tmp.checkerboard = !in.checkerboard;
+      tmp.Checkerboard() = !in.Checkerboard();
-	//std::cout <<"grid pointers: in._grid="<< in._grid << " out._grid=" << out._grid << "  _Mat.Grid=" << _Mat.Grid() << " _Mat.RedBlackGrid=" << _Mat.RedBlackGrid() << std::endl;
       
       _Mat.Meooe(in,tmp);
       _Mat.MooeeInv(tmp,out);
       _Mat.Meooe(out,tmp);
-
-      //std::cout << "cb in " << in.checkerboard << "  cb out " << out.checkerboard << std::endl;
       _Mat.Mooee(in,out);
-	return axpy_norm(out,-1.0,tmp,out);
+      axpy(out,-1.0,tmp,out);
     }
-      virtual  RealD MpcDag   (const Field &in, Field &out){
+    virtual void MpcDag   (const Field &in, Field &out){
-	Field tmp(in._grid);
+      Field tmp(in.Grid());
 	
       _Mat.MeooeDag(in,tmp);
       _Mat.MooeeInvDag(tmp,out);
       _Mat.MeooeDag(out,tmp);
-
       _Mat.MooeeDag(in,out);
-	return axpy_norm(out,-1.0,tmp,out);
+      axpy(out,-1.0,tmp,out);
     }
 };
 template<class Matrix,class Field>
@@ -247,25 +283,23 @@ namespace Grid {
  public:
     SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
     
-      virtual  RealD Mpc      (const Field &in, Field &out) {
+    virtual void Mpc      (const Field &in, Field &out) {
-	Field tmp(in._grid);
+      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);
-	return axpy_norm(out,-1.0,tmp,in);
     }
-      virtual  RealD MpcDag   (const Field &in, Field &out){
+    virtual void MpcDag   (const Field &in, Field &out){
-	Field tmp(in._grid);
+      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);
-	return axpy_norm(out,-1.0,tmp,in);
     }
 };
 template<class Matrix,class Field>
@@ -275,30 +309,159 @@ namespace Grid {
  public:
     SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
     
-      virtual  RealD Mpc      (const Field &in, Field &out) {
+    virtual void Mpc      (const Field &in, Field &out) {
-	Field tmp(in._grid);
+      Field tmp(in.Grid());
       
       _Mat.MooeeInv(in,out);
       _Mat.Meooe(out,tmp);
       _Mat.MooeeInv(tmp,out);
       _Mat.Meooe(out,tmp);
       
-	return axpy_norm(out,-1.0,tmp,in);
+      axpy(out,-1.0,tmp,in);
     }
-      virtual  RealD MpcDag   (const Field &in, Field &out){
+    virtual  void MpcDag   (const Field &in, Field &out){
-	Field tmp(in._grid);
+      Field tmp(in.Grid());
 
       _Mat.MeooeDag(in,out);
       _Mat.MooeeInvDag(out,tmp);
       _Mat.MeooeDag(tmp,out);
       _Mat.MooeeInvDag(out,tmp);
 
-	return axpy_norm(out,-1.0,tmp,in);
+      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
+// 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> ;
@@ -311,61 +474,38 @@ namespace Grid {
   Matrix &_Mat;
   Field tmp;
   RealD mass;
-      double tMpc;
-      double tIP;
-      double tMeo;
-      double taxpby_norm;
-      uint64_t ncall;
  public:
-      void Report(void)
-      {
-	std::cout << GridLogMessage << " HermOpAndNorm.Mpc "<< tMpc/ncall<<" usec "<<std::endl;
-	std::cout << GridLogMessage << " HermOpAndNorm.IP  "<< tIP /ncall<<" usec "<<std::endl;
-	std::cout << GridLogMessage << " Mpc.MeoMoe        "<< tMeo/ncall<<" usec "<<std::endl;
-	std::cout << GridLogMessage << " Mpc.axpby_norm    "<< taxpby_norm/ncall<<" usec "<<std::endl;
-      }
   SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
   { 
     assert( _Mat.isTrivialEE() );
     mass = _Mat.Mass();
-	tMpc=0;
-	tIP =0;
-        tMeo=0;
-        taxpby_norm=0;
-	ncall=0;
   }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	ncall++;
+    Mpc(in,out);
-	tMpc-=usecond();
-	n2 = Mpc(in,out);
-	tMpc+=usecond();
-	tIP-=usecond();
     ComplexD dot= innerProduct(in,out);
-	tIP+=usecond();
     n1 = real(dot);
+    n2 =0.0;
   }
   virtual void HermOp(const Field &in, Field &out){
-	ncall++;
+    Mpc(in,out);
-	tMpc-=usecond();
+    //    _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);
-	tMpc+=usecond();
-	taxpby_norm-=usecond();
     axpby(out,-1.0,mass*mass,tmp,in);
-	taxpby_norm+=usecond();
   }
-      virtual  RealD Mpc      (const Field &in, Field &out) {
+  virtual  void MpcDag   (const Field &in, Field &out){
-	tMeo-=usecond();
+    Mpc(in,out);
-	_Mat.Meooe(in,out);
-	_Mat.Meooe(out,tmp);
-	tMeo+=usecond();
-	taxpby_norm-=usecond();
-	RealD nn=axpby_norm(out,-1.0,mass*mass,tmp,in);
-	taxpby_norm+=usecond();
-	return nn;
-      }
-      virtual  RealD MpcDag   (const Field &in, Field &out){
-	return Mpc(in,out);
   }
   virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
     assert(0);// Never need with staggered
@@ -373,7 +513,6 @@ namespace Grid {
 };
 template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
 
-
 /////////////////////////////////////////////////////////////
 // Base classes for functions of operators
 /////////////////////////////////////////////////////////////
@@ -465,13 +604,15 @@ namespace Grid {
 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 AtoN(in.Grid());
-      Field Mtmp(in._grid);
+    Field Mtmp(in.Grid());
     AtoN = in;
     out = AtoN*Coeffs[0];
     for(int n=1;n<Coeffs.size();n++){
@@ -482,6 +623,4 @@ namespace Grid {
   };
 };
 
-}
+NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/Preconditioner.h

@@ -28,7 +28,7 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 #ifndef GRID_PRECONDITIONER_H
 #define GRID_PRECONDITIONER_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class Field> class Preconditioner :  public LinearFunction<Field> { 
   virtual void operator()(const Field &src, Field & psi)=0;
@@ -42,5 +42,5 @@ namespace Grid {
   TrivialPrecon(void){};
 };
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/SparseMatrix.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #define  GRID_ALGORITHM_SPARSE_MATRIX_H
 
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////////////////////////////////////
 // Interface defining what I expect of a general sparse matrix, such as a Fermion action
@@ -38,15 +38,16 @@ namespace Grid {
 public:
   virtual GridBase *Grid(void) =0;
   // Full checkerboar operations
-      virtual RealD M    (const Field &in, Field &out)=0;
+  virtual void  M    (const Field &in, Field &out)=0;
-      virtual RealD Mdag (const Field &in, Field &out)=0;
+  virtual void  Mdag (const Field &in, Field &out)=0;
-      virtual void  MdagM(const Field &in, Field &out,RealD &ni,RealD &no) {
+  virtual void  MdagM(const Field &in, Field &out) {
-	Field tmp (in._grid);
+    Field tmp (in.Grid());
-	ni=M(in,tmp);
+    M(in,tmp);
-	no=Mdag(tmp,out);
+    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;
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -60,7 +61,7 @@ namespace Grid {
   // Query the even even properties to make algorithmic decisions
   //////////////////////////////////////////////////////////////////////
   virtual RealD  Mass(void)        { return 0.0; };
-      virtual int    ConstEE(void)     { return 0; }; // Disable assumptions unless overridden
+  virtual int    ConstEE(void)     { return 1; }; // Disable assumptions unless overridden
   virtual int    isTrivialEE(void) { return 0; }; // by a derived class that knows better
 
   // half checkerboard operaions
@@ -74,6 +75,6 @@ namespace Grid {
 
 };
 
-}
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/approx/Chebyshev.h

@@ -32,7 +32,7 @@ Author: Christoph Lehner <clehner@bnl.gov>
 
 #include <Grid/algorithms/LinearOperator.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 struct ChebyParams : Serializable {
   GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
@@ -47,6 +47,8 @@ struct ChebyParams : Serializable {
 template<class Field>
 class Chebyshev : public OperatorFunction<Field> {
 private:
+  using OperatorFunction<Field>::operator();
+
   std::vector<RealD> Coeffs;
   int order;
   RealD hi;
@@ -55,7 +57,7 @@ struct ChebyParams : Serializable {
 public:
   void csv(std::ostream &out){
     RealD diff = hi-lo;
-      RealD delta = (hi-lo)*1.0e-9;
+    RealD delta = diff*1.0e-9;
     for (RealD x=lo; x<hi; x+=delta) {
       delta*=1.1;
       RealD f = approx(x);
@@ -93,6 +95,24 @@ struct ChebyParams : Serializable {
     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;
@@ -212,12 +232,10 @@ struct ChebyParams : Serializable {
   // Implement the required interface
   void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 
-      GridBase *grid=in._grid;
+    GridBase *grid=in.Grid();
-
-      // std::cout << "Chevyshef(): in._grid="<<in._grid<<std::endl;
-      //std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
 
     int vol=grid->gSites();
+    typedef typename Field::vector_type vector_type;
 
     Field T0(grid); T0 = in;  
     Field T1(grid); 
@@ -232,20 +250,34 @@ struct ChebyParams : Serializable {
     RealD xscale = 2.0/(hi-lo);
     RealD mscale = -(hi+lo)/(hi-lo);
     Linop.HermOp(T0,y);
-      T1=y*xscale+in*mscale;
+    axpby(T1,xscale,mscale,y,in);
 
     // sum = .5 c[0] T0 + c[1] T1
-      out = (0.5*Coeffs[0])*T0 + Coeffs[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);
-
+#if 0
-	y=xscale*y+mscale*(*Tn);
+      auto y_v = y.View();
-
+      auto Tn_v = Tn->View();
-	*Tnp=2.0*y-(*Tnm);
+      auto Tnp_v = Tnp->View();
-
+      auto Tnm_v = Tnm->View();
-	out=out+Coeffs[n]* (*Tnp);
+      constexpr int Nsimd = vector_type::Nsimd();
-
+      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
+	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
+	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
+      });
+      if ( Coeffs[n] != 0.0) {
+	axpy(out,Coeffs[n],*Tnp,out);
+      }
+#else
+      axpby(y,xscale,mscale,y,(*Tn));
+      axpby(*Tnp,2.0,-1.0,y,(*Tnm));
+      if ( Coeffs[n] != 0.0) {
+	axpy(out,Coeffs[n],*Tnp,out);
+      }
+#endif
       // Cycle pointers to avoid copies
       Field *swizzle = Tnm;
       Tnm    =Tn;
@@ -321,7 +353,7 @@ struct ChebyParams : Serializable {
   // shift_Multiply in Rudy's code
   void AminusMuSq(LinearOperatorBase<Field> &Linop, const Field &in, Field &out) 
   {
-      GridBase *grid=in._grid;
+    GridBase *grid=in.Grid();
     Field tmp(grid);
 
     RealD aa= alpha*alpha;
@@ -338,7 +370,7 @@ struct ChebyParams : Serializable {
   // Implement the required interface
   void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 
-      GridBase *grid=in._grid;
+    GridBase *grid=in.Grid();
 
     int vol=grid->gSites();
 
@@ -373,5 +405,5 @@ struct ChebyParams : Serializable {
     }
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/approx/Forecast.h

@@ -31,7 +31,7 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #ifndef INCLUDED_FORECAST_H
 #define INCLUDED_FORECAST_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 // Abstract base class.
 // Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
@@ -57,10 +57,10 @@ namespace Grid {
     Field chi(phi); // forecasted solution
 
     // Trivial cases
-        if(degree == 0){ chi = zero; return chi; }
+    if(degree == 0){ chi = Zero(); return chi; }
     else if(degree == 1){ return prev_solns[0]; }
 
-        RealD dot;
+    //    RealD dot;
     ComplexD xp;
     Field r(phi); // residual
     Field Mv(phi);
@@ -92,7 +92,7 @@ namespace Grid {
     for(int j=0; j<degree; j++){
       for(int k=j+1; k<degree; k++){
 	G[j][k] = innerProduct(v[j],MdagMv[k]);
-          G[k][j] = std::conj(G[j][k]);
+	G[k][j] = conjugate(G[j][k]);
       }}
 
     // Gauss-Jordan elimination with partial pivoting
@@ -100,7 +100,7 @@ namespace Grid {
 
       // Perform partial pivoting
       int k = i;
-          for(int j=i+1; j<degree; j++){ if(std::abs(G[j][j]) > std::abs(G[k][k])){ k = j; } }
+      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];
@@ -121,7 +121,7 @@ namespace Grid {
     }
 
     // Use Gaussian elimination to solve equations and calculate initial guess
-        chi = zero;
+    chi = Zero();
     r = phi;
     for(int i=degree-1; i>=0; i--){
       a[i] = 0.0;
@@ -136,7 +136,7 @@ namespace Grid {
     for(int i=0; i<degree; i++){
       tmp = -b[i];
       for(int j=0; j<degree; j++){ tmp += G[i][j]*a[j]; }
-          tmp = std::conj(tmp)*tmp;
+      tmp = conjugate(tmp)*tmp;
       true_r += std::sqrt(tmp.real());
     }
 
@@ -147,6 +147,6 @@ namespace Grid {
   };
 };
 
-}
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/approx/JacobiPolynomial.h

@@ -0,0 +1,129 @@
+#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/MultiShiftFunction.cc

@@ -27,7 +27,8 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
     /*  END LEGAL */
 #include <Grid/GridCore.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
+
 double MultiShiftFunction::approx(double x)
 {
   double a = norm;
@@ -53,4 +54,4 @@ void MultiShiftFunction::csv(std::ostream &out)
   }
   return;
 }
-}
+NAMESPACE_END(Grid);

Grid/algorithms/approx/MultiShiftFunction.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef MULTI_SHIFT_FUNCTION
 #define MULTI_SHIFT_FUNCTION
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 class MultiShiftFunction {
 public:
@@ -63,5 +63,5 @@ public:
   }
 
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/approx/Remez.cc

@@ -298,7 +298,7 @@ void AlgRemez::stpini(bigfloat *step) {
 // Search for error maxima and minima
 void AlgRemez::search(bigfloat *step) {
   bigfloat a, q, xm, ym, xn, yn, xx0, xx1;
-  int i, j, meq, emsign, ensign, steps;
+  int i, meq, emsign, ensign, steps;
 
   meq = neq + 1;
   bigfloat *yy = new bigfloat[meq];
@@ -306,7 +306,6 @@ void AlgRemez::search(bigfloat *step) {
   bigfloat eclose = 1.0e30;
   bigfloat farther = 0l;
 
-  j = 1;
   xx0 = apstrt;
 
   for (i = 0; i < meq; i++) {

Grid/algorithms/approx/RemezGeneral.cc

@@ -0,0 +1,473 @@
+#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

@@ -0,0 +1,170 @@
+/*
+  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

@@ -0,0 +1,183 @@
+/*************************************************************************************
+
+    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

@@ -0,0 +1,57 @@
+/*************************************************************************************
+
+    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/approx/Zolotarev.cc

@@ -58,8 +58,8 @@
 
 /* Compute the partial fraction expansion coefficients (alpha) from the
  * factored form */
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
-namespace Approx {
+NAMESPACE_BEGIN(Approx);
 
 static void construct_partfrac(izd *z) {
   int dn = z -> dn, dd = z -> dd, type = z -> type;
@@ -516,7 +516,9 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
   free(d);
   return zd;
 }
-}}
+
+NAMESPACE_END(Approx);
+NAMESPACE_END(Grid);
 
 #ifdef TEST
 
@@ -585,6 +587,7 @@ static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) {
   return (ONE - T) / (ONE + T);
 }
 
+
 /* Test program. Apart from printing out the parameters for R(x) it produces
  * the following data files for plotting (unless NPLOT is defined):
  *
@@ -723,5 +726,5 @@ int main(int argc, char** argv) {
   return EXIT_SUCCESS;
 }
 
-
 #endif /* TEST */
+

Grid/algorithms/approx/Zolotarev.h

@@ -1,13 +1,13 @@
 /* -*- Mode: C; comment-column: 22; fill-column: 79; -*- */
 
 #ifdef __cplusplus
-namespace Grid {
+#include <Grid/Namespace.h>
-namespace Approx {
+NAMESPACE_BEGIN(Grid);
+NAMESPACE_BEGIN(Approx);
 #endif
 
 #define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY>
 
-
 #ifndef ZOLOTAREV_INTERNAL
 #ifndef PRECISION
 #define PRECISION double
@@ -83,5 +83,6 @@ void zolotarev_free(zolotarev_data *zdata);
 #endif
 
 #ifdef __cplusplus
-}}
+NAMESPACE_END(Approx);
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/approx/bigfloat.h

@@ -10,10 +10,12 @@
 #ifndef INCLUDED_BIGFLOAT_H
 #define INCLUDED_BIGFLOAT_H
 
-
+#define __GMP_WITHIN_CONFIGURE
 #include <gmp.h>
 #include <mpf2mpfr.h>
 #include <mpfr.h>
+#undef  __GMP_WITHIN_CONFIGURE
+
 class bigfloat {
 
 private:

Grid/algorithms/approx/bigfloat_double.h

@@ -25,6 +25,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+
+#ifndef INCLUDED_BIGFLOAT_DOUBLE_H
+#define INCLUDED_BIGFLOAT_DOUBLE_H
+
 #include <math.h>
 
 typedef double mfloat; 
@@ -186,4 +190,6 @@ public:
   //  friend bigfloat& random(void);
 };
 
+#endif
+
 

Grid/algorithms/iterative/AdefGeneric.h

@@ -90,8 +90,8 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
   void operator() (const Field &src, Field &psi){
   void operator() (const Field &src, Field &psi){
 
-    psi.checkerboard = src.checkerboard;
+    psi.Checkerboard() = src.Checkerboard();
-    grid             = src._grid;
+    grid             = src.Grid();
 
     RealD f;
     RealD rtzp,rtz,a,d,b;

Grid/algorithms/iterative/BiCGSTAB.h

@@ -0,0 +1,234 @@
+/*************************************************************************************
+
+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

@@ -0,0 +1,158 @@
+/*************************************************************************************
+
+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:                                                
+    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

@@ -27,11 +27,9 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_BLOCK_CONJUGATE_GRADIENT_H
+#pragma once
-#define GRID_BLOCK_CONJUGATE_GRADIENT_H
 
-
+NAMESPACE_BEGIN(Grid);
-namespace Grid {
 
 enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
 
@@ -54,6 +52,7 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
   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)
@@ -154,12 +153,12 @@ virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Fiel
 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];
+  Nblock = B.Grid()->_fdimensions[Orthog];
 /* FAKE */
   Nblock=8;
   std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
 
-  X.checkerboard = B.checkerboard;
+  X.Checkerboard() = B.Checkerboard();
   conformable(X, B);
 
   Field tmp(B);
@@ -308,7 +307,8 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
 
       Linop.HermOp(X, AD);
       AD = AD-B;
-      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AD)/norm2(B)) <<std::endl;
+      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;
@@ -334,11 +334,11 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
 void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
 {
   int Orthog = blockDim; // First dimension is block dim
-  Nblock = Src._grid->_fdimensions[Orthog];
+  Nblock = Src.Grid()->_fdimensions[Orthog];
 
   std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
 
-  Psi.checkerboard = Src.checkerboard;
+  Psi.Checkerboard() = Src.Checkerboard();
   conformable(Psi, Src);
 
   Field P(Src);
@@ -444,7 +444,8 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
 
       Linop.HermOp(Psi, AP);
       AP = AP-Src;
-      std::cout <<GridLogMessage << "\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
+      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;
@@ -478,7 +479,7 @@ void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<
   for(int b=0;b<Nblock;b++){
     tmp[b]   = Y[b];
     for(int bp=0;bp<Nblock;bp++) {
-      tmp[b] = tmp[b] + (scale*m(bp,b))*X[bp]; 
+      tmp[b] = tmp[b] + scomplex(scale*m(bp,b))*X[bp]; 
     }
   }
   for(int b=0;b<Nblock;b++){
@@ -488,9 +489,9 @@ void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<
 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;
+    AP[b] = Zero();
     for(int bp=0;bp<Nblock;bp++) {
-      AP[b] += (m(bp,b))*X[bp]; 
+      AP[b] += scomplex(m(bp,b))*X[bp]; 
     }
   }
 }
@@ -517,7 +518,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   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;
+    X[b].Checkerboard() = B[b].Checkerboard();
     conformable(X[b], B[b]);
     conformable(X[b], X[0]); 
   }
@@ -655,7 +656,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
       if ( rr > max_resid ) max_resid = rr;
     }
 
-    std::cout << GridLogIterative << "\t Block Iteration "<<k<<" ave resid "<< sqrt(rrsum/sssum) << " max "<< sqrt(max_resid) <<std::endl;
+    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 ) { 
 
@@ -670,7 +671,8 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
 
       for(int b=0;b<Nblock;b++) Linop.HermOp(X[b], AD[b]);
       for(int b=0;b<Nblock;b++) AD[b] = AD[b]-B[b];
-      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(normv(AD)/normv(B)) <<std::endl;
+      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;
@@ -690,9 +692,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   IterationsToComplete = k;
 }
 
-
-
 };
 
-}
+NAMESPACE_END(Grid);
-#endif
+

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,248 @@
+/*************************************************************************************
+
+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

@@ -31,7 +31,7 @@ directory
 #ifndef GRID_CONJUGATE_GRADIENT_H
 #define GRID_CONJUGATE_GRADIENT_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -41,11 +41,15 @@ namespace Grid {
 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),
@@ -54,11 +58,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
+    psi.Checkerboard() = src.Checkerboard();
 
-    psi.checkerboard = src.checkerboard;
     conformable(psi, src);
 
-    RealD cp, c, a, d, b, ssq, qq, b_pred;
+    RealD cp, c, a, d, b, ssq, qq;
+    //RealD b_pred;
 
     Field p(src);
     Field mmp(src);
@@ -68,7 +73,6 @@ class ConjugateGradient : public OperatorFunction<Field> {
     RealD guess = norm2(psi);
     assert(std::isnan(guess) == 0);
     
-    
     Linop.HermOpAndNorm(psi, mmp, d, b);
     
     r = src - mmp;
@@ -78,6 +82,14 @@ class ConjugateGradient : public OperatorFunction<Field> {
     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;
@@ -89,6 +101,9 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
     // 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;
     }
 
@@ -104,7 +119,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
     SolverTimer.Start();
     int k;
-    for (k = 1; k <= MaxIterations*1000; k++) {
+    for (k = 1; k <= MaxIterations; k++) {
       c = cp;
 
       MatrixTimer.Start();
@@ -125,15 +140,20 @@ class ConjugateGradient : public OperatorFunction<Field> {
       b = cp / c;
 
       LinearCombTimer.Start();
-      parallel_for(int ss=0;ss<src._grid->oSites();ss++){
+      {
-	vstream(psi[ss], a      *  p[ss] + psi[ss]);
+	autoView( psi_v , psi, AcceleratorWrite);
-	vstream(p  [ss], b      *  p[ss] + r[ss]);
+	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();
 
       std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
-                << " residual^2 " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
+                << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -141,37 +161,43 @@ class ConjugateGradient : public OperatorFunction<Field> {
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
 
-        RealD srcnorm = sqrt(norm2(src));
+        RealD srcnorm = std::sqrt(norm2(src));
-        RealD resnorm = sqrt(norm2(p));
+        RealD resnorm = std::sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
 
-        std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k << std::endl;
+        std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k 
-        std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+		  << "\tComputed residual " << std::sqrt(cp / ssq)
-	std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
+		  << "\tTrue residual " << true_residual
-	std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
+		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
+        std::cout << GridLogIterative << "Time breakdown "<<std::endl;
-	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 
 	IterationsToComplete = k;	
+	TrueResidual = true_residual;
 
         return;
       }
     }
-    std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
+    // Failed. Calculate true residual before giving up                                                         
-              << std::endl;
+    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

@@ -28,10 +28,12 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
 #ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
 #define GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
 
-namespace Grid {
+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> 
+  template<class FieldD,class FieldF, 
+    typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+    typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
   public:                                                
     RealD   Tolerance;
@@ -50,7 +52,12 @@ namespace Grid {
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
     
-    MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) :
+    MixedPrecisionConjugateGradient(RealD tol, 
+				    Integer maxinnerit, 
+				    Integer maxouterit, 
+				    GridBase* _sp_grid, 
+				    LinearOperatorBase<FieldF> &_Linop_f, 
+				    LinearOperatorBase<FieldD> &_Linop_d) :
       Linop_f(_Linop_f), Linop_d(_Linop_d),
       Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
       OuterLoopNormMult(100.), guesser(NULL){ };
@@ -65,18 +72,18 @@ namespace Grid {
     GridStopWatch TotalTimer;
     TotalTimer.Start();
     
-      int cb = src_d_in.checkerboard;
+    int cb = src_d_in.Checkerboard();
-      sol_d.checkerboard = cb;
+    sol_d.Checkerboard() = cb;
     
     RealD src_norm = norm2(src_d_in);
     RealD stop = src_norm * Tolerance*Tolerance;
 
-      GridBase* DoublePrecGrid = src_d_in._grid;
+    GridBase* DoublePrecGrid = src_d_in.Grid();
     FieldD tmp_d(DoublePrecGrid);
-      tmp_d.checkerboard = cb;
+    tmp_d.Checkerboard() = cb;
     
     FieldD tmp2_d(DoublePrecGrid);
-      tmp2_d.checkerboard = cb;
+    tmp2_d.Checkerboard() = cb;
     
     FieldD src_d(DoublePrecGrid);
     src_d = src_d_in; //source for next inner iteration, computed from residual during operation
@@ -84,10 +91,10 @@ namespace Grid {
     RealD inner_tol = InnerTolerance;
     
     FieldF src_f(SinglePrecGrid);
-      src_f.checkerboard = cb;
+    src_f.Checkerboard() = cb;
     
     FieldF sol_f(SinglePrecGrid);
-      sol_f.checkerboard = cb;
+    sol_f.Checkerboard() = cb;
     
     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
     CG_f.ErrorOnNoConverge = false;
@@ -115,7 +122,7 @@ namespace Grid {
       precisionChange(src_f, src_d);
       PrecChangeTimer.Stop();
       
-	zeroit(sol_f);
+      sol_f = Zero();
 
       //Optionally improve inner solver guess (eg using known eigenvectors)
       if(guesser != NULL)
@@ -149,6 +156,6 @@ namespace Grid {
   }
 };
 
-}
+NAMESPACE_END(Grid);
 
 #endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
 #define GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -41,22 +41,29 @@ namespace Grid {
 				    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,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;
+    GridBase *grid = src.Grid();
     int nshift = shifts.order;
     std::vector<Field> results(nshift,grid);
     (*this)(Linop,src,results,psi);
@@ -78,7 +85,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
   
-  GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
   
     ////////////////////////////////////////////////////////////////////////
     // Convenience references to the info stored in "MultiShiftFunction"
@@ -122,6 +129,17 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     // 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
@@ -267,6 +285,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
       for(int s=0;s<nshift;s++){
       
 	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
 	
 	  RealD css  = c * z[s][iz]* z[s][iz];
 	
@@ -296,7 +315,8 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	  axpy(r,-alpha[s],src,tmp);
 	  RealD rn = norm2(r);
 	  RealD cn = norm2(src);
-	std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
+	  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;
@@ -318,5 +338,5 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   }
 
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -28,9 +28,11 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
 #ifndef GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
 #define GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
 
-namespace Grid {
+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> 
+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.
@@ -74,7 +76,7 @@ namespace Grid {
     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
     bool using_fallback = false;
       
-      psi.checkerboard = src.checkerboard;
+    psi.Checkerboard() = src.Checkerboard();
     conformable(psi, src);
 
     RealD cp, c, a, d, b, ssq, qq, b_pred;
@@ -108,17 +110,17 @@ namespace Grid {
     // Check if guess is really REALLY good :)
     if (cp <= rsq) {
       std::cout << GridLogMessage << "ConjugateGradientReliableUpdate guess was REALLY good\n";
-	std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+      std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
       return;
     }
 
     //Single prec initialization
     FieldF r_f(SinglePrecGrid);
-      r_f.checkerboard = r.checkerboard;
+    r_f.Checkerboard() = r.Checkerboard();
     precisionChange(r_f, r);
 
     FieldF psi_f(r_f);
-      psi_f = zero;
+    psi_f = Zero();
 
     FieldF p_f(r_f);
     FieldF mmp_f(r_f);
@@ -178,12 +180,12 @@ namespace Grid {
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
 	p = mmp - src;
 
-	  RealD srcnorm = sqrt(norm2(src));
+	RealD srcnorm = std::sqrt(norm2(src));
-	  RealD resnorm = sqrt(norm2(p));
+	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 " << sqrt(cp / ssq)<<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;
 
@@ -217,7 +219,7 @@ namespace Grid {
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
 	r = src - mmp;
 
-	  psi_f = zero;
+	psi_f = Zero();
 	precisionChange(r_f, r);
 	cp = norm2(r);
 	MaxResidSinceLastRelUp = cp;
@@ -249,7 +251,7 @@ namespace Grid {
 };
 
 
-};
+NAMESPACE_END(Grid);
 
 
 

Grid/algorithms/iterative/ConjugateResidual.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CONJUGATE_RESIDUAL_H
 #define GRID_CONJUGATE_RESIDUAL_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -39,6 +39,8 @@ namespace Grid {
 template<class Field> 
 class ConjugateResidual : public OperatorFunction<Field> {
 public:                                                
+  using OperatorFunction<Field>::operator();
+
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;
@@ -49,14 +51,14 @@ namespace Grid {
 
   void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
 
-      RealD a, b, c, d;
+    RealD a, b; // c, d;
     RealD cp, ssq,rsq;
       
     RealD rAr, rAAr, rArp;
     RealD pAp, pAAp;
 
-      GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
-      psi=zero;
+    psi=Zero();
     Field r(grid),  p(grid), Ap(grid), Ar(grid);
       
     r=src;
@@ -95,8 +97,8 @@ namespace Grid {
 	axpy(r,-1.0,src,Ap);
 	RealD true_resid = norm2(r)/ssq;
 	std::cout<<GridLogMessage<<"ConjugateResidual: Converged on iteration " <<k
-		   << " computed residual "<<sqrt(cp/ssq)
+		 << " computed residual "<<std::sqrt(cp/ssq)
-	           << " true residual "<<sqrt(true_resid)
+		 << " true residual "<<std::sqrt(true_resid)
 		 << " target "       <<Tolerance <<std::endl;
 	return;
       }
@@ -107,5 +109,5 @@ namespace Grid {
     assert(0);
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/Deflation.h

@@ -33,9 +33,13 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
-  virtual void operator()(const Field &src, Field &guess) { guess = zero; };
+    virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
+};
+template<class Field>
+class DoNothingGuesser: public LinearFunction<Field> {
+public:
+  virtual void operator()(const Field &src, Field &guess) {  };
 };
-
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
@@ -56,14 +60,14 @@ public:
   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
 
   virtual void operator()(const Field &src,Field &guess) {
-    guess = zero;
+    guess = Zero();
     assert(evec.size()==eval.size());
     auto N = evec.size();
     for (int i=0;i<N;i++) {
       const Field& tmp = evec[i];
       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
     }
-    guess.checkerboard = src.checkerboard;
+    guess.Checkerboard() = src.Checkerboard();
   }
 };
 
@@ -86,15 +90,15 @@ public:
   
   void operator()(const FineField &src,FineField &guess) { 
     int N = (int)evec_coarse.size();
-    CoarseField src_coarse(evec_coarse[0]._grid);
+    CoarseField src_coarse(evec_coarse[0].Grid());
-    CoarseField guess_coarse(evec_coarse[0]._grid);    guess_coarse = zero;
+    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;
+    guess.Checkerboard() = src.Checkerboard();
   };
 };
 

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,258 @@
+/*************************************************************************************
+
+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

@@ -0,0 +1,256 @@
+/*************************************************************************************
+
+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

@@ -0,0 +1,244 @@
+/*************************************************************************************
+
+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

@@ -35,120 +35,7 @@ Author: Christoph Lehner <clehner@bnl.gov>
 //#include <zlib.h>
 #include <sys/stat.h>
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
-
-  ////////////////////////////////////////////////////////
-  // Move following 100 LOC to lattice/Lattice_basis.h
-  ////////////////////////////////////////////////////////
-template<class Field>
-void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k) 
-{
-  for(int j=0; j<k; ++j){
-    auto ip = innerProduct(basis[j],w);
-    w = w - ip*basis[j];
-  }
-}
-
-template<class Field>
-void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm) 
-{
-  typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0]._grid;
-      
-  parallel_region
-  {
-
-    std::vector < vobj , commAllocator<vobj> > B(Nm); // Thread private
-       
-    parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
-      for(int j=j0; j<j1; ++j) B[j]=0.;
-      
-      for(int j=j0; j<j1; ++j){
-	for(int k=k0; k<k1; ++k){
-	  B[j] +=Qt(j,k) * basis[k]._odata[ss];
-	}
-      }
-      for(int j=j0; j<j1; ++j){
-	  basis[j]._odata[ss] = B[j];
-      }
-    }
-  }
-}
-
-// Extract a single rotated vector
-template<class Field>
-void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
-{
-  typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0]._grid;
-
-  result.checkerboard = basis[0].checkerboard;
-  parallel_for(int ss=0;ss < grid->oSites();ss++){
-    vobj B = zero;
-    for(int k=k0; k<k1; ++k){
-      B +=Qt(j,k) * basis[k]._odata[ss];
-    }
-    result._odata[ss] = B;
-  }
-}
-
-template<class Field>
-void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx) 
-{
-  int vlen = idx.size();
-
-  assert(vlen>=1);
-  assert(vlen<=sort_vals.size());
-  assert(vlen<=_v.size());
-
-  for (size_t i=0;i<vlen;i++) {
-
-    if (idx[i] != i) {
-
-      //////////////////////////////////////
-      // idx[i] is a table of desired sources giving a permutation.
-      // Swap v[i] with v[idx[i]].
-      // Find  j>i for which _vnew[j] = _vold[i],
-      // track the move idx[j] => idx[i]
-      // track the move idx[i] => i
-      //////////////////////////////////////
-      size_t j;
-      for (j=i;j<idx.size();j++)
-	if (idx[j]==i)
-	  break;
-
-      assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
-
-      std::swap(_v[i]._odata,_v[idx[i]]._odata); // should use vector move constructor, no data copy
-      std::swap(sort_vals[i],sort_vals[idx[i]]);
-
-      idx[j] = idx[i];
-      idx[i] = i;
-    }
-  }
-}
-
-inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals) 
-{
-  std::vector<int> idx(sort_vals.size());
-  std::iota(idx.begin(), idx.end(), 0);
-
-  // sort indexes based on comparing values in v
-  std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
-    return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
-  });
-  return idx;
-}
-
-template<class Field>
-void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse) 
-{
-  std::vector<int> idx = basisSortGetIndex(sort_vals);
-  if (reverse)
-    std::reverse(idx.begin(), idx.end());
-  
-  basisReorderInPlace(_v,sort_vals,idx);
-}
 
 /////////////////////////////////////////////////////////////
 // Implicitly restarted lanczos
@@ -259,7 +146,7 @@ public:
 			    RealD _eresid, // resid in lmdue deficit 
 			    int _MaxIter, // Max iterations
 			    RealD _betastp=0.0, // if beta(k) < betastp: converged
-			    int _MinRestart=1, int _orth_period = 1,
+			    int _MinRestart=0, int _orth_period = 1,
 			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
     SimpleTester(HermOp), _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(Tester),
     Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
@@ -275,7 +162,7 @@ public:
 			       RealD _eresid, // resid in lmdue deficit 
 			       int _MaxIter, // Max iterations
 			       RealD _betastp=0.0, // if beta(k) < betastp: converged
-			       int _MinRestart=1, int _orth_period = 1,
+			       int _MinRestart=0, int _orth_period = 1,
 			       IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
     SimpleTester(HermOp),  _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(SimpleTester),
     Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
@@ -289,7 +176,7 @@ public:
   template<typename T>  static RealD normalise(T& v) 
   {
     RealD nn = norm2(v);
-    nn = sqrt(nn);
+    nn = std::sqrt(nn);
     v = v * (1.0/nn);
     return nn;
   }
@@ -321,10 +208,10 @@ until convergence
 */
   void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
   {
-    GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
-    assert(grid == evec[0]._grid);
+    assert(grid == evec[0].Grid());
     
-    GridLogIRL.TimingMode(1);
+    //    GridLogIRL.TimingMode(1);
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
     std::cout << GridLogIRL <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
@@ -349,14 +236,17 @@ until convergence
     {
       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) < 0.05)
+	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;
@@ -446,7 +336,7 @@ until convergence
       assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
 
       basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
-      std::cout<<GridLogIRL <<"basisRotated  by Qt"<<std::endl;
+      std::cout<<GridLogIRL <<"basisRotated  by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
       
       ////////////////////////////////////////////////////
       // Compressed vector f and beta(k2)
@@ -454,7 +344,7 @@ until convergence
       f *= Qt(k2-1,Nm-1);
       f += lme[k2-1] * evec[k2];
       beta_k = norm2(f);
-      beta_k = sqrt(beta_k);
+      beta_k = std::sqrt(beta_k);
       std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
 	  
       RealD betar = 1.0/beta_k;
@@ -477,7 +367,7 @@ until convergence
 
 	std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
 
-	Field B(grid); B.checkerboard = evec[0].checkerboard;
+	Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
 
 	//  power of two search pattern;  not every evalue in eval2 is assessed.
 	int allconv =1;
@@ -515,7 +405,7 @@ until convergence
 	
   converged:
     {
-      Field B(grid); B.checkerboard = evec[0].checkerboard;
+      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;
@@ -554,11 +444,11 @@ until convergence
 /* Saad PP. 195
 1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
 2. For k = 1,2,...,m Do:
-3. wk:=Avk−βkv_{k−1}      
+3. wk:=Avk - b_k v_{k-1}      
-4. αk:=(wk,vk)       // 
+4. ak:=(wk,vk)       // 
-5. wk:=wk−αkvk       // wk orthog vk 
+5. wk:=wk-akvk       // wk orthog vk 
-6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+6. bk+1 := ||wk||_2. If b_k+1 = 0 then Stop
-7. vk+1 := wk/βk+1
+7. vk+1 := wk/b_k+1
 8. EndDo
  */
   void step(std::vector<RealD>& lmd,
@@ -566,6 +456,7 @@ until convergence
 	    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 );
 
@@ -577,20 +468,20 @@ until convergence
 
     if(k>0) w -= lme[k-1] * evec[k-1];
 
-    ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
+    ComplexD zalph = innerProduct(evec_k,w);
     RealD     alph = real(zalph);
 
-    w = w - alph * evec_k;// 5. wk:=wk−αkvk
+    w = w - alph * evec_k;
 
-    RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+    RealD beta = normalise(w); 
-    // 7. vk+1 := wk/βk+1
 
     lmd[k] = alph;
     lme[k] = beta;
 
-    if (k>0 && k % orth_period == 0) {
+    if ( (k>0) && ( (k % orth_period) == 0 )) {
+      std::cout<<GridLogIRL << "Orthogonalising " <<k<<std::endl;
       orthogonalize(w,evec,k); // orthonormalise
-      std::cout<<GridLogIRL << "Orthogonalised " <<std::endl;
+      std::cout<<GridLogIRL << "Orthogonalised " <<k<<std::endl;
     }
 
     if(k < Nm-1) evec[k+1] = w;
@@ -598,6 +489,8 @@ until convergence
     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, 
@@ -807,7 +700,7 @@ void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
     
     // determination of 2x2 leading submatrix
     RealD dsub = lmd[kmax-1]-lmd[kmax-2];
-    RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
+    RealD 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)
     
@@ -838,5 +731,6 @@ void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
   abort();
 }
 };
-}
+
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -29,8 +29,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LOCAL_COHERENCE_IRL_H
 #define GRID_LOCAL_COHERENCE_IRL_H
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
-
 
 struct LanczosParams : Serializable {
 public:
@@ -59,7 +58,7 @@ struct LocalCoherenceLanczosParams : Serializable {
 				  RealD        , coarse_relax_tol,
 				  std::vector<int>, blockSize,
 				  std::string, config,
-				  std::vector < std::complex<double>  >, omega,
+				  std::vector < ComplexD  >, omega,
 				  RealD, mass,
 				  RealD, M5);
 };
@@ -83,11 +82,11 @@ public:
   };
 
   void operator()(const CoarseField& in, CoarseField& out) {
-    GridBase *FineGrid = subspace[0]._grid;    
+    GridBase *FineGrid = subspace[0].Grid();    
-    int   checkerboard = subspace[0].checkerboard;
+    int   checkerboard = subspace[0].Checkerboard();
 
-    FineField fin (FineGrid);     fin.checkerboard= checkerboard;
+    FineField fin (FineGrid);     fin.Checkerboard()= checkerboard;
-    FineField fout(FineGrid);   fout.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;
@@ -118,11 +117,11 @@ public:
 
   void operator()(const CoarseField& in, CoarseField& out) {
 
-    GridBase *FineGrid = subspace[0]._grid;    
+    GridBase *FineGrid = subspace[0].Grid();    
-    int   checkerboard = subspace[0].checkerboard;
+    int   checkerboard = subspace[0].Checkerboard();
 
-    FineField fin (FineGrid); fin.checkerboard =checkerboard;
+    FineField fin (FineGrid); fin.Checkerboard() =checkerboard;
-    FineField fout(FineGrid);fout.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;
@@ -182,10 +181,10 @@ class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanc
   }
   int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
-    GridBase *FineGrid = _subspace[0]._grid;    
+    GridBase *FineGrid = _subspace[0].Grid();    
-    int checkerboard   = _subspace[0].checkerboard;
+    int checkerboard   = _subspace[0].Checkerboard();
-    FineField fB(FineGrid);fB.checkerboard =checkerboard;
+    FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
-    FineField fv(FineGrid);fv.checkerboard =checkerboard;
+    FineField fv(FineGrid);fv.Checkerboard() =checkerboard;
 
     blockPromote(B,fv,_subspace);  
     
@@ -305,11 +304,11 @@ public:
     int Nk = nbasis;
     subspace.resize(Nk,_FineGrid);
     subspace[0]=1.0;
-    subspace[0].checkerboard=_checkerboard;
+    subspace[0].Checkerboard()=_checkerboard;
     normalise(subspace[0]);
     PlainHermOp<FineField>    Op(_FineOp);
     for(int k=1;k<Nk;k++){
-      subspace[k].checkerboard=_checkerboard;
+      subspace[k].Checkerboard()=_checkerboard;
       Op(subspace[k-1],subspace[k]);
       normalise(subspace[k]);
     }
@@ -360,7 +359,11 @@ public:
 
     ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
 
-    FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
+    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);
@@ -402,5 +405,5 @@ public:
   }
 };
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/MinimalResidual.h

@@ -0,0 +1,157 @@
+/*************************************************************************************
+
+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

@@ -0,0 +1,276 @@
+/*************************************************************************************
+
+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

@@ -28,33 +28,85 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_NORMAL_EQUATIONS_H
 #define GRID_NORMAL_EQUATIONS_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
 // Take a matrix and form an NE solver calling a Herm solver
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field> class NormalEquations : public OperatorFunction<Field>{
+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) 
+ NormalEquations(SparseMatrixBase<Field> &Matrix, OperatorFunction<Field> &HermitianSolver,
-    :  _Matrix(Matrix), _HermitianSolver(HermitianSolver) {}; 
+		 LinearFunction<Field> &Guess) 
+   :  _Matrix(Matrix), _HermitianSolver(HermitianSolver), _Guess(Guess) {}; 
 
   void operator() (const Field &in, Field &out){
  
-      Field src(in._grid);
+    Field src(in.Grid());
+    Field tmp(in.Grid());
 
+    MdagMLinearOperator<SparseMatrixBase<Field>,Field> MdagMOp(_Matrix);
     _Matrix.Mdag(in,src);
-      _HermitianSolver(src,out);  // Mdag M out = Mdag in
+    _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

@@ -0,0 +1,45 @@
+#pragma once
+namespace Grid {
+template<class Field> class PowerMethod  
+{ 
+ public: 
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  RealD operator()(LinearOperatorBase<Field> &HermOp, const Field &src) 
+  { 
+    GridBase *grid = src.Grid(); 
+    
+    // quickly get an idea of the largest eigenvalue to more properly normalize the residuum 
+    RealD evalMaxApprox = 0.0; 
+    auto src_n = src; 
+    auto tmp = src; 
+    const int _MAX_ITER_EST_ = 50; 
+
+    for (int i=0;i<_MAX_ITER_EST_;i++) { 
+      
+      normalise(src_n); 
+      HermOp.HermOp(src_n,tmp); 
+      RealD vnum = real(innerProduct(src_n,tmp)); // HermOp. 
+      RealD vden = norm2(src_n); 
+      RealD na = vnum/vden; 
+      
+      if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (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

@@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PREC_CONJUGATE_RESIDUAL_H
 #define GRID_PREC_CONJUGATE_RESIDUAL_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////////////
 // Base classes for iterative processes based on operators
@@ -56,7 +56,7 @@ namespace Grid {
     RealD rAr, rAAr, rArp;
     RealD pAp, pAAp;
 
-      GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
     Field r(grid),  p(grid), Ap(grid), Ar(grid), z(grid);
       
     psi=zero;
@@ -115,5 +115,5 @@ namespace Grid {
     assert(0);
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -36,41 +36,50 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 //NB. Likely not original reference since they are focussing on a preconditioner variant.
 //    but VPGCR was nicely written up in their paper
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
+
+#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" " 
 
 template<class Field>
-    class PrecGeneralisedConjugateResidual : public OperatorFunction<Field> {
+class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
 public:                                                
+
   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;
 
-   PrecGeneralisedConjugateResidual(RealD tol,Integer maxit,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
+  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() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
+  void operator() (const Field &src, Field &psi){
 
-      psi=zero;
+    psi=Zero();
     RealD cp, ssq,rsq;
     ssq=norm2(src);
     rsq=Tolerance*Tolerance*ssq;
       
-      Field r(src._grid);
+    Field r(src.Grid());
 
     PrecTimer.Reset();
     MatTimer.Reset();
@@ -82,9 +91,9 @@ namespace Grid {
     steps=0;
     for(int k=0;k<MaxIterations;k++){
 
-	cp=GCRnStep(Linop,src,psi,rsq);
+      cp=GCRnStep(src,psi,rsq);
 
-	std::cout<<GridLogMessage<<"VPGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<std::endl;
+      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
 
       if(cp<rsq) {
 
@@ -93,31 +102,33 @@ namespace Grid {
 	Linop.HermOp(psi,r);
 	axpy(r,-1.0,src,r);
 	RealD tr = norm2(r);
-	  std::cout<<GridLogMessage<<"PrecGeneralisedConjugateResidual: Converged on iteration " <<steps
+	GCRLogLevel<<"PGCR: Converged on iteration " <<steps
 		 << " computed residual "<<sqrt(cp/ssq)
 		 << " true residual "    <<sqrt(tr/ssq)
 		 << " target "           <<Tolerance <<std::endl;
 
-	  std::cout<<GridLogMessage<<"VPGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
+	GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
-	  std::cout<<GridLogMessage<<"VPGCR Time elapsed: Precon "<<   PrecTimer.Elapsed() <<std::endl;
+	/*
-	  std::cout<<GridLogMessage<<"VPGCR Time elapsed: Matrix "<<    MatTimer.Elapsed() <<std::endl;
+	  GCRLogLevel<<"PGCR Time elapsed: Precon "<<   PrecTimer.Elapsed() <<std::endl;
-	  std::cout<<GridLogMessage<<"VPGCR Time elapsed: Linalg "<< LinalgTimer.Elapsed() <<std::endl;
+	  GCRLogLevel<<"PGCR Time elapsed: Matrix "<<    MatTimer.Elapsed() <<std::endl;
+	  GCRLogLevel<<"PGCR Time elapsed: Linalg "<< LinalgTimer.Elapsed() <<std::endl;
+	*/
 	return;
       }
 
     }
-      std::cout<<GridLogMessage<<"Variable Preconditioned GCR did not converge"<<std::endl;
+    GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-      assert(0);
+    //    assert(0);
   }
 
-    RealD GCRnStep(LinearOperatorBase<Field> &Linop,const Field &src, Field &psi,RealD rsq){
+  RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
 
     RealD cp;
-      RealD a, b, c, d;
+    RealD a, b;
     RealD zAz, zAAz;
-      RealD rAq, rq;
+    RealD rq;
 
-      GridBase *grid = src._grid;
+    GridBase *grid = src.Grid();
 
     Field r(grid);
     Field z(grid);
@@ -132,6 +143,8 @@ namespace 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
@@ -139,39 +152,34 @@ namespace Grid {
     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.HermOp(z,tmp); 
-      MatTimer.Stop();
-
-      ttmp=tmp;
-      tmp=tmp-r;
-
-      /*
-      std::cout<<GridLogMessage<<r<<std::endl;
-      std::cout<<GridLogMessage<<z<<std::endl;
-      std::cout<<GridLogMessage<<ttmp<<std::endl;
-      std::cout<<GridLogMessage<<tmp<<std::endl;
-      */
-
     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++){
 
@@ -181,18 +189,21 @@ namespace Grid {
       int peri_k = k %mmax;
       int peri_kp= kp%mmax;
 
+      LinalgTimer.Start();
       rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
       a = rq/qq[peri_k];
 
       axpy(psi,a,p[peri_k],psi);         
 
       cp = axpy_norm(r,-a,q[peri_k],r);
+      LinalgTimer.Stop();
+
+      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
 
       if((k==nstep-1)||(cp<rsq)){
 	return cp;
       }
 
-	std::cout<<GridLogMessage<< " VPGCR_step["<<steps<<"]  resid " <<sqrt(cp/rsq)<<std::endl; 
 
       PrecTimer.Start();
       Preconditioner(r,z);// solve Az = r
@@ -200,10 +211,9 @@ namespace Grid {
 
       MatTimer.Start();
       Linop.HermOpAndNorm(z,Az,zAz,zAAz);
-	Linop.HermOp(z,tmp);
       MatTimer.Stop();
-        tmp=tmp-r;
+
-	std::cout<<GridLogMessage<< " Preconditioner resid " <<sqrt(norm2(tmp)/norm2(r))<<std::endl; 
+      LinalgTimer.Start();
 
       q[peri_kp]=Az;
       p[peri_kp]=z;
@@ -219,12 +229,11 @@ namespace Grid {
 
       }
       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

@@ -0,0 +1,241 @@
+/*************************************************************************************
+
+    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:                                                
+
+  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, 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

@@ -0,0 +1,371 @@
+/*************************************************************************************
+
+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

@@ -99,10 +99,13 @@ namespace Grid {
     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)  :
+    SchurRedBlackBase(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-    _HermitianRBSolver(HermitianRBSolver) 
+        const bool _solnAsInitGuess = false)  :
+    _HermitianRBSolver(HermitianRBSolver),
+    useSolnAsInitGuess(_solnAsInitGuess)
     { 
       CBfactorise = 0;
       subtractGuess(initSubGuess);
@@ -156,7 +159,11 @@ namespace Grid {
       if ( subGuess ) guess_save.resize(nblock,grid);
 
       for(int b=0;b<nblock;b++){
+        if(useSolnAsInitGuess) {
+          pickCheckerboard(Odd, sol_o[b], out[b]);
+        } else {
           guess(src_o[b],sol_o[b]); 
+        }
 
 	if ( subGuess ) { 
 	  guess_save[b] = sol_o[b];
@@ -216,8 +223,11 @@ namespace Grid {
       ////////////////////////////////
       // Construct the guess
       ////////////////////////////////
-      Field   tmp(grid);
+      if(useSolnAsInitGuess) {
+        pickCheckerboard(Odd, sol_o, out);
+      } else {
         guess(src_o,sol_o);
+      }
 
       Field  guess_save(grid);
       guess_save = sol_o;
@@ -251,7 +261,7 @@ namespace Grid {
     }     
     
     /////////////////////////////////////////////////////////////
-    // Override in derived. Not virtual as template methods
+    // 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;
@@ -264,8 +274,9 @@ namespace Grid {
   public:
     typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
 
-    SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false) 
+    SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-      :    SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess) 
+        const bool _solnAsInitGuess = false) 
+      :    SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) 
     {
     }
 
@@ -286,9 +297,9 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.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.
     }
@@ -306,17 +317,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.Checkerboard()   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      src_e = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_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_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.checkerboard ==Odd );
+      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);
+      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)
     {
@@ -333,8 +344,9 @@ namespace Grid {
   public:
     typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
 
-    SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false)  
+    SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-      : SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess) {};
+        const bool _solnAsInitGuess = false)  
+      : SchurRedBlackBase<Field> (HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
 
 
     //////////////////////////////////////////////////////
@@ -354,13 +366,13 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.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);       
+      _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)
@@ -374,17 +386,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.checkerboard   ==Even);
+      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.Checkerboard()   ==Even);
-      src_e_i = src_e-tmp;               assert(  src_e_i.checkerboard ==Even);
+      src_e_i = src_e-tmp;               assert(  src_e_i.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,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 );
+      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);
+      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)
     {
@@ -393,6 +405,70 @@ namespace Grid {
     }
   };
 
+  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
@@ -405,8 +481,9 @@ namespace Grid {
     /////////////////////////////////////////////////////
     // Wrap the usual normal equations Schur trick
     /////////////////////////////////////////////////////
-  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false)  
+  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
-    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess) {};
+      const bool _solnAsInitGuess = false)  
+    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
 
     virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
     {
@@ -424,12 +501,12 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
+      _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)
@@ -450,12 +527,12 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.checkerboard   ==Even);
+      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      tmp = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e);   assert(  sol_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_e);    assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.checkerboard ==Odd );
+      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.Checkerboard() ==Odd );
     };
 
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
@@ -469,5 +546,76 @@ namespace Grid {
       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

@@ -26,107 +26,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_ALIGNED_ALLOCATOR_H
+#pragma once
-#define GRID_ALIGNED_ALLOCATOR_H
 
-#ifdef HAVE_MALLOC_MALLOC_H
+NAMESPACE_BEGIN(Grid);
-#include <malloc/malloc.h>
-#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_MM_MALLOC_H
-#include <mm_malloc.h>
-#endif
-
-namespace Grid {
-
-  class PointerCache {
-  private:
-
-    static const int Ncache=8;
-    static int victim;
-
-    typedef struct { 
-      void *address;
-      size_t bytes;
-      int valid;
-    } PointerCacheEntry;
-    
-    static PointerCacheEntry Entries[Ncache];
-
-  public:
-
-
-    static void *Insert(void *ptr,size_t bytes) ;
-    static void *Lookup(size_t bytes) ;
-
-  };
-  
-  std::string sizeString(size_t bytes);
-
-  struct MemoryStats
-  {
-    size_t totalAllocated{0}, maxAllocated{0}, 
-           currentlyAllocated{0}, totalFreed{0};
-  };
-    
-  class MemoryProfiler
-  {
-  public:
-    static MemoryStats *stats;
-    static bool        debug;
-  };
-
-  #define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
-  #define profilerDebugPrint \
-  if (MemoryProfiler::stats)\
-  {\
-    auto s = MemoryProfiler::stats;\
-    std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl;\
-    std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
-              << std::endl;\
-    std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
-              << std::endl;\
-    std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
-              << std::endl;\
-    std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
-              << std::endl;\
-  }
-
-  #define profilerAllocate(bytes)\
-  if (MemoryProfiler::stats)\
-  {\
-    auto s = MemoryProfiler::stats;\
-    s->totalAllocated     += (bytes);\
-    s->currentlyAllocated += (bytes);\
-    s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated);\
-  }\
-  if (MemoryProfiler::debug)\
-  {\
-    std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl;\
-    profilerDebugPrint;\
-  }
-
-  #define profilerFree(bytes)\
-  if (MemoryProfiler::stats)\
-  {\
-    auto s = MemoryProfiler::stats;\
-    s->totalFreed         += (bytes);\
-    s->currentlyAllocated -= (bytes);\
-  }\
-  if (MemoryProfiler::debug)\
-  {\
-    std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl;\
-    profilerDebugPrint;\
-  }
-
-  void check_huge_pages(void *Buf,uint64_t BYTES);
-
-////////////////////////////////////////////////////////////////////
-// A lattice of something, but assume the something is SIMDized.
-////////////////////////////////////////////////////////////////////
 
 template<typename _Tp>
 class alignedAllocator {
@@ -151,68 +53,31 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
-
+    _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
-    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-    //    if ( ptr != NULL ) 
-    //      std::cout << "alignedAllocator "<<__n << " cache hit "<< std::hex << ptr <<std::dec <<std::endl;
-
-    //////////////////
-    // Hack 2MB align; could make option probably doesn't need configurability
-    //////////////////
-//define GRID_ALLOC_ALIGN (128)
-#define GRID_ALLOC_ALIGN (2*1024*1024)
-#ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
-#else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
-#endif
-    //    std::cout << "alignedAllocator " << std::hex << ptr <<std::dec <<std::endl;
-    // First touch optimise in threaded loop
-    uint8_t *cp = (uint8_t *)ptr;
-#ifdef GRID_OMP
-#pragma omp parallel for
-#endif
-    for(size_type n=0;n<bytes;n+=4096){
-      cp[n]=0;
-    }
     return ptr;
   }
 
-  void deallocate(pointer __p, size_type __n) { 
+  void deallocate(pointer __p, size_type __n) 
+  { 
     size_type bytes = __n * sizeof(_Tp);
-
     profilerFree(bytes);
-
+    MemoryManager::CpuFree((void *)__p,bytes);
-    pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
-
-#ifdef HAVE_MM_MALLOC_H
-    if ( __freeme ) _mm_free((void *)__freeme); 
-#else
-    if ( __freeme ) free((void *)__freeme);
-#endif
   }
-  void construct(pointer __p, const _Tp& __val) { };
+
+  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
+  void construct(pointer __p, const _Tp& __val) { 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; }
 
-//////////////////////////////////////////////////////////////////////////////////////////
+//////////////////////////////////////////////////////////////////////////////////////
-// MPI3 : comms must use shm region
+// Unified virtual memory
-// SHMEM: comms must use symmetric heap
+//////////////////////////////////////////////////////////////////////////////////////
-//////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_COMMS_SHMEM
-extern "C" { 
-#include <mpp/shmem.h>
-extern void * shmem_align(size_t, size_t);
-extern void  shmem_free(void *);
-}
-#define PARANOID_SYMMETRIC_HEAP
-#endif
-
 template<typename _Tp>
-class commAllocator {
+class uvmAllocator {
 public: 
   typedef std::size_t     size_type;
   typedef std::ptrdiff_t  difference_type;
@@ -222,94 +87,97 @@ public:
   typedef const _Tp& const_reference;
   typedef _Tp        value_type;
 
-  template<typename _Tp1>  struct rebind { typedef commAllocator<_Tp1> other; };
+  template<typename _Tp1>  struct rebind { typedef uvmAllocator<_Tp1> other; };
-  commAllocator() throw() { }
+  uvmAllocator() throw() { }
-  commAllocator(const commAllocator&) throw() { }
+  uvmAllocator(const uvmAllocator&) throw() { }
-  template<typename _Tp1> commAllocator(const commAllocator<_Tp1>&) throw() { }
+  template<typename _Tp1> uvmAllocator(const uvmAllocator<_Tp1>&) throw() { }
-  ~commAllocator() throw() { }
+  ~uvmAllocator() throw() { }
   pointer       address(reference __x)       const { return &__x; }
   size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
 
-#ifdef GRID_COMMS_SHMEM
   pointer allocate(size_type __n, const void* _p= 0)
   { 
     size_type bytes = __n*sizeof(_Tp);
-
     profilerAllocate(bytes);
-#ifdef CRAY
+    _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
-    _Tp *ptr = (_Tp *) shmem_align(bytes,64);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-#else
-    _Tp *ptr = (_Tp *) shmem_align(64,bytes);
-#endif
-#ifdef PARANOID_SYMMETRIC_HEAP
-    static void * bcast;
-    static long  psync[_SHMEM_REDUCE_SYNC_SIZE];
-
-    bcast = (void *) ptr;
-    shmem_broadcast32((void *)&bcast,(void *)&bcast,sizeof(void *)/4,0,0,0,shmem_n_pes(),psync);
-
-    if ( bcast != ptr ) {
-      std::printf("inconsistent alloc pe %d %lx %lx \n",shmem_my_pe(),bcast,ptr);std::fflush(stdout);
-      //      BACKTRACEFILE();
-      exit(0);
-    }
-    assert( bcast == (void *) ptr);
-#endif 
     return ptr;
   }
-  void deallocate(pointer __p, size_type __n) { 
-    size_type bytes = __n*sizeof(_Tp);
 
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
     profilerFree(bytes);
-    shmem_free((void *)__p);
+    MemoryManager::SharedFree((void *)__p,bytes);
   }
-#else
+
+  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);
-#ifdef HAVE_MM_MALLOC_H
+    _Tp *ptr = (_Tp*) MemoryManager::AcceleratorAllocate(bytes);
-    _Tp * ptr = (_Tp *) _mm_malloc(bytes, GRID_ALLOC_ALIGN);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
-#else
-    _Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN, bytes);
-#endif
-    uint8_t *cp = (uint8_t *)ptr;
-    if ( ptr ) { 
-    // One touch per 4k page, static OMP loop to catch same loop order
-#ifdef GRID_OMP
-#pragma omp parallel for schedule(static)
-#endif
-      for(size_type n=0;n<bytes;n+=4096){
-	cp[n]=0;
-      }
-    }
     return ptr;
   }
-  void deallocate(pointer __p, size_type __n) {
-    size_type bytes = __n*sizeof(_Tp);
 
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
     profilerFree(bytes);
-#ifdef HAVE_MM_MALLOC_H
+    MemoryManager::AcceleratorFree((void *)__p,bytes);
-    _mm_free((void *)__p); 
-#else
-    free((void *)__p);
-#endif
   }
-#endif
   void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
-template<typename _Tp>  inline bool operator==(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator==(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return true; }
-template<typename _Tp>  inline bool operator!=(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return false; }
+template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return false; }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
+#ifdef ACCELERATOR_CSHIFT
-template<class T> using commVector = std::vector<T,commAllocator<T> >;              
+// Cshift on device
-template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
+template<class T> using cshiftAllocator = devAllocator<T>;
-    
+#else
-}; // namespace Grid
+// 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 cshiftVector = std::vector<T,cshiftAllocator<T> >;
+
+NAMESPACE_END(Grid);
+
+

Grid/allocator/Allocator.h

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

Grid/allocator/MemoryManager.cc

@@ -0,0 +1,254 @@
+#include <Grid/GridCore.h>
+
+NAMESPACE_BEGIN(Grid);
+
+/*Allocation types, saying which pointer cache should be used*/
+#define Cpu      (0)
+#define CpuSmall (1)
+#define Acc      (2)
+#define AccSmall (3)
+#define Shared   (4)
+#define SharedSmall (5)
+uint64_t total_shared;
+uint64_t total_device;
+uint64_t total_host;;
+void MemoryManager::PrintBytes(void)
+{
+  std::cout << " MemoryManager : "<<total_shared<<" shared      bytes "<<std::endl;
+  std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
+  std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<std::endl;
+}
+
+//////////////////////////////////////////////////////////////////////
+// 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] = { 8, 32, 8, 32, 8, 32 };
+
+//////////////////////////////////////////////////////////////////////
+// Actual allocation and deallocation utils
+//////////////////////////////////////////////////////////////////////
+void *MemoryManager::AcceleratorAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Acc);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocDevice(bytes);
+    total_device+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes)
+{
+  void *__freeme = Insert(ptr,bytes,Acc);
+  if ( __freeme ) {
+    acceleratorFreeDevice(__freeme);
+    total_device-=bytes;
+    //    PrintBytes();
+  }
+}
+void *MemoryManager::SharedAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Shared);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+    total_shared+=bytes;
+    //    std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
+    //    PrintBytes();
+  }
+  return ptr;
+}
+void  MemoryManager::SharedFree    (void *ptr,size_t bytes)
+{
+  void *__freeme = Insert(ptr,bytes,Shared);
+  if ( __freeme ) {
+    acceleratorFreeShared(__freeme);
+    total_shared-=bytes;
+    //    PrintBytes();
+  }
+}
+#ifdef GRID_UVM
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocShared(bytes);
+    total_host+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeShared(__freeme);
+    total_host-=bytes;
+  }
+}
+#else
+void *MemoryManager::CpuAllocate(size_t bytes)
+{
+  void *ptr = (void *) Lookup(bytes,Cpu);
+  if ( ptr == (void *) NULL ) {
+    ptr = (void *) acceleratorAllocCpu(bytes);
+    total_host+=bytes;
+  }
+  return ptr;
+}
+void  MemoryManager::CpuFree    (void *_ptr,size_t bytes)
+{
+  NotifyDeletion(_ptr);
+  void *__freeme = Insert(_ptr,bytes,Cpu);
+  if ( __freeme ) { 
+    acceleratorFreeCpu(__freeme);
+    total_host-=bytes;
+  }
+}
+#endif
+
+//////////////////////////////////////////
+// call only once
+//////////////////////////////////////////
+void MemoryManager::Init(void)
+{
+
+  char * str;
+  int Nc;
+  int NcS;
+  
+  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_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 allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<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
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type + small;
+  return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);  
+#else
+  return ptr;
+#endif
+}
+
+void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) 
+{
+  assert(ncache>0);
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  void * ret = NULL;
+  int v = -1;
+
+  for(int e=0;e<ncache;e++) {
+    if ( entries[e].valid==0 ) {
+      v=e; 
+      break;
+    }
+  }
+
+  if ( v==-1 ) {
+    v=victim;
+    victim = (victim+1)%ncache;
+  }
+
+  if ( entries[v].valid ) {
+    ret = entries[v].address;
+    entries[v].valid = 0;
+    entries[v].address = NULL;
+    entries[v].bytes = 0;
+  }
+
+  entries[v].address=ptr;
+  entries[v].bytes  =bytes;
+  entries[v].valid  =1;
+
+  return ret;
+}
+
+void *MemoryManager::Lookup(size_t bytes,int type)
+{
+#ifdef ALLOCATION_CACHE
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type+small;
+  return Lookup(bytes,Entries[cache],Ncache[cache]);
+#else
+  return NULL;
+#endif
+}
+
+void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) 
+{
+  assert(ncache>0);
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+  for(int e=0;e<ncache;e++){
+    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
+      entries[e].valid = 0;
+      return entries[e].address;
+    }
+  }
+  return NULL;
+}
+
+
+NAMESPACE_END(Grid);
+

Grid/allocator/MemoryManager.h

@@ -0,0 +1,180 @@
+/*************************************************************************************
+
+    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)
+
+/*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
+};
+
+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=6;
+  static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
+  static int Victim[NallocType];
+  static int Ncache[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) ;
+  static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
+
+  static void PrintBytes(void);
+ public:
+  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;
+ 
+ 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 int   isOpen   (void* CpuPtr);
+  static void  ViewClose(void* CpuPtr,ViewMode mode);
+  static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
+
+};
+
+NAMESPACE_END(Grid);
+
+

Grid/allocator/MemoryManagerCache.cc

@@ -0,0 +1,478 @@
+#include <Grid/GridCore.h>
+#ifndef GRID_UVM
+
+#warning "Using explicit device memory copies"
+NAMESPACE_BEGIN(Grid);
+//define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
+#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;
+
+////////////////////////////////////
+// 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);
+  
+   dprintf("MemoryManager: Discard(%llx) %llx\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);
+    DeviceBytes   -=AccCache.bytes;
+    LRUremove(AccCache);
+    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\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 entry
+  // Cannot be locked. If allocated must be in LRU pool.
+  ///////////////////////////////////////////////////////////////////////////
+  assert(AccCache.state!=Empty);
+  
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  assert(AccCache.accLock==0);
+  assert(AccCache.cpuLock==0);
+  if(AccCache.state==AccDirty) {
+    Flush(AccCache);
+  }
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  if(AccCache.AccPtr) {
+    AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceBytes   -=AccCache.bytes;
+    LRUremove(AccCache);
+    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+  }
+  uint64_t CpuPtr = AccCache.CpuPtr;
+  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);
+  dprintf("MemoryManager: Flush  %llx -> %llx\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;
+  }
+  dprintf("MemoryManager: Clone %llx <- %llx\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) ){
+    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) ){
+    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)
+{
+  while(bytes+DeviceLRUBytes > DeviceMaxBytes){
+    if ( DeviceLRUBytes > 0){
+      assert(LRU.size()>0);
+      uint64_t victim = LRU.back();
+      auto AccCacheIterator = EntryLookup(victim);
+      auto & AccCache = AccCacheIterator->second;
+      Evict(AccCache);
+    }
+  }
+}
+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 %llx %llx : %lld %lld\n",
+		    (uint64_t)AccCache.CpuPtr,
+		    (uint64_t)CpuPtr,
+		    (uint64_t)AccCache.bytes,
+		    (uint64_t)bytes);
+    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;
+  } 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("Copied 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 into device accLock %d\n",AccCache.accLock);
+  } else {
+    assert(0);
+  }
+
+  // If view is opened on device remove from LRU
+  if(AccCache.LRU_valid==1){
+    // must possibly remove from LRU as now locked on GPU
+    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) {
+    LRUinsert(AccCache);
+  }
+}
+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;
+
+  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)
+{
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << 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 << GridLogDebug << "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 << GridLogDebug << "--------------------------------------------" << 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;
+  }
+}
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/allocator/MemoryManagerShared.cc

@@ -0,0 +1,23 @@
+#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;
+
+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::Print(void){};
+void  MemoryManager::NotifyDeletion(void *ptr){};
+
+NAMESPACE_END(Grid);
+#endif

Grid/allocator/MemoryStats.cc

@@ -1,70 +1,11 @@
 #include <Grid/GridCore.h>
 #include <fcntl.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 MemoryStats *MemoryProfiler::stats = nullptr;
 bool         MemoryProfiler::debug = false;
 
-int PointerCache::victim;
-
-PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
-
-void *PointerCache::Insert(void *ptr,size_t bytes) {
-
-  if (bytes < 4096 ) return ptr;
-
-#ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
-#endif 
-
-  void * ret = NULL;
-  int v = -1;
-
-  for(int e=0;e<Ncache;e++) {
-    if ( Entries[e].valid==0 ) {
-      v=e; 
-      break;
-    }
-  }
-
-  if ( v==-1 ) {
-    v=victim;
-    victim = (victim+1)%Ncache;
-  }
-
-  if ( Entries[v].valid ) {
-    ret = Entries[v].address;
-    Entries[v].valid = 0;
-    Entries[v].address = NULL;
-    Entries[v].bytes = 0;
-  }
-
-  Entries[v].address=ptr;
-  Entries[v].bytes  =bytes;
-  Entries[v].valid  =1;
-
-  return ret;
-}
-
-void *PointerCache::Lookup(size_t bytes) {
-
- if (bytes < 4096 ) return NULL;
-
-#ifdef _OPENMP
-  assert(omp_in_parallel()==0);
-#endif 
-
-  for(int e=0;e<Ncache;e++){
-    if ( Entries[e].valid && ( Entries[e].bytes == bytes ) ) {
-      Entries[e].valid = 0;
-      return Entries[e].address;
-    }
-  }
-  return NULL;
-}
-
-
 void check_huge_pages(void *Buf,uint64_t BYTES)
 {
 #ifdef __linux__
@@ -122,4 +63,5 @@ std::string sizeString(const size_t bytes)
   return std::string(buf);
 }
 
-}
+NAMESPACE_END(Grid);
+

Grid/allocator/MemoryStats.h

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

Grid/cartesian/Cartesian_base.h

@@ -30,16 +30,15 @@
 #ifndef GRID_CARTESIAN_BASE_H
 #define GRID_CARTESIAN_BASE_H
 
-
+NAMESPACE_BEGIN(Grid);
-namespace Grid{
 
 //////////////////////////////////////////////////////////////////////
 // Commicator provides information on the processor grid
 //////////////////////////////////////////////////////////////////////
 //    unsigned long _ndimension;
-  //    std::vector<int> _processors; // processor grid
+//    Coordinate _processors; // processor grid
 //    int              _processor;  // linear processor rank
-  //    std::vector<int> _processor_coor;  // linear processor rank
+//    Coordinate _processor_coor;  // linear processor rank
 //////////////////////////////////////////////////////////////////////
 class GridBase : public CartesianCommunicator , public GridThread {
 
@@ -48,38 +47,41 @@ public:
   // Give Lattice access
   template<class object> friend class Lattice;
 
-    GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
+  GridBase(const Coordinate & processor_grid) : CartesianCommunicator(processor_grid) { LocallyPeriodic=0;}; 
-    GridBase(const std::vector<int> & processor_grid,
+
+  GridBase(const Coordinate & processor_grid,
 	   const CartesianCommunicator &parent,
 	   int &split_rank) 
-      : CartesianCommunicator(processor_grid,parent,split_rank) {};
+    : CartesianCommunicator(processor_grid,parent,split_rank) {LocallyPeriodic=0;};
-    GridBase(const std::vector<int> & processor_grid,
+
+  GridBase(const Coordinate & processor_grid,
 	   const CartesianCommunicator &parent) 
-      : CartesianCommunicator(processor_grid,parent,dummy) {};
+    : CartesianCommunicator(processor_grid,parent,dummy) {LocallyPeriodic=0;};
 
   virtual ~GridBase() = default;
 
-
   // Physics Grid information.
-    std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
+  Coordinate _simd_layout;// Which dimensions get relayed out over simd lanes.
-    std::vector<int> _fdimensions;// (full) Global dimensions of array prior to cb removal
+  Coordinate _fdimensions;// (full) Global dimensions of array prior to cb removal
-    std::vector<int> _gdimensions;// Global dimensions of array after cb removal
+  Coordinate _gdimensions;// Global dimensions of array after cb removal
-    std::vector<int> _ldimensions;// local dimensions of array with processor images removed
+  Coordinate _ldimensions;// local dimensions of array with processor images removed
-    std::vector<int> _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
+  Coordinate _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
-    std::vector<int> _ostride;    // Outer stride for each dimension
+  Coordinate _ostride;    // Outer stride for each dimension
-    std::vector<int> _istride;    // Inner stride i.e. within simd lane
+  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;
-    std::vector<int> _slice_block;// subslice information
+  Coordinate _slice_block;// subslice information
-    std::vector<int> _slice_stride;
+  Coordinate _slice_stride;
-    std::vector<int> _slice_nblock;
+  Coordinate _slice_nblock;
 
-    std::vector<int> _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
+  Coordinate _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
-    std::vector<int> _lend  ;     // local end of array in gcoors   _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
+  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:
 
@@ -88,7 +90,7 @@ public:
   // GridCartesian / GridRedBlackCartesian
   ////////////////////////////////////////////////////////////////
   virtual int CheckerBoarded(int dim)=0;
-    virtual int CheckerBoard(const std::vector<int> &site)=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;
@@ -107,20 +109,20 @@ public:
   // coordinate. Note, however, for data parallel operations the "inner" indexing cost is not paid and all
   // lanes are operated upon simultaneously.
   
-    virtual int oIndex(std::vector<int> &coor)
+  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(std::vector<int> &lcoor)
+  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(std::vector<int> &ocoor)
+  inline int oIndexReduced(Coordinate &ocoor)
   {
     int idx=0; 
     // ocoor is already reduced so can eliminate the modulo operation
@@ -128,11 +130,11 @@ public:
     for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*ocoor[d];
     return idx;
   }
-    inline void oCoorFromOindex (std::vector<int>& coor,int Oindex){
+  inline void oCoorFromOindex (Coordinate& coor,int Oindex){
     Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
   }
 
-    inline void InOutCoorToLocalCoor (std::vector<int> &ocoor, std::vector<int> &icoor, std::vector<int> &lcoor) {
+  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];
@@ -141,7 +143,7 @@ public:
   //////////////////////////////////////////////////////////
   // SIMD lane addressing
   //////////////////////////////////////////////////////////
-    inline void iCoorFromIindex(std::vector<int> &coor,int lane)
+  inline void iCoorFromIindex(Coordinate &coor,int lane)
   {
     Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
   }
@@ -152,8 +154,6 @@ public:
   inline int PermuteType(int dimension){
     int permute_type=0;
     //
-      // FIXME:
-      //
     // 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 
@@ -186,11 +186,11 @@ public:
   inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
-    inline const std::vector<int> LocalStarts(void)             { return _lstart;    };
+  inline const Coordinate LocalStarts(void)             { return _lstart;    };
-    inline const std::vector<int> &FullDimensions(void)         { return _fdimensions;};
+  inline const Coordinate &FullDimensions(void)         { return _fdimensions;};
-    inline const std::vector<int> &GlobalDimensions(void)       { return _gdimensions;};
+  inline const Coordinate &GlobalDimensions(void)       { return _gdimensions;};
-    inline const std::vector<int> &LocalDimensions(void)        { return _ldimensions;};
+  inline const Coordinate &LocalDimensions(void)        { return _ldimensions;};
-    inline const std::vector<int> &VirtualLocalDimensions(void) { return _ldimensions;};
+  inline const Coordinate &VirtualLocalDimensions(void) { return _ldimensions;};
 
   ////////////////////////////////////////////////////////////////
   // Utility to print the full decomposition details 
@@ -214,15 +214,15 @@ public:
   ////////////////////////////////////////////////////////////////
   // Global addressing
   ////////////////////////////////////////////////////////////////
-    void GlobalIndexToGlobalCoor(int gidx,std::vector<int> &gcoor){
+  void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
     assert(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
-    void LocalIndexToLocalCoor(int lidx,std::vector<int> &lcoor){
+  void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
     assert(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
-    void GlobalCoorToGlobalIndex(const std::vector<int> & gcoor,int & gidx){
+  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
     gidx=0;
     int mult=1;
     for(int mu=0;mu<_ndimension;mu++) {
@@ -230,7 +230,7 @@ public:
       mult*=_gdimensions[mu];
     }
   }
-    void GlobalCoorToProcessorCoorLocalCoor(std::vector<int> &pcoor,std::vector<int> &lcoor,const std::vector<int> &gcoor)
+  void GlobalCoorToProcessorCoorLocalCoor(Coordinate &pcoor,Coordinate &lcoor,const Coordinate &gcoor)
   {
     pcoor.resize(_ndimension);
     lcoor.resize(_ndimension);
@@ -240,14 +240,14 @@ public:
       lcoor[mu] = gcoor[mu]%_fld;
     }
   }
-    void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const std::vector<int> &gcoor)
+  void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const Coordinate &gcoor)
   {
-      std::vector<int> pcoor;
+    Coordinate pcoor;
-      std::vector<int> lcoor;
+    Coordinate lcoor;
     GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
     rank = RankFromProcessorCoor(pcoor);
     /*
-      std::vector<int> cblcoor(lcoor);
+      Coordinate cblcoor(lcoor);
       for(int d=0;d<cblcoor.size();d++){
       if( this->CheckerBoarded(d) ) {
       cblcoor[d] = lcoor[d]/2;
@@ -258,10 +258,10 @@ public:
     o_idx= oIndex(lcoor);
   }
 
-    void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , std::vector<int> &gcoor)
+  void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , Coordinate &gcoor)
   {
     gcoor.resize(_ndimension);
-      std::vector<int> coor(_ndimension);
+    Coordinate coor(_ndimension);
 
     ProcessorCoorFromRank(rank,coor);
     for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = _ldimensions[mu]*coor[mu];
@@ -273,20 +273,19 @@ public:
     for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += coor[mu];
       
   }
-    void RankIndexCbToFullGlobalCoor(int rank, int o_idx, int i_idx, int cb,std::vector<int> &fcoor)
+  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(std::vector<int> &Pcoor,std::vector<int> &Lcoor,std::vector<int> &gcoor)
+  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

@@ -28,17 +28,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CARTESIAN_FULL_H
 #define GRID_CARTESIAN_FULL_H
 
-namespace Grid{
+NAMESPACE_BEGIN(Grid);
     
 /////////////////////////////////////////////////////////////////////////////////////////
 // Grid Support.
 /////////////////////////////////////////////////////////////////////////////////////////
 
-
 class GridCartesian: public GridBase {
 
 public:
   int dummy;
+  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -49,7 +49,7 @@ public:
   virtual int CheckerBoarded(int dim){
     return 0;
   }
-    virtual int CheckerBoard(const std::vector<int> &site){
+  virtual int CheckerBoard(const Coordinate &site){
     return 0;
   }
   virtual int CheckerBoardDestination(int cb,int shift,int dim){
@@ -64,16 +64,16 @@ public:
   /////////////////////////////////////////////////////////////////////////
   // Constructor takes a parent grid and possibly subdivides communicator.
   /////////////////////////////////////////////////////////////////////////
-    GridCartesian(const std::vector<int> &dimensions,
+  GridCartesian(const Coordinate &dimensions,
-		  const std::vector<int> &simd_layout,
+		const Coordinate &simd_layout,
-		  const std::vector<int> &processor_grid,
+		const Coordinate &processor_grid,
 		const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
   {
     Init(dimensions,simd_layout,processor_grid);
   }
-    GridCartesian(const std::vector<int> &dimensions,
+  GridCartesian(const Coordinate &dimensions,
-		  const std::vector<int> &simd_layout,
+		const Coordinate &simd_layout,
-		  const std::vector<int> &processor_grid,
+		const Coordinate &processor_grid,
 		const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
   {
     Init(dimensions,simd_layout,processor_grid);
@@ -81,18 +81,18 @@ public:
   /////////////////////////////////////////////////////////////////////////
   // Construct from comm world
   /////////////////////////////////////////////////////////////////////////
-    GridCartesian(const std::vector<int> &dimensions,
+  GridCartesian(const Coordinate &dimensions,
-		  const std::vector<int> &simd_layout,
+		const Coordinate &simd_layout,
-		  const std::vector<int> &processor_grid) : GridBase(processor_grid)
+		const Coordinate &processor_grid) : GridBase(processor_grid)
   {
     Init(dimensions,simd_layout,processor_grid);
   }
 
   virtual ~GridCartesian() = default;
 
-    void Init(const std::vector<int> &dimensions,
+  void Init(const Coordinate &dimensions,
-	      const std::vector<int> &simd_layout,
+	    const Coordinate &simd_layout,
-	      const std::vector<int> &processor_grid)
+	    const Coordinate &processor_grid)
   {
     ///////////////////////
     // Grid information
@@ -105,6 +105,7 @@ public:
     _ldimensions.resize(_ndimension);
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
+    _checker_dim_mask.resize(_ndimension);;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -115,6 +116,8 @@ public:
 
     for (int d = 0; d < _ndimension; d++)
       {
+	_checker_dim_mask[d]=0;
+
         _fdimensions[d] = dimensions[d];   // Global dimensions
         _gdimensions[d] = _fdimensions[d]; // Global dimensions
         _simd_layout[d] = simd_layout[d];
@@ -170,5 +173,6 @@ public:
   };
 
 };
-}
+
+NAMESPACE_END(Grid);
 #endif

Grid/cartesian/Cartesian_red_black.h

@@ -29,19 +29,34 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CARTESIAN_RED_BLACK_H
 #define GRID_CARTESIAN_RED_BLACK_H
 
-
+NAMESPACE_BEGIN(Grid);
-namespace 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, Coordinate &rdim, 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:
-    std::vector<int> _checker_dim_mask;
+  //  Coordinate _checker_dim_mask;
   int              _checker_dim;
   std::vector<int> _checker_board;
 
@@ -49,7 +64,7 @@ public:
     if( dim==_checker_dim) return 1;
     else return 0;
   }
-    virtual int CheckerBoard(const std::vector<int> &site){
+  virtual int CheckerBoard(const Coordinate &site){
     int linear=0;
     assert(site.size()==_ndimension);
     for(int d=0;d<_ndimension;d++){ 
@@ -59,7 +74,6 @@ public:
     return (linear&0x1);
   }
 
-
   // Depending on the cb of site, we toggle source cb.
   // for block #b, element #e = (b, e)
   // we need 
@@ -83,7 +97,7 @@ public:
   }
   virtual int  CheckerBoardFromOindex (int Oindex)
   {
-      std::vector<int> ocoor;
+    Coordinate ocoor;
     oCoorFromOindex(ocoor,Oindex);
     return CheckerBoard(ocoor);
   }
@@ -118,7 +132,7 @@ public:
   GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
   {
     int dims = base->_ndimension;
-      std::vector<int> checker_dim_mask(dims,1);
+    Coordinate checker_dim_mask(dims,1);
     int checker_dim = 0;
     Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
   };
@@ -127,7 +141,7 @@ public:
   // Create redblack from original grid, with non-trivial checker dim mask
   ////////////////////////////////////////////////////////////
   GridRedBlackCartesian(const GridBase *base,
-			  const std::vector<int> &checker_dim_mask,
+			const Coordinate &checker_dim_mask,
 			int checker_dim
 			) :  GridBase(base->_processors,*base) 
   {
@@ -135,40 +149,11 @@ public:
   }
 
   virtual ~GridRedBlackCartesian() = default;
-#if 0
-    ////////////////////////////////////////////////////////////
-    // Create redblack grid ;; deprecate these. Should not
-    // need direct creation of redblack without a full grid to base on
-    ////////////////////////////////////////////////////////////
-    GridRedBlackCartesian(const GridBase *base,
-			  const std::vector<int> &dimensions,
-			  const std::vector<int> &simd_layout,
-			  const std::vector<int> &processor_grid,
-			  const std::vector<int> &checker_dim_mask,
-			  int checker_dim
-			  ) :  GridBase(processor_grid,*base) 
-    {
-      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
-    }
 
-    ////////////////////////////////////////////////////////////
+  void Init(const Coordinate &dimensions,
-    // Create redblack grid
+	    const Coordinate &simd_layout,
-    ////////////////////////////////////////////////////////////
+	    const Coordinate &processor_grid,
-    GridRedBlackCartesian(const GridBase *base,
+	    const Coordinate &checker_dim_mask,
-			  const std::vector<int> &dimensions,
-			  const std::vector<int> &simd_layout,
-			  const std::vector<int> &processor_grid) : GridBase(processor_grid,*base) 
-    {
-      std::vector<int> checker_dim_mask(dimensions.size(),1);
-      int checker_dim = 0;
-      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
-    }
-#endif
-
-    void Init(const std::vector<int> &dimensions,
-              const std::vector<int> &simd_layout,
-              const std::vector<int> &processor_grid,
-              const std::vector<int> &checker_dim_mask,
 	    int checker_dim)
   {
 
@@ -282,7 +267,7 @@ public:
   };
 
 protected:
-    virtual int oIndex(std::vector<int> &coor)
+  virtual int oIndex(Coordinate &coor)
   {
     int idx = 0;
     for (int d = 0; d < _ndimension; d++)
@@ -299,7 +284,7 @@ public:
     return idx;
   };
 
-    virtual int iIndex(std::vector<int> &lcoor)
+  virtual int iIndex(Coordinate &lcoor)
   {
     int idx = 0;
     for (int d = 0; d < _ndimension; d++)
@@ -316,5 +301,5 @@ public:
     return idx;
   }
 };
-}
+NAMESPACE_END(Grid);
 #endif

Grid/communicator/Communicator.h

@@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_COMMUNICATOR_H
 #define GRID_COMMUNICATOR_H
 
+#include <Grid/util/Coordinate.h>
 #include <Grid/communicator/SharedMemory.h>
 #include <Grid/communicator/Communicator_base.h>
 

Grid/communicator/Communicator_base.cc

@@ -31,7 +31,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <limits.h>
 #include <sys/mman.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
@@ -47,8 +47,8 @@ int                      CartesianCommunicator::Dimensions(void)        { return
 int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
 int                      CartesianCommunicator::BossRank(void)          { return 0; };
 int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
-const std::vector<int> & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
+const Coordinate & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
-const std::vector<int> & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
+const Coordinate & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
 int                      CartesianCommunicator::ProcessorCount(void)    { return _Nprocessors; };
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -72,5 +72,6 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
   GlobalSumVector((double *)c,2*N);
 }
   
-}
+NAMESPACE_END(Grid);
+
 

Grid/communicator/Communicator_base.h

@@ -34,7 +34,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 class CartesianCommunicator : public SharedMemory {
 
@@ -52,9 +52,9 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  std::vector<int> _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  std::vector<int> _processor_coor;  // linear processor coordinate
+  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
@@ -69,8 +69,8 @@ public:
   // Constructors to sub-divide a parent communicator
   // and default to comm world
   ////////////////////////////////////////////////
-  CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank);
+  CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank);
-  CartesianCommunicator(const std::vector<int> &pdimensions_in);
+  CartesianCommunicator(const Coordinate &pdimensions_in);
   virtual ~CartesianCommunicator();
 
 private:
@@ -79,7 +79,7 @@ public:
   // Private initialise from an MPI communicator
   // Can use after an MPI_Comm_split, but hidden from user so private
   ////////////////////////////////////////////////
-  void InitFromMPICommunicator(const std::vector<int> &processors, Grid_MPI_Comm communicator_base);
+  void InitFromMPICommunicator(const Coordinate &processors, Grid_MPI_Comm communicator_base);
 
 public:
   
@@ -88,15 +88,15 @@ public:
   // Wraps MPI_Cart routines, or implements equivalent on other impls
   ////////////////////////////////////////////////////////////////////////////////////////
   void ShiftedRanks(int dim,int shift,int & source, int & dest);
-  int  RankFromProcessorCoor(std::vector<int> &coor);
+  int  RankFromProcessorCoor(Coordinate &coor);
-  void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
+  void ProcessorCoorFromRank(int rank,Coordinate &coor);
   
   int                      Dimensions(void)        ;
   int                      IsBoss(void)            ;
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
-  const std::vector<int> & ThisProcessorCoor(void) ;
+  const Coordinate & ThisProcessorCoor(void) ;
-  const std::vector<int> & ProcessorGrid(void)     ;
+  const Coordinate & ProcessorGrid(void)     ;
   int                      ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -114,6 +114,7 @@ public:
   void GlobalSumVector(RealD *,int N);
   void GlobalSum(uint32_t &);
   void GlobalSum(uint64_t &);
+  void GlobalSumVector(uint64_t*,int N);
   void GlobalSum(ComplexF &c);
   void GlobalSumVector(ComplexF *c,int N);
   void GlobalSum(ComplexD &c);
@@ -137,21 +138,6 @@ public:
 		      int recv_from_rank,
 		      int bytes);
   
-  void SendRecvPacket(void *xmit,
-		      void *recv,
-		      int xmit_to_rank,
-		      int recv_from_rank,
-		      int bytes);
-  
-  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-			   void *xmit,
-			   int xmit_to_rank,
-			   void *recv,
-			   int recv_from_rank,
-			   int bytes);
-  
-  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
-
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,
 			       void *recv,
@@ -199,9 +185,10 @@ public:
   template<class obj> void Broadcast(int root,obj &data)
   {
     Broadcast(root,(void *)&data,sizeof(data));
-    };
-
-}; 
   }
 
+}; 
+
+NAMESPACE_END(Grid);
+
 #endif

Grid/communicator/Communicator_mpi3.cc

@@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/communicator/SharedMemory.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
@@ -43,22 +43,35 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 
   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) ||
+    if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
-        (nCommThreads > 1 && provided != MPI_THREAD_MULTIPLE) )
       assert(0);
     }
 
-  Grid_quiesce_nodes();
+    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();
 }
 
 ///////////////////////////////////////////////////////////////////////////
@@ -69,14 +82,14 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
   assert(ierr==0);
 }
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
+int 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, std::vector<int> &coor)
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 {
   coor.resize(_ndimension);
   int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
@@ -86,7 +99,7 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Initialises from communicator_world
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   MPI_Comm optimal_comm;
   ////////////////////////////////////////////////////
@@ -105,13 +118,12 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)    
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
 {
-  _ndimension = processors.size();
+  _ndimension = processors.size();  assert(_ndimension>=1);
-
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
-  std::vector<int> parent_processor_coor(_ndimension,0);
+  Coordinate parent_processor_coor(_ndimension,0);
-  std::vector<int> parent_processors    (_ndimension,1);
+  Coordinate parent_processors    (_ndimension,1);
 
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
@@ -124,10 +136,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   //  int Nparent = parent._processors ;
-  //  std::cout << " splitting from communicator "<<parent.communicator <<std::endl;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
-  //  std::cout << " Parent size  "<<Nparent <<std::endl;
 
   int childsize=1;
   for(int d=0;d<processors.size();d++) {
@@ -136,11 +146,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   int Nchild = Nparent/childsize;
   assert (childsize * Nchild == Nparent);
 
-  //  std::cout << " child size  "<<childsize <<std::endl;
+  Coordinate ccoor(_ndimension); // coor within subcommunicator
-
+  Coordinate scoor(_ndimension); // coor of split within parent
-  std::vector<int> ccoor(_ndimension); // coor within subcommunicator
+  Coordinate ssize(_ndimension); // coor of split within parent
-  std::vector<int> scoor(_ndimension); // coor of split within parent
-  std::vector<int> ssize(_ndimension); // coor of split within parent
 
   for(int d=0;d<_ndimension;d++){
     ccoor[d] = parent_processor_coor[d] % processors[d];
@@ -157,36 +165,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   MPI_Comm comm_split;
   if ( Nchild > 1 ) {
 
-    if(0){
-      std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
-      std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"]    ";
-      for(int d=0;d<parent._ndimension;d++)  std::cout << parent._processors[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" child grid["<< _ndimension <<"]    ";
-      for(int d=0;d<processors.size();d++)  std::cout << processors[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"]    ";
-      for(int d=0;d<parent._ndimension;d++)  std::cout << parent._processor_coor[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"]    ";
-      for(int d=0;d<processors.size();d++)  std::cout << scoor[d] << " ";
-      std::cout<<std::endl;
-      
-      std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"]    ";
-      for(int d=0;d<processors.size();d++)  std::cout << ccoor[d] << " ";
-      std::cout<<std::endl;
-
-      //////////////////////////////////////////////////////////////////////////////////////////////////////
-      // Declare victory
-      //////////////////////////////////////////////////////////////////////////////////////////////////////
-      std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
-		<< Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
-      std::cout << " Split communicator " <<comm_split <<std::endl;
-    }
-
     ////////////////////////////////////////////////////////////////
     // Split the communicator
     ////////////////////////////////////////////////////////////////
@@ -225,7 +203,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   }
 }
 
-void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
+void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors, MPI_Comm communicator_base)
 {
   ////////////////////////////////////////////////////
   // Creates communicator, and the communicator_halo
@@ -242,7 +220,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &proc
     _Nprocessors*=_processors[i];
   }
 
-  std::vector<int> periodic(_ndimension,1);
+  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]);
@@ -285,6 +263,10 @@ 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);
@@ -320,60 +302,28 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int bytes)
 {
   std::vector<CommsRequest_t> reqs(0);
-  //    unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  //    unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
-  //    xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
+
-  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
-  SendToRecvFromComplete(reqs);
-  //    rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //    printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
-}
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int sender,
-					   int receiver,
-					   int bytes)
-{
-  MPI_Status stat;
-  assert(sender != receiver);
-  int tag = sender;
-  if ( _processor == sender ) {
-    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
-  }
-  if ( _processor == receiver ) { 
-    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
-  }
-}
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
   int myrank = _processor;
   int ierr;
 
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+  // Enforce no UVM in comms, device or host OK
-    MPI_Request xrq;
+  assert(acceleratorIsCommunicable(xmit));
-    MPI_Request rrq;
+  assert(acceleratorIsCommunicable(recv));
 
-    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-    
-    assert(ierr==0);
-    list.push_back(xrq);
-    list.push_back(rrq);
-  } else { 
   // Give the CPU to MPI immediately; can use threads to overlap optionally
+  //  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,
 						     void *recv,
@@ -408,16 +358,19 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
+  int tag;
 
   if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
+    tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
   if ( gdest == MPI_UNDEFINED ) {
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
+    tag= dir+_processor*32;
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
@@ -429,15 +382,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 
   return off_node_bytes;
 }
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
-{
-  SendToRecvFromComplete(waitall);
-}
-void CartesianCommunicator::StencilBarrier(void)
-{
-  MPI_Barrier  (ShmComm);
-}
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
   int nreq=list.size();
 
@@ -448,6 +393,13 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
   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);
@@ -479,7 +431,7 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
-  std::vector<int> row(_ndimension,1);
+  Coordinate row(_ndimension,1);
   assert(dim>=0 && dim<_ndimension);
 
   //  Split the communicator
@@ -508,7 +460,4 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   MPI_Type_free(&object);
 }
 
-
+NAMESPACE_END(Grid);
-
-}
-

Grid/communicator/Communicator_none.cc

@@ -27,7 +27,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #include <Grid/GridCore.h>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
@@ -42,17 +42,17 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 					   GlobalSharedMemory::Hugepages);
 }
 
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
   srank=0;
   SetCommunicator(communicator_world);
 }
 
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   _processors = processors;
-  _ndimension = processors.size();
+  _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
   
   // Require 1^N processor grid for fake
@@ -70,21 +70,13 @@ CartesianCommunicator::~CartesianCommunicator(){}
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}
+void CartesianCommunicator::GlobalSumVector(double *,int N){}
 void CartesianCommunicator::GlobalSum(uint32_t &){}
 void CartesianCommunicator::GlobalSum(uint64_t &){}
-void CartesianCommunicator::GlobalSumVector(double *,int N){}
+void CartesianCommunicator::GlobalSumVector(uint64_t *,int N){}
 void CartesianCommunicator::GlobalXOR(uint32_t &){}
 void CartesianCommunicator::GlobalXOR(uint64_t &){}
 
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int xmit_to_rank,
-					   int recv_from_rank,
-					   int bytes)
-{
-  assert(0);
-}
-
 
 // Basic Halo comms primitive -- should never call in single node
 void CartesianCommunicator::SendToRecvFrom(void *xmit,
@@ -95,20 +87,6 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
-  assert(0);
-}
-
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  assert(0);
-}
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   bcopy(in,out,bytes*words);
@@ -122,8 +100,8 @@ int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
-int  CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) {  return 0;}
+int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
-void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){  coor = _processor_coor; }
+void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   source =0;
@@ -136,10 +114,6 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int recv_from_rank,
 						     int bytes, int dir)
 {
-  std::vector<CommsRequest_t> list;
-  // Discard the "dir"
-  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
-  SendToRecvFromComplete(list);
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
@@ -149,17 +123,14 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int recv_from_rank,
 							 int bytes, int dir)
 {
-  // Discard the "dir"
-  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
   return 2.0*bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {
-  SendToRecvFromComplete(waitall);
 }
 
 void CartesianCommunicator::StencilBarrier(void){};
 
+NAMESPACE_END(Grid);
 
-}
 

Grid/communicator/SharedMemory.cc

@@ -28,10 +28,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 #include <Grid/GridCore.h>
 
-namespace Grid { 
+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;
@@ -73,9 +74,12 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
   if (heap_bytes >= heap_size) {
     std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
     std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
-    std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
+    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) { 
@@ -84,9 +88,9 @@ void SharedMemory::ShmBufferFreeAll(void) {
 }
 void *SharedMemory::ShmBufferSelf(void)
 {
+  //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
   return ShmCommBufs[ShmRank];
 }
 
+NAMESPACE_END(Grid); 
 
-
-}

Grid/communicator/SharedMemory.h

@@ -25,18 +25,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-
-
-// TODO
-// 1) move includes into SharedMemory.cc
-//
-// 2) split shared memory into a) optimal communicator creation from comm world
-// 
-//                             b) shared memory buffers container
-//                                -- static globally shared; init once
-//                                -- per instance set of buffers.
-//                                   
-
 #pragma once 
 
 #include <Grid/GridCore.h>
@@ -53,11 +41,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <sys/shm.h>
 #include <sys/mman.h>
 #include <zlib.h>
-#ifdef HAVE_NUMAIF_H
-#include <numaif.h>
-#endif
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 #if defined (GRID_COMMS_MPI3) 
 typedef MPI_Comm    Grid_MPI_Comm;
@@ -71,12 +56,18 @@ 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; }
@@ -102,12 +93,17 @@ class GlobalSharedMemory {
   // Create an optimal reordered communicator that makes MPI_Cart_create get it right
   //////////////////////////////////////////////////////////////////////////////////////
   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  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);
 
 };
 
@@ -148,6 +144,7 @@ class SharedMemory
   // Call on any instance
   ///////////////////////////////////////////////////
   void SharedMemoryTest(void);
+  
   void *ShmBufferSelf(void);
   void *ShmBuffer    (int rank);
   void *ShmBufferTranslate(int rank,void * local_p);
@@ -162,4 +159,5 @@ class SharedMemory
 
 };
 
-}
+NAMESPACE_END(Grid);
+

Grid/communicator/SharedMemoryMPI.cc

@@ -29,8 +29,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
-namespace Grid { 
+#ifdef GRID_CUDA
+#include <cuda_runtime_api.h>
+#endif
+#ifdef GRID_HIP
+#include <hip/hip_runtime_api.h>
+#endif
 
+NAMESPACE_BEGIN(Grid); 
+#define header "SharedMemoryMpi: "
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -43,9 +50,19 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
+#ifndef GRID_MPI3_SHM_NONE
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
+#else
+  MPI_Comm_split(comm, WorldRank, 0, &WorldShmComm);
+#endif
+
   MPI_Comm_rank(WorldShmComm     ,&WorldShmRank);
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
+
+  if ( WorldRank == 0) {
+    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+  }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
   // WorldNodes
@@ -130,9 +147,60 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+{
+  //////////////////////////////////////////////////////////////////////////////
+  // Look and see if it looks like an HPE 8600 based on hostname conventions
+  //////////////////////////////////////////////////////////////////////////////
+  const int namelen = _POSIX_HOST_NAME_MAX;
+  char name[namelen];
+  int R;
+  int I;
+  int N;
+  gethostname(name,namelen);
+  int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
+
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
+}
+static inline int divides(int a,int b)
+{
+  return ( b == ( (b/a)*a ) );
+}
+void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
+{
+  ////////////////////////////////////////////////////////////////
+  // 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;
+  }
+}
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
-#ifdef HYPERCUBE
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
   ////////////////////////////////////////////////////////////////
@@ -173,9 +241,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   }
 
   std::string hname(name);
-  std::cout << "hostname "<<hname<<std::endl;
+  //  std::cout << "hostname "<<hname<<std::endl;
-  std::cout << "R " << R << " I " << I << " N "<< N
+  //  std::cout << "R " << R << " I " << I << " N "<< N
-            << " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
+  //            << " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
 
   //////////////////////////////////////////////////////////////////
   // broadcast node 0's base coordinate for this partition.
@@ -197,16 +265,13 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   // in a maximally symmetrical way
   ////////////////////////////////////////////////////////////////
   int ndimension              = processors.size();
-  std::vector<int> processor_coor(ndimension);
+  Coordinate processor_coor(ndimension);
-  std::vector<int> WorldDims = processors;   std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
+  Coordinate WorldDims = processors;
-  std::vector<int> ShmCoor  (ndimension);    std::vector<int> NodeCoor (ndimension);    std::vector<int> WorldCoor(ndimension);
+  Coordinate ShmDims  (ndimension);  Coordinate NodeDims (ndimension);
-  std::vector<int> HyperCoor(ndimension);
+  Coordinate ShmCoor  (ndimension);    Coordinate NodeCoor (ndimension);    Coordinate WorldCoor(ndimension);
-  int dim = 0;
+  Coordinate HyperCoor(ndimension);
-  for(int l2=0;l2<log2size;l2++){
+
-    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
+  GetShmDims(WorldDims,ShmDims);
-    ShmDims[dim]*=2;
-    dim=(dim+1)%ndimension;
-  }
 
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
@@ -253,28 +318,19 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
-#else 
+}
-  ////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
-  // Assert power of two shm_size.
+{
-  ////////////////////////////////////////////////////////////////
-  int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
-  assert(log2size != -1);
-
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
   // in a maximally symmetrical way
   ////////////////////////////////////////////////////////////////
   int ndimension              = processors.size();
-  std::vector<int> processor_coor(ndimension);
+  Coordinate processor_coor(ndimension);
-  std::vector<int> WorldDims = processors;   std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
+  Coordinate WorldDims = processors; Coordinate ShmDims(ndimension);  Coordinate NodeDims (ndimension);
-  std::vector<int> ShmCoor  (ndimension);    std::vector<int> NodeCoor (ndimension);    std::vector<int> WorldCoor(ndimension);
+  Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
-  int dim = 0;
-  for(int l2=0;l2<log2size;l2++){
-    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
-    ShmDims[dim]*=2;
-    dim=(dim+1)%ndimension;
-  }
 
+  GetShmDims(WorldDims,ShmDims);
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -306,7 +362,6 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
-#endif
 }
 ////////////////////////////////////////////////////////////////////////////////////////////
 // SHMGET
@@ -314,7 +369,7 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -337,7 +392,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
         int errsv = errno;
         printf("Errno %d\n",errsv);
         printf("key   %d\n",key);
-        printf("size  %lld\n",size);
+        printf("size  %ld\n",size);
         printf("flags %d\n",flags);
         perror("shmget");
         exit(1);
@@ -373,10 +428,122 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
+#if defined(GRID_CUDA) ||defined(GRID_HIP)
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  void * ShmCommBuf ; 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the pointer array for shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // TODO/FIXME : NOT ALL NVLINK BOARDS have full Peer to peer connectivity.
+  // The annoyance is that they have partial peer 2 peer. This occurs on the 8 GPU blades.
+  // e.g. DGX1, supermicro board, 
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  //  cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Each MPI rank should allocate our own buffer
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  ShmCommBuf = acceleratorAllocDevice(bytes);
+
+  if (ShmCommBuf == (void *)NULL ) {
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    exit(EXIT_FAILURE);  
+  }
+  //  if ( WorldRank == 0 ){
+  if ( 1 ){
+    std::cout << WorldRank << header " SharedMemoryMPI.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
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+  for(int r=0;r<WorldShmSize;r++){
+
+#ifndef GRID_MPI3_SHM_NONE
+    //////////////////////////////////////////////////
+    // If it is me, pass around the IPC access key
+    //////////////////////////////////////////////////
+#ifdef GRID_CUDA
+    cudaIpcMemHandle_t handle;
+    if ( r==WorldShmRank ) { 
+      auto err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      if ( err !=  cudaSuccess) {
+	std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+#endif
+#ifdef GRID_HIP
+    hipIpcMemHandle_t handle;    
+    if ( r==WorldShmRank ) { 
+      auto err = hipIpcGetMemHandle(&handle,ShmCommBuf);
+      if ( err !=  hipSuccess) {
+	std::cerr << " SharedMemoryMPI.cc hipIpcGetMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+#endif
+    //////////////////////////////////////////////////
+    // Share this IPC handle across the Shm Comm
+    //////////////////////////////////////////////////
+    { 
+      int ierr=MPI_Bcast(&handle,
+			 sizeof(handle),
+			 MPI_BYTE,
+			 r,
+			 WorldShmComm);
+      assert(ierr==0);
+    }
+    
+    ///////////////////////////////////////////////////////////////
+    // If I am not the source, overwrite thisBuf with remote buffer
+    ///////////////////////////////////////////////////////////////
+    void * thisBuf = ShmCommBuf;
+#ifdef GRID_CUDA
+    if ( r!=WorldShmRank ) { 
+      auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      if ( err !=  cudaSuccess) {
+	std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+#endif
+#ifdef GRID_HIP
+    if ( r!=WorldShmRank ) { 
+      auto err = hipIpcOpenMemHandle(&thisBuf,handle,hipIpcMemLazyEnablePeerAccess);
+      if ( err !=  hipSuccess) {
+	std::cerr << " SharedMemoryMPI.cc hipIpcOpenMemHandle failed for rank" << r <<" "<<hipGetErrorString(err)<< std::endl;
+	exit(EXIT_FAILURE);
+      }
+    }
+#endif
+    ///////////////////////////////////////////////////////////////
+    // Save a copy of the device buffers
+    ///////////////////////////////////////////////////////////////
+    WorldShmCommBufs[r] = thisBuf;
+#else
+    WorldShmCommBufs[r] = ShmCommBuf;
+#endif
+  }
+
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+}
+#else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -413,7 +580,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -423,7 +590,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -455,7 +622,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -470,7 +637,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
@@ -499,7 +666,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    
@@ -536,10 +702,27 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _ShmAllocBytes = bytes;
 }
 #endif
+#endif // End NVCC case for GPU device buffers
 
-
+/////////////////////////////////////////////////////////////////////////
-
+// Routines accessing shared memory should route through for GPU safety
-
+/////////////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
+{
+#ifdef GRID_CUDA
+  cudaMemset(dest,0,bytes);
+#else
+  bzero(dest,bytes);
+#endif
+}
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+{
+#ifdef GRID_CUDA
+  cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
+#else   
+  bcopy(src,dest,bytes);
+#endif
+}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
@@ -554,7 +737,11 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
+#ifndef GRID_MPI3_SHM_NONE
   MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
+#else
+  MPI_Comm_split(comm, rank, 0, &ShmComm);
+#endif
   MPI_Comm_rank(ShmComm     ,&ShmRank);
   MPI_Comm_size(ShmComm     ,&ShmSize);
   ShmCommBufs.resize(ShmSize);
@@ -571,7 +758,6 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
     MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
 
     ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
-    //    std::cout << "SetCommunicator ShmCommBufs ["<< r<< "] = "<< ShmCommBufs[r]<< "  wsr = "<<wsr<<std::endl;
   }
   ShmBufferFreeAll();
 
@@ -584,6 +770,19 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
 
   std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
   MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
+
+#ifdef GRID_SHM_FORCE_MPI
+  // Hide the shared memory path between ranks
+  {
+    for(int r=0;r<size;r++){
+      if ( r!=rank ) {
+	ShmRanks[r] = MPI_UNDEFINED;
+      }
+    }
+  }
+#endif
+
+  SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier
@@ -598,25 +797,27 @@ void SharedMemory::ShmBarrier(void)
 void SharedMemory::SharedMemoryTest(void)
 {
   ShmBarrier();
+  uint64_t check[3];
+  uint64_t magic = 0x5A5A5A;
   if ( ShmRank == 0 ) {
-    for(int r=0;r<ShmSize;r++){
+    for(uint64_t r=0;r<ShmSize;r++){
-      uint64_t * check = (uint64_t *) ShmCommBufs[r];
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
-      check[2] = 0x5A5A5A;
+       check[2]=magic;
+       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
-  for(int r=0;r<ShmSize;r++){
+  for(uint64_t r=0;r<ShmSize;r++){
-    uint64_t * check = (uint64_t *) ShmCommBufs[r];
+    ShmBarrier();
-    
+    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
+    ShmBarrier();
     assert(check[0]==GlobalSharedMemory::WorldNode);
     assert(check[1]==r);
-    assert(check[2]==0x5A5A5A);
+    assert(check[2]==magic);
-    
-  }
     ShmBarrier();
   }
+}
 
 void *SharedMemory::ShmBuffer(int rank)
 {
@@ -629,7 +830,6 @@ void *SharedMemory::ShmBuffer(int rank)
 }
 void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 {
-  static int count =0;
   int gpeer = ShmRanks[rank];
   assert(gpeer!=ShmRank); // never send to self
   if (gpeer == MPI_UNDEFINED){
@@ -648,4 +848,5 @@ SharedMemory::~SharedMemory()
   }
 };
 
-}
+NAMESPACE_END(Grid); 
+

Grid/communicator/SharedMemoryNone.cc

@@ -28,7 +28,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 #include <Grid/GridCore.h>
 
-namespace Grid { 
+NAMESPACE_BEGIN(Grid); 
+#define header "SharedMemoryNone: "
 
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@@ -47,7 +48,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   _ShmSetup=1;
 }
 
-void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   optimal_comm = WorldComm;
 }
@@ -55,6 +56,38 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
 ////////////////////////////////////////////////////////////////////////////////////////////
 // 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 ; 
@@ -83,7 +116,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _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
@@ -125,4 +166,5 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 SharedMemory::~SharedMemory()
 {};
 
-}
+NAMESPACE_END(Grid); 
+

Grid/cshift/Cshift.h

@@ -49,4 +49,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_COMMS_SHMEM
 #include <Grid/cshift/Cshift_mpi.h> // uses same implementation of communicator
 #endif 
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
+auto Cshift(const Expression &expr,int dim,int shift)  -> decltype(closure(expr)) 
+{
+  return Cshift(closure(expr),dim,shift);
+}
+NAMESPACE_END(Grid);
+
 #endif

Grid/cshift/Cshift_common.h

@@ -25,37 +25,39 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef _GRID_CSHIFT_COMMON_H_
+#pragma once
-#define _GRID_CSHIFT_COMMON_H_
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
+
+extern Vector<std::pair<int,int> > Cshift_table; 
 
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+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];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask = 0x3;
   }
   
-  int so=plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so=plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
-  int e1=rhs._grid->_slice_nblock[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
   int ent = 0;
 
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
+
+  int stride=rhs.Grid()->_slice_stride[dimension];
 
-  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;
-	table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
+	Cshift_table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
       }
     }
   } else { 
@@ -63,15 +65,27 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
      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);
+	 int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	 if ( ocb &cbmask ) {
-	   table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
+	   Cshift_table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
 	 }
        }
      }
   }
-  parallel_for(int i=0;i<ent;i++){
+  {
-    buffer[table[i].first]=rhs._odata[table[i].second];
+    auto buffer_p = & buffer[0];
+    auto table = &Cshift_table[0];
+#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
   }
 }
 
@@ -79,80 +93,116 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
 // Gather for when there *is* need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
+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];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask = 0x3;
   }
 
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
 
-  int e1=rhs._grid->_slice_nblock[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
-  int n1=rhs._grid->_slice_stride[dimension];
+  int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
-    parallel_for_nest2(int n=0;n<e1;n++){
+#ifdef ACCELERATOR_CSHIFT    
-      for(int b=0;b<e2;b++){
+    autoView(rhs_v , rhs, AcceleratorRead);
-
+    accelerator_for2d(n,e1,b,e2,1,{
 	int o      =   n*n1;
 	int offset = b+n*e2;
 	
-	vobj temp =rhs._odata[so+o+b];
+	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_for2d(n,e1,b,e2,1,{
 
-    // Case of SIMD split AND checker dim cannot currently be hit, except in 
+	Coordinate coor;
-    // Test_cshift_red_black code.
-    std::cout << " Dense packed buffer WARNING " <<std::endl;
-    parallel_for_nest2(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
 
 	int o=n*n1;
-	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
+	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._odata[so+o+b];
+	  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,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
+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];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
     
-  int e1=rhs._grid->_slice_nblock[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
-  int stride=rhs._grid->_slice_stride[dimension];
+  int stride=rhs.Grid()->_slice_stride[dimension];
+
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
 
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
   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 o   =n*rhs.Grid()->_slice_stride[dimension];
-	int bo  =n*rhs._grid->_slice_block[dimension];
+	int bo  =n*rhs.Grid()->_slice_block[dimension];
-	table[ent++] = std::pair<int,int>(so+o+b,bo+b);
+	Cshift_table[ent++] = std::pair<int,int>(so+o+b,bo+b);
       }
     }
 
@@ -160,57 +210,81 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
     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 o   =n*rhs.Grid()->_slice_stride[dimension];
-	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
+	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	if ( ocb & cbmask ) {
-	  table[ent++]=std::pair<int,int> (so+o+b,bo++);
+	  Cshift_table[ent++]=std::pair<int,int> (so+o+b,bo++);
 	}
       }
     }
   }
   
-  parallel_for(int i=0;i<ent;i++){
+  {
-    rhs._odata[table[i].first]=buffer[table[i].second];
+    auto buffer_p = & buffer[0];
+    auto table = &Cshift_table[0];
+#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,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
+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];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane 
     
-  int e1=rhs._grid->_slice_nblock[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block[dimension];
+  int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
-    parallel_for_nest2(int n=0;n<e1;n++){
+    int _slice_stride = rhs.Grid()->_slice_stride[dimension];
-      for(int b=0;b<e2;b++){
+    int _slice_block = rhs.Grid()->_slice_block[dimension];
-	int o      = n*rhs._grid->_slice_stride[dimension];
+#ifdef ACCELERATOR_CSHIFT    
-	int offset = b+n*rhs._grid->_slice_block[dimension];
+    autoView( rhs_v , rhs, AcceleratorWrite);
-	merge(rhs._odata[so+o+b],pointers,offset);
+    accelerator_for2d(n,e1,b,e2,1,{
-      }
+	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 o      = n*rhs.Grid()->_slice_stride[dimension];
-	int offset = b+n*rhs._grid->_slice_block[dimension];
+	int offset = b+n*rhs.Grid()->_slice_block[dimension];
-	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
+	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
 	if ( ocb&cbmask ) {
-	  merge(rhs._odata[so+o+b],pointers,offset);
+	  merge(rhs_v[so+o+b],pointers,offset);
 	}
       }
     }
@@ -220,85 +294,112 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<typ
 //////////////////////////////////////////////////////
 // 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];
+  int rd = rhs.Grid()->_rdimensions[dimension];
 
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int ro  = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  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 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 e1=rhs._grid->_slice_nblock[dimension]; // clearly loop invariant for icpc
-  int e2=rhs._grid->_slice_block[dimension];
-  int stride = rhs._grid->_slice_stride[dimension];
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
   int ent=0;
 
   if(cbmask == 0x3 ){
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int o =n*stride+b;
-	table[ent++] = std::pair<int,int>(lo+o,ro+o);
+	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);
+        int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o);
         if ( ocb&cbmask ) {
-	  table[ent++] = std::pair<int,int>(lo+o,ro+o);
+	  Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
 	}
       }
     }
   }
 
-  parallel_for(int i=0;i<ent;i++){
+  {
-    lhs._odata[table[i].first]=rhs._odata[table[i].second];
+    auto table = &Cshift_table[0];
+#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];
 
-  int rd = rhs._grid->_rdimensions[dimension];
+  if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
-
-  if ( !rhs._grid->CheckerBoarded(dimension) ) {
     cbmask=0x3;
   }
 
-  int ro  = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
+  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 lo  = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane 
 
-  int e1=rhs._grid->_slice_nblock[dimension];
+  int e1=rhs.Grid()->_slice_nblock[dimension];
-  int e2=rhs._grid->_slice_block [dimension];
+  int e2=rhs.Grid()->_slice_block [dimension];
-  int stride = rhs._grid->_slice_stride[dimension];
+  int stride = rhs.Grid()->_slice_stride[dimension];
+
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
 
-  static std::vector<std::pair<int,int> > table;  table.resize(e1*e2);
   int ent=0;
 
-  double t_tab,t_perm;
   if ( cbmask == 0x3 ) {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
-      table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+      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);
+      int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o+b);
-      if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
+      if ( ocb&cbmask ) Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
   }
 
-  parallel_for(int i=0;i<ent;i++){
+  {
-    permute(lhs._odata[table[i].first],rhs._odata[table[i].second],permute_type);
+    auto table = &Cshift_table[0];
+#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
   }
 }
 
@@ -309,10 +410,8 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
 {
   int sshift[2];
 
-  sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
+  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
-  sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
+  sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
-
-  double t_local;
 
   if ( sshift[0] == sshift[1] ) {
     Cshift_local(ret,rhs,dimension,shift,0x3);
@@ -324,7 +423,7 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
 
 template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
-  GridBase *grid = rhs._grid;
+  GridBase *grid = rhs.Grid();
   int fd = grid->_fdimensions[dimension];
   int rd = grid->_rdimensions[dimension];
   int ld = grid->_ldimensions[dimension];
@@ -335,18 +434,18 @@ template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &r
   shift = (shift+fd)%fd;
 
   // the permute type
-  ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
+  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 o   = 0;
     int bo  = x * grid->_ostride[dimension];
     int cb= (cbmask==0x2)? Odd : Even;
 
-    int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
+    int sshift = grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
     int sx     = (x+sshift)%rd;
     
     // wrap is whether sshift > rd.
@@ -387,5 +486,5 @@ template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &r
   
   }
 }
-}
+NAMESPACE_END(Grid);
-#endif
+

Grid/cshift/Cshift_mpi.h

@@ -30,27 +30,27 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define _GRID_CSHIFT_MPI_H_
 
 
-namespace Grid { 
+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); 
+  Lattice<vobj> ret(rhs.Grid()); 
   
-  int fd = rhs._grid->_fdimensions[dimension];
+  int fd = rhs.Grid()->_fdimensions[dimension];
-  int rd = rhs._grid->_rdimensions[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);
+  ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
         
   // the permute type
-  int simd_layout     = rhs._grid->_simd_layout[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs._grid->_processors[dimension] >1 ;
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  int splice_dim      = rhs._grid->_simd_layout[dimension]>1 && (comm_dim);
+  int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
 
 
   if ( !comm_dim ) {
@@ -70,10 +70,10 @@ template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &r
 {
   int sshift[2];
 
-  sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
+  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
-  sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
+  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;
+  //  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);
@@ -88,8 +88,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
 {
   int sshift[2];
 
-  sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
+  sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
-  sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
+  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] ) {
@@ -101,31 +101,32 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
     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;
+  GridBase *grid=rhs.Grid();
-  Lattice<vobj> temp(rhs._grid);
+  Lattice<vobj> temp(rhs.Grid());
 
-  int fd              = rhs._grid->_fdimensions[dimension];
+  int fd              = rhs.Grid()->_fdimensions[dimension];
-  int rd              = rhs._grid->_rdimensions[dimension];
+  int rd              = rhs.Grid()->_rdimensions[dimension];
-  int pd              = rhs._grid->_processors[dimension];
+  int pd              = rhs.Grid()->_processors[dimension];
-  int simd_layout     = rhs._grid->_simd_layout[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs._grid->_processors[dimension] >1 ;
+  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];
+  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  commVector<vobj> send_buf(buffer_size);
+  cshiftVector<vobj> send_buf(buffer_size);
-  commVector<vobj> recv_buf(buffer_size);
+  cshiftVector<vobj> recv_buf(buffer_size);
     
   int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
+  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
 
   for(int x=0;x<rd;x++){       
 
@@ -138,24 +139,26 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
     } else {
 
-      int words = send_buf.size();
+      int words = buffer_size;
       if (cbmask != 0x3) words=words>>1;
 
       int bytes = words * sizeof(vobj);
 
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
 
-      int rank           = grid->_processor;
+      //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
+      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+
       grid->Barrier();
 
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
@@ -165,7 +168,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
-  GridBase *grid=rhs._grid;
+  GridBase *grid=rhs.Grid();
   const int Nsimd = grid->Nsimd();
   typedef typename vobj::vector_type vector_type;
   typedef typename vobj::scalar_object scalar_object;
@@ -193,21 +196,28 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   // 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);
+  //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  std::vector<commVector<scalar_object> >   send_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
+  std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd);
-  std::vector<commVector<scalar_object> >   recv_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
+  std::vector<cshiftVector<scalar_object> >  recv_buf_extract(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);
 
-  std::vector<scalar_object *>  pointers(Nsimd); // 
+  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
-  std::vector<scalar_object *> rpointers(Nsimd); // received pointers
+  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);
+  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
 
   // loop over outer coord planes orthog to dim
   for(int x=0;x<rd;x++){       
@@ -242,11 +252,204 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	grid->SendToRecvFrom((void *)&send_buf_extract[nbr_lane][0],
+	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[i][0],
+			     (void *)recv_buf_extract_mpi,
 			     recv_from_rank,
 			     bytes);
+
+	grid->Barrier();
+
+	rpointers[i] = &recv_buf_extract[i][0];
+      } else { 
+	rpointers[i] = &send_buf_extract[nbr_lane][0];
+      }
+
+    }
+    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
+  }
+
+}
+#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);
+  
+  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
+  cshiftVector<vobj> send_buf_v(buffer_size);
+  cshiftVector<vobj> recv_buf_v(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) {
+
+      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
+
+    } else {
+
+      int words = buffer_size;
+      if (cbmask != 0x3) words=words>>1;
+
+      int bytes = words * sizeof(vobj);
+
+      Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
+
+      //      int rank           = grid->_processor;
+      int recv_from_rank;
+      int xmit_to_rank;
+      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+
+      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);
+      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
+
+      grid->Barrier();
+
+      Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
+    }
+  }
+}
+
+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);
+
+  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);
+
+  std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd);
+  std::vector<cshiftVector<scalar_object> >  recv_buf_extract(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];
+    }
+    int sx   = (x+sshift)%rd;
+    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+
+    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); 
+
+	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);
+
 	grid->Barrier();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -258,5 +461,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   }
 
 }
-}
+#endif
+NAMESPACE_END(Grid); 
+
 #endif

Grid/cshift/Cshift_none.h

@@ -27,13 +27,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #ifndef _GRID_CSHIFT_NONE_H_
 #define _GRID_CSHIFT_NONE_H_
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
-  Lattice<vobj> ret(rhs._grid);
+  Lattice<vobj> ret(rhs.Grid());
-  ret.checkerboard = rhs._grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
+  ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
   Cshift_local(ret,rhs,dimension,shift);
   return ret;
 }
-}
+NAMESPACE_END(Grid);
+
 #endif

Grid/cshift/Cshift_table.cc

@@ -0,0 +1,4 @@
+#include <Grid/GridCore.h>       
+NAMESPACE_BEGIN(Grid);
+Vector<std::pair<int,int> > Cshift_table; 
+NAMESPACE_END(Grid);

Grid/json/json.hpp

@@ -1,3 +1,4 @@
+#ifndef __NVCC__
 /*
     __ _____ _____ _____
  __|  |   __|     |   | |  JSON for Modern C++
@@ -18918,3 +18919,4 @@ inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std
 
 
 #endif
+#endif

Grid/lattice/Lattice.h

@@ -25,9 +25,24 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_LATTICE_H
+#pragma once
-#define GRID_LATTICE_H
+#include <Grid/lattice/Lattice_view.h>
-
 #include <Grid/lattice/Lattice_base.h>
-
+#include <Grid/lattice/Lattice_conformable.h>
-#endif
+#include <Grid/lattice/Lattice_ET.h>
+#include <Grid/lattice/Lattice_arith.h>
+#include <Grid/lattice/Lattice_trace.h>
+#include <Grid/lattice/Lattice_transpose.h>
+#include <Grid/lattice/Lattice_local.h>
+#include <Grid/lattice/Lattice_reduction.h>
+#include <Grid/lattice/Lattice_peekpoke.h>
+#include <Grid/lattice/Lattice_reality.h>
+#include <Grid/lattice/Lattice_real_imag.h>
+#include <Grid/lattice/Lattice_comparison_utils.h>
+#include <Grid/lattice/Lattice_comparison.h>
+#include <Grid/lattice/Lattice_coordinate.h>
+//#include <Grid/lattice/Lattice_where.h>
+#include <Grid/lattice/Lattice_rng.h>
+#include <Grid/lattice/Lattice_unary.h>
+#include <Grid/lattice/Lattice_transfer.h>
+#include <Grid/lattice/Lattice_basis.h>

Grid/lattice/Lattice_ET.h

@@ -9,6 +9,7 @@ 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
@@ -36,14 +37,29 @@ directory
 #include <typeinfo>
 #include <vector>
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 ////////////////////////////////////////////////////
 // Predicated where support
 ////////////////////////////////////////////////////
+#ifdef GRID_SIMT
+// drop to scalar in SIMT; cleaner in fact
 template <class iobj, class vobj, class robj>
-inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
+accelerator_inline vobj predicatedWhere(const iobj &predicate, 
-                            const robj &iffalse) {
+					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;
@@ -51,11 +67,10 @@ inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
   typedef typename vobj::vector_type vector_type;
 
   const int Nsimd = vobj::vector_type::Nsimd();
-  const int words = sizeof(vobj) / sizeof(vector_type);
 
-  std::vector<Integer> mask(Nsimd);
+  ExtractBuffer<Integer> mask(Nsimd);
-  std::vector<scalar_object> truevals(Nsimd);
+  ExtractBuffer<scalar_object> truevals(Nsimd);
-  std::vector<scalar_object> falsevals(Nsimd);
+  ExtractBuffer<scalar_object> falsevals(Nsimd);
 
   extract(iftrue, truevals);
   extract(iffalse, falsevals);
@@ -68,178 +83,273 @@ inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
   merge(ret, falsevals);
   return ret;
 }
+#endif
 
-////////////////////////////////////////////
+/////////////////////////////////////////////////////
-// recursive evaluation of expressions; Could
-// switch to generic approach with variadics, a la
-// Antonin's Lat Sim but the repack to variadic with popped
-// from tuple is hideous; C++14 introduces std::make_index_sequence for this
-////////////////////////////////////////////
-
-// leaf eval of lattice ; should enable if protect using traits
-
-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 <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
-
 //Specialization of getVectorType for lattices
+/////////////////////////////////////////////////////
 template<typename T>
 struct getVectorType<Lattice<T> >{
   typedef typename Lattice<T>::vector_object type;
 };
 
-template<class sobj>
+////////////////////////////////////////////
-inline sobj eval(const unsigned int ss, const sobj &arg)
+//--  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>
+template <class lobj> accelerator_inline 
-inline const lobj &eval(const unsigned int ss, const Lattice<lobj> &arg) {
+auto eval(const uint64_t ss, const LatticeView<lobj> &arg) -> decltype(arg(ss))
-  return arg._odata[ss];
+{
+  return arg(ss);
 }
 
-// handle nodes in syntax tree
+////////////////////////////////////////////
-template <typename Op, typename T1>
+//--  recursive evaluation of expressions; --
-auto inline eval(
+// whole vector return, used only for expression return type inference
-    const unsigned int ss,
+///////////////////////////////////////////////////
-    const LatticeUnaryExpression<Op, T1> &expr)  // eval one operand
+template<class sobj> accelerator_inline 
-    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)))) {
+sobj vecEval(const uint64_t ss, const sobj &arg)
-  return expr.first.func(eval(ss, std::get<0>(expr.second)));
+{
+  return arg;
+}
+template <class lobj> accelerator_inline 
+const lobj & vecEval(const uint64_t ss, const LatticeView<lobj> &arg) 
+{
+  return arg[ss];
 }
 
-template <typename Op, typename T1, typename T2>
+///////////////////////////////////////////////////
-auto inline eval(
+// handle nodes in syntax tree- eval one operand
-    const unsigned int ss,
+// vecEval needed (but never called as all expressions offloaded) to infer the return type
-    const LatticeBinaryExpression<Op, T1, T2> &expr)  // eval two operands
+// in SIMT contexts of closure.
-    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
+///////////////////////////////////////////////////
-                                eval(ss, std::get<1>(expr.second)))) {
+template <typename Op, typename T1> accelerator_inline 
-  return expr.first.func(eval(ss, std::get<0>(expr.second)),
+auto vecEval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
-                         eval(ss, std::get<1>(expr.second)));
+  -> 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));
 }
 
-template <typename Op, typename T1, typename T2, typename T3>
+///////////////////////////////////////////////////
-auto inline eval(const unsigned int ss,
+// handle nodes in syntax tree- eval one operand coalesced
-                 const LatticeTrinaryExpression<Op, T1, T2, T3>
+///////////////////////////////////////////////////
-                     &expr)  // eval three operands
+template <typename Op, typename T1> accelerator_inline 
-    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
+auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
-                                eval(ss, std::get<1>(expr.second)),
+  -> decltype(expr.op.func( eval(ss, expr.arg1)))
-                                eval(ss, std::get<2>(expr.second)))) {
+{
-  return expr.first.func(eval(ss, std::get<0>(expr.second)),
+  return expr.op.func( eval(ss, expr.arg1) );
-                         eval(ss, std::get<1>(expr.second)),
+}
-                         eval(ss, std::get<2>(expr.second)));
+// 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
+// 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,
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-          typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+accelerator_inline void GridFromExpression(GridBase *&grid, const T1 &lat)  // Lattice leaf
-inline void GridFromExpression(GridBase *&grid, const T1 &lat)  // Lattice leaf
 {
-  if (grid) {
+  lat.Conformable(grid);
-    conformable(grid, lat._grid);
 }
-  grid = lat._grid;
+
-}
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-template <class T1,
+accelerator_inline 
-          typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+void GridFromExpression(GridBase *&grid,const T1 &notlat)  // non-lattice leaf
-inline void GridFromExpression(GridBase *&grid,
-                               const T1 &notlat)  // non-lattice leaf
 {}
+
 template <typename Op, typename T1>
-inline void GridFromExpression(GridBase *&grid,
+accelerator_inline 
-                               const LatticeUnaryExpression<Op, T1> &expr) {
+void GridFromExpression(GridBase *&grid,const LatticeUnaryExpression<Op, T1> &expr) 
-  GridFromExpression(grid, std::get<0>(expr.second));  // recurse
+{
+  GridFromExpression(grid, expr.arg1);  // recurse
 }
 
 template <typename Op, typename T1, typename T2>
-inline void GridFromExpression(
+accelerator_inline 
-    GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) {
+void GridFromExpression(GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) 
-  GridFromExpression(grid, std::get<0>(expr.second));  // recurse
+{
-  GridFromExpression(grid, std::get<1>(expr.second));
+  GridFromExpression(grid, expr.arg1);  // recurse
+  GridFromExpression(grid, expr.arg2);
 }
 template <typename Op, typename T1, typename T2, typename T3>
-inline void GridFromExpression(
+accelerator_inline 
-    GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
+void GridFromExpression(GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
-  GridFromExpression(grid, std::get<0>(expr.second));  // recurse
+{
-  GridFromExpression(grid, std::get<1>(expr.second));
+  GridFromExpression(grid, expr.arg1);  // recurse
-  GridFromExpression(grid, std::get<2>(expr.second));
+  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,
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
-          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);
+    assert(cb == lat.Checkerboard());
   }
-  cb = lat.checkerboard;
+  cb = lat.Checkerboard();
-  //  std::cout<<GridLogMessage<<"Lattice leaf cb "<<cb<<std::endl;
 }
-template <class T1,
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-          typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+inline void CBFromExpression(int &cb, const T1 &notlat) {} // non-lattice leaf
-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) 
 {
-  //  std::cout<<GridLogMessage<<"Non lattice leaf cb"<<cb<<std::endl;
+  CBFromExpression(cb, expr.arg1);  // recurse AST
 }
-template <typename Op, typename T1>
+template <typename Op, typename T1, typename T2> inline 
-inline void CBFromExpression(int &cb,
+void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr) 
-                             const LatticeUnaryExpression<Op, T1> &expr) {
+{
-  CBFromExpression(cb, std::get<0>(expr.second));  // recurse
+  CBFromExpression(cb, expr.arg1);  // recurse AST
-  //  std::cout<<GridLogMessage<<"Unary node cb "<<cb<<std::endl;
+  CBFromExpression(cb, expr.arg2);  // recurse AST
-}
-
-template <typename Op, typename T1, typename T2>
-inline void CBFromExpression(int &cb,
-                             const LatticeBinaryExpression<Op, T1, T2> &expr) {
-  CBFromExpression(cb, std::get<0>(expr.second));  // recurse
-  CBFromExpression(cb, std::get<1>(expr.second));
-  //  std::cout<<GridLogMessage<<"Binary node cb "<<cb<<std::endl;
 }
 template <typename Op, typename T1, typename T2, typename T3>
-inline void CBFromExpression(
+inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
-    int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
+{
-  CBFromExpression(cb, std::get<0>(expr.second));  // recurse
+  CBFromExpression(cb, expr.arg1);  // recurse AST
-  CBFromExpression(cb, std::get<1>(expr.second));
+  CBFromExpression(cb, expr.arg2);  // recurse AST
-  CBFromExpression(cb, std::get<2>(expr.second));
+  CBFromExpression(cb, expr.arg3);  // recurse AST
-  //  std::cout<<GridLogMessage<<"Trinary node cb "<<cb<<std::endl;
+}
+
+
+//////////////////////////////////////////////////////////////////////////
+// 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)					\
-  template <class arg>                                                    \
   struct name {								\
-    static auto inline func(const arg a) -> decltype(ret) { return ret; } \
+    template<class _arg> static auto accelerator_inline func(const _arg a) -> decltype(ret) { return ret; } \
   };
 
 GridUnopClass(UnarySub, -a);
 GridUnopClass(UnaryNot, Not(a));
-GridUnopClass(UnaryAdj, adj(a));
-GridUnopClass(UnaryConj, conjugate(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
-GridUnopClass(UnaryReal, real(a));
-GridUnopClass(UnaryImag, imag(a));
-GridUnopClass(UnaryToReal, toReal(a));
-GridUnopClass(UnaryToComplex, toComplex(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
 GridUnopClass(UnarySqrt, sqrt(a));
-GridUnopClass(UnaryRsqrt, rsqrt(a));
 GridUnopClass(UnarySin, sin(a));
 GridUnopClass(UnaryCos, cos(a));
 GridUnopClass(UnaryAsin, asin(a));
@@ -251,18 +361,20 @@ GridUnopClass(UnaryExp, exp(a));
 // Binary operators
 ////////////////////////////////////////////
 #define GridBinOpClass(name, combination)			\
-  template <class left, class right>                           \
   struct name {							\
-    static auto inline func(const left &lhs, const right &rhs) \
+    template <class _left, class _right>			\
-        -> decltype(combination) const {                       \
+    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);
@@ -272,70 +384,53 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
 // Trinary conditional op
 ////////////////////////////////////////////////////
 #define GridTrinOpClass(name, combination)				\
-  template <class predicate, class left, class right>                          \
   struct name {								\
-    static auto inline func(const predicate &pred, const left &lhs,            \
+    template <class _predicate,class _left, class _right>		\
-                            const right &rhs) -> decltype(combination) const { \
+    static auto accelerator_inline					\
+    func(const _predicate &pred, const _left &lhs, const _right &rhs)	\
+      -> decltype(combination) const					\
+    {									\
       return combination;						\
     }									\
-  }
+  };
 
-GridTrinOpClass(
+GridTrinOpClass(TrinaryWhere,
-    TrinaryWhere,
+		(predicatedWhere<
-    (predicatedWhere<predicate, typename std::remove_reference<left>::type,
+		 typename std::remove_reference<_predicate>::type, 
-                     typename std::remove_reference<right>::type>(pred, lhs,
+		 typename std::remove_reference<_left>::type,
-                                                                  rhs)));
+		 typename std::remove_reference<_right>::type>(pred, lhs,rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
 ////////////////////////////////////////////
-
+#define GRID_UNOP(name)   name
-#define GRID_UNOP(name) name<decltype(eval(0, arg))>
+#define GRID_BINOP(name)  name
-#define GRID_BINOP(name) name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
+#define GRID_TRINOP(name) name
-#define GRID_TRINOP(name) \
-  name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
 
 #define GRID_DEF_UNOP(op, name)						\
-  template <typename T1,                                                    \
+  template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
-            typename std::enable_if<is_lattice<T1>::value ||                \
+  inline auto op(const T1 &arg) ->decltype(LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg)) \
-                                        is_lattice_expr<T1>::value,         \
+  {									\
-                                    T1>::type * = nullptr>                  \
+    return     LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg); \
-  inline auto op(const T1 &arg)                                             \
-      ->decltype(LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(       \
-          std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)))) { \
-    return LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(             \
-        std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)));     \
   }
 
 #define GRID_BINOP_LEFT(op, name)					\
   template <typename T1, typename T2,					\
-            typename std::enable_if<is_lattice<T1>::value ||                  \
+            typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
-                                        is_lattice_expr<T1>::value,           \
-                                    T1>::type * = nullptr>                    \
   inline auto op(const T1 &lhs, const T2 &rhs)				\
-      ->decltype(                                                             \
+    ->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs)) \
-          LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(  \
+  {									\
-              std::make_pair(GRID_BINOP(name)(),                              \
+    return     LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs);\
-                             std::forward_as_tuple(lhs, rhs)))) {             \
-    return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
-        std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
   }
 
 #define GRID_BINOP_RIGHT(op, name)					\
   template <typename T1, typename T2,					\
-            typename std::enable_if<!is_lattice<T1>::value &&                 \
+            typename std::enable_if<!is_lattice<T1>::value&&!is_lattice_expr<T1>::value,T1>::type * = nullptr, \
-                                        !is_lattice_expr<T1>::value,          \
+            typename std::enable_if< is_lattice<T2>::value|| is_lattice_expr<T2>::value,T2>::type * = nullptr> \
-                                    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(                                                             \
+    ->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs)) \
-          LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(  \
+  {									\
-              std::make_pair(GRID_BINOP(name)(),                              \
+    return     LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs); \
-                             std::forward_as_tuple(lhs, rhs)))) {             \
-    return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
-        std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
   }
 
 #define GRID_DEF_BINOP(op, name)		\
@@ -345,37 +440,28 @@ GridTrinOpClass(
 #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(                                                              \
+    ->decltype(LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs)) \
-          LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &,  \
+  {									\
-                                   const T3 &>(std::make_pair(                 \
+    return LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs); \
-              GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs)))) {  \
-    return LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &, \
-                                    const T3 &>(std::make_pair(                \
-        GRID_TRINOP(name)(), std::forward_as_tuple(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(adj, UnaryAdj);
-GRID_DEF_UNOP(conjugate, UnaryConj);
+//GRID_DEF_UNOP(conjugate, UnaryConj);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
-GRID_DEF_UNOP(real, UnaryReal);
-GRID_DEF_UNOP(imag, UnaryImag);
-GRID_DEF_UNOP(toReal, UnaryToReal);
-GRID_DEF_UNOP(toComplex, UnaryToComplex);
 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(rsqrt, UnaryRsqrt);
 GRID_DEF_UNOP(sin, UnarySin);
 GRID_DEF_UNOP(cos, UnaryCos);
 GRID_DEF_UNOP(asin, UnaryAsin);
@@ -400,31 +486,36 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-    -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> {
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
-  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> ret(
+{
-      expr);
+  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<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type >
-                                        eval(0, std::get<1>(expr.second))))> {
+{
-  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
-                                   eval(0, std::get<1>(expr.second))))>
-      ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-    -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
-                                        eval(0, std::get<1>(expr.second)),
+				   vecEval(0, expr.arg2),
-                                        eval(0, std::get<2>(expr.second))))> {
+				   vecEval(0, expr.arg3)))>::type >
-  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
+{
-                                   eval(0, std::get<1>(expr.second)),
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
-                                   eval(0, std::get<2>(expr.second))))>
+				vecEval(0, expr.arg2),
-      ret(expr);
+			        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
@@ -433,34 +524,7 @@ auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
 #undef GRID_DEF_UNOP
 #undef GRID_DEF_BINOP
 #undef GRID_DEF_TRINOP
-}
 
-#if 0
+NAMESPACE_END(Grid);
-using namespace Grid;
-        
- int main(int argc,char **argv){
-   
-   Lattice<double> v1(16);
-   Lattice<double> v2(16);
-   Lattice<double> v3(16);
-
-   BinaryAdd<double,double> tmp;
-   LatticeBinaryExpression<BinaryAdd<double,double>,Lattice<double> &,Lattice<double> &> 
-     expr(std::make_pair(tmp,
-    std::forward_as_tuple(v1,v2)));
-   tmp.func(eval(0,v1),eval(0,v2));
-
-   auto var = v1+v2;
-   std::cout<<GridLogMessage<<typeid(var).name()<<std::endl;
-
-   v3=v1+v2;
-   v3=v1+v2+v1*v2;
- };
-
-void testit(Lattice<double> &v1,Lattice<double> &v2,Lattice<double> &v3)
-{
-   v3=v1+v2+v1*v2;
-}
-#endif
 
 #endif

Grid/lattice/Lattice_arith.h

@@ -7,6 +7,7 @@
     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
@@ -28,228 +29,230 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_ARITH_H
 #define GRID_LATTICE_ARITH_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
-
 
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 //  avoid copy back routines for mult, mac, sub, add
 //////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  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);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-#ifdef STREAMING_STORES
+    decltype(coalescedRead(obj1())) tmp;
-      obj1 tmp;
+    auto lhs_t = lhs_v(ss);
-      mult(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
+    auto rhs_t = rhs_v(ss);
-      vstream(ret._odata[ss],tmp);
+    mult(&tmp,&lhs_t,&rhs_t);
-#else
+    coalescedWrite(ret_v[ss],tmp);
-      mult(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
+  });
-#endif
-    }
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( lhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  autoView( rhs_v , rhs, AcceleratorRead);
-      mac(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      vstream(ret._odata[ss],tmp);
+    auto lhs_t=lhs_v(ss);
-#else
+    auto rhs_t=rhs_v(ss);
-      mac(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
+    auto tmp  =ret_v(ss);
-#endif
+    mac(&tmp,&lhs_t,&rhs_t);
-    }
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( lhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  autoView( rhs_v , rhs, AcceleratorRead);
-      sub(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      vstream(ret._odata[ss],tmp);
+    decltype(coalescedRead(obj1())) tmp;
-#else
+    auto lhs_t=lhs_v(ss);
-      sub(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
+    auto rhs_t=rhs_v(ss);
-#endif
+    sub(&tmp,&lhs_t,&rhs_t);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
-  }
+template<class obj1,class obj2,class obj3> inline
-  template<class obj1,class obj2,class obj3> strong_inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( lhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  autoView( rhs_v , rhs, AcceleratorRead);
-      add(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      vstream(ret._odata[ss],tmp);
+    decltype(coalescedRead(obj1())) tmp;
-#else
+    auto lhs_t=lhs_v(ss);
-      add(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
+    auto rhs_t=rhs_v(ss);
-#endif
+    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> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-      obj1 tmp;
+  autoView( lhs_v , lhs, AcceleratorRead);
-      mult(&tmp,&lhs._odata[ss],&rhs);
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      vstream(ret._odata[ss],tmp);
+    decltype(coalescedRead(obj1())) tmp;
-    }
+    mult(&tmp,&lhs_v(ss),&rhs);
+    coalescedWrite(ret_v[ss],tmp);
+  });
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-      obj1 tmp;
+  autoView( lhs_v , lhs, AcceleratorRead);
-      mac(&tmp,&lhs._odata[ss],&rhs);
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      vstream(ret._odata[ss],tmp);
+    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> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( lhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      sub(&tmp,&lhs._odata[ss],&rhs);
+    decltype(coalescedRead(obj1())) tmp;
-      vstream(ret._odata[ss],tmp);
+    auto lhs_t=lhs_v(ss);
-#else 
+    sub(&tmp,&lhs_t,&rhs);
-      sub(&ret._odata[ss],&lhs._odata[ss],&rhs);
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
 }
-  }
+template<class obj1,class obj2,class obj3> inline
-  template<class obj1,class obj2,class obj3> strong_inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-    ret.checkerboard = lhs.checkerboard;
+  ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( lhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-      add(&tmp,&lhs._odata[ss],&rhs);
+    decltype(coalescedRead(obj1())) tmp;
-      vstream(ret._odata[ss],tmp);
+    auto lhs_t=lhs_v(ss);
-#else 
+    add(&tmp,&lhs_t,&rhs);
-      add(&ret._odata[ss],&lhs._odata[ss],&rhs);
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
-    }
 }
 
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 //  avoid copy back routines for mult, mac, sub, add
 //////////////////////////////////////////////////////////////////////////////////////////////////////
-    template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( rhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-      mult(&tmp,&lhs,&rhs._odata[ss]);
+    decltype(coalescedRead(obj1())) tmp;
-      vstream(ret._odata[ss],tmp);
+    auto rhs_t=rhs_v(ss);
-#else 
+    mult(&tmp,&lhs,&rhs_t);
-      mult(&ret._odata[ss],&lhs,&rhs._odata[ss]);
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
-    }
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( rhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-      mac(&tmp,&lhs,&rhs._odata[ss]);
+    auto tmp  =ret_v(ss);
-      vstream(ret._odata[ss],tmp);
+    auto rhs_t=rhs_v(ss);
-#else 
+    mac(&tmp,&lhs,&rhs_t);
-      mac(&ret._odata[ss],&lhs,&rhs._odata[ss]);
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
-    }
 }
   
-  template<class obj1,class obj2,class obj3> strong_inline
+template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( rhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-      sub(&tmp,&lhs,&rhs._odata[ss]);
+    decltype(coalescedRead(obj1())) tmp;
-      vstream(ret._odata[ss],tmp);
+    auto rhs_t=rhs_v(ss);
-#else 
+    sub(&tmp,&lhs,&rhs_t);
-      sub(&ret._odata[ss],&lhs,&rhs._odata[ss]);
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
 }
-  }
+template<class obj1,class obj2,class obj3> inline
-  template<class obj1,class obj2,class obj3> strong_inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-    ret.checkerboard = rhs.checkerboard;
+  ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( rhs_v , lhs, AcceleratorRead);
-      obj1 tmp;
+  accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-      add(&tmp,&lhs,&rhs._odata[ss]);
+    decltype(coalescedRead(obj1())) tmp;
-      vstream(ret._odata[ss],tmp);
+    auto rhs_t=rhs_v(ss);
-#else 
+    add(&tmp,&lhs,&rhs_t);
-      add(&ret._odata[ss],&lhs,&rhs._odata[ss]);
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
-    }
 }
   
-  template<class sobj,class vobj> strong_inline
+template<class sobj,class vobj> inline
 void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
+  ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-    parallel_for(int ss=0;ss<x._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( x_v , x, AcceleratorRead);
-      vobj tmp = a*x._odata[ss]+y._odata[ss];
+  autoView( y_v , y, AcceleratorRead);
-      vstream(ret._odata[ss],tmp);
+  accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-#else
+    auto tmp = a*x_v(ss)+y_v(ss);
-      ret._odata[ss]=a*x._odata[ss]+y._odata[ss];
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
 }
-  }
+template<class sobj,class vobj> inline
-  template<class sobj,class vobj> strong_inline
 void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
+  ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-    parallel_for(int ss=0;ss<x._grid->oSites();ss++){
+  autoView( ret_v , ret, AcceleratorWrite);
-#ifdef STREAMING_STORES
+  autoView( x_v , x, AcceleratorRead);
-      vobj tmp = a*x._odata[ss]+b*y._odata[ss];
+  autoView( y_v , y, AcceleratorRead);
-      vstream(ret._odata[ss],tmp);
+  accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
-#else
+    auto tmp = a*x_v(ss)+b*y_v(ss);
-      ret._odata[ss]=a*x._odata[ss]+b*y._odata[ss];
+    coalescedWrite(ret_v[ss],tmp);
-#endif
+  });
-    }
 }
 
-  template<class sobj,class vobj> strong_inline
+template<class sobj,class vobj> inline
-  RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
+RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
+{
     return axpy_norm_fast(ret,a,x,y);
 }
-  template<class sobj,class vobj> strong_inline
+template<class sobj,class vobj> inline
-  RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
+RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
+{
     return axpby_norm_fast(ret,a,b,x,y);
 }
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_base.h

@@ -9,6 +9,7 @@ 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
@@ -28,311 +29,333 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
-#ifndef GRID_LATTICE_BASE_H
+
-#define GRID_LATTICE_BASE_H
+#pragma once 
 
 #define STREAMING_STORES
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
-
-// TODO: 
-//       mac,real,imag
-
-// Functionality:
-//     -=,+=,*=,()
-//     add,+,sub,-,mult,mac,*
-//     adj,conjugate
-//     real,imag
-//     transpose,transposeIndex  
-//     trace,traceIndex
-//     peekIndex
-//     innerProduct,outerProduct,
-//     localNorm2
-//     localInnerProduct
 
 extern int GridCshiftPermuteMap[4][16];
 
-////////////////////////////////////////////////
+/////////////////////////////////////////////////////////////////////////////////////////
-// Basic expressions used in Expression Template
+// 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
-
+/////////////////////////////////////////////////////////////////////////////////////////
-class LatticeBase
-{
-public:
-    virtual ~LatticeBase(void) = default;
-    GridBase *_grid;
-};
-    
-class LatticeExpressionBase {};
-
-template <typename Op, typename T1>                           
-class LatticeUnaryExpression  : public std::pair<Op,std::tuple<T1> > , public LatticeExpressionBase {
- public:
- LatticeUnaryExpression(const std::pair<Op,std::tuple<T1> > &arg): std::pair<Op,std::tuple<T1> >(arg) {};
-};
-
-template <typename Op, typename T1, typename T2>              
-class LatticeBinaryExpression : public std::pair<Op,std::tuple<T1,T2> > , public LatticeExpressionBase {
- public:
- LatticeBinaryExpression(const std::pair<Op,std::tuple<T1,T2> > &arg): std::pair<Op,std::tuple<T1,T2> >(arg) {};
-};
-
-template <typename Op, typename T1, typename T2, typename T3> 
-class LatticeTrinaryExpression :public std::pair<Op,std::tuple<T1,T2,T3> >, public LatticeExpressionBase {
- public:
- LatticeTrinaryExpression(const std::pair<Op,std::tuple<T1,T2,T3> > &arg): std::pair<Op,std::tuple<T1,T2,T3> >(arg) {};
-};
-
-void inline conformable(GridBase *lhs,GridBase *rhs)
-{
-  assert(lhs == rhs);
-}
-
 template<class vobj>
-class Lattice : public LatticeBase
+class Lattice : public LatticeAccelerator<vobj>
 {
 public:
-    int checkerboard;
+  GridBase *Grid(void) const { return this->_grid; }
-    Vector<vobj> _odata;
+  ///////////////////////////////////////////////////
-    
+  // Member types
-    // to pthread need a computable loop where loop induction is not required
+  ///////////////////////////////////////////////////
-    int begin(void) { return 0;};
-    int end(void)   { return _odata.size(); }
-    vobj & operator[](int i) { return _odata[i]; };
-    const vobj & operator[](int i) const { return _odata[i]; };
-
-public:
   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();
+  }
+  /////////////////////////////////////////////////////////////////////////////////
+  // 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>                         strong_inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
+  template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
   {
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
-    conformable(_grid,egrid);
+    conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
     
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+    auto exprCopy = expr;
-#ifdef STREAMING_STORES
+    ExpressionViewOpen(exprCopy);
-      vobj tmp = eval(ss,expr);
+    auto me  = View(AcceleratorWriteDiscard);
-      vstream(_odata[ss] ,tmp);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-#else
+      auto tmp = eval(ss,exprCopy);
-      _odata[ss]=eval(ss,expr);
+      coalescedWrite(me[ss],tmp);
-#endif
+    });
-    }
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
-  template <typename Op, typename T1,typename T2> strong_inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
+  template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
   {
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
-    conformable(_grid,egrid);
+    conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+    auto exprCopy = expr;
-#ifdef STREAMING_STORES
+    ExpressionViewOpen(exprCopy);
-      vobj tmp = eval(ss,expr);
+    auto me  = View(AcceleratorWriteDiscard);
-      vstream(_odata[ss] ,tmp);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-#else
+      auto tmp = eval(ss,exprCopy);
-      _odata[ss]=eval(ss,expr);
+      coalescedWrite(me[ss],tmp);
-#endif
+    });
-    }
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
-  template <typename Op, typename T1,typename T2,typename T3> strong_inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
+  template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
   {
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
-    conformable(_grid,egrid);
+    conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
-
+    auto exprCopy = expr;
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+    ExpressionViewOpen(exprCopy);
-#ifdef STREAMING_STORES
+    auto me  = View(AcceleratorWriteDiscard);
-      //vobj tmp = eval(ss,expr);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-      vstream(_odata[ss] ,eval(ss,expr));
+      auto tmp = eval(ss,exprCopy);
-#else
+      coalescedWrite(me[ss],tmp);
-      _odata[ss] = eval(ss,expr);
+    });
-#endif
+    me.ViewClose();
-    }
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   //GridFromExpression is tricky to do
   template<class Op,class T1>
   Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
-    _grid = nullptr;
+    this->_grid = nullptr;
-    GridFromExpression(_grid,expr);
+    GridFromExpression(this->_grid,expr);
-    assert(_grid!=nullptr);
+    assert(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    _odata.resize(_grid->oSites());
+    resize(this->_grid->oSites());
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+
-#ifdef STREAMING_STORES
+    *this = expr;
-      vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,tmp);
-#else
-      _odata[ss]=eval(ss,expr);
-#endif
   }
-  };
   template<class Op,class T1, class T2>
   Lattice(const LatticeBinaryExpression<Op,T1,T2> & expr) {
-    _grid = nullptr;
+    this->_grid = nullptr;
-    GridFromExpression(_grid,expr);
+    GridFromExpression(this->_grid,expr);
-    assert(_grid!=nullptr);
+    assert(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    _odata.resize(_grid->oSites());
+    resize(this->_grid->oSites());
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+
-#ifdef STREAMING_STORES
+    *this = expr;
-      vobj tmp = eval(ss,expr);
-      vstream(_odata[ss] ,tmp);
-#else
-      _odata[ss]=eval(ss,expr);
-#endif
   }
-  };
   template<class Op,class T1, class T2, class T3>
   Lattice(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr) {
-    _grid = nullptr;
+    this->_grid = nullptr;
-    GridFromExpression(_grid,expr);
+    GridFromExpression(this->_grid,expr);
-    assert(_grid!=nullptr);
+    assert(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
-    checkerboard=cb;
+    this->checkerboard=cb;
 
-    _odata.resize(_grid->oSites());
+    resize(this->_grid->oSites());
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+
-      vstream(_odata[ss] ,eval(ss,expr));
+    *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;
   }
-  };
 
   //////////////////////////////////////////////////////////////////
-  // Constructor requires "grid" passed.
+  // Follow rule of five, with Constructor requires "grid" passed
-  // what about a default grid?
+  // to user defined constructor
-  //////////////////////////////////////////////////////////////////
+  ///////////////////////////////////////////
-  Lattice(GridBase *grid) : _odata(grid->oSites()) {
+  // user defined constructor
-    _grid = grid;
+  ///////////////////////////////////////////
-    //        _odata.reserve(_grid->oSites());
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
-    //        _odata.resize(_grid->oSites());
+    this->_grid = grid;
-    //      std::cout << "Constructing lattice object with Grid pointer "<<_grid<<std::endl;
+    resize(this->_grid->oSites());
-    assert((((uint64_t)&_odata[0])&0xF) ==0);
+    assert((((uint64_t)&this->_odata[0])&0xF) ==0);
-    checkerboard=0;
+    this->checkerboard=0;
+    SetViewMode(mode);
   }
   
-  Lattice(const Lattice& r){ // copy constructor
+  //  virtual ~Lattice(void) = default;
-    _grid = r._grid;
-    checkerboard = r.checkerboard;
-    _odata.resize(_grid->oSites());// essential
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      _odata[ss]=r._odata[ss];
-    }  	
-  }
-
-  Lattice(Lattice&& r){ // move constructor
-    _grid = r._grid;
-    checkerboard = r.checkerboard;
-    _odata=std::move(r._odata);
-  }
-  
-  inline Lattice<vobj> & operator = (Lattice<vobj> && r)
-  {
-    _grid        = r._grid;
-    checkerboard = r.checkerboard;
-    _odata       =std::move(r._odata);
-    return *this;
-  }
-
-  inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
-    _grid        = r._grid;
-    checkerboard = r.checkerboard;
-    _odata.resize(_grid->oSites());// essential
-    
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      _odata[ss]=r._odata[ss];
-    }  	
-    return *this;
-  }
-
-  template<class robj> strong_inline Lattice<vobj> & operator = (const Lattice<robj> & r){
-    this->checkerboard = r.checkerboard;
-    conformable(*this,r);
-    
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
-      this->_odata[ss]=r._odata[ss];
-    }
-    return *this;
-  }
-
-  virtual ~Lattice(void) = default;
     
   void reset(GridBase* grid) {
-    if (_grid != grid) {
+    if (this->_grid != grid) {
-      _grid = grid;
+      this->_grid = grid;
-      _odata.resize(grid->oSites());
+      this->resize(grid->oSites());
-      checkerboard = 0;
+      this->checkerboard = 0;
     }
   }
-  
+  ///////////////////////////////////////////
-
+  // copy constructor
-  template<class sobj> strong_inline Lattice<vobj> & operator = (const sobj & r){
+  ///////////////////////////////////////////
-    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+  Lattice(const Lattice& r){ 
-      this->_odata[ss]=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 me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
+    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 me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
+    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> strong_inline Lattice<vobj> &operator *=(const T &r) {
+  /////////////////////////////////////////////////////////////////////////////
+  template<class T> inline Lattice<vobj> &operator *=(const T &r) {
     *this = (*this)*r;
     return *this;
   }
   
-  template<class T> strong_inline Lattice<vobj> &operator -=(const T &r) {
+  template<class T> inline Lattice<vobj> &operator -=(const T &r) {
     *this = (*this)-r;
     return *this;
   }
-  template<class T> strong_inline Lattice<vobj> &operator +=(const T &r) {
+  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){
-    std::vector<int> gcoor;
   typedef typename vobj::scalar_object sobj;
+  for(int g=0;g<o.Grid()->_gsites;g++){
+
+    Coordinate gcoor;
+    o.Grid()->GlobalIndexToGlobalCoor(g,gcoor);
+
     sobj ss;
-    for(int g=0;g<o._grid->_gsites;g++){
-      o._grid->GlobalIndexToGlobalCoor(g,gcoor);
     peekSite(ss,o,gcoor);
     stream<<"[";
     for(int d=0;d<gcoor.size();d++){
@@ -345,31 +368,5 @@ public:
   return stream;
 }
   
-}
+NAMESPACE_END(Grid);
 
-
-
-#include "Lattice_conformable.h"
-#define GRID_LATTICE_EXPRESSION_TEMPLATES
-#ifdef  GRID_LATTICE_EXPRESSION_TEMPLATES
-#include "Lattice_ET.h"
-#else 
-#include "Lattice_overload.h"
-#endif
-#include "Lattice_arith.h"
-#include "Lattice_trace.h"
-#include "Lattice_transpose.h"
-#include "Lattice_local.h"
-#include "Lattice_reduction.h"
-#include "Lattice_peekpoke.h"
-#include "Lattice_reality.h"
-#include "Lattice_comparison_utils.h"
-#include "Lattice_comparison.h"
-#include "Lattice_coordinate.h"
-#include "Lattice_where.h"
-#include "Lattice_rng.h"
-#include "Lattice_unary.h"
-#include "Lattice_transfer.h"
-
-
-#endif

Grid/lattice/Lattice_basis.h

@@ -0,0 +1,248 @@
+/*************************************************************************************
+
+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)) )
+  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_v[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_v[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

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_COMPARISON_H
 #define GRID_LATTICE_COMPARISON_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////////
 // relational operators
@@ -40,40 +40,50 @@ namespace Grid {
 //Query supporting logical &&, ||, 
 //////////////////////////////////////////////////////////////////////////
 
+typedef iScalar<vInteger> vPredicate ;
+
 //////////////////////////////////////////////////////////////////////////
 // compare lattice to lattice
 //////////////////////////////////////////////////////////////////////////
+
 template<class vfunctor,class lobj,class robj>  
-    inline Lattice<vInteger> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
+inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
 {
-    Lattice<vInteger> ret(rhs._grid);
+  Lattice<vPredicate> ret(rhs.Grid());
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+  autoView( lhs_v, lhs, CpuRead);
-      ret._odata[ss]=op(lhs._odata[ss],rhs._odata[ss]);
+  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<vInteger> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
+inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
 {
-    Lattice<vInteger> ret(lhs._grid);
+  Lattice<vPredicate> ret(lhs.Grid());
-    parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
+  autoView( lhs_v, lhs, CpuRead);
-      ret._odata[ss]=op(lhs._odata[ss],rhs);
+  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<vInteger> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
+inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
 {
-    Lattice<vInteger> ret(rhs._grid);
+  Lattice<vPredicate> ret(rhs.Grid());
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+  autoView( rhs_v, rhs, CpuRead);
-      ret._odata[ss]=op(lhs._odata[ss],rhs);
+  autoView( ret_v, ret, CpuWrite);
-    }
+  thread_for( ss, rhs_v.size(), {
+    ret_v[ss]=op(lhs,rhs_v[ss]);
+  });
   return ret;
 }
   
@@ -82,88 +92,88 @@ namespace Grid {
 //////////////////////////////////////////////////////////////////////////
 // Less than
 template<class lobj,class robj>
-    inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
+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<vInteger> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
+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<vInteger> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
+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<vInteger> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
+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<vInteger> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
+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<vInteger> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+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<vInteger> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
+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<vInteger> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
+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_comparison_utils.h

@@ -26,10 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_COMPARISON_H
-#define GRID_COMPARISON_H
 
-namespace Grid {
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
 
   /////////////////////////////////////////
   // This implementation is a bit poor.
@@ -44,42 +44,42 @@ namespace Grid {
   //
   template<class lobj,class robj> class veq {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) == (rhs);
     }
   };
   template<class lobj,class robj> class vne {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) != (rhs);
     }
   };
   template<class lobj,class robj> class vlt {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) < (rhs);
     }
   };
   template<class lobj,class robj> class vle {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) <= (rhs);
     }
   };
   template<class lobj,class robj> class vgt {
   public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) > (rhs);
     }
   };
   template<class lobj,class robj> class vge {
     public:
-    vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) >= (rhs);
     }
@@ -88,42 +88,42 @@ namespace Grid {
   // Generic list of functors
   template<class lobj,class robj> class seq {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) == (rhs);
     }
   };
   template<class lobj,class robj> class sne {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) != (rhs);
     }
   };
   template<class lobj,class robj> class slt {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) < (rhs);
     }
   };
   template<class lobj,class robj> class sle {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) <= (rhs);
     }
   };
   template<class lobj,class robj> class sgt {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) > (rhs);
     }
   };
   template<class lobj,class robj> class sge {
   public:
-    Integer operator()(const lobj &lhs, const robj &rhs)
+    accelerator Integer operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) >= (rhs);
     }
@@ -133,12 +133,12 @@ namespace Grid {
   // Integer and real get extra relational functions.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
+    accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
-      std::vector<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
+      ExtractBuffer<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
-      std::vector<scalar> vrhs(vsimd::Nsimd());
+      ExtractBuffer<scalar> vrhs(vsimd::Nsimd());
-      std::vector<Integer> vpred(vsimd::Nsimd());
+      ExtractBuffer<Integer> vpred(vsimd::Nsimd());
       vInteger ret;
       extract<vsimd,scalar>(lhs,vlhs);
       extract<vsimd,scalar>(rhs,vrhs);
@@ -150,11 +150,11 @@ namespace Grid {
     }
 
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
+    accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
-      std::vector<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
+      ExtractBuffer<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
-      std::vector<Integer> vpred(vsimd::Nsimd());
+      ExtractBuffer<Integer> vpred(vsimd::Nsimd());
       vInteger ret;
       extract<vsimd,scalar>(lhs,vlhs);
       for(int s=0;s<vsimd::Nsimd();s++){
@@ -165,11 +165,11 @@ namespace Grid {
     }
 
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
+    accelerator_inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
-      std::vector<scalar> vrhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
+      ExtractBuffer<scalar> vrhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
-      std::vector<Integer> vpred(vsimd::Nsimd());
+      ExtractBuffer<Integer> vpred(vsimd::Nsimd());
       vInteger ret;
       extract<vsimd,scalar>(rhs,vrhs);
       for(int s=0;s<vsimd::Nsimd();s++){
@@ -181,30 +181,30 @@ namespace Grid {
 
 #define DECLARE_RELATIONAL_EQ(op,functor) \
   template<class vsimd,IfSimd<vsimd> = 0>\
-    inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
+    accelerator_inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd,IfSimd<vsimd> = 0>\
-    inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
+    accelerator_inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd,IfSimd<vsimd> = 0>\
-    inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
+    accelerator_inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd>\
-    inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
+    accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
     {									\
       return lhs._internal op rhs;					\
     }									\
   template<class vsimd>\
-    inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
+    accelerator_inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
     {									\
       return lhs op rhs._internal;					\
     }									\
@@ -212,7 +212,7 @@ namespace Grid {
 #define DECLARE_RELATIONAL(op,functor) \
   DECLARE_RELATIONAL_EQ(op,functor)    \
   template<class vsimd>\
-    inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
+    accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
     {									\
       return lhs._internal op rhs._internal;				\
     }									
@@ -226,7 +226,7 @@ DECLARE_RELATIONAL(!=,sne);
 
 #undef DECLARE_RELATIONAL
 
-}
+NAMESPACE_END(Grid);
+
 
 
-#endif

Grid/lattice/Lattice_conformable.h

@@ -28,13 +28,13 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_CONFORMABLE_H
 #define GRID_LATTICE_CONFORMABLE_H
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class obj1,class obj2> void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
 {
-        assert(lhs._grid == rhs._grid);
+  assert(lhs.Grid() == rhs.Grid());
-        assert(lhs.checkerboard == rhs.checkerboard);
+  assert(lhs.Checkerboard() == rhs.Checkerboard());
 }
 
-}
+NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_coordinate.h

@@ -25,32 +25,31 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_LATTICE_COORDINATE_H
+#pragma once 
-#define GRID_LATTICE_COORDINATE_H
 
-namespace Grid {
+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;
+  GridBase *grid = l.Grid();
   int Nsimd = grid->iSites();
 
-      std::vector<int> gcoor;
+  autoView(l_v, l, CpuWrite);
-      std::vector<scalar_type> mergebuf(Nsimd);
+  thread_for( o, grid->oSites(), {
-
     vector_type vI;
-      for(int o=0;o<grid->oSites();o++){
+    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._odata[o]=vI;
+    l_v[o]=vI;
-      }
+  });
 };
 
-}
+NAMESPACE_END(Grid);
-#endif
+

Grid/lattice/Lattice_local.h

@@ -32,7 +32,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 // localInner, localNorm, outerProduct
 ///////////////////////////////////////////////
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 /////////////////////////////////////////////////////
 // Non site, reduced locally reduced routines
@@ -42,10 +42,12 @@ namespace Grid {
 template<class vobj>
 inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
 {
-      Lattice<typename vobj::tensor_reduced> ret(rhs._grid);
+  Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+  autoView( rhs_v , rhs, AcceleratorRead);
-	ret._odata[ss]=innerProduct(rhs._odata[ss],rhs._odata[ss]);
+  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;
 }
   
@@ -53,23 +55,33 @@ namespace Grid {
 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);
+  Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+  autoView( lhs_v , lhs, AcceleratorRead);
-	ret._odata[ss]=innerProduct(lhs._odata[ss],rhs._odata[ss]);
+  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(lhs._odata[0],rhs._odata[0]))>
+inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(ll(),rr()))>
 {
-    Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))> ret(rhs._grid);
+  typedef decltype(coalescedRead(ll())) sll;
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+  typedef decltype(coalescedRead(rr())) srr;
-      ret._odata[ss]=outerProduct(lhs._odata[ss],rhs._odata[ss]);
+  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

@@ -0,0 +1,202 @@
+/*************************************************************************************
+    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::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
+
+  GridBase *FullGrid  = X.Grid();
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+
+  //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::scalar_type scalar_type;
+  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::scalar_type scalar_type;
+  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_overload.h

@@ -1,138 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_overload.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_OVERLOAD_H
-#define GRID_LATTICE_OVERLOAD_H
-
-namespace Grid {
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // unary negation
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class vobj>
-  inline Lattice<vobj> operator -(const Lattice<vobj> &r)
-  {
-    Lattice<vobj> ret(r._grid);
-    parallel_for(int ss=0;ss<r._grid->oSites();ss++){
-      vstream(ret._odata[ss], -r._odata[ss]);
-    }
-    return ret;
-  } 
-  /////////////////////////////////////////////////////////////////////////////////////
-  // Lattice BinOp Lattice,
-  //NB mult performs conformable check. Do not reapply here for performance.
-  /////////////////////////////////////////////////////////////////////////////////////
-  template<class left,class right>
-    inline auto operator * (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]*rhs._odata[0])>
-  {
-    Lattice<decltype(lhs._odata[0]*rhs._odata[0])> ret(rhs._grid);
-    mult(ret,lhs,rhs);
-    return ret;
-  }
-  template<class left,class right>
-    inline auto operator + (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]+rhs._odata[0])>
-  {
-    Lattice<decltype(lhs._odata[0]+rhs._odata[0])> ret(rhs._grid);
-    add(ret,lhs,rhs);
-    return ret;
-  }
-  template<class left,class right>
-    inline auto operator - (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]-rhs._odata[0])>
-  {
-    Lattice<decltype(lhs._odata[0]-rhs._odata[0])> ret(rhs._grid);
-    sub(ret,lhs,rhs);
-    return ret;
-  }
-  
-  // Scalar BinOp Lattice ;generate return type
-  template<class left,class right>
-  inline auto operator * (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs*rhs._odata[0])>
-  {
-    Lattice<decltype(lhs*rhs._odata[0])> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      decltype(lhs*rhs._odata[0]) tmp=lhs*rhs._odata[ss]; 
-      vstream(ret._odata[ss],tmp);
-	   //      ret._odata[ss]=lhs*rhs._odata[ss];
-    }
-    return ret;
-  }
-  template<class left,class right>
-    inline auto operator + (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs+rhs._odata[0])>
-    {
-      Lattice<decltype(lhs+rhs._odata[0])> ret(rhs._grid);
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	decltype(lhs+rhs._odata[0]) tmp =lhs-rhs._odata[ss];  
-	vstream(ret._odata[ss],tmp);
-	//	ret._odata[ss]=lhs+rhs._odata[ss];
-      }
-        return ret;
-    }
-  template<class left,class right>
-    inline auto operator - (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs-rhs._odata[0])>
-  {
-    Lattice<decltype(lhs-rhs._odata[0])> ret(rhs._grid);
-    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-      decltype(lhs-rhs._odata[0]) tmp=lhs-rhs._odata[ss];  
-      vstream(ret._odata[ss],tmp);
-    }
-    return ret;
-  }
-    template<class left,class right>
-      inline auto operator * (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]*rhs)>
-    {
-      Lattice<decltype(lhs._odata[0]*rhs)> ret(lhs._grid);
-      parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
-	decltype(lhs._odata[0]*rhs) tmp =lhs._odata[ss]*rhs;
-	vstream(ret._odata[ss],tmp);
-	//            ret._odata[ss]=lhs._odata[ss]*rhs;
-      }
-      return ret;
-    }
-    template<class left,class right>
-      inline auto operator + (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]+rhs)>
-    {
-        Lattice<decltype(lhs._odata[0]+rhs)> ret(lhs._grid);
-	parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  decltype(lhs._odata[0]+rhs) tmp=lhs._odata[ss]+rhs; 
-	  vstream(ret._odata[ss],tmp);
-	  //	  ret._odata[ss]=lhs._odata[ss]+rhs;
-        }
-        return ret;
-    }
-    template<class left,class right>
-      inline auto operator - (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]-rhs)>
-    {
-      Lattice<decltype(lhs._odata[0]-rhs)> ret(lhs._grid);
-      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  decltype(lhs._odata[0]-rhs) tmp=lhs._odata[ss]-rhs;
-	  vstream(ret._odata[ss],tmp);
-	  //	ret._odata[ss]=lhs._odata[ss]-rhs;
-      }
-      return ret;
-    }
-}
-#endif

Grid/lattice/Lattice_peekpoke.h

@@ -34,29 +34,35 @@ Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 // Peeking and poking around
 ///////////////////////////////////////////////
 
-namespace Grid {
+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>(lhs._odata[0],i))>
+auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Index>(vobj(),i))>
 {
-      Lattice<decltype(peekIndex<Index>(lhs._odata[0],i))> ret(lhs._grid);
+  Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
-      ret.checkerboard=lhs.checkerboard;
+  ret.Checkerboard()=lhs.Checkerboard();
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v, ret, AcceleratorWrite);
-	ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i);
+  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>(lhs._odata[0],i,j))>
+auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(vobj(),i,j))>
 {
-      Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))> ret(lhs._grid);
+  Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
-      ret.checkerboard=lhs.checkerboard;
+  ret.Checkerboard()=lhs.Checkerboard();
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( ret_v, ret, AcceleratorWrite);
-	ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i,j);
+  autoView( lhs_v, lhs, AcceleratorRead);
-      }
+  accelerator_for( ss, lhs_v.size(), 1, {
+    ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
+  });
   return ret;
 };
 
@@ -64,34 +70,38 @@ namespace Grid {
 // Poke internal indices of a Lattice object
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<int Index,class vobj>  
-    void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0))> & rhs,int i)
+void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
 {
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( rhs_v, rhs, AcceleratorRead);
-	pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i);
+  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>(lhs._odata[0],0,0))> & rhs,int i,int j)
+void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
 {
-      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+  autoView( rhs_v, rhs, AcceleratorRead);
-	pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i,j);
+  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 std::vector<int> &site){
+void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
-      GridBase *grid=l._grid;
+  GridBase *grid=l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-      assert( l.checkerboard== l._grid->CheckerBoard(site));
+  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   int rank,odx,idx;
@@ -99,13 +109,13 @@ namespace Grid {
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
   grid->Broadcast(grid->BossRank(),s);
 
-      std::vector<sobj> buf(Nsimd);
-
   // 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._odata[odx],buf);
+    extract(l_v[odx],buf);
     buf[idx] = s;
-	merge(l._odata[odx],buf);
+    merge(l_v[odx],buf);
   }
 
   return;
@@ -116,22 +126,23 @@ namespace Grid {
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
 template<class vobj,class sobj>
-      void peekSite(sobj &s,const Lattice<vobj> &l,const std::vector<int> &site){
+void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         
-      GridBase *grid=l._grid;
+  GridBase *grid=l.Grid();
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-      assert( l.checkerboard == l._grid->CheckerBoard(site));
+  assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
 
   int rank,odx,idx;
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
 
-      std::vector<sobj> buf(Nsimd);
+  ExtractBuffer<sobj> buf(Nsimd);
-      extract(l._odata[odx],buf);
+  autoView( l_v , l, CpuWrite);
+  extract(l_v[odx],buf);
 
   s = buf[idx];
 
@@ -140,21 +151,21 @@ namespace Grid {
   return;
 };
 
-
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+// Must be CPU read view
 template<class vobj,class sobj>
-    void peekLocalSite(sobj &s,const Lattice<vobj> &l,std::vector<int> &site){
+inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
-        
+{
-      GridBase *grid = l._grid;
+  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== l._grid->CheckerBoard(site));
+  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -162,7 +173,7 @@ namespace Grid {
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-      scalar_type * vp = (scalar_type *)&l._odata[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
@@ -171,18 +182,27 @@ namespace Grid {
       
   return;
 };
-
 template<class vobj,class sobj>
-    void pokeLocalSite(const sobj &s,Lattice<vobj> &l,std::vector<int> &site){
+inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
+{
+  autoView(lv,l,CpuRead);
+  peekLocalSite(s,lv,site);
+  return;
+};
 
-      GridBase *grid=l._grid;
+// 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== l._grid->CheckerBoard(site));
+  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -190,16 +210,22 @@ namespace Grid {
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-      scalar_type * vp = (scalar_type *)&l._odata[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
-      
   for(int w=0;w<words;w++){
     vp[idx+w*Nsimd] = pt[w];
   }
-
   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

@@ -0,0 +1,79 @@
+/*************************************************************************************
+
+    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

@@ -36,22 +36,81 @@ Author: neo <cossu@post.kek.jp>
 // The choice of burying complex in the SIMD
 // is making the use of "real" and "imag" very cumbersome
 
-namespace Grid {
+NAMESPACE_BEGIN(Grid);
 
 template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
-        Lattice<vobj> ret(lhs._grid);
+  Lattice<vobj> ret(lhs.Grid());
-	parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+
-            ret._odata[ss] = adj(lhs._odata[ss]);
+  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);
+  Lattice<vobj> ret(lhs.Grid());
-	parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+
-	  ret._odata[ss] = conjugate(lhs._odata[ss]);
+  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

@@ -5,6 +5,7 @@
 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
@@ -19,58 +20,172 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_LATTICE_REDUCTION_H
+#pragma once
-#define GRID_LATTICE_REDUCTION_H
 
 #include <Grid/Grid_Eigen_Dense.h>
 
-namespace Grid {
+
-#ifdef GRID_WARN_SUBOPTIMAL
+#if defined(GRID_CUDA)||defined(GRID_HIP)
-#warning "Optimisation alert all these reduction loops are NOT threaded "
+#include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
 
+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];
+  } 
+  
+  typedef typename vobj::scalar_object ssobj;
+  ssobj ret = ssum;
+  return ret;
+}
+
+
+template<class vobj>
+inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  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)
+  return sumD_gpu(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  autoView( arg_v, arg, AcceleratorRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_gpu(&arg_v[0],osites);
+#else
+  autoView(arg_v, arg, CpuRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif  
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Deterministic Reduction operations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
-  auto nrm = innerProduct(arg,arg);
+  ComplexD nrm = innerProduct(arg,arg);
-  return std::real(nrm); 
+  return real(nrm); 
 }
 
 // Double inner product
 template<class vobj>
-inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
+inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
 {
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
-  GridBase *grid = left._grid;
+  ComplexD  nrm;
-  const int pad = 8;
   
-  ComplexD  inner;
+  GridBase *grid = left.Grid();
-  Vector<ComplexD> sumarray(grid->SumArraySize()*pad);
 
-  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
+  const uint64_t nsimd = grid->Nsimd();
-    int nwork, mywork, myoff;
+  const uint64_t sites = grid->oSites();
-    GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
   
-    decltype(innerProductD(left._odata[0],right._odata[0])) vinner=zero; // private to thread; sub summation
+  // Might make all code paths go this way.
-    for(int ss=myoff;ss<mywork+myoff; ss++){
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
-      vinner = vinner + innerProductD(left._odata[ss],right._odata[ss]);
+  Vector<inner_t> inner_tmp(sites);
-    }
+  auto inner_tmp_v = &inner_tmp[0];
-    // All threads sum across SIMD; reduce serial work at end
+    
-    // one write per cacheline with streaming store
+  {
-    ComplexD tmp = Reduce(TensorRemove(vinner)) ;
+    autoView( left_v , left, AcceleratorRead);
-    vstream(sumarray[thr*pad],tmp);
+    autoView( right_v,right, AcceleratorRead);
+
+    // GPU - SIMT lane compliance...
+    accelerator_for( ss, sites, 1,{
+	auto x_l = left_v[ss];
+	auto y_l = right_v[ss];
+	inner_tmp_v[ss]=innerProductD(x_l,y_l);
+    });
   }
 
-  inner=0.0;
+  // This is in single precision and fails some tests
-  for(int i=0;i<grid->SumArraySize();i++){
+  auto anrm = sum(inner_tmp_v,sites);  
-    inner = inner+sumarray[i*pad];
+  nrm = anrm;
+  return nrm;
 }
-  right._grid->GlobalSum(inner);
+
-  return inner;
+template<class vobj>
+inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
+  GridBase *grid = left.Grid();
+  ComplexD nrm = rankInnerProduct(left,right);
+  grid->GlobalSum(nrm);
+  return nrm;
 }
 
+
 /////////////////////////
 // Fast axpby_norm
 // z = a x + b y
@@ -86,8 +201,7 @@ axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj
 template<class sobj,class vobj> strong_inline RealD 
 axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) 
 {
-  const int pad = 8;
+  z.Checkerboard() = x.Checkerboard();
-  z.checkerboard = x.checkerboard;
   conformable(z,x);
   conformable(x,y);
 
@@ -95,43 +209,79 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
-  GridBase *grid = x._grid;
+  GridBase *grid = x.Grid();
 
-  Vector<RealD> sumarray(grid->SumArraySize()*pad);
+  const uint64_t nsimd = grid->Nsimd();
+  const uint64_t sites = grid->oSites();
   
-  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
+  // GPU
-    int nwork, mywork, myoff;
+  autoView( x_v, x, AcceleratorRead);
-    GridThread::GetWork(x._grid->oSites(),thr,mywork,myoff);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
 
-    // private to thread; sub summation
+  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
-    decltype(innerProductD(z._odata[0],z._odata[0])) vnrm=zero; 
+  Vector<inner_t> inner_tmp(sites);
-    for(int ss=myoff;ss<mywork+myoff; ss++){
+  auto inner_tmp_v = &inner_tmp[0];
-      vobj tmp = a*x._odata[ss]+b*y._odata[ss];
-      vnrm = vnrm + innerProductD(tmp,tmp);
-      vstream(z._odata[ss],tmp);
-    }
-    vstream(sumarray[thr*pad],real(Reduce(TensorRemove(vnrm)))) ;
-  }
 
-  nrm = 0.0; // sum across threads; linear in thread count but fast
+  accelerator_for( ss, sites, 1,{
-  for(int i=0;i<grid->SumArraySize();i++){
+      auto tmp = a*x_v[ss]+b*y_v[ss];
-    nrm = nrm+sumarray[i*pad];
+      inner_tmp_v[ss]=innerProductD(tmp,tmp);
-  } 
+      z_v[ss]=tmp;
-  z._grid->GlobalSum(nrm);
+  });
+  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
+  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::scalar_type scalar_type;
+  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.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
+  ->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.first.func(eval(0,std::get<0>(expr.second)),eval(0,std::get<1>(expr.second))))::scalar_object
+      ->typename decltype(expr.op.func(eval(0,expr.arg1),eval(0,expr.arg2)))::scalar_object
 {
   return sum(closure(expr));
 }
@@ -139,54 +289,14 @@ inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
 
 template<class Op,class T1,class T2,class T3>
 inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
-  ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
+  ->typename decltype(expr.op.func(eval(0,expr.arg1),
-				      eval(0,std::get<1>(expr.second)),
+				      eval(0,expr.arg2),
-				      eval(0,std::get<2>(expr.second))
+				      eval(0,expr.arg3)
 				      ))::scalar_object
 {
   return sum(closure(expr));
 }
 
-template<class vobj>
-inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
-{
-  GridBase *grid=arg._grid;
-  int Nsimd = grid->Nsimd();
-  
-  std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
-  for(int i=0;i<grid->SumArraySize();i++){
-    sumarray[i]=zero;
-  }
-  
-  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
-    int nwork, mywork, myoff;
-    GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
-    
-    vobj vvsum=zero;
-    for(int ss=myoff;ss<mywork+myoff; ss++){
-      vvsum = vvsum + arg._odata[ss];
-    }
-    sumarray[thr]=vvsum;
-  }
-  
-  vobj vsum=zero;  // sum across threads
-  for(int i=0;i<grid->SumArraySize();i++){
-    vsum = vsum+sumarray[i];
-  } 
-  
-  typedef typename vobj::scalar_object sobj;
-  sobj ssum=zero;
-  
-  std::vector<sobj>               buf(Nsimd);
-  extract(vsum,buf);
-  
-  for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
-  arg._grid->GlobalSum(ssum);
-  
-  return ssum;
-}
-
-
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -199,7 +309,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   // But easily avoided by using double precision fields
   ///////////////////////////////////////////////////////
   typedef typename vobj::scalar_object sobj;
-  GridBase  *grid = Data._grid;
+  GridBase  *grid = Data.Grid();
   assert(grid!=NULL);
 
   const int    Nd = grid->_ndimension;
@@ -212,13 +322,13 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vobj,alignedAllocator<vobj> > lvSum(rd); // will locally sum vectors first
+  Vector<vobj> lvSum(rd); // will locally sum vectors first
-  std::vector<sobj> lsSum(ld,zero);                    // sum across these down to scalars
+  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
-  std::vector<sobj> extracted(Nsimd);                  // splitting the SIMD
+  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;
+    lvSum[r]=Zero();
   }
 
   int e1=    grid->_slice_nblock[orthogdim];
@@ -227,20 +337,19 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
 
   // sum over reduced dimension planes, breaking out orthog dir
   // Parallel over orthog direction
-  parallel_for(int r=0;r<rd;r++){
+  autoView( Data_v, Data, 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;
-	lvSum[r]=lvSum[r]+Data._odata[ss];
+	lvSum[r]=lvSum[r]+Data_v[ss];
-      }
       }
     }
+  });
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
-  std::vector<int> icoor(Nd);
+  Coordinate icoor(Nd);
 
   for(int rt=0;rt<rd;rt++){
 
@@ -265,7 +374,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
     if ( pt == grid->_processor_coor[orthogdim] ) {
       gsum=lsSum[lt];
     } else {
-      gsum=zero;
+      gsum=Zero();
     }
 
     grid->GlobalSum(gsum);
@@ -274,123 +383,14 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   }
 }
 
-template<class vobj>
-static void mySliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)
-{
-  // std::cout << GridLogMessage << "Start mySliceInnerProductVector" << std::endl;
-
-  typedef typename vobj::scalar_type scalar_type;
-  std::vector<scalar_type> lsSum;
-  localSliceInnerProductVector(result, lhs, rhs, lsSum, orthogdim);
-  globalSliceInnerProductVector(result, lhs, lsSum, orthogdim);
-  // std::cout << GridLogMessage << "End mySliceInnerProductVector" << std::endl;
-}
-
-template <class vobj>
-static void localSliceInnerProductVector(std::vector<ComplexD> &result, const Lattice<vobj> &lhs, const Lattice<vobj> &rhs, std::vector<typename vobj::scalar_type> &lsSum, int orthogdim)
-{
-  // std::cout << GridLogMessage << "Start prep" << std::endl;
-  typedef typename vobj::vector_type   vector_type;
-  typedef typename vobj::scalar_type   scalar_type;
-  GridBase  *grid = lhs._grid;
-  assert(grid!=NULL);
-  conformable(grid,rhs._grid);
-
-  const int    Nd = grid->_ndimension;
-  const int Nsimd = grid->Nsimd();
-
-  assert(orthogdim >= 0);
-  assert(orthogdim < Nd);
-
-  int fd=grid->_fdimensions[orthogdim];
-  int ld=grid->_ldimensions[orthogdim];
-  int rd=grid->_rdimensions[orthogdim];
-  // std::cout << GridLogMessage << "Start alloc" << std::endl;
-
-  std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
-  lsSum.resize(ld,scalar_type(0.0));                    // sum across these down to scalars
-  std::vector<iScalar<scalar_type>> extracted(Nsimd);   // splitting the SIMD  
-  // std::cout << GridLogMessage << "End alloc" << std::endl;
-
-  result.resize(fd); // And then global sum to return the same vector to every node for IO to file
-  for(int r=0;r<rd;r++){
-    lvSum[r]=zero;
-  }
-
-  int e1=    grid->_slice_nblock[orthogdim];
-  int e2=    grid->_slice_block [orthogdim];
-  int stride=grid->_slice_stride[orthogdim];
-  // std::cout << GridLogMessage << "End prep" << std::endl;
-  // std::cout << GridLogMessage << "Start parallel inner product, _rd = " << rd << std::endl;
-  vector_type vv;
-  parallel_for(int r=0;r<rd;r++)
-  {
-
-    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-        int ss = so + n * stride + b;
-        vv = TensorRemove(innerProduct(lhs._odata[ss], rhs._odata[ss]));
-        lvSum[r] = lvSum[r] + vv;
-      }
-    }
-  }
-  // std::cout << GridLogMessage << "End parallel inner product" << std::endl;
-
-  // Sum across simd lanes in the plane, breaking out orthog dir.
-  std::vector<int> icoor(Nd);
-  for(int rt=0;rt<rd;rt++){
-
-    iScalar<vector_type> temp; 
-    temp._internal = lvSum[rt];
-    extract(temp,extracted);
-
-    for(int idx=0;idx<Nsimd;idx++){
-
-      grid->iCoorFromIindex(icoor,idx);
-
-      int ldx =rt+icoor[orthogdim]*rd;
-
-      lsSum[ldx]=lsSum[ldx]+extracted[idx]._internal;
-
-    }
-  }
-  // std::cout << GridLogMessage << "End sum over simd lanes" << std::endl;
-}
-template <class vobj>
-static void globalSliceInnerProductVector(std::vector<ComplexD> &result, const Lattice<vobj> &lhs, std::vector<typename vobj::scalar_type> &lsSum, int orthogdim)
-{
-  typedef typename vobj::scalar_type scalar_type;
-  GridBase *grid = lhs._grid;
-  int fd = result.size();
-  int ld = lsSum.size();
-  // sum over nodes.
-  std::vector<scalar_type> gsum;
-  gsum.resize(fd, scalar_type(0.0));
-  // std::cout << GridLogMessage << "Start of gsum[t] creation:" << std::endl;
-  for(int t=0;t<fd;t++){
-    int pt = t/ld; // processor plane
-    int lt = t%ld;
-    if ( pt == grid->_processor_coor[orthogdim] ) {
-      gsum[t]=lsSum[lt];
-    }
-  }
-  // std::cout << GridLogMessage << "End of gsum[t] creation:" << std::endl;
-  // std::cout << GridLogMessage << "Start of GlobalSumVector:" << std::endl;
-  grid->GlobalSumVector(&gsum[0], fd);
-  // std::cout << GridLogMessage << "End of GlobalSumVector:" << std::endl;
-
-  result = gsum;
-}
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {
   typedef typename vobj::vector_type   vector_type;
   typedef typename vobj::scalar_type   scalar_type;
-  GridBase  *grid = lhs._grid;
+  GridBase  *grid = lhs.Grid();
   assert(grid!=NULL);
-  conformable(grid,rhs._grid);
+  conformable(grid,rhs.Grid());
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
@@ -402,34 +402,36 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
+  Vector<vector_type> lvSum(rd); // will locally sum vectors first
-  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
-  std::vector<iScalar<scalar_type> > extracted(Nsimd);                  // splitting the SIMD
+  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;
+    lvSum[r]=Zero();
   }
 
   int e1=    grid->_slice_nblock[orthogdim];
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
 
-  parallel_for(int r=0;r<rd;r++){
+  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(lhs._odata[ss],rhs._odata[ss]));
+	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.
-  std::vector<int> icoor(Nd);
+  Coordinate icoor(Nd);
   for(int rt=0;rt<rd;rt++){
 
     iScalar<vector_type> temp; 
@@ -470,7 +472,7 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
-  int Nblock = rhs._grid->GlobalDimensions()[Orthog];
+  int Nblock = rhs.Grid()->GlobalDimensions()[Orthog];
   std::vector<ComplexD> ip(Nblock);
   sn.resize(Nblock);
   
@@ -492,7 +494,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   
   scalar_type zscale(scale);
 
-  GridBase *grid  = X._grid;
+  GridBase *grid  = X.Grid();
 
   int Nsimd  =grid->Nsimd();
   int Nblock =grid->GlobalDimensions()[orthogdim];
@@ -505,8 +507,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   int e2     =grid->_slice_block [orthogdim];
   int stride =grid->_slice_stride[orthogdim];
 
-  std::vector<int> icoor;
+  Coordinate icoor;
-
   for(int r=0;r<rd;r++){
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
@@ -522,12 +523,13 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
 
     tensor_reduced at; at=av;
 
-    parallel_for_nest2(int n=0;n<e1;n++){
+    autoView( Rv, R, CpuWrite);
-      for(int b=0;b<e2;b++){
+    autoView( Xv, X, CpuRead);
+    autoView( Yv, Y, CpuRead);
+    thread_for2d( n, e1, b,e2, {
 	int ss= so+n*stride+b;
-	R._odata[ss] = at*X._odata[ss]+Y._odata[ss];
+	Rv[ss] = at*Xv[ss]+Yv[ss];
-      }
+    });
-    }
   }
 };
 
@@ -559,18 +561,18 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
-  int Nblock = X._grid->GlobalDimensions()[Orthog];
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
 
-  GridBase *FullGrid  = X._grid;
+  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 nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
-  int nl = nh-1;
+  //  int nl = nh-1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -578,28 +580,31 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int block =FullGrid->_slice_block [Orthog];
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
-#pragma omp parallel 
-  {
-    std::vector<vobj> s_x(Nblock);
 
-#pragma omp for collapse(2)
+  autoView( X_v, X, CpuRead);
-    for(int n=0;n<nblock;n++){
+  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[o+i*ostride];
+	s_x[i] = X_v[o+i*ostride];
       }
 
       vobj dot;
       for(int i=0;i<Nblock;i++){
-	dot = Y[o+i*ostride];
+	dot = Y_v[o+i*ostride];
 	for(int j=0;j<Nblock;j++){
 	  dot = dot + s_x[j]*(scale*aa(j,i));
 	}
-	R[o+i*ostride]=dot;
+	R_v[o+i*ostride]=dot;
       }
-    }}
+    }});
   }
 };
 
@@ -610,35 +615,38 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
-  int Nblock = X._grid->GlobalDimensions()[Orthog];
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
 
-  GridBase *FullGrid  = X._grid;
+  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 nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
-  int nl=1;
+  //  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];
-#pragma omp parallel 
+  autoView( R_v, R, CpuWrite);
+  autoView( X_v, X, CpuRead);
+  thread_region
   {
     std::vector<vobj> s_x(Nblock);
 
-#pragma omp for collapse(2)
+
-    for(int n=0;n<nblock;n++){
+    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[o+i*ostride];
+	s_x[i] = X_v[o+i*ostride];
       }
 
       vobj dot;
@@ -647,11 +655,10 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
 	for(int j=1;j<Nblock;j++){
 	  dot = dot + s_x[j]*(scale*aa(j,i));
 	}
-	R[o+i*ostride]=dot;
+	R_v[o+i*ostride]=dot;
       }
-    }}
+    }});
   }
-
 };
 
 
@@ -662,7 +669,7 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
-  GridBase *FullGrid  = lhs._grid;
+  GridBase *FullGrid  = lhs.Grid();
   //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
   
   int Nblock = FullGrid->GlobalDimensions()[Orthog];
@@ -673,9 +680,9 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
-  int nh =  FullGrid->_ndimension;
+  //  int nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
-  int nl = nh-1;
+  //  int nl = nh-1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -686,31 +693,33 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
 
   typedef typename vobj::vector_typeD vector_typeD;
 
-#pragma omp parallel 
+  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);
 
-#pragma omp for collapse(2)
+    thread_for_collapse_in_region( 2, n,nblock,{
-    for(int n=0;n<nblock;n++){
     for(int b=0;b<block;b++){
 
       int o  = n*stride + b;
 
       for(int i=0;i<Nblock;i++){
-	Left [i] = lhs[o+i*ostride];
+	Left [i] = lhs_v[o+i*ostride];
-	Right[i] = rhs[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);
-	mat_thread(i,j) += Reduce(rtmp);
+	auto red  =  Reduce(rtmp);
+	mat_thread(i,j) += std::complex<double>(real(red),imag(red));
       }}
-    }}
+    }});
-#pragma omp critical
+    thread_critical
     {
       mat += mat_thread;
     }  
@@ -726,8 +735,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   return;
 }
 
-} /*END NAMESPACE GRID*/
+NAMESPACE_END(Grid);
-#endif
+
 
 
 

Grid/lattice/Lattice_reduction_gpu.h

@@ -0,0 +1,232 @@
+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>
+void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+  
+  int device;
+#ifdef GRID_CUDA
+  cudaGetDevice(&device);
+#endif
+#ifdef GRID_HIP
+  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 << "\tmaxThreadsPerBlock  = " << warpSize << 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;
+  while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
+  // keep all the streaming multiprocessors busy
+  blocks = nextPow2(multiProcessorCount);
+  
+}
+
+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(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;
+  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  Integer smemSize = numThreads * sizeof(sobj);
+
+  Vector<sobj> buffer(numBlocks);
+  sobj *buffer_v = &buffer[0];
+  
+  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  accelerator_barrier();
+  auto result = buffer_v[0];
+  return result;
+}
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// 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;
+}
+
+
+
+NAMESPACE_END(Grid);