Fixed test for very much non-unit det

Revised dslash/dwf kernels for A64FX

.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

Grid/DisableWarnings.h

@@ -37,7 +37,9 @@ directory
 #endif
 
  //disables and intel compiler specific warning (in json.hpp)
+#ifdef __ICC
 #pragma warning disable 488  
+#endif
 
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp

Grid/GridCore.h

@@ -47,9 +47,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
-#include <Grid/allocator/AlignedAllocator.h>
+#include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
-#include <Grid/threads/Threads.h>
+#include <Grid/threads/ThreadReduction.h>
 #include <Grid/serialisation/Serialisation.h>
 #include <Grid/util/Sha.h>
 #include <Grid/communicator/Communicator.h> 

Grid/GridStd.h

@@ -6,6 +6,7 @@
 ///////////////////
 #include <cassert>
 #include <complex>
+#include <memory>
 #include <vector>
 #include <array>
 #include <string>
@@ -27,4 +28,7 @@
 ///////////////////
 #include "Config.h"
 
+#ifdef TOFU
+#undef GRID_COMMS_THREADS
+#endif
 #endif /* GRID_STD_H */

Grid/Grid_Eigen_Dense.h

@@ -18,12 +18,29 @@
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")
+#undef __CUDA_ARCH__
 #undef __NVCC__
 #undef __CUDACC__
-#undef __CUDA_ARCH__
 #define __NVCC__REDEFINE__
+#endif 
+
+/* SYCL save and restore compile environment*/
+#ifdef 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>
 
@@ -35,7 +52,20 @@
 #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/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/algorithms/Algorithms.h

@@ -29,9 +29,11 @@ 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>
@@ -39,14 +41,20 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #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>
@@ -58,7 +66,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #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>
 
 #endif

Grid/algorithms/FFT.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -37,7 +36,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif
 #endif
 
-
 NAMESPACE_BEGIN(Grid);
 
 template<class scalar> struct FFTW { };
@@ -191,7 +189,7 @@ public:
     typedef typename sobj::scalar_type   scalar;
       
     Lattice<sobj> pgbuf(&pencil_g);
-    auto pgbuf_v = pgbuf.View();
+    autoView(pgbuf_v , pgbuf, CpuWrite);
 
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -232,15 +230,18 @@ public:
     result = source;
     int pc = processor_coor[dim];
     for(int p=0;p<processors[dim];p++) {
-      thread_for(idx, sgrid->lSites(),{
+      {
+	autoView(r_v,result,CpuRead);
+	autoView(p_v,pgbuf,CpuWrite);
+	thread_for(idx, sgrid->lSites(),{
           Coordinate cbuf(Nd);
           sobj s;
 	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,result,cbuf);
+	  peekLocalSite(s,r_v,cbuf);
 	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  //            cbuf[dim]+=p*L;
-	  pokeLocalSite(s,pgbuf,cbuf);
-      });
+	  pokeLocalSite(s,p_v,cbuf);
+        });
+      }
       if (p != processors[dim] - 1) {
 	result = Cshift(result,dim,L);
       }
@@ -269,15 +270,19 @@ public:
     flops+= flops_call*NN;
       
     // writing out result
-    thread_for(idx,sgrid->lSites(),{
+    {
+      autoView(pgbuf_v,pgbuf,CpuRead);
+      autoView(result_v,result,CpuWrite);
+      thread_for(idx,sgrid->lSites(),{
 	Coordinate clbuf(Nd), cgbuf(Nd);
 	sobj s;
 	sgrid->LocalIndexToLocalCoor(idx,clbuf);
 	cgbuf = clbuf;
 	cgbuf[dim] = clbuf[dim]+L*pc;
-	peekLocalSite(s,pgbuf,cgbuf);
-	pokeLocalSite(s,result,clbuf);
-    });
+	peekLocalSite(s,pgbuf_v,cgbuf);
+	pokeLocalSite(s,result_v,clbuf);
+      });
+    }
     result = result*div;
       
     // destroying plan

Grid/algorithms/LinearOperator.h

@@ -43,7 +43,6 @@ NAMESPACE_BEGIN(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
@@ -94,7 +93,10 @@ public:
     _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){
     _Mat.MdagM(in,out);
@@ -131,17 +133,14 @@ public:
     assert(0);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    _Mat.MdagM(in,out,n1,n2);
-    out = out + _shift*in;
-
-    ComplexD dot;	
-    dot= innerProduct(in,out);
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
     n1=real(dot);
     n2=norm2(out);
   }
   void HermOp(const Field &in, Field &out){
-    RealD n1,n2;
-    HermOpAndNorm(in,out,n1,n2);
+    _Mat.MdagM(in,out);
+    out = out + _shift*in;
   }
 };
 
@@ -170,7 +169,7 @@ public:
     _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);
   }
@@ -208,212 +207,305 @@ public:
   }
 };
 
-    //////////////////////////////////////////////////////////
-    // Even Odd Schur decomp operators; there are several
-    // ways to introduce the even odd checkerboarding
-    //////////////////////////////////////////////////////////
+//////////////////////////////////////////////////////////
+// Even Odd Schur decomp operators; there are several
+// ways to introduce the even odd checkerboarding
+//////////////////////////////////////////////////////////
 
-    template<class Field>
-    class SchurOperatorBase :  public LinearOperatorBase<Field> {
-    public:
-      virtual  RealD Mpc      (const Field &in, Field &out) =0;
-      virtual  RealD MpcDag   (const Field &in, Field &out) =0;
-      virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
-      Field tmp(in.Grid());
-      tmp.Checkerboard() = in.Checkerboard();
-	ni=Mpc(in,tmp);
-	no=MpcDag(tmp,out);
-      }
-      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-      out.Checkerboard() = in.Checkerboard();
-	MpcDagMpc(in,out,n1,n2);
-      }
-      virtual void HermOp(const Field &in, Field &out){
-	RealD n1,n2;
-	HermOpAndNorm(in,out,n1,n2);
-      }
-      void Op     (const Field &in, Field &out){
-	Mpc(in,out);
-      }
-      void AdjOp     (const Field &in, Field &out){ 
-	MpcDag(in,out);
-      }
-      // Support for coarsening to a multigrid
-      void OpDiag (const Field &in, Field &out) {
-	assert(0); // must coarsen the unpreconditioned system
-      }
-      void OpDir  (const Field &in, Field &out,int dir,int disp) {
-	assert(0);
-      }
-      void OpDirAll  (const Field &in, std::vector<Field> &out){
-	assert(0);
-      };
-    };
-    template<class Matrix,class Field>
-    class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
-    public:
-      Matrix &_Mat;
-      SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
-      virtual  RealD Mpc      (const Field &in, Field &out) {
-      Field tmp(in.Grid());
-      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);
-      }
-      virtual  RealD 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);
-	return axpy_norm(out,-1.0,tmp,out);
-      }
-    };
-    template<class Matrix,class Field>
-      class SchurDiagOneOperator :  public SchurOperatorBase<Field> {
-    protected:
-      Matrix &_Mat;
-    public:
-      SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
-
-      virtual  RealD 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);
-
-	return axpy_norm(out,-1.0,tmp,in);
-      }
-      virtual  RealD 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);
-
-	return axpy_norm(out,-1.0,tmp,in);
-      }
-    };
-    template<class Matrix,class Field>
-      class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
-    protected:
-      Matrix &_Mat;
-    public:
-      SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
-
-      virtual  RealD 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);
-
-	return axpy_norm(out,-1.0,tmp,in);
-      }
-      virtual  RealD 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);
-
-	return axpy_norm(out,-1.0,tmp,in);
-      }
-    };
-    ///////////////////////////////////////////////////////////////////////////////////////////////////
-    // Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
-    // Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi
-    ///////////////////////////////////////////////////////////////////////////////////////////////////
-    template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
-    template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
-    ///////////////////////////////////////////////////////////////////////////////////////////////////
-    //  Staggered use
-    ///////////////////////////////////////////////////////////////////////////////////////////////////
-    template<class Matrix,class Field>
-      class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
-    protected:
-      Matrix &_Mat;
-      Field tmp;
-      RealD mass;
-      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;
-      }
+template<class Field>
+class SchurOperatorBase :  public LinearOperatorBase<Field> {
+ public:
+  virtual  void Mpc      (const Field &in, Field &out) =0;
+  virtual  void MpcDag   (const Field &in, Field &out) =0;
+  virtual  void MpcDagMpc(const Field &in, Field &out) {
+    Field tmp(in.Grid());
+    tmp.Checkerboard() = in.Checkerboard();
+    Mpc(in,tmp);
+    MpcDag(tmp,out);
+  }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	ncall++;
-	tMpc-=usecond();
-    n2 = Mpc(in,out);
-	tMpc+=usecond();
-	tIP-=usecond();
-    ComplexD dot= innerProduct(in,out);
-	tIP+=usecond();
-    n1 = real(dot);
+    out.Checkerboard() = in.Checkerboard();
+    MpcDagMpc(in,out);
+    ComplexD dot= innerProduct(in,out); 
+    n1=real(dot);
+    n2=norm2(out);
   }
   virtual void HermOp(const Field &in, Field &out){
-	ncall++;
-	tMpc-=usecond();
-	_Mat.Meooe(in,out);
-	_Mat.Meooe(out,tmp);
-	tMpc+=usecond();
-	taxpby_norm-=usecond();
-	axpby(out,-1.0,mass*mass,tmp,in);
-	taxpby_norm+=usecond();
+    out.Checkerboard() = in.Checkerboard();
+    MpcDagMpc(in,out);
   }
-  virtual  RealD Mpc      (const Field &in, Field &out) 
-  {
+  void Op     (const Field &in, Field &out){
+    Mpc(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){ 
+    MpcDag(in,out);
+  }
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    assert(0); // must coarsen the unpreconditioned system
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    assert(0);
+  };
+};
+template<class Matrix,class Field>
+  class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
+ public:
+    Matrix &_Mat;
+    SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
+    virtual  void Mpc      (const Field &in, Field &out) {
+      Field tmp(in.Grid());
+      tmp.Checkerboard() = !in.Checkerboard();
+      
+      _Mat.Meooe(in,tmp);
+      _Mat.MooeeInv(tmp,out);
+      _Mat.Meooe(out,tmp);
+      _Mat.Mooee(in,out);
+      axpy(out,-1.0,tmp,out);
+    }
+    virtual void MpcDag   (const Field &in, Field &out){
+      Field tmp(in.Grid());
+	
+      _Mat.MeooeDag(in,tmp);
+      _Mat.MooeeInvDag(tmp,out);
+      _Mat.MeooeDag(out,tmp);
+      _Mat.MooeeDag(in,out);
+      axpy(out,-1.0,tmp,out);
+    }
+};
+template<class Matrix,class Field>
+  class SchurDiagOneOperator :  public SchurOperatorBase<Field> {
+ protected:
+    Matrix &_Mat;
+ public:
+    SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
+    
+    virtual void Mpc      (const Field &in, Field &out) {
+      Field tmp(in.Grid());
 
+      _Mat.Meooe(in,out);
+      _Mat.MooeeInv(out,tmp);
+      _Mat.Meooe(tmp,out);
+      _Mat.MooeeInv(out,tmp);
+      axpy(out,-1.0,tmp,in);
+    }
+    virtual void MpcDag   (const Field &in, Field &out){
+      Field tmp(in.Grid());
+      
+      _Mat.MooeeInvDag(in,out);
+      _Mat.MeooeDag(out,tmp);
+      _Mat.MooeeInvDag(tmp,out);
+      _Mat.MeooeDag(out,tmp);
+      axpy(out,-1.0,tmp,in);
+    }
+};
+template<class Matrix,class Field>
+  class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
+ protected:
+    Matrix &_Mat;
+ public:
+    SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
+    
+    virtual void Mpc      (const Field &in, Field &out) {
+      Field tmp(in.Grid());
+      
+      _Mat.MooeeInv(in,out);
+      _Mat.Meooe(out,tmp);
+      _Mat.MooeeInv(tmp,out);
+      _Mat.Meooe(out,tmp);
+      
+      axpy(out,-1.0,tmp,in);
+    }
+    virtual  void MpcDag   (const Field &in, Field &out){
+      Field tmp(in.Grid());
+
+      _Mat.MeooeDag(in,out);
+      _Mat.MooeeInvDag(out,tmp);
+      _Mat.MeooeDag(tmp,out);
+      _Mat.MooeeInvDag(out,tmp);
+
+      axpy(out,-1.0,tmp,in);
+    }
+};
+
+template<class Field>
+class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field> 
+{
+ public:
+  virtual void  Mpc      (const Field& in, Field& out) = 0;
+  virtual void  MpcDag   (const Field& in, Field& out) = 0;
+  virtual void  MpcDagMpc(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+    tmp.Checkerboard() = in.Checkerboard();
+    Mpc(in,tmp);
+    MpcDag(tmp,out);
+  }
+  virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
+    assert(0);
+  }
+  virtual void HermOp(const Field& in, Field& out) {
+    assert(0);
+  }
+  void Op(const Field& in, Field& out) {
+    Mpc(in, out);
+  }
+  void AdjOp(const Field& in, Field& out) { 
+    MpcDag(in, out);
+  }
+  // Support for coarsening to a multigrid
+  void OpDiag(const Field& in, Field& out) {
+    assert(0); // must coarsen the unpreconditioned system
+  }
+  void OpDir(const Field& in, Field& out, int dir, int disp) {
+    assert(0);
+  }
+  void OpDirAll(const Field& in, std::vector<Field>& out){
+    assert(0);
+  };
+};
+
+template<class Matrix, class Field>
+class NonHermitianSchurDiagMooeeOperator :  public NonHermitianSchurOperatorBase<Field> 
+{
+ public:
+  Matrix& _Mat;
+ NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
+  virtual void Mpc(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+    tmp.Checkerboard() = !in.Checkerboard();
+    
+    _Mat.Meooe(in, tmp);
+    _Mat.MooeeInv(tmp, out);
+    _Mat.Meooe(out, tmp);
+    
+    _Mat.Mooee(in, out);
+    
+    axpy(out, -1.0, tmp, out);
+  }
+  virtual void MpcDag(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+    
+    _Mat.MeooeDag(in, tmp);
+    _Mat.MooeeInvDag(tmp, out);
+    _Mat.MeooeDag(out, tmp);
+	  
+    _Mat.MooeeDag(in, out);
+    
+    axpy(out, -1.0, tmp, out);
+  }
+};
+    
+template<class Matrix,class Field>
+class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Field> 
+{
+ protected:
+  Matrix &_Mat;
+  
+ public:
+  NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
+  virtual void Mpc(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+	  
+    _Mat.Meooe(in, out);
+    _Mat.MooeeInv(out, tmp);
+    _Mat.Meooe(tmp, out);
+    _Mat.MooeeInv(out, tmp);
+
+    axpy(out, -1.0, tmp, in);
+  }
+  virtual void MpcDag(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+    
+    _Mat.MooeeInvDag(in, out);
+    _Mat.MeooeDag(out, tmp);
+    _Mat.MooeeInvDag(tmp, out);
+    _Mat.MeooeDag(out, tmp);
+    
+    axpy(out, -1.0, tmp, in);
+  }
+};
+
+template<class Matrix, class Field>
+class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Field> 
+{
+ protected:
+  Matrix& _Mat;
+  
+ public:
+ NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
+
+  virtual void Mpc(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+    
+    _Mat.MooeeInv(in, out);
+    _Mat.Meooe(out, tmp);
+    _Mat.MooeeInv(tmp, out);
+    _Mat.Meooe(out, tmp);
+
+    axpy(out, -1.0, tmp, in);
+  }
+  virtual void MpcDag(const Field& in, Field& out) {
+    Field tmp(in.Grid());
+    
+    _Mat.MeooeDag(in, out);
+    _Mat.MooeeInvDag(out, tmp);
+    _Mat.MeooeDag(tmp, out);
+    _Mat.MooeeInvDag(out, tmp);
+
+    axpy(out, -1.0, tmp, in);
+  }
+};
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
+// Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi
+///////////////////////////////////////////////////////////////////////////////////////////////////
+template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
+template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
+///////////////////////////////////////////////////////////////////////////////////////////////////
+//  Staggered use
+///////////////////////////////////////////////////////////////////////////////////////////////////
+template<class Matrix,class Field>
+class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
+ protected:
+  Matrix &_Mat;
+  Field tmp;
+  RealD mass;
+ public:
+  SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
+  { 
+    assert( _Mat.isTrivialEE() );
+    mass = _Mat.Mass();
+  }
+  virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    Mpc(in,out);
+    ComplexD dot= innerProduct(in,out);
+    n1 = real(dot);
+    n2 =0.0;
+  }
+  virtual void HermOp(const Field &in, Field &out){
+    Mpc(in,out);
+    //    _Mat.Meooe(in,out);
+    //    _Mat.Meooe(out,tmp);
+    //    axpby(out,-1.0,mass*mass,tmp,in);
+  }
+  virtual  void Mpc      (const Field &in, Field &out) 
+  {
     Field tmp(in.Grid());
     Field tmp2(in.Grid());
+	
+    //    _Mat.Mooee(in,out);
+    //    _Mat.Mooee(out,tmp);
 
-    //    std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
-    _Mat.Mooee(in,out);
-    _Mat.Mooee(out,tmp);
-    //    std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
-
-    tMeo-=usecond();
     _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;
+    axpby(out,-1.0,mass*mass,tmp,in);
   }
-  virtual  RealD MpcDag   (const Field &in, Field &out){
-    return Mpc(in,out);
+  virtual  void MpcDag   (const Field &in, Field &out){
+    Mpc(in,out);
   }
   virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
     assert(0);// Never need with staggered
@@ -421,7 +513,6 @@ public:
 };
 template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
 
-
 /////////////////////////////////////////////////////////////
 // Base classes for functions of operators
 /////////////////////////////////////////////////////////////

Grid/algorithms/SparseMatrix.h

@@ -38,16 +38,12 @@ template<class Field> class SparseMatrixBase {
 public:
   virtual GridBase *Grid(void) =0;
   // Full checkerboar operations
-  virtual RealD M    (const Field &in, Field &out)=0;
-  virtual RealD Mdag (const Field &in, Field &out)=0;
-  virtual void  MdagM(const Field &in, Field &out,RealD &ni,RealD &no) {
-    Field tmp (in.Grid());
-    ni=M(in,tmp);
-    no=Mdag(tmp,out);
-  }
+  virtual void  M    (const Field &in, Field &out)=0;
+  virtual void  Mdag (const Field &in, Field &out)=0;
   virtual void  MdagM(const Field &in, Field &out) {
-    RealD ni, no;
-    MdagM(in,out,ni,no);
+    Field tmp (in.Grid());
+    M(in,tmp);
+    Mdag(tmp,out);
   }
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;

Grid/algorithms/approx/Chebyshev.h

@@ -234,10 +234,8 @@ public:
 
     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); 
@@ -258,14 +256,28 @@ public:
     //    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);
-      //     y=xscale*y+mscale*(*Tn);
-      //      *Tnp=2.0*y-(*Tnm);
-      //      out=out+Coeffs[n]* (*Tnp);
+#if 0
+      auto y_v = y.View();
+      auto Tn_v = Tn->View();
+      auto Tnp_v = Tnp->View();
+      auto Tnm_v = Tnm->View();
+      constexpr int Nsimd = vector_type::Nsimd();
+      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
+	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
+	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
+      });
+      if ( Coeffs[n] != 0.0) {
+	axpy(out,Coeffs[n],*Tnp,out);
+      }
+#else
       axpby(y,xscale,mscale,y,(*Tn));
       axpby(*Tnp,2.0,-1.0,y,(*Tnm));
-      axpy(out,Coeffs[n],*Tnp,out);
+      if ( Coeffs[n] != 0.0) {
+	axpy(out,Coeffs[n],*Tnp,out);
+      }
+#endif
       // Cycle pointers to avoid copies
       Field *swizzle = Tnm;
       Tnm    =Tn;

Grid/algorithms/approx/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/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/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

@@ -52,6 +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)
@@ -306,7 +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;
@@ -442,7 +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;
@@ -653,7 +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 ) { 
 
@@ -668,7 +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;

Grid/algorithms/iterative/ConjugateGradient.h

@@ -49,6 +49,7 @@ public:
   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),
@@ -81,6 +82,14 @@ public:
     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;
@@ -92,6 +101,7 @@ public:
 
     // 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;
@@ -130,18 +140,20 @@ public:
       b = cp / c;
 
       LinearCombTimer.Start();
-      auto psi_v = psi.View();
-      auto p_v   = p.View();
-      auto r_v   = r.View();
-      accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
-	  coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
-	  coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
-      });
+      {
+	autoView( psi_v , psi, AcceleratorWrite);
+	autoView( p_v   , p,   AcceleratorWrite);
+	autoView( r_v   , r,   AcceleratorWrite);
+	accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+	    coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+	    coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+	});
+      }
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
       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) {
@@ -169,10 +181,17 @@ public:
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 
 	IterationsToComplete = k;	
+	TrueResidual = true_residual;
 
         return;
       }
     }
+    // Failed. Calculate true residual before giving up                                                         
+    Linop.HermOpAndNorm(psi, mmp, d, qq);
+    p = mmp - src;
+
+    TrueResidual = sqrt(norm2(p)/ssq);
+
     std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
 
     if (ErrorOnNoConverge) assert(0);

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -46,15 +46,19 @@ public:
 
   RealD   Tolerance;
   Integer MaxIterations;
-    Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  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)
@@ -125,6 +129,17 @@ public:
     // 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
@@ -270,6 +285,7 @@ public:
       for(int s=0;s<nshift;s++){
       
 	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
 	
 	  RealD css  = c * z[s][iz]* z[s][iz];
 	
@@ -299,7 +315,8 @@ public:
 	  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;

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -37,211 +37,6 @@ Author: Christoph Lehner <clehner@bnl.gov>
 
 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) 
-{
-  // 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 Field>
-void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm) 
-{
-  typedef decltype(basis[0].View()) View;
-  auto tmp_v = basis[0].View();
-  Vector<View> basis_v(basis.size(),tmp_v);
-  typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0].Grid();
-
-  for(int k=0;k<basis.size();k++){
-    basis_v[k] = basis[k].View();
-  }
-#if 0
-  std::vector < vobj , commAllocator<vobj> > Bt(thread_max() * Nm); // Thread private
-  thread_region
-  {
-    vobj* B = Bt.data() + 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
-
-  int nrot = j1-j0;
-
-
-  uint64_t oSites   =grid->oSites();
-  uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
-
-  //  printf("BasisRotate %d %d nrot %d siteBlock %d\n",j0,j1,nrot,siteBlock);
-
-  Vector <vobj> Bt(siteBlock * nrot); 
-  auto Bp=&Bt[0];
-
-  // GPU readable copy of Eigen matrix
-  Vector<double> Qt_jv(Nm*Nm);
-  double *Qt_p = & Qt_jv[0];
-  for(int k=0;k<Nm;++k){
-    for(int j=0;j<Nm;++j){
-      Qt_p[j*Nm+k]=Qt(j,k);
-    }
-  }
-
-  // Block the loop to keep storage footprint down
-  vobj zz=Zero();
-  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(),{
-	auto z=coalescedRead(zz);
-	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
-}
-
-// 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()) View;
-  typedef typename Field::vector_object vobj;
-  GridBase* grid = basis[0].Grid();
-
-  result.Checkerboard() = basis[0].Checkerboard();
-  auto result_v=result.View();
-  Vector<View> basis_v(basis.size(),result_v);
-  for(int k=0;k<basis.size();k++){
-    basis_v[k] = basis[k].View();
-  }
-  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);
-  accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    auto B=coalescedRead(zz);
-    for(int k=k0; k<k1; ++k){
-      B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
-    }
-    coalescedWrite(result_v[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);
-
-      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);
-  }
-}
-
 /////////////////////////////////////////////////////////////
 // Implicitly restarted lanczos
 /////////////////////////////////////////////////////////////

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/SchurRedBlack.h

@@ -405,6 +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
@@ -482,5 +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

@@ -26,118 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-#ifndef GRID_ALIGNED_ALLOCATOR_H
-#define GRID_ALIGNED_ALLOCATOR_H
-
-#ifdef HAVE_MALLOC_MALLOC_H
-#include <malloc/malloc.h>
-#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_MM_MALLOC_H
-#include <mm_malloc.h>
-#endif
-
-#define POINTER_CACHE
-#define GRID_ALLOC_ALIGN (2*1024*1024)
+#pragma once
 
 NAMESPACE_BEGIN(Grid);
 
-// Move control to configure.ac and Config.h?
-#ifdef POINTER_CACHE
-class PointerCache {
-private:
-/*Pinning pages is costly*/
-/*Could maintain separate large and small allocation caches*/
-#ifdef GRID_NVCC 
-  static const int Ncache=128;
-#else
-  static const int Ncache=8;
-#endif
-  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) ;
-
-};
-#endif  
-
-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 {
 public: 
@@ -161,89 +53,131 @@ public:
   { 
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
-
-
-#ifdef POINTER_CACHE
-    _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
-#else
-    pointer ptr = nullptr;
-#endif
-
-#ifdef GRID_NVCC
-    ////////////////////////////////////
-    // Unified (managed) memory
-    ////////////////////////////////////
-    if ( ptr == (_Tp *) NULL ) {
-      //      printf(" alignedAllocater cache miss %ld bytes ",bytes);      BACKTRACEFP(stdout);
-      auto err = cudaMallocManaged((void **)&ptr,bytes);
-      if( err != cudaSuccess ) {
-	ptr = (_Tp *) NULL;
-	std::cerr << " cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
-	assert(0);
-      }
-    } 
-    assert( ptr != (_Tp *)NULL);
-#else 
-    //////////////////////////////////////////////////////////////////////////////////////////
-    // 2MB align; could make option probably doesn't need configurability
-    //////////////////////////////////////////////////////////////////////////////////////////
-  #ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
-  #else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
-  #endif
-    assert( ptr != (_Tp *)NULL);
-
-    //////////////////////////////////////////////////
-    // First touch optimise in threaded loop 
-    //////////////////////////////////////////////////
-    uint64_t *cp = (uint64_t *)ptr;
-    thread_for(n,bytes/sizeof(uint64_t), { // need only one touch per page
-      cp[n]=0;
-    });
-#endif
+    _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
-  void deallocate(pointer __p, size_type __n) { 
+  void deallocate(pointer __p, size_type __n) 
+  { 
     size_type bytes = __n * sizeof(_Tp);
-
     profilerFree(bytes);
-
-#ifdef POINTER_CACHE
-    pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
-#else 
-    pointer __freeme = __p;
-#endif
-
-#ifdef GRID_NVCC
-    if ( __freeme ) cudaFree((void *)__freeme);
-#else 
-  #ifdef HAVE_MM_MALLOC_H
-    if ( __freeme ) _mm_free((void *)__freeme); 
-  #else
-    if ( __freeme ) free((void *)__freeme);
-  #endif
-#endif
+    MemoryManager::CpuFree((void *)__p,bytes);
   }
 
-  // FIXME: hack for the copy constructor, eventually it must be avoided
-  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
-  //void construct(pointer __p, const _Tp& __val) { };
+  // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
+  void construct(pointer __p, const _Tp& __val) { assert(0);};
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
 template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
 template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
 
+//////////////////////////////////////////////////////////////////////////////////////
+// Unified virtual memory
+//////////////////////////////////////////////////////////////////////////////////////
+template<typename _Tp>
+class uvmAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef uvmAllocator<_Tp1> other; };
+  uvmAllocator() throw() { }
+  uvmAllocator(const uvmAllocator&) throw() { }
+  template<typename _Tp1> uvmAllocator(const uvmAllocator<_Tp1>&) throw() { }
+  ~uvmAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+  pointer allocate(size_type __n, const void* _p= 0)
+  { 
+    size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
+    _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
+    profilerFree(bytes);
+    MemoryManager::SharedFree((void *)__p,bytes);
+  }
+
+  void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
+
+////////////////////////////////////////////////////////////////////////////////
+// Device memory
+////////////////////////////////////////////////////////////////////////////////
+template<typename _Tp>
+class devAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef devAllocator<_Tp1> other; };
+  devAllocator() throw() { }
+  devAllocator(const devAllocator&) throw() { }
+  template<typename _Tp1> devAllocator(const devAllocator<_Tp1>&) throw() { }
+  ~devAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+  pointer allocate(size_type __n, const void* _p= 0)
+  { 
+    size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
+    _Tp *ptr = (_Tp*) MemoryManager::AcceleratorAllocate(bytes);
+    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type __n) 
+  { 
+    size_type bytes = __n * sizeof(_Tp);
+    profilerFree(bytes);
+    MemoryManager::AcceleratorFree((void *)__p,bytes);
+  }
+  void construct(pointer __p, const _Tp& __val) { };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return false; }
+
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using commAllocator = alignedAllocator<T>;
-template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
-template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
+#ifdef ACCELERATOR_CSHIFT
+// Cshift on device
+template<class T> using cshiftAllocator = devAllocator<T>;
+#else
+// Cshift on host
+template<class T> using cshiftAllocator = std::allocator<T>;
+#endif
+
+template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
+template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
+template<class T> using commVector = std::vector<T,devAllocator<T> >;
+template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);
 
-#endif
+

Grid/allocator/Allocator.h

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

Grid/allocator/MemoryManager.cc

@@ -0,0 +1,254 @@
+#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

@@ -6,72 +6,6 @@ NAMESPACE_BEGIN(Grid);
 MemoryStats *MemoryProfiler::stats = nullptr;
 bool         MemoryProfiler::debug = false;
 
-#ifdef GRID_NVCC
-#define SMALL_LIMIT (0)
-#else
-#define SMALL_LIMIT (4096)
-#endif
-
-#ifdef POINTER_CACHE
-int PointerCache::victim;
-
-PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
-
-void *PointerCache::Insert(void *ptr,size_t bytes) {
-
-  if (bytes < SMALL_LIMIT ) 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 < SMALL_LIMIT ) return NULL;
-
-#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;
-}
-#endif
-
 void check_huge_pages(void *Buf,uint64_t BYTES)
 {
 #ifdef __linux__

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

@@ -81,6 +81,7 @@ public:
 
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
+  Coordinate _checker_dim_mask;
 
 public:
 

Grid/cartesian/Cartesian_full.h

@@ -38,6 +38,7 @@ class GridCartesian: public GridBase {
 
 public:
   int dummy;
+  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -104,6 +105,7 @@ public:
     _ldimensions.resize(_ndimension);
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
+    _checker_dim_mask.resize(_ndimension);;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -114,6 +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];

Grid/cartesian/Cartesian_red_black.h

@@ -35,12 +35,28 @@ 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:
-  Coordinate _checker_dim_mask;
+  //  Coordinate _checker_dim_mask;
   int              _checker_dim;
   std::vector<int> _checker_board;
 

Grid/communicator/Communicator_base.h

@@ -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,

Grid/communicator/Communicator_mpi3.cc

@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/communicator/Communicator_mpi.cc
 
@@ -35,7 +35,7 @@ Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 ////////////////////////////////////////////
 // First initialise of comms system
 ////////////////////////////////////////////
-void CartesianCommunicator::Init(int *argc, char ***argv) 
+void CartesianCommunicator::Init(int *argc, char ***argv)
 {
 
   int flag;
@@ -43,8 +43,16 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
 
+#ifndef GRID_COMMS_THREADS
+    nCommThreads=1;
+    // wrong results here too
+    // For now: comms-overlap leads to wrong results in Benchmark_wilson even on single node MPI runs
+    // other comms schemes are ok
+    MPI_Init_thread(argc,argv,MPI_THREAD_SERIALIZED,&provided);
+#else
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+#endif
     //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
     if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
       assert(0);
@@ -91,7 +99,7 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Initialises from communicator_world
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) 
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   MPI_Comm optimal_comm;
   ////////////////////////////////////////////////////
@@ -110,7 +118,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)    
+CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
 {
   _ndimension = processors.size();  assert(_ndimension>=1);
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
@@ -127,7 +135,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
-  //  int Nparent = parent._processors ; 
+  //  int Nparent = parent._processors ;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
 
@@ -149,13 +157,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
-  int crank;  
+  int crank;
   // Mpi uses the reverse Lexico convention to us; so reversed routines called
   Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
   Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
 
   MPI_Comm comm_split;
-  if ( Nchild > 1 ) { 
+  if ( Nchild > 1 ) {
 
     ////////////////////////////////////////////////////////////////
     // Split the communicator
@@ -180,11 +188,11 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   SetCommunicator(comm_split);
 
   ///////////////////////////////////////////////
-  // Free the temp communicator 
+  // Free the temp communicator
   ///////////////////////////////////////////////
   MPI_Comm_free(&comm_split);
 
-  if(0){ 
+  if(0){
     std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
     for(int d=0;d<processors.size();d++){
       std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
@@ -245,7 +253,7 @@ CartesianCommunicator::~CartesianCommunicator()
     for(int i=0;i<communicator_halo.size();i++){
       MPI_Comm_free(&communicator_halo[i]);
     }
-  }  
+  }
 }
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
@@ -255,6 +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);
@@ -290,60 +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  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)
-{
+  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+
   int myrank = _processor;
   int ierr;
 
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    MPI_Request xrq;
-    MPI_Request rrq;
+  // Enforce no UVM in comms, device or host OK
+  assert(acceleratorIsCommunicable(xmit));
+  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
-    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		      recv,bytes,MPI_CHAR,from, from,
-		      communicator,MPI_STATUS_IGNORE);
-    assert(ierr==0);
-  }
+  // 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,
@@ -363,7 +343,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int from,
 							 int bytes,int dir)
 {
-  int ncomm  =communicator_halo.size(); 
+  int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
 
   MPI_Request xrq;
@@ -378,36 +358,31 @@ 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;
   }
 
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);
   }
 
   return off_node_bytes;
 }
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
-{
-  SendToRecvFromComplete(waitall);
-}
-void CartesianCommunicator::StencilBarrier(void)
-{
-  MPI_Barrier  (ShmComm);
-}
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
   int nreq=list.size();
 
@@ -418,6 +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);
@@ -432,8 +414,8 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-int CartesianCommunicator::RankWorld(void){ 
-  int r; 
+int CartesianCommunicator::RankWorld(void){
+  int r;
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
@@ -466,7 +448,7 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
   // (Turns up on 32^3 x 64 Gparity too)
   MPI_Datatype object;
-  int iwords; 
+  int iwords;
   int ibytes;
   iwords = words;
   ibytes = bytes;
@@ -479,5 +461,3 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 }
 
 NAMESPACE_END(Grid);
-
-

Grid/communicator/Communicator_none.cc

@@ -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);
@@ -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,
@@ -149,13 +123,10 @@ 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){};

Grid/communicator/SharedMemory.cc

@@ -74,7 +74,9 @@ 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;

Grid/communicator/SharedMemory.h

@@ -102,7 +102,7 @@ public:
   ///////////////////////////////////////////////////
   static void SharedMemoryAllocate(uint64_t bytes, int flags);
   static void SharedMemoryFree(void);
-  static void SharedMemoryCopy(void *dest,const void *src,size_t bytes);
+  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
   static void SharedMemoryZero(void *dest,size_t bytes);
 
 };

Grid/communicator/SharedMemoryMPI.cc

@@ -29,9 +29,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
-#ifdef GRID_NVCC
+#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: "
@@ -47,7 +50,12 @@ 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);
 
@@ -170,17 +178,24 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
   std::vector<int> primes({2,3,5});
 
   int dim = 0;
+  int last_dim = ndimension - 1;
   int AutoShmSize = 1;
   while(AutoShmSize != WorldShmSize) {
-    for(int p=0;p<primes.size();p++) {
+    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;
   }
 }
@@ -413,7 +428,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA) ||defined(GRID_HIP)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -433,27 +448,19 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   //  cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
 
-#ifdef GRID_IBM_SUMMIT
-  // IBM Jsrun makes cuda Device numbering screwy and not match rank
-    std::cout << "IBM Summit or similar - NOT setting device to WorldShmRank"<<std::endl;
-#else
-    std::cout << "setting device to WorldShmRank"<<std::endl;
-    cudaSetDevice(WorldShmRank);
-#endif
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  auto err =  cudaMalloc(&ShmCommBuf, bytes);
-  if ( err !=  cudaSuccess) {
-    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
-    exit(EXIT_FAILURE);  
-  }
+  ShmCommBuf = acceleratorAllocDevice(bytes);
+
   if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
-  if ( WorldRank == 0 ){
-    std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+  //  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);
 
@@ -461,19 +468,31 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // 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 ) { 
-      err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
+      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
     //////////////////////////////////////////////////
@@ -490,17 +509,31 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // If I am not the source, overwrite thisBuf with remote buffer
     ///////////////////////////////////////////////////////////////
     void * thisBuf = ShmCommBuf;
+#ifdef GRID_CUDA
     if ( r!=WorldShmRank ) { 
-      err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
+      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;
@@ -633,7 +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);    
@@ -677,15 +709,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 /////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
   cudaMemset(dest,0,bytes);
 #else
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
+void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
 {
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
   cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
 #else   
   bcopy(src,dest,bytes);
@@ -705,7 +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);
@@ -735,19 +771,12 @@ 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_IBM_SUMMIT
-  // Hide the shared memory path between sockets 
-  // if even number of nodes
-  if ( (ShmSize & 0x1)==0 ) {
-    int SocketSize = ShmSize/2;
-    int mySocket = ShmRank/SocketSize; 
+#ifdef GRID_SHM_FORCE_MPI
+  // Hide the shared memory path between ranks
+  {
     for(int r=0;r<size;r++){
-      int hisRank=ShmRanks[r];
-      if ( hisRank!= MPI_UNDEFINED ) {
-	int hisSocket=hisRank/SocketSize;
-	if ( hisSocket != mySocket ) {
-	  ShmRanks[r] = MPI_UNDEFINED;
-	}
+      if ( r!=rank ) {
+	ShmRanks[r] = MPI_UNDEFINED;
       }
     }
   }

Grid/communicator/SharedMemoryNone.cc

@@ -29,6 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 
 NAMESPACE_BEGIN(Grid); 
+#define header "SharedMemoryNone: "
 
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@@ -55,6 +56,38 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
 ////////////////////////////////////////////////////////////////////////////////////////////
 // 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

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

@@ -29,11 +29,13 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 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];
 
@@ -46,16 +48,16 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
   int e2=rhs.Grid()->_slice_block[dimension];
   int ent = 0;
 
-  static 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];
 
-  auto rhs_v = rhs.View();
   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 { 
@@ -65,14 +67,26 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
 	 int o  = n*stride;
 	 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);
 	 }
        }
      }
   }
-  thread_for(i,ent,{
-    buffer[table[i].first]=rhs_v[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
+  }
 }
 
 ///////////////////////////////////////////////////////////////////
@@ -95,43 +109,76 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int e2=rhs.Grid()->_slice_block[dimension];
   int n1=rhs.Grid()->_slice_stride[dimension];
 
-  auto rhs_v = rhs.View();
   if ( cbmask ==0x3){
-    thread_for_collapse(2,n,e1,{
-      for(int b=0;b<e2;b++){
-
+#ifdef ACCELERATOR_CSHIFT    
+    autoView(rhs_v , rhs, AcceleratorRead);
+    accelerator_for2d(n,e1,b,e2,1,{
 	int o      =   n*n1;
 	int offset = b+n*e2;
 	
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
-      }
-    });
+      });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+	int o      =   n*n1;
+	int offset = b+n*e2;
+	
+	vobj temp =rhs_v[so+o+b];
+	extract<vobj>(temp,pointers,offset);
+      });
+#endif
   } else { 
+    Coordinate rdim=rhs.Grid()->_rdimensions;
+    Coordinate cdm =rhs.Grid()->_checker_dim_mask;
+    std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
+#ifdef ACCELERATOR_CSHIFT    
+    autoView(rhs_v , rhs, AcceleratorRead);
+    accelerator_for2d(n,e1,b,e2,1,{
 
-    // Case of SIMD split AND checker dim cannot currently be hit, except in 
-    // Test_cshift_red_black code.
-    std::cout << " Dense packed buffer WARNING " <<std::endl;
-    thread_for_collapse(2,n,e1,{
-      for(int b=0;b<e2;b++){
+	Coordinate coor;
 
 	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_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];
 
@@ -145,7 +192,8 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
   int e2=rhs.Grid()->_slice_block[dimension];
   int stride=rhs.Grid()->_slice_stride[dimension];
 
-  static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
+  if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
+
   int ent    =0;
 
   if ( cbmask ==0x3 ) {
@@ -154,7 +202,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
       for(int b=0;b<e2;b++){
 	int o   =n*rhs.Grid()->_slice_stride[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);
       }
     }
 
@@ -165,16 +213,27 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
 	int o   =n*rhs.Grid()->_slice_stride[dimension];
 	int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	if ( ocb & cbmask ) {
-	  table[ent++]=std::pair<int,int> (so+o+b,bo++);
+	  Cshift_table[ent++]=std::pair<int,int> (so+o+b,bo++);
 	}
       }
     }
   }
   
-  auto rhs_v = rhs.View();
-  thread_for(i,ent,{
-    rhs_v[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
+  }
 }
 
 //////////////////////////////////////////////////////
@@ -194,21 +253,31 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   int e2=rhs.Grid()->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
-    auto rhs_v = rhs.View();
-    thread_for_collapse(2,n,e1,{
-      for(int b=0;b<e2;b++){
-	int o      = n*rhs.Grid()->_slice_stride[dimension];
-	int offset = b+n*rhs.Grid()->_slice_block[dimension];
+    int _slice_stride = rhs.Grid()->_slice_stride[dimension];
+    int _slice_block = rhs.Grid()->_slice_block[dimension];
+#ifdef ACCELERATOR_CSHIFT    
+    autoView( rhs_v , rhs, AcceleratorWrite);
+    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;
-    auto rhs_v = rhs.View();
+    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];
@@ -225,6 +294,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 //////////////////////////////////////////////////////
 // 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];
@@ -239,14 +309,16 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int e1=rhs.Grid()->_slice_nblock[dimension]; // clearly loop invariant for icpc
   int e2=rhs.Grid()->_slice_block[dimension];
   int stride = rhs.Grid()->_slice_stride[dimension];
-  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 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 { 
@@ -255,23 +327,32 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
         int o =n*stride+b;
         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);
 	}
       }
     }
   }
 
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for(i,ent,{
-    lhs_v[table[i].first]=rhs_v[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];
 
   if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
@@ -285,29 +366,41 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int e2=rhs.Grid()->_slice_block [dimension];
   int stride = rhs.Grid()->_slice_stride[dimension];
 
-  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 ent=0;
 
   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);
-      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);
     }}
   }
 
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for(i,ent,{
-    permute(lhs_v[table[i].first],rhs_v[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
+  }
 }
 
 //////////////////////////////////////////////////////

Grid/cshift/Cshift_mpi.h

@@ -101,7 +101,8 @@ 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;
@@ -121,9 +122,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  commVector<vobj> send_buf(buffer_size);
-  commVector<vobj> recv_buf(buffer_size);
-
+  cshiftVector<vobj> send_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);
 
@@ -138,7 +139,7 @@ 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);
@@ -150,12 +151,14 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       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);
@@ -195,8 +198,15 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  std::vector<commVector<scalar_object> >   send_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
-  std::vector<commVector<scalar_object> >   recv_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
+  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;
+ 
+  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);
 
@@ -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 { 
@@ -258,7 +461,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   }
 
 }
-
+#endif
 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/lattice/Lattice.h

@@ -26,6 +26,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     *************************************************************************************/
     /*  END LEGAL */
 #pragma once
+#include <Grid/lattice/Lattice_view.h>
 #include <Grid/lattice/Lattice_base.h>
 #include <Grid/lattice/Lattice_conformable.h>
 #include <Grid/lattice/Lattice_ET.h>
@@ -36,6 +37,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #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>
@@ -43,4 +45,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #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
@@ -41,9 +42,24 @@ NAMESPACE_BEGIN(Grid);
 ////////////////////////////////////////////////////
 // Predicated where support
 ////////////////////////////////////////////////////
+#ifdef GRID_SIMT
+// drop to scalar in SIMT; cleaner in fact
 template <class iobj, class vobj, class robj>
-accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
-                            const robj &iffalse) {
+accelerator_inline vobj predicatedWhere(const iobj &predicate, 
+					const vobj &iftrue, 
+					const robj &iffalse) 
+{
+  Integer mask = TensorRemove(predicate);
+  typename std::remove_const<vobj>::type ret= iffalse;
+  if (mask) ret=iftrue;
+  return ret;
+}
+#else
+template <class iobj, class vobj, class robj>
+accelerator_inline vobj predicatedWhere(const iobj &predicate, 
+					const vobj &iftrue, 
+					const robj &iffalse) 
+{
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
@@ -67,6 +83,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftru
   merge(ret, falsevals);
   return ret;
 }
+#endif
 
 /////////////////////////////////////////////////////
 //Specialization of getVectorType for lattices
@@ -80,26 +97,62 @@ struct getVectorType<Lattice<T> >{
 //--  recursive evaluation of expressions; --
 // handle leaves of syntax tree
 ///////////////////////////////////////////////////
-template<class sobj> accelerator_inline 
+template<class sobj,
+  typename std::enable_if<!is_lattice<sobj>::value&&!is_lattice_expr<sobj>::value,sobj>::type * = nullptr> 
+accelerator_inline 
 sobj eval(const uint64_t ss, const sobj &arg)
 {
   return arg;
 }
-
 template <class lobj> accelerator_inline 
-const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg) 
+auto eval(const uint64_t ss, const LatticeView<lobj> &arg) -> decltype(arg(ss))
 {
-  return arg[ss];
+  return arg(ss);
+}
+
+////////////////////////////////////////////
+//--  recursive evaluation of expressions; --
+// whole vector return, used only for expression return type inference
+///////////////////////////////////////////////////
+template<class sobj> accelerator_inline 
+sobj vecEval(const uint64_t ss, const sobj &arg)
+{
+  return arg;
 }
 template <class lobj> accelerator_inline 
-const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg) 
+const lobj & vecEval(const uint64_t ss, const LatticeView<lobj> &arg) 
 {
-  auto view = arg.View();
-  return view[ss];
+  return arg[ss];
 }
 
 ///////////////////////////////////////////////////
 // handle nodes in syntax tree- eval one operand
+// vecEval needed (but never called as all expressions offloaded) to infer the return type
+// in SIMT contexts of closure.
+///////////////////////////////////////////////////
+template <typename Op, typename T1> accelerator_inline 
+auto vecEval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
+  -> decltype(expr.op.func( vecEval(ss, expr.arg1)))
+{
+  return expr.op.func( vecEval(ss, expr.arg1) );
+}
+// vecEval two operands
+template <typename Op, typename T1, typename T2> accelerator_inline
+auto vecEval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
+  -> decltype(expr.op.func( vecEval(ss,expr.arg1),vecEval(ss,expr.arg2)))
+{
+  return expr.op.func( vecEval(ss,expr.arg1), vecEval(ss,expr.arg2) );
+}
+// vecEval three operands
+template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
+auto vecEval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
+  -> decltype(expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3)))
+{
+  return expr.op.func(vecEval(ss, expr.arg1), vecEval(ss, expr.arg2), vecEval(ss, expr.arg3));
+}
+
+///////////////////////////////////////////////////
+// handle nodes in syntax tree- eval one operand coalesced
 ///////////////////////////////////////////////////
 template <typename Op, typename T1> accelerator_inline 
 auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)  
@@ -107,23 +160,41 @@ auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)
 {
   return expr.op.func( eval(ss, expr.arg1) );
 }
-///////////////////////
 // eval two operands
-///////////////////////
 template <typename Op, typename T1, typename T2> accelerator_inline
 auto eval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)  
   -> decltype(expr.op.func( eval(ss,expr.arg1),eval(ss,expr.arg2)))
 {
   return expr.op.func( eval(ss,expr.arg1), eval(ss,expr.arg2) );
 }
-///////////////////////
 // eval three operands
-///////////////////////
 template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
 auto eval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)  
-  -> decltype(expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3)))
+  -> decltype(expr.op.func(eval(ss, expr.arg1), 
+			   eval(ss, expr.arg2), 
+			   eval(ss, expr.arg3)))
 {
-  return 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
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -179,16 +250,12 @@ inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
   cb = lat.Checkerboard();
 }
 template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-inline void CBFromExpression(int &cb, const T1 &notlat)  // non-lattice leaf
-{
-}
-
+inline void CBFromExpression(int &cb, const T1 &notlat) {} // non-lattice leaf
 template <typename Op, typename T1> inline 
 void CBFromExpression(int &cb,const LatticeUnaryExpression<Op, T1> &expr) 
 {
   CBFromExpression(cb, expr.arg1);  // recurse AST
 }
-
 template <typename Op, typename T1, typename T2> inline 
 void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr) 
 {
@@ -203,32 +270,86 @@ inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2,
   CBFromExpression(cb, expr.arg3);  // recurse AST
 }
 
+
+//////////////////////////////////////////////////////////////////////////
+// ViewOpen
+//////////////////////////////////////////////////////////////////////////
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+inline void ExpressionViewOpen(T1 &lat)  // Lattice leaf
+{
+  lat.ViewOpen(AcceleratorRead);
+}
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+  inline void ExpressionViewOpen(T1 &notlat) {}
+
+template <typename Op, typename T1> inline 
+void ExpressionViewOpen(LatticeUnaryExpression<Op, T1> &expr) 
+{  
+  ExpressionViewOpen(expr.arg1); // recurse AST
+}
+
+template <typename Op, typename T1, typename T2> inline 
+void ExpressionViewOpen(LatticeBinaryExpression<Op, T1, T2> &expr) 
+{
+  ExpressionViewOpen(expr.arg1);  // recurse AST
+  ExpressionViewOpen(expr.arg2);  // rrecurse AST
+}
+template <typename Op, typename T1, typename T2, typename T3>
+inline void ExpressionViewOpen(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
+{
+  ExpressionViewOpen(expr.arg1);  // recurse AST
+  ExpressionViewOpen(expr.arg2);  // recurse AST
+  ExpressionViewOpen(expr.arg3);  // recurse AST
+}
+
+//////////////////////////////////////////////////////////////////////////
+// ViewClose
+//////////////////////////////////////////////////////////////////////////
+template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
+inline void ExpressionViewClose( T1 &lat)  // Lattice leaf
+{
+  lat.ViewClose();
+}
+template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
+inline void ExpressionViewClose(T1 &notlat) {}
+
+template <typename Op, typename T1> inline 
+void ExpressionViewClose(LatticeUnaryExpression<Op, T1> &expr) 
+{  
+  ExpressionViewClose(expr.arg1); // recurse AST
+}
+template <typename Op, typename T1, typename T2> inline 
+void ExpressionViewClose(LatticeBinaryExpression<Op, T1, T2> &expr) 
+{
+  ExpressionViewClose(expr.arg1);  // recurse AST
+  ExpressionViewClose(expr.arg2);  // recurse AST
+}
+template <typename Op, typename T1, typename T2, typename T3>
+inline void ExpressionViewClose(LatticeTrinaryExpression<Op, T1, T2, T3> &expr) 
+{
+  ExpressionViewClose(expr.arg1);  // recurse AST
+  ExpressionViewClose(expr.arg2);  // recurse AST
+  ExpressionViewClose(expr.arg3);  // recurse AST
+}
+
 ////////////////////////////////////////////
 // Unary operators and funcs
 ////////////////////////////////////////////
 #define GridUnopClass(name, ret)					\
-  template <class arg>							\
   struct name {								\
-    static auto accelerator_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));
@@ -240,10 +361,10 @@ GridUnopClass(UnaryExp, exp(a));
 // Binary operators
 ////////////////////////////////////////////
 #define GridBinOpClass(name, combination)			\
-  template <class left, class right>				\
   struct name {							\
+    template <class _left, class _right>			\
     static auto accelerator_inline				\
-    func(const left &lhs, const right &rhs)			\
+    func(const _left &lhs, const _right &rhs)			\
       -> decltype(combination) const				\
     {								\
       return combination;					\
@@ -263,10 +384,10 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
 // Trinary conditional op
 ////////////////////////////////////////////////////
 #define GridTrinOpClass(name, combination)				\
-  template <class predicate, class left, class right>			\
   struct name {								\
+    template <class _predicate,class _left, class _right>		\
     static auto accelerator_inline					\
-    func(const predicate &pred, const left &lhs, const right &rhs)	\
+    func(const _predicate &pred, const _left &lhs, const _right &rhs)	\
       -> decltype(combination) const					\
     {									\
       return combination;						\
@@ -274,17 +395,17 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
   };
 
 GridTrinOpClass(TrinaryWhere,
-		(predicatedWhere<predicate, 
-		 typename std::remove_reference<left>::type,
-		 typename std::remove_reference<right>::type>(pred, lhs,rhs)));
+		(predicatedWhere<
+		 typename std::remove_reference<_predicate>::type, 
+		 typename std::remove_reference<_left>::type,
+		 typename std::remove_reference<_right>::type>(pred, lhs,rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
 ////////////////////////////////////////////
-
-#define GRID_UNOP(name)   name<decltype(eval(0, arg))>
-#define GRID_BINOP(name)  name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
-#define GRID_TRINOP(name) name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
+#define GRID_UNOP(name)   name
+#define GRID_BINOP(name)  name
+#define GRID_TRINOP(name) name
 
 #define GRID_DEF_UNOP(op, name)						\
   template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
@@ -330,22 +451,17 @@ GridTrinOpClass(TrinaryWhere,
 GRID_DEF_UNOP(operator-, UnarySub);
 GRID_DEF_UNOP(Not, UnaryNot);
 GRID_DEF_UNOP(operator!, UnaryNot);
-GRID_DEF_UNOP(adj, UnaryAdj);
-GRID_DEF_UNOP(conjugate, UnaryConj);
+//GRID_DEF_UNOP(adj, UnaryAdj);
+//GRID_DEF_UNOP(conjugate, UnaryConj);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
 GRID_DEF_UNOP(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);
@@ -370,29 +486,36 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1)))>::type > 
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> 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.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type >
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),vecEval(0, expr.arg2)))>::type > ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-  -> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
-				   eval(0, expr.arg2),
-				   eval(0, expr.arg3)))> 
+  -> Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+				   vecEval(0, expr.arg2),
+				   vecEval(0, expr.arg3)))>::type >
 {
-  Lattice<decltype(expr.op.func(eval(0, expr.arg1),
-				eval(0, expr.arg2),
-				eval(0, expr.arg3)))>  ret(expr);
+  Lattice<typename std::remove_const<decltype(expr.op.func(vecEval(0, expr.arg1),
+				vecEval(0, expr.arg2),
+			        vecEval(0, expr.arg3)))>::type >  ret(expr);
   return ret;
 }
+#define EXPRESSION_CLOSURE(function)					\
+  template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> \
+    auto function(Expression &expr) -> decltype(function(closure(expr))) \
+  {									\
+    return function(closure(expr));					\
+  }
+
 
 #undef GRID_UNOP
 #undef GRID_BINOP

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
@@ -36,9 +37,9 @@ NAMESPACE_BEGIN(Grid);
 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();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   conformable(ret,rhs);
   conformable(lhs,rhs);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
@@ -55,13 +56,13 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     auto rhs_t=rhs_v(ss);
+    auto tmp  =ret_v(ss);
     mac(&tmp,&lhs_t,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
   });
@@ -72,9 +73,9 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -88,9 +89,9 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -107,8 +108,8 @@ template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     mult(&tmp,&lhs_v(ss),&rhs);
@@ -120,10 +121,10 @@ template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     auto lhs_t=lhs_v(ss);
     mac(&tmp,&lhs_t,&rhs);
     coalescedWrite(ret_v[ss],tmp);
@@ -134,8 +135,8 @@ template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -147,8 +148,8 @@ template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
@@ -164,8 +165,8 @@ template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.View();
-  auto rhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -178,10 +179,10 @@ template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.View();
-  auto rhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     auto rhs_t=rhs_v(ss);
     mac(&tmp,&lhs,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
@@ -192,8 +193,8 @@ template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.View();
-  auto rhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -205,8 +206,8 @@ template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
-  auto ret_v = ret.View();
-  auto rhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
     decltype(coalescedRead(obj1())) tmp;
     auto rhs_t=rhs_v(ss);
@@ -220,9 +221,9 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-  auto ret_v = ret.View();
-  auto x_v = x.View();
-  auto y_v = y.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( x_v , x, AcceleratorRead);
+  autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
     auto tmp = a*x_v(ss)+y_v(ss);
     coalescedWrite(ret_v[ss],tmp);
@@ -233,9 +234,9 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
-  auto ret_v = ret.View();
-  auto x_v = x.View();
-  auto y_v = y.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( x_v , x, AcceleratorRead);
+  autoView( y_v , y, AcceleratorRead);
   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
     auto tmp = a*x_v(ss)+b*y_v(ss);
     coalescedWrite(ret_v[ss],tmp);

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
@@ -28,6 +29,7 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
+
 #pragma once 
 
 #define STREAMING_STORES
@@ -36,129 +38,6 @@ NAMESPACE_BEGIN(Grid);
 
 extern int GridCshiftPermuteMap[4][16];
 
-///////////////////////////////////////////////////////////////////
-// Base class which can be used by traits to pick up behaviour
-///////////////////////////////////////////////////////////////////
-class LatticeBase {};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Conformable checks; same instance of Grid required
-/////////////////////////////////////////////////////////////////////////////////////////
-void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
-{
-  assert(lhs == rhs);
-}
-
-////////////////////////////////////////////////////////////////////////////
-// Minimal base class containing only data valid to access from accelerator
-// _odata will be a managed pointer in CUDA
-////////////////////////////////////////////////////////////////////////////
-// Force access to lattice through a view object.
-// prevents writing of code that will not offload to GPU, but perhaps annoyingly
-// strict since host could could in principle direct access through the lattice object
-// Need to decide programming model.
-#define LATTICE_VIEW_STRICT
-template<class vobj> class LatticeAccelerator : public LatticeBase
-{
-protected:
-  GridBase *_grid;
-  int checkerboard;
-  vobj     *_odata;    // A managed pointer
-  uint64_t _odata_size;    
-public:
-  accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr) { }; 
-  accelerator_inline uint64_t oSites(void) const { return _odata_size; };
-  accelerator_inline int  Checkerboard(void) const { return checkerboard; };
-  accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
-  accelerator_inline void Conformable(GridBase * &grid) const
-  { 
-    if (grid) conformable(grid, _grid);
-    else      grid = _grid;
-  };
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// A View class which provides accessor to the data.
-// This will be safe to call from accelerator_for and is trivially copy constructible
-// The copy constructor for this will need to be used by device lambda functions
-/////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> 
-class LatticeView : public LatticeAccelerator<vobj>
-{
-public:
-
-
-  // Rvalue
-#ifdef __CUDA_ARCH__
-  accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { return coalescedRead(this->_odata[i]); }
-#else 
-  accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
-#endif
-
-  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
-  accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
-
-  accelerator_inline uint64_t begin(void) const { return 0;};
-  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
-  accelerator_inline uint64_t size(void)  const { return this->_odata_size; };
-
-  LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me)
-  {
-  }
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////
-// Lattice expression types used by ET to assemble the AST
-// 
-// Need to be able to detect code paths according to the whether a lattice object or not
-// so introduce some trait type things
-/////////////////////////////////////////////////////////////////////////////////////////
-
-class LatticeExpressionBase {};
-
-template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
-template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
-
-template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
-template<class T>                 struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
-template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
-
-template <typename Op, typename _T1>                           
-class LatticeUnaryExpression : public  LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  Op op;
-  T1 arg1;
-  LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
-};
-
-template <typename Op, typename _T1, typename _T2>              
-class LatticeBinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
-};
-
-template <typename Op, typename _T1, typename _T2, typename _T3> 
-class LatticeTrinaryExpression : public LatticeExpressionBase 
-{
-public:
-  typedef typename ViewMap<_T1>::Type T1;
-  typedef typename ViewMap<_T2>::Type T2;
-  typedef typename ViewMap<_T3>::Type T3;
-  Op op;
-  T1 arg1;
-  T2 arg2;
-  T3 arg3;
-  LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
-};
-
 /////////////////////////////////////////////////////////////////////////////////////////
 // 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
@@ -194,24 +73,33 @@ private:
       dealloc();
       
       this->_odata_size = size;
-      if ( 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 (void) const 
+
+  LatticeView<vobj> View (ViewMode mode) const 
   {
-    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
+    LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
     return accessor;
   }
-  
+
   ~Lattice() { 
     if ( this->_odata_size ) {
       dealloc();
@@ -231,12 +119,16 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-
-    auto me  = View();
-    accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
-      vstream(me[ss],tmp);
+    
+    auto exprCopy = expr;
+    ExpressionViewOpen(exprCopy);
+    auto me  = View(AcceleratorWriteDiscard);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+      auto tmp = eval(ss,exprCopy);
+      coalescedWrite(me[ss],tmp);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
@@ -251,11 +143,15 @@ public:
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
-    auto me  = View();
-    accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
-      vstream(me[ss],tmp);
+    auto exprCopy = expr;
+    ExpressionViewOpen(exprCopy);
+    auto me  = View(AcceleratorWriteDiscard);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+      auto tmp = eval(ss,exprCopy);
+      coalescedWrite(me[ss],tmp);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
@@ -269,11 +165,15 @@ public:
     CBFromExpression(cb,expr);
     assert( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
-    auto me  = View();
-    accelerator_for(ss,me.size(),1,{
-      auto tmp = eval(ss,expr);
-      vstream(me[ss],tmp);
+    auto exprCopy = expr;
+    ExpressionViewOpen(exprCopy);
+    auto me  = View(AcceleratorWriteDiscard);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+      auto tmp = eval(ss,exprCopy);
+      coalescedWrite(me[ss],tmp);
     });
+    me.ViewClose();
+    ExpressionViewClose(exprCopy);
     return *this;
   }
   //GridFromExpression is tricky to do
@@ -324,10 +224,11 @@ public:
   }
 
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
-    auto me  = View();
+    auto me  = View(CpuWrite);
     thread_for(ss,me.size(),{
-      me[ss] = r;
+	me[ss]= r;
     });
+    me.ViewClose();
     return *this;
   }
 
@@ -337,11 +238,12 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid) { 
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
     this->checkerboard=0;
+    SetViewMode(mode);
   }
   
   //  virtual ~Lattice(void) = default;
@@ -357,7 +259,6 @@ public:
   // copy constructor
   ///////////////////////////////////////////
   Lattice(const Lattice& r){ 
-    //    std::cout << "Lattice constructor(const Lattice &) "<<this<<std::endl; 
     this->_grid = r.Grid();
     resize(this->_grid->oSites());
     *this = r;
@@ -380,11 +281,12 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View();
-    auto him= r.View();
+    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;
   }
 
@@ -394,11 +296,12 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View();
-    auto him= r.View();
+    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;
   }
   ///////////////////////////////////////////

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

@@ -42,34 +42,6 @@ NAMESPACE_BEGIN(Grid);
 
 typedef iScalar<vInteger> vPredicate ;
 
-/*
-template <class iobj, class vobj, class robj> accelerator_inline 
-vobj predicatedWhere(const iobj &predicate, const vobj &iftrue, const robj &iffalse) 
-{
-  typename std::remove_const<vobj>::type ret;
-
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
-  const int Nsimd = vobj::vector_type::Nsimd();
-
-  ExtractBuffer<Integer> mask(Nsimd);
-  ExtractBuffer<scalar_object> truevals(Nsimd);
-  ExtractBuffer<scalar_object> falsevals(Nsimd);
-
-  extract(iftrue, truevals);
-  extract(iffalse, falsevals);
-  extract<vInteger, Integer>(TensorRemove(predicate), mask);
-
-  for (int s = 0; s < Nsimd; s++) {
-    if (mask[s]) falsevals[s] = truevals[s];
-  }
-
-  merge(ret, falsevals);
-  return ret;
-}
-*/
 //////////////////////////////////////////////////////////////////////////
 // compare lattice to lattice
 //////////////////////////////////////////////////////////////////////////
@@ -78,9 +50,9 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
 {
   Lattice<vPredicate> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
       ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
   });
@@ -93,8 +65,8 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
 {
   Lattice<vPredicate> ret(lhs.Grid());
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, lhs_v.size(), {
     ret_v[ss]=op(lhs_v[ss],rhs);
   });
@@ -107,8 +79,8 @@ template<class vfunctor,class lobj,class robj>
 inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
 {
   Lattice<vPredicate> ret(rhs.Grid());
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( rhs_v, rhs, CpuRead);
+  autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
     ret_v[ss]=op(lhs,rhs_v[ss]);
   });

Grid/lattice/Lattice_coordinate.h

@@ -37,7 +37,7 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
   GridBase *grid = l.Grid();
   int Nsimd = grid->iSites();
 
-  auto l_v = l.View();
+  autoView(l_v, l, CpuWrite);
   thread_for( o, grid->oSites(), {
     vector_type vI;
     Coordinate gcoor;
@@ -51,23 +51,5 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
   });
 };
 
-// LatticeCoordinate();
-// FIXME for debug; deprecate this; made obscelete by 
-template<class vobj> void lex_sites(Lattice<vobj> &l){
-  auto l_v = l.View();
-  Real *v_ptr = (Real *)&l_v[0];
-  size_t o_len = l.Grid()->oSites();
-  size_t v_len = sizeof(vobj)/sizeof(vRealF);
-  size_t vec_len = vRealF::Nsimd();
-
-  for(int i=0;i<o_len;i++){
-    for(int j=0;j<v_len;j++){
-      for(int vv=0;vv<vec_len;vv+=2){
-	v_ptr[i*v_len*vec_len+j*vec_len+vv  ]= i+vv*500;
-	v_ptr[i*v_len*vec_len+j*vec_len+vv+1]= i+vv*500;
-      }
-    }}
-}
-
 NAMESPACE_END(Grid);
 

Grid/lattice/Lattice_local.h

@@ -43,8 +43,8 @@ template<class vobj>
 inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
 {
   Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
   });
@@ -56,9 +56,9 @@ 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());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
+  autoView( ret_v , ret, AcceleratorWrite);
   accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
   });
@@ -73,9 +73,9 @@ inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Latt
   typedef decltype(coalescedRead(ll())) sll;
   typedef decltype(coalescedRead(rr())) srr;
   Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
-  auto ret_v = ret.View();
+  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

Grid/lattice/Lattice_matrix_reduction.h

@@ -51,9 +51,9 @@ 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];
-  auto X_v = X.View();
-  auto Y_v = Y.View();
-  auto R_v = R.View();
+  autoView( X_v , X, CpuRead);
+  autoView( Y_v , Y, CpuRead);
+  autoView( R_v , R, CpuWrite);
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -97,8 +97,8 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  auto X_v = X.View();
-  auto R_v = R.View();
+  autoView( X_v , X, CpuRead);
+  autoView( R_v , R, CpuWrite);
 
   thread_region
   {
@@ -156,8 +156,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   int ostride=FullGrid->_ostride[Orthog];
 
   typedef typename vobj::vector_typeD vector_typeD;
-  auto lhs_v = lhs.View();
-  auto rhs_v = rhs.View();
+  autoView( lhs_v , lhs, CpuRead);
+  autoView( rhs_v , rhs, CpuRead);
   thread_region {
     std::vector<vobj> Left(Nblock);
     std::vector<vobj> Right(Nblock);

Grid/lattice/Lattice_peekpoke.h

@@ -46,9 +46,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Ind
 {
   Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
   ret.Checkerboard()=lhs.Checkerboard();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
+  accelerator_for( ss, lhs_v.size(), 1, {
     ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
   });
   return ret;
@@ -58,9 +58,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
 {
   Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
   ret.Checkerboard()=lhs.Checkerboard();
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
+  accelerator_for( ss, lhs_v.size(), 1, {
     ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
   });
   return ret;
@@ -72,18 +72,18 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
 template<int Index,class vobj>  
 void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
 {
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( rhs_v, rhs, AcceleratorRead);
+  autoView( lhs_v, lhs, AcceleratorWrite);
+  accelerator_for( ss, lhs_v.size(), 1, {
     pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
   });
 }
 template<int Index,class vobj> 
 void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
 {
-  auto rhs_v = rhs.View();
-  auto lhs_v = lhs.View();
-  thread_for( ss, lhs_v.size(), {
+  autoView( rhs_v, rhs, AcceleratorRead);
+  autoView( lhs_v, lhs, AcceleratorWrite);
+  accelerator_for( ss, lhs_v.size(), 1, {
     pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
   });
 }
@@ -111,7 +111,7 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   // extract-modify-merge cycle is easiest way and this is not perf critical
   ExtractBuffer<sobj> buf(Nsimd);
-  auto l_v = l.View();
+  autoView( l_v , l, CpuWrite);
   if ( rank == grid->ThisRank() ) {
     extract(l_v[odx],buf);
     buf[idx] = s;
@@ -141,7 +141,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
 
   ExtractBuffer<sobj> buf(Nsimd);
-  auto l_v = l.View();
+  autoView( l_v , l, CpuWrite);
   extract(l_v[odx],buf);
 
   s = buf[idx];
@@ -151,21 +151,21 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
   return;
 };
 
-
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+// Must be CPU read view
 template<class vobj,class sobj>
-accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
-        
-  GridBase *grid = l.Grid();
-
+inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
+{
+  GridBase *grid = l.getGrid();
+  assert(l.mode==CpuRead);
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== 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);
@@ -173,8 +173,7 @@ accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  auto l_v = l.View();
-  scalar_type * vp = (scalar_type *)&l_v[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
@@ -183,18 +182,27 @@ accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate
       
   return;
 };
-
 template<class vobj,class sobj>
-accelerator_inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &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);
@@ -202,13 +210,19 @@ accelerator_inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  auto l_v = l.View();
-  scalar_type * vp = (scalar_type *)&l_v[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;
 };
 

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

@@ -40,24 +40,77 @@ NAMESPACE_BEGIN(Grid);
 
 template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
-  accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
-    coalescedWrite(ret_v[ss], adj(lhs_v(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());
-  auto lhs_v = lhs.View();
-  auto ret_v = ret.View();
+
+  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
@@ -24,7 +25,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/Grid_Eigen_Dense.h>
 
 
-#ifdef GRID_NVCC
+#if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
 
@@ -38,7 +39,36 @@ 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 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);
@@ -62,23 +92,43 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
     ssum = ssum+sumarray[i];
   } 
   
-  return ssum;
+  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)
 {
-#ifdef GRID_NVCC
+#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)
 {
-  auto arg_v = arg.View();
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum(&arg_v[0],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;
 }
@@ -93,55 +143,49 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
 
 // 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;
   ComplexD  nrm;
   
   GridBase *grid = left.Grid();
-  
-  // Might make all code paths go this way.
-  auto left_v = left.View();
-  auto right_v=right.View();
 
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-#ifdef GRID_NVCC
-  // GPU - SIMT lane compliance...
-  typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
+  // Might make all code paths go this way.
+  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  
+    
+  {
+    autoView( left_v , left, AcceleratorRead);
+    autoView( right_v,right, AcceleratorRead);
 
-  accelerator_for( ss, sites, nsimd,{
-      auto x_l = left_v(ss);
-      auto y_l = right_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
-  })
+    // 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);
+    });
+  }
 
   // This is in single precision and fails some tests
-  // Need a sumD that sums in double
-  nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));  
-#else
-  // CPU 
-  typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  
-  accelerator_for( ss, sites, nsimd,{
-      auto x_l = left_v[ss];
-      auto y_l = right_v[ss];
-      inner_tmp_v[ss]=innerProductD(x_l,y_l);
-  })
-  nrm = TensorRemove(sum(inner_tmp_v,sites));
-#endif
-  grid->GlobalSum(nrm);
-
+  auto anrm = sum(inner_tmp_v,sites);  
+  nrm = anrm;
   return nrm;
 }
 
+template<class vobj>
+inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
+  GridBase *grid = left.Grid();
+  ComplexD nrm = rankInnerProduct(left,right);
+  grid->GlobalSum(nrm);
+  return nrm;
+}
+
+
 /////////////////////////
 // Fast axpby_norm
 // z = a x + b y
@@ -167,45 +211,67 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   
   GridBase *grid = x.Grid();
 
-  auto x_v=x.View();
-  auto y_v=y.View();
-  auto z_v=z.View();
-
   const uint64_t nsimd = grid->Nsimd();
   const uint64_t sites = grid->oSites();
   
-#ifdef GRID_NVCC
   // GPU
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
+  autoView( x_v, x, AcceleratorRead);
+  autoView( y_v, y, AcceleratorRead);
+  autoView( z_v, z, AcceleratorWrite);
 
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-
-  nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
-#else
-  // CPU 
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
+
+  accelerator_for( ss, sites, 1,{
+      auto tmp = a*x_v[ss]+b*y_v[ss];
       inner_tmp_v[ss]=innerProductD(tmp,tmp);
       z_v[ss]=tmp;
   });
-  // Already promoted to double
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
-
  
+template<class vobj> strong_inline void
+innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Lattice<vobj> &right)
+{
+  conformable(left,right);
+
+  typedef typename vobj::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.op.func(eval(0,expr.arg1)))::scalar_object
@@ -271,7 +337,7 @@ 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
-  auto Data_v=Data.View();
+  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++){
@@ -349,8 +415,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
 
-  auto lhv=lhs.View();
-  auto rhv=rhs.View();
+  autoView( lhv, lhs, CpuRead);
+  autoView( rhv, rhs, CpuRead);
   thread_for( r,rd,{
 
     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
@@ -457,14 +523,12 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
 
     tensor_reduced at; at=av;
 
-    auto Rv=R.View();
-    auto Xv=X.View();
-    auto Yv=Y.View();
-    thread_for_collapse(2, n, e1, {
-      for(int b=0;b<e2;b++){
+    autoView( Rv, R, CpuWrite);
+    autoView( Xv, X, CpuRead);
+    autoView( Yv, Y, CpuRead);
+    thread_for2d( n, e1, b,e2, {
 	int ss= so+n*stride+b;
 	Rv[ss] = at*Xv[ss]+Yv[ss];
-      }
     });
   }
 };
@@ -517,9 +581,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int nblock=FullGrid->_slice_nblock[Orthog];
   int ostride=FullGrid->_ostride[Orthog];
 
-  auto X_v=X.View();
-  auto Y_v=Y.View();
-  auto R_v=R.View();
+  autoView( X_v, X, CpuRead);
+  autoView( Y_v, Y, CpuRead);
+  autoView( R_v, R, CpuWrite);
   thread_region
   {
     Vector<vobj> s_x(Nblock);
@@ -564,13 +628,14 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   //  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];
-  auto R_v = R.View();
-  auto X_v = X.View();
+  autoView( R_v, R, CpuWrite);
+  autoView( X_v, X, CpuRead);
   thread_region
   {
     std::vector<vobj> s_x(Nblock);
@@ -628,8 +693,8 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
 
   typedef typename vobj::vector_typeD vector_typeD;
 
-  auto lhs_v=lhs.View();
-  auto rhs_v=rhs.View();
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
   thread_region
   {
     std::vector<vobj> Left(Nblock);

Grid/lattice/Lattice_reduction_gpu.h

@@ -1,7 +1,14 @@
 NAMESPACE_BEGIN(Grid);
 
-#define WARP_SIZE 32
+#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>
@@ -19,7 +26,12 @@ 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;
@@ -53,7 +65,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
   
   // cannot use overloaded operators for sobj as they are not volatile-qualified
   memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
-  __syncwarp();
+  acceleratorSynchronise();
   
   const Iterator VEC = WARP_SIZE;
   const Iterator vid = tid & (VEC-1);
@@ -67,9 +79,9 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
       beta += temp;
       memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
     }
-    __syncwarp();
+    acceleratorSynchronise();
   }
-  __syncthreads();
+  acceleratorSynchroniseAll();
   
   if (threadIdx.x == 0) {
     beta  = Zero();
@@ -79,7 +91,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
     }
     memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
   }
-  __syncthreads();
+  acceleratorSynchroniseAll();
 }
 
 
@@ -147,7 +159,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     sobj *smem = (sobj *)shmem_pointer;
     
     // wait until all outstanding memory instructions in this thread are finished
-    __threadfence();
+    acceleratorFence();
     
     if (tid==0) {
       unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
@@ -156,8 +168,8 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
     }
     
     // each thread must read the correct value of amLast
-    __syncthreads();
-    
+    acceleratorSynchroniseAll();
+
     if (amLast) {
       // reduce buffer[0], ..., buffer[gridDim.x-1]
       Iterator i = tid;
@@ -199,13 +211,7 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   sobj *buffer_v = &buffer[0];
   
   reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
-  cudaDeviceSynchronize();
-  
-  cudaError err = cudaGetLastError();
-  if ( cudaSuccess != err ) {
-    printf("Cuda error %s\n",cudaGetErrorString( err ));
-    exit(0);
-  }
+  accelerator_barrier();
   auto result = buffer_v[0];
   return result;
 }

Grid/lattice/Lattice_rng.h

@@ -375,7 +375,7 @@ public:
     int osites = _grid->oSites();  // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
     int words  = sizeof(scalar_object) / sizeof(scalar_type);
 
-    auto l_v = l.View();
+    autoView(l_v, l, CpuWrite);
     thread_for( ss, osites, {
       ExtractBuffer<scalar_object> buf(Nsimd);
       for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
@@ -461,8 +461,8 @@ public:
     }
 
     {
-      // Obtain one reseeded generator per thread
-      int Nthread = GridThread::GetThreads();
+      // Obtain one reseeded generator per thread      
+      int Nthread = 32; // Hardwire a good level or parallelism
       std::vector<RngEngine> seeders(Nthread);
       for(int t=0;t<Nthread;t++){
 	seeders[t] = Reseed(master_engine);

Grid/lattice/Lattice_trace.h

@@ -37,17 +37,19 @@ NAMESPACE_BEGIN(Grid);
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Trace
 ////////////////////////////////////////////////////////////////////////////////////////////////////
+/*
 template<class vobj>
 inline auto trace(const Lattice<vobj> &lhs)  -> Lattice<decltype(trace(vobj()))>
 {
   Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView(ret_v , ret, AcceleratorWrite);
+  autoView(lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
   });
   return ret;
 };
+*/
     
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Trace Index level dependent operation
@@ -56,8 +58,8 @@ template<int Index,class vobj>
 inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
 {
   Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v , ret, AcceleratorWrite);
+  autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
     coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
   });

Grid/lattice/Lattice_transfer.h

@@ -6,6 +6,7 @@
     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
@@ -46,11 +47,12 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // remove and insert a half checkerboard
 ////////////////////////////////////////////////////////////////////////////////////////////
-template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
+template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full)
+{
   half.Checkerboard() = cb;
 
-  auto half_v = half.View();
-  auto full_v = full.View();
+  autoView( half_v, half, CpuWrite);
+  autoView( full_v, full, CpuRead);
   thread_for(ss, full.Grid()->oSites(),{
     int cbos;
     Coordinate coor;
@@ -63,10 +65,11 @@ template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,con
     }
   });
 }
-template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
+template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half)
+{
   int cb = half.Checkerboard();
-  auto half_v = half.View();
-  auto full_v = full.View();
+  autoView( half_v , half, CpuRead);
+  autoView( full_v , full, CpuWrite);
   thread_for(ss,full.Grid()->oSites(),{
 
     Coordinate coor;
@@ -81,96 +84,143 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
     }
   });
 }
-  
-template<class vobj,class CComplex,int nbasis>
-inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
-			  const             Lattice<vobj>   &fineData,
-			  const std::vector<Lattice<vobj> > &Basis)
-{
-  GridBase * fine  = fineData.Grid();
-  GridBase * coarse= coarseData.Grid();
 
-  Lattice<CComplex> ip(coarse); 
-
-  //  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
-  auto ip_         = ip.View();
-  for(int v=0;v<nbasis;v++) {
-    blockInnerProduct(ip,Basis[v],fineData);
-    accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
-	coalescedWrite(coarseData_[sc](v),ip_(sc));
-      });
-  }
+////////////////////////////////////////////////////////////////////////////////////////////
+// Flexible Type Conversion for internal promotion to double as well as graceful
+// treatment of scalar-compatible types
+////////////////////////////////////////////////////////////////////////////////////////////
+accelerator_inline void convertType(ComplexD & out, const std::complex<double> & in) {
+  out = in;
 }
 
-template<class vobj,class CComplex,int nbasis>
-inline void blockProject1(Lattice<iVector<CComplex,nbasis > > &coarseData,
-			 const             Lattice<vobj>   &fineData,
-			 const std::vector<Lattice<vobj> > &Basis)
-{
-  typedef iVector<CComplex,nbasis > coarseSiteData;
-  coarseSiteData elide;
-  typedef decltype(coalescedRead(elide)) ScalarComplex;
-  GridBase * fine  = fineData.Grid();
-  GridBase * coarse= coarseData.Grid();
-  int  _ndimension = coarse->_ndimension;
+accelerator_inline void convertType(ComplexF & out, const std::complex<float> & in) {
+  out = in;
+}
 
-  // checks
-  assert( nbasis == Basis.size() );
-  subdivides(coarse,fine); 
-  for(int i=0;i<nbasis;i++){
-    conformable(Basis[i],fineData);
-  }
+#ifdef GRID_SIMT
+accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
+  ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
+}
+accelerator_inline void convertType(vComplexD & out, const ComplexD & in) {
+  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd())] = in;
+}
+accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
+  ((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd()*2)] = in;
+}
+#endif
 
-  Coordinate block_r      (_ndimension);
-  
-  for(int d=0 ; d<_ndimension;d++){
-    block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
-    assert(block_r[d]*coarse->_rdimensions[d] == fine->_rdimensions[d]);
-  }
-  int blockVol = fine->oSites()/coarse->oSites();
+accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) {
+  out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v);
+}
 
-  coarseData=Zero();
+accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
+  Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
+}
 
-  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
-  ////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // To make this lock free, loop over coars parallel, and then loop over fine associated with coarse.
-  // Otherwise do fine inner product per site, and make the update atomic
-  ////////////////////////////////////////////////////////////////////////////////////////////////////////
-  accelerator_for( sci, nbasis*coarse->oSites(), vobj::Nsimd(), {
+template<typename T1,typename T2,int N>
+  accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in);
+template<typename T1,typename T2,int N>
+  accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in);
 
-    auto sc=sci/nbasis;
-    auto i=sci%nbasis;
-    auto Basis_      = Basis[i].View();
+template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
+accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
+  convertType(out,in._internal);
+}
 
-    Coordinate coor_c(_ndimension);
-    Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions);  // Block coordinate
+template<typename T1, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
+accelerator_inline void convertType(T1 & out, const iScalar<T1> & in) {
+  convertType(out,in._internal);
+}
 
-    int sf;
-    decltype(innerProduct(Basis_(sf),fineData_(sf))) reduce=Zero();
+template<typename T1,typename T2>
+accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
+  convertType(out._internal,in);
+}
 
-    for(int sb=0;sb<blockVol;sb++){
+template<typename T1,typename T2,int N>
+accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in) {
+  for (int i=0;i<N;i++)
+    for (int j=0;j<N;j++)
+      convertType(out._internal[i][j],in._internal[i][j]);
+}
 
-      Coordinate coor_b(_ndimension);
-      Coordinate coor_f(_ndimension);
+template<typename T1,typename T2,int N>
+accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in) {
+  for (int i=0;i<N;i++)
+    convertType(out._internal[i],in._internal[i]);
+}
 
-      Lexicographic::CoorFromIndex(coor_b,sb,block_r);
-      for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d]+coor_b[d];
-      Lexicographic::IndexFromCoor(coor_f,sf,fine->_rdimensions);
-      
-      reduce=reduce+innerProduct(Basis_(sf),fineData_(sf));
-    }
-    coalescedWrite(coarseData_[sc](i),reduce);
+template<typename T, typename std::enable_if<isGridFundamental<T>::value, T>::type* = nullptr>
+accelerator_inline void convertType(T & out, const T & in) {
+  out = in;
+}
+
+template<typename T1,typename T2>
+accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
+  autoView( out_v , out,AcceleratorWrite);
+  autoView( in_v  , in ,AcceleratorRead);
+  accelerator_for(ss,out_v.size(),T1::Nsimd(),{
+      convertType(out_v[ss],in_v(ss));
   });
-  return;
 }
 
-template<class vobj,class CComplex>
-inline void blockZAXPY(Lattice<vobj> &fineZ,
-		       const Lattice<CComplex> &coarseA,
-		       const Lattice<vobj> &fineX,
-		       const Lattice<vobj> &fineY)
+////////////////////////////////////////////////////////////////////////////////////////////
+// precision-promoted local inner product
+////////////////////////////////////////////////////////////////////////////////////////////
+template<class vobj>
+inline auto localInnerProductD(const Lattice<vobj> &lhs,const Lattice<vobj> &rhs)
+-> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View(CpuRead)[0],rhs.View(CpuRead)[0])))>>
+{
+  autoView( lhs_v , lhs, AcceleratorRead);
+  autoView( rhs_v , rhs, AcceleratorRead);
+
+  typedef decltype(TensorRemove(innerProductD2(lhs_v[0],rhs_v[0]))) t_inner;
+  Lattice<iScalar<t_inner>> ret(lhs.Grid());
+
+  {
+    autoView(ret_v, ret,AcceleratorWrite);
+    accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
+      convertType(ret_v[ss],innerProductD2(lhs_v(ss),rhs_v(ss)));
+    });
+  }
+  return ret;
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////
+// block routines
+////////////////////////////////////////////////////////////////////////////////////////////
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
+			   const             Lattice<vobj>   &fineData,
+			   const VLattice &Basis)
+{
+  GridBase * fine  = fineData.Grid();
+  GridBase * coarse= coarseData.Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  Lattice<vobj>     fineDataRed = fineData;
+
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( ip_         , ip,         AcceleratorWrite);
+  for(int v=0;v<nbasis;v++) {
+    blockInnerProductD(ip,Basis[v],fineDataRed); // ip = <basis|fine>
+    accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+	convertType(coarseData_[sc](v),ip_[sc]);
+    });
+
+    // improve numerical stability of projection
+    // |fine> = |fine> - <basis|fine> |basis>
+    ip=-ip;
+    blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
+  }
+}
+
+
+template<class vobj,class vobj2,class CComplex>
+  inline void blockZAXPY(Lattice<vobj> &fineZ,
+			 const Lattice<CComplex> &coarseA,
+			 const Lattice<vobj2> &fineX,
+			 const Lattice<vobj> &fineY)
 {
   GridBase * fine  = fineZ.Grid();
   GridBase * coarse= coarseA.Grid();
@@ -182,7 +232,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
   conformable(fineX,fineZ);
 
   int _ndimension = coarse->_ndimension;
-  
+
   Coordinate  block_r      (_ndimension);
 
   // FIXME merge with subdivide checking routine as this is redundant
@@ -191,29 +241,68 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
     assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
-  auto fineZ_  = fineZ.View();
-  auto fineX_  = fineX.View();
-  auto fineY_  = fineY.View();
-  auto coarseA_= coarseA.View();
+  autoView( fineZ_  , fineZ, AcceleratorWrite);
+  autoView( fineX_  , fineX, AcceleratorRead);
+  autoView( fineY_  , fineY, AcceleratorRead);
+  autoView( coarseA_, coarseA, AcceleratorRead);
+  Coordinate fine_rdimensions = fine->_rdimensions;
+  Coordinate coarse_rdimensions = coarse->_rdimensions;
 
   accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
-    
-    int sc;
-    Coordinate coor_c(_ndimension);
-    Coordinate coor_f(_ndimension);
 
-    Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
-    for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-    Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+      int sc;
+      Coordinate coor_c(_ndimension);
+      Coordinate coor_f(_ndimension);
 
-    // z = A x + y
-    coalescedWrite(fineZ_[sf],coarseA_(sc)*fineX_(sf)+fineY_(sf));
+      Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+      for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
+      Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
 
-  });
+      // z = A x + y
+#ifdef GRID_SIMT
+      typename vobj2::tensor_reduced::scalar_object cA;
+      typename vobj::scalar_object cAx;
+#else
+      typename vobj2::tensor_reduced cA;
+      vobj cAx;
+#endif
+      convertType(cA,TensorRemove(coarseA_(sc)));
+      auto prod = cA*fineX_(sf);
+      convertType(cAx,prod);
+      coalescedWrite(fineZ_[sf],cAx+fineY_(sf));
+
+    });
 
   return;
 }
+
 template<class vobj,class CComplex>
+  inline void blockInnerProductD(Lattice<CComplex> &CoarseInner,
+				 const Lattice<vobj> &fineX,
+				 const Lattice<vobj> &fineY)
+{
+  typedef iScalar<decltype(TensorRemove(innerProductD2(vobj(),vobj())))> dotp;
+
+  GridBase *coarse(CoarseInner.Grid());
+  GridBase *fine  (fineX.Grid());
+
+  Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
+  Lattice<dotp> coarse_inner(coarse);
+
+  // Precision promotion
+  fine_inner = localInnerProductD<vobj>(fineX,fineY);
+  blockSum(coarse_inner,fine_inner);
+  {
+    autoView( CoarseInner_  , CoarseInner,AcceleratorWrite);
+    autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
+    accelerator_for(ss, coarse->oSites(), 1, {
+      convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
+    });
+  }
+ 
+}
+
+template<class vobj,class CComplex> // deprecate
 inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
 			      const Lattice<vobj> &fineX,
 			      const Lattice<vobj> &fineY)
@@ -227,15 +316,17 @@ inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
   Lattice<dotp> coarse_inner(coarse);
 
   // Precision promotion?
-  auto CoarseInner_  = CoarseInner.View();
-  auto coarse_inner_ = coarse_inner.View();
-
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-  accelerator_for(ss, coarse->oSites(), 1, {
-    CoarseInner_[ss] = coarse_inner_[ss];
-  });
+  {
+    autoView( CoarseInner_  , CoarseInner, AcceleratorWrite);
+    autoView( coarse_inner_ , coarse_inner, AcceleratorRead);
+    accelerator_for(ss, coarse->oSites(), 1, {
+	CoarseInner_[ss] = coarse_inner_[ss];
+    });
+  }
 }
+
 template<class vobj,class CComplex>
 inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 {
@@ -248,7 +339,7 @@ inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 // useful in multigrid project;
 // Generic name : Coarsen?
 template<class vobj>
-inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
+inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData) 
 {
   GridBase * fine  = fineData.Grid();
   GridBase * coarse= coarseData.Grid();
@@ -256,9 +347,9 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   subdivides(coarse,fine); // require they map
 
   int _ndimension = coarse->_ndimension;
-  
+
   Coordinate  block_r      (_ndimension);
-  
+
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
@@ -266,32 +357,36 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 
   // Turn this around to loop threaded over sc and interior loop 
   // over sf would thread better
-  auto coarseData_ = coarseData.View();
-  auto fineData_   = fineData.View();
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( fineData_   , fineData, AcceleratorRead);
 
+  Coordinate fine_rdimensions = fine->_rdimensions;
+  Coordinate coarse_rdimensions = coarse->_rdimensions;
+  
   accelerator_for(sc,coarse->oSites(),1,{
 
-    // One thread per sub block
-    Coordinate coor_c(_ndimension);
-    Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions);  // Block coordinate
-    coarseData_[sc]=Zero();
+      // One thread per sub block
+      Coordinate coor_c(_ndimension);
+      Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
+      coarseData_[sc]=Zero();
 
-    for(int sb=0;sb<blockVol;sb++){
-      
-      int sf;
-      Coordinate coor_b(_ndimension);
-      Coordinate coor_f(_ndimension);
-      Lexicographic::CoorFromIndex(coor_b,sb,block_r);               // Block sub coordinate
-      for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
-      Lexicographic::IndexFromCoor(coor_f,sf,fine->_rdimensions);
+      for(int sb=0;sb<blockVol;sb++){
 
-      coarseData_[sc]=coarseData_[sc]+fineData_[sf];
-    }
+	int sf;
+	Coordinate coor_b(_ndimension);
+	Coordinate coor_f(_ndimension);
+	Lexicographic::CoorFromIndex(coor_b,sb,block_r);               // Block sub coordinate
+	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
+	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
 
-  });
+	coarseData_[sc]=coarseData_[sc]+fineData_[sf];
+      }
+
+    });
   return;
 }
 
+
 template<class vobj>
 inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vobj> &picked,Coordinate coor)
 {
@@ -313,8 +408,8 @@ inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vob
   }
 }
 
-template<class vobj,class CComplex>
-inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> > &Basis)
+template<class CComplex,class VLattice>
+inline void blockOrthonormalize(Lattice<CComplex> &ip,VLattice &Basis)
 {
   GridBase *coarse = ip.Grid();
   GridBase *fine   = Basis[0].Grid();
@@ -322,23 +417,30 @@ inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> >
   int       nbasis = Basis.size() ;
 
   // checks
-  subdivides(coarse,fine); 
+  subdivides(coarse,fine);
   for(int i=0;i<nbasis;i++){
     conformable(Basis[i].Grid(),fine);
   }
 
   for(int v=0;v<nbasis;v++) {
     for(int u=0;u<v;u++) {
-      //Inner product & remove component 
-      blockInnerProduct(ip,Basis[u],Basis[v]);
+      //Inner product & remove component
+      blockInnerProductD(ip,Basis[u],Basis[v]);
       ip = -ip;
-      blockZAXPY<vobj,CComplex> (Basis[v],ip,Basis[u],Basis[v]);
+      blockZAXPY(Basis[v],ip,Basis[u],Basis[v]);
     }
     blockNormalise(ip,Basis[v]);
   }
 }
 
+template<class vobj,class CComplex>
+inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> > &Basis) // deprecated inaccurate naming
+{
+  blockOrthonormalize(ip,Basis);
+}
+
 #if 0
+// TODO: CPU optimized version here
 template<class vobj,class CComplex,int nbasis>
 inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 			 Lattice<vobj>   &fineData,
@@ -360,8 +462,8 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
-  auto fineData_   = fineData.View();
-  auto coarseData_ = coarseData.View();
+  autoView( fineData_   , fineData, AcceleratorWrite);
+  autoView( coarseData_ , coarseData, AcceleratorRead);
 
   // Loop with a cache friendly loop ordering
   accelerator_for(sf,fine->oSites(),1,{
@@ -374,7 +476,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
     Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
     for(int i=0;i<nbasis;i++) {
-      auto basis_ = Basis[i].View();
+      /*      auto basis_ = Basis[i],  );*/
       if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
       else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
     }
@@ -383,24 +485,25 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   
 }
 #else
-template<class vobj,class CComplex,int nbasis>
+template<class vobj,class CComplex,int nbasis,class VLattice>
 inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 			 Lattice<vobj>   &fineData,
-			 const std::vector<Lattice<vobj> > &Basis)
+			 const VLattice &Basis)
 {
   GridBase * fine  = fineData.Grid();
   GridBase * coarse= coarseData.Grid();
-
   fineData=Zero();
   for(int i=0;i<nbasis;i++) {
     Lattice<iScalar<CComplex> > ip = PeekIndex<0>(coarseData,i);
-    Lattice<CComplex> cip(coarse);
-    auto cip_ = cip.View();
-    auto  ip_ =  ip.View();
-    accelerator_forNB(sc,coarse->oSites(),CComplex::Nsimd(),{
-	coalescedWrite(cip_[sc], ip_(sc)());
-    });
-    blockZAXPY<vobj,CComplex >(fineData,cip,Basis[i],fineData);
+
+    //Lattice<CComplex> cip(coarse);
+    //autoView( cip_ , cip, AcceleratorWrite);
+    //autoView(  ip_ ,  ip, AcceleratorRead);
+    //accelerator_forNB(sc,coarse->oSites(),CComplex::Nsimd(),{
+    //	coalescedWrite(cip_[sc], ip_(sc)());
+    //  });
+    //blockZAXPY<vobj,CComplex >(fineData,cip,Basis[i],fineData);
+    blockZAXPY(fineData,ip,Basis[i],fineData);
   }
 }
 #endif
@@ -427,15 +530,17 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
     assert(ig->lSites() == og->lSites());
   }
 
+  autoView(in_v,in,CpuRead);
+  autoView(out_v,out,CpuWrite);
   thread_for(idx, ig->lSites(),{
     sobj s;
     ssobj ss;
 
     Coordinate lcoor(ni);
     ig->LocalIndexToLocalCoor(idx,lcoor);
-    peekLocalSite(s,in,lcoor);
+    peekLocalSite(s,in_v,lcoor);
     ss=s;
-    pokeLocalSite(ss,out,lcoor);
+    pokeLocalSite(ss,out_v,lcoor);
   });
 }
 
@@ -470,8 +575,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate rdt = Tg->_rdimensions;
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
-  auto t_v = To.View();
-  auto f_v = From.View();
+
+  autoView( t_v , To, AcceleratorWrite);
+  autoView( f_v , From, AcceleratorRead);
   accelerator_for(idx,Fg->lSites(),1,{
     sobj s;
     Coordinate Fcoor(nd);
@@ -494,8 +600,6 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       for(int w=0;w<words;w++){
 	tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd];  // FIXME IF RRII layout, type pun no worke
       }
-      //      peekLocalSite(s,From,Fcoor);
-      //      pokeLocalSite(s,To  ,Tcoor);
     }
   });
 }
@@ -526,6 +630,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuRead);
+  autoView(higherDimv,higherDim,CpuWrite);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -538,8 +644,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,lowDim,lcoor);
-    pokeLocalSite(s,higherDim,hcoor);
+    peekLocalSite(s,lowDimv,lcoor);
+    pokeLocalSite(s,higherDimv,hcoor);
   });
 }
 
@@ -567,6 +673,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     }
   }
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuWrite);
+  autoView(higherDimv,higherDim,CpuRead);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -579,8 +687,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 	hcoor[d]=lcoor[ddl++];
       }
     }
-    peekLocalSite(s,higherDim,hcoor);
-    pokeLocalSite(s,lowDim,lcoor);
+    peekLocalSite(s,higherDimv,hcoor);
+    pokeLocalSite(s,lowDimv,lcoor);
   });
 
 }
@@ -608,6 +716,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuRead);
+  autoView(higherDimv,higherDim,CpuWrite);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -616,8 +726,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,lowDim,lcoor);
-      pokeLocalSite(s,higherDim,hcoor);
+      peekLocalSite(s,lowDimv,lcoor);
+      pokeLocalSite(s,higherDimv,hcoor);
     }
   });
 }
@@ -645,6 +755,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
   }
 
   // the above should guarantee that the operations are local
+  autoView(lowDimv,lowDim,CpuWrite);
+  autoView(higherDimv,higherDim,CpuRead);
   thread_for(idx,lg->lSites(),{
     sobj s;
     Coordinate lcoor(nl);
@@ -653,8 +765,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
     if( lcoor[orthog] == slice_lo ) { 
       hcoor=lcoor;
       hcoor[orthog] = slice_hi;
-      peekLocalSite(s,higherDim,hcoor);
-      pokeLocalSite(s,lowDim,lcoor);
+      peekLocalSite(s,higherDimv,hcoor);
+      pokeLocalSite(s,lowDimv,lcoor);
     }
   });
 }
@@ -718,7 +830,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
   }
 
   //loop over outer index
-  auto in_v  = in.View();
+  autoView( in_v  , in, CpuRead);
   thread_for(in_oidx,in_grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> out_ptrs(in_nsimd);
@@ -811,7 +923,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
     icoor[lane].resize(ndim);
     grid->iCoorFromIindex(icoor[lane],lane);
   }
-  auto out_v = out.View();
+  autoView( out_v , out, CpuWrite);
   thread_for(oidx, grid->oSites(),{
     //Assemble vector of pointers to output elements
     ExtractPointerArray<sobj> ptrs(nsimd);
@@ -914,7 +1026,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   std::vector<SobjOut> in_slex_conv(in_grid->lSites());
   unvectorizeToLexOrdArray(in_slex_conv, in);
     
-  auto out_v = out.View();
+  autoView( out_v , out, CpuWrite);
   thread_for(out_oidx,out_grid->oSites(),{
     Coordinate out_ocoor(ndim);
     out_grid->oCoorFromOindex(out_ocoor, out_oidx);

Grid/lattice/Lattice_transpose.h

@@ -38,17 +38,19 @@ NAMESPACE_BEGIN(Grid);
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Transpose
 ////////////////////////////////////////////////////////////////////////////////////////////////////
+/*
 template<class vobj>
 inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
   Lattice<vobj> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
   });
   return ret;
 };
-    
+*/    
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Index level dependent transpose
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -56,8 +58,8 @@ template<int Index,class vobj>
 inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
 {
   Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
-  auto ret_v = ret.View();
-  auto lhs_v = lhs.View();
+  autoView( ret_v, ret, AcceleratorWrite);
+  autoView( lhs_v, lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
     coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
   });

Grid/lattice/Lattice_unary.h

@@ -35,8 +35,8 @@ NAMESPACE_BEGIN(Grid);
 
 template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs, rhs_i, AcceleratorRead);
+  autoView( ret, ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),1,{
       ret[ss]=pow(rhs[ss],y);
@@ -45,8 +45,8 @@ template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
 }
 template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs , rhs_i, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],mod(rhs(ss),y));
@@ -56,8 +56,8 @@ template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
 
 template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto ret = ret_i.View();
-  auto rhs = rhs_i.View();
+  autoView( ret , ret_i, AcceleratorWrite);
+  autoView( rhs , rhs_i, AcceleratorRead);
   ret.Checkerboard() = rhs_i.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],div(rhs(ss),y));
@@ -67,8 +67,8 @@ template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
 
 template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
   Lattice<obj> ret_i(rhs_i.Grid());
-  auto rhs = rhs_i.View();
-  auto ret = ret_i.View();
+  autoView( rhs , rhs_i, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
   ret.Checkerboard() = rhs.Checkerboard();
   accelerator_for(ss,rhs.size(),obj::Nsimd(),{
     coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));

Grid/lattice/Lattice_view.h

@@ -0,0 +1,168 @@
+#pragma once
+NAMESPACE_BEGIN(Grid);
+///////////////////////////////////////////////////////////////////
+// Base class which can be used by traits to pick up behaviour
+///////////////////////////////////////////////////////////////////
+class LatticeBase {};
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// Conformable checks; same instance of Grid required
+/////////////////////////////////////////////////////////////////////////////////////////
+void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
+{
+  assert(lhs == rhs);
+}
+
+////////////////////////////////////////////////////////////////////////////
+// Minimal base class containing only data valid to access from accelerator
+// _odata will be a managed pointer in CUDA
+////////////////////////////////////////////////////////////////////////////
+// Force access to lattice through a view object.
+// prevents writing of code that will not offload to GPU, but perhaps annoyingly
+// strict since host could could in principle direct access through the lattice object
+// Need to decide programming model.
+#define LATTICE_VIEW_STRICT
+template<class vobj> class LatticeAccelerator : public LatticeBase
+{
+protected:
+  //public:
+  GridBase *_grid;
+  int checkerboard;
+  vobj     *_odata;    // A managed pointer
+  uint64_t _odata_size;    
+  ViewAdvise advise;
+public:
+  accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr), advise(AdviseDefault) { }; 
+  accelerator_inline uint64_t oSites(void) const { return _odata_size; };
+  accelerator_inline int  Checkerboard(void) const { return checkerboard; };
+  accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
+  accelerator_inline ViewAdvise Advise(void) const { return advise; };
+  accelerator_inline ViewAdvise &Advise(void) { return this->advise; }; // can assign advise on a container, not a view
+  accelerator_inline void Conformable(GridBase * &grid) const
+  { 
+    if (grid) conformable(grid, _grid);
+    else      grid = _grid;
+  };
+  // Host only
+  GridBase * getGrid(void) const { return _grid; };
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// A View class which provides accessor to the data.
+// This will be safe to call from accelerator_for and is trivially copy constructible
+// The copy constructor for this will need to be used by device lambda functions
+/////////////////////////////////////////////////////////////////////////////////////////
+template<class vobj> 
+class LatticeView : public LatticeAccelerator<vobj>
+{
+public:
+  // Rvalue
+  ViewMode mode;
+  void * cpu_ptr;
+#ifdef GRID_SIMT
+  accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { 
+    return coalescedRead(this->_odata[i]); 
+  }
+#else 
+  accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
+#endif
+
+  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
+  accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; };
+
+  accelerator_inline uint64_t begin(void) const { return 0;};
+  accelerator_inline uint64_t end(void)   const { return this->_odata_size; };
+  accelerator_inline uint64_t size(void)  const { return this->_odata_size; };
+
+  LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me){}
+  LatticeView(const LatticeView<vobj> &refer_to_me) = default; // Trivially copyable
+  LatticeView(const LatticeAccelerator<vobj> &refer_to_me,ViewMode mode) : LatticeAccelerator<vobj> (refer_to_me)
+  {
+    this->ViewOpen(mode);
+  }
+
+  // Host functions
+  void ViewOpen(ViewMode mode)
+  { // Translate the pointer, could save a copy. Could use a "Handle" and not save _odata originally in base
+    //    std::cout << "View Open"<<std::hex<<this->_odata<<std::dec <<std::endl;
+    this->cpu_ptr = (void *)this->_odata;
+    this->mode    = mode;
+    this->_odata  =(vobj *)
+      MemoryManager::ViewOpen(this->cpu_ptr,
+				this->_odata_size*sizeof(vobj),
+				mode,
+				this->advise);    
+  }
+  void ViewClose(void)
+  { // Inform the manager
+    //    std::cout << "View Close"<<std::hex<<this->cpu_ptr<<std::dec <<std::endl;
+    MemoryManager::ViewClose(this->cpu_ptr,this->mode);    
+  }
+
+};
+// Little autoscope assister
+template<class View> 
+class ViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+ public:
+  ViewCloser(View &_v) : v(_v) {};
+  ~ViewCloser() { v.ViewClose(); }
+};
+
+#define autoView(l_v,l,mode)				\
+	  auto l_v = l.View(mode);			\
+	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v);
+
+/////////////////////////////////////////////////////////////////////////////////////////
+// Lattice expression types used by ET to assemble the AST
+// 
+// Need to be able to detect code paths according to the whether a lattice object or not
+// so introduce some trait type things
+/////////////////////////////////////////////////////////////////////////////////////////
+
+class LatticeExpressionBase {};
+
+template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
+template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
+
+template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
+template<class T>                 struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
+template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
+
+template <typename Op, typename _T1>                           
+class LatticeUnaryExpression : public  LatticeExpressionBase 
+{
+public:
+  typedef typename ViewMap<_T1>::Type T1;
+  Op op;
+  T1 arg1;
+  LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
+};
+
+template <typename Op, typename _T1, typename _T2>              
+class LatticeBinaryExpression : public LatticeExpressionBase 
+{
+public:
+  typedef typename ViewMap<_T1>::Type T1;
+  typedef typename ViewMap<_T2>::Type T2;
+  Op op;
+  T1 arg1;
+  T2 arg2;
+  LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
+};
+
+template <typename Op, typename _T1, typename _T2, typename _T3> 
+class LatticeTrinaryExpression : public LatticeExpressionBase 
+{
+public:
+  typedef typename ViewMap<_T1>::Type T1;
+  typedef typename ViewMap<_T2>::Type T2;
+  typedef typename ViewMap<_T3>::Type T3;
+  Op op;
+  T1 arg1;
+  T2 arg2;
+  T3 arg3;
+  LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
+};
+NAMESPACE_END(Grid);

Grid/log/Log.h

@@ -130,6 +130,8 @@ public:
   friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
     if ( log.active ) {
+      std::ios_base::fmtflags f(stream.flags());
+
       stream << log.background()<<  std::left;
       if (log.topWidth > 0)
       {
@@ -152,6 +154,8 @@ public:
 	       << now	       << log.background() << " : " ;
       }
       stream << log.colour();
+      stream.flags(f);
+
       return stream;
     } else { 
       return devnull;

Grid/parallelIO/BinaryIO.cc

@@ -1,3 +1,4 @@
 #include <Grid/GridCore.h>
 
-int Grid::BinaryIO::latticeWriteMaxRetry = -1;
+int                    Grid::BinaryIO::latticeWriteMaxRetry = -1;
+Grid::BinaryIO::IoPerf Grid::BinaryIO::lastPerf;

Grid/parallelIO/BinaryIO.h

@@ -79,6 +79,13 @@ inline void removeWhitespace(std::string &key)
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 class BinaryIO {
  public:
+  struct IoPerf
+  {
+    uint64_t size{0},time{0};
+    double   mbytesPerSecond{0.};
+  };
+
+  static IoPerf lastPerf;
   static int latticeWriteMaxRetry;
 
   /////////////////////////////////////////////////////////////////////////////
@@ -341,7 +348,7 @@ class BinaryIO {
     int ieee32big = (format == std::string("IEEE32BIG"));
     int ieee32    = (format == std::string("IEEE32"));
     int ieee64big = (format == std::string("IEEE64BIG"));
-    int ieee64    = (format == std::string("IEEE64"));
+    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
     assert(ieee64||ieee32|ieee64big||ieee32big);
     assert((ieee64+ieee32+ieee64big+ieee32big)==1);
     //////////////////////////////////////////////////////////////////////////////
@@ -502,12 +509,15 @@ class BinaryIO {
       timer.Stop();
     }
     
+    lastPerf.size            = sizeof(fobj)*iodata.size()*nrank;
+    lastPerf.time            = timer.useconds();
+    lastPerf.mbytesPerSecond = lastPerf.size/1024./1024./(lastPerf.time/1.0e6);
     std::cout<<GridLogMessage<<"IOobject: ";
     if ( control & BINARYIO_READ) std::cout << " read  ";
     else                          std::cout << " write ";
     uint64_t bytes = sizeof(fobj)*iodata.size()*nrank;
-    std::cout<< bytes <<" bytes in "<<timer.Elapsed() <<" "
-	     << (double)bytes/ (double)timer.useconds() <<" MB/s "<<std::endl;
+    std::cout<< lastPerf.size <<" bytes in "<< timer.Elapsed() <<" "
+	     << lastPerf.mbytesPerSecond <<" MB/s "<<std::endl;
 
     std::cout<<GridLogMessage<<"IOobject: endian and checksum overhead "<<bstimer.Elapsed()  <<std::endl;
 
@@ -663,10 +673,15 @@ class BinaryIO {
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     timer.Start();
-    thread_for(lidx,lsites,{
+    thread_for(lidx,lsites,{  // FIX ME, suboptimal implementation
       std::vector<RngStateType> tmp(RngStateCount);
       std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
-      parallel_rng.SetState(tmp,lidx);
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(lidx, lcoor);
+      int o_idx=grid->oIndex(lcoor);
+      int i_idx=grid->iIndex(lcoor);
+      int gidx=parallel_rng.generator_idx(o_idx,i_idx);
+      parallel_rng.SetState(tmp,gidx);
       });
     timer.Stop();
 
@@ -723,7 +738,12 @@ class BinaryIO {
     std::vector<RNGstate> iodata(lsites);
     thread_for(lidx,lsites,{
       std::vector<RngStateType> tmp(RngStateCount);
-      parallel_rng.GetState(tmp,lidx);
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(lidx, lcoor);
+      int o_idx=grid->oIndex(lcoor);
+      int i_idx=grid->iIndex(lcoor);
+      int gidx=parallel_rng.generator_idx(o_idx,i_idx);
+      parallel_rng.GetState(tmp,gidx);
       std::copy(tmp.begin(),tmp.end(),iodata[lidx].begin());
     });
     timer.Stop();

Grid/parallelIO/IldgIO.h

@@ -123,7 +123,7 @@ assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
  ////////////////////////////////////////////////////////////
  // Helper to fill out metadata
  ////////////////////////////////////////////////////////////
- template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
+template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
 					  FieldMetaData &header,
 					  scidacRecord & _scidacRecord,
 					  scidacFile   & _scidacFile) 
@@ -619,12 +619,12 @@ class IldgWriter : public ScidacWriter {
   // Don't require scidac records EXCEPT checksum
   // Use Grid MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
-  void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description) 
+  template <class stats = PeriodicGaugeStatistics>
+  void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description) 
   {
     GridBase * grid = Umu.Grid();
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
-    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef Lattice<vLorentzColourMatrixD> GaugeField;
+    typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     ////////////////////////////////////////
@@ -636,6 +636,9 @@ class IldgWriter : public ScidacWriter {
 
     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
 
+    stats Stats;
+    Stats(Umu,header);
+    
     std::string format = header.floating_point;
     header.ensemble_id    = description;
     header.ensemble_label = description;
@@ -705,10 +708,10 @@ class IldgReader : public GridLimeReader {
   // Else use ILDG MetaData object if present.
   // Else use SciDAC MetaData object if present.
   ////////////////////////////////////////////////////////////////
-  template <class vsimd>
-  void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
+  template <class stats = PeriodicGaugeStatistics>
+  void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) {
 
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
+    typedef Lattice<vLorentzColourMatrixD > GaugeField;
     typedef typename GaugeField::vector_object  vobj;
     typedef typename vobj::scalar_object sobj;
 
@@ -921,7 +924,8 @@ class IldgReader : public GridLimeReader {
 
     if ( found_FieldMetaData || found_usqcdInfo ) {
       FieldMetaData checker;
-      GaugeStatistics(Umu,checker);
+      stats Stats;
+      Stats(Umu,checker);
       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5);
       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;

Grid/parallelIO/MetaData.h

@@ -176,29 +176,18 @@ template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMet
   GridMetaData(grid,header); 
   MachineCharacteristics(header);
 }
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
+template<class Impl>
+class GaugeStatistics
 {
-  // How to convert data precision etc...
-  header.link_trace=WilsonLoops<PeriodicGimplF>::linkTrace(data);
-  header.plaquette =WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
-}
-inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
-{
-  // How to convert data precision etc...
-  header.link_trace=WilsonLoops<PeriodicGimplD>::linkTrace(data);
-  header.plaquette =WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
-}
-template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header)
-{
-   
-  GridBase *grid = field.Grid();
-  std::string format = getFormatString<vLorentzColourMatrixF>();
-  header.floating_point = format;
-  header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
-  GridMetaData(grid,header); 
-  GaugeStatistics(field,header);
-  MachineCharacteristics(header);
-}
+public:
+  void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
+  {
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
+  }
+};
+typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
+typedef GaugeStatistics<ConjugateGimplD> ConjugateGaugeStatistics;
 template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
 {
   GridBase *grid = field.Grid();
@@ -206,7 +195,6 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
   header.floating_point = format;
   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
   GridMetaData(grid,header); 
-  GaugeStatistics(field,header);
   MachineCharacteristics(header);
 }
 
@@ -301,6 +289,30 @@ struct GaugeSimpleUnmunger {
   };
 };
 
+template<class fobj,class sobj>
+struct GaugeDoubleStoredMunger{
+  void operator()(fobj &in, sobj &out) {
+    for (int mu = 0; mu < Nds; mu++) {
+      for (int i = 0; i < Nc; i++) {
+        for (int j = 0; j < Nc; j++) {
+          out(mu)()(i, j) = in(mu)()(i, j);
+        }}
+    }
+  };
+};
+
+template <class fobj, class sobj>
+struct GaugeDoubleStoredUnmunger {
+  void operator()(sobj &in, fobj &out) {
+    for (int mu = 0; mu < Nds; mu++) {
+      for (int i = 0; i < Nc; i++) {
+        for (int j = 0; j < Nc; j++) {
+          out(mu)()(i, j) = in(mu)()(i, j);
+        }}
+    }
+  };
+};
+
 template<class fobj,class sobj>
 struct Gauge3x2munger{
   void operator() (fobj &in,sobj &out){

Grid/parallelIO/NerscIO.h

@@ -40,6 +40,8 @@ using namespace Grid;
 class NerscIO : public BinaryIO { 
 public:
 
+  typedef Lattice<vLorentzColourMatrixD> GaugeField;
+
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -129,12 +131,12 @@ public:
   // Now the meat: the object readers
   /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-  template<class vsimd>
-  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+  template<class GaugeStats=PeriodicGaugeStatistics>
+  static inline void readConfiguration(GaugeField &Umu,
 				       FieldMetaData& header,
-				       std::string file)
+				       std::string file,
+				       GaugeStats GaugeStatisticsCalculator=GaugeStats())
   {
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
 
     GridBase *grid = Umu.Grid();
     uint64_t offset = readHeader(file,Umu.Grid(),header);
@@ -146,30 +148,30 @@ public:
     int ieee32big = (format == std::string("IEEE32BIG"));
     int ieee32    = (format == std::string("IEEE32"));
     int ieee64big = (format == std::string("IEEE64BIG"));
-    int ieee64    = (format == std::string("IEEE64"));
+    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     // depending on datatype, set up munger;
     // munger is a function of <floating point, Real, data_type>
     if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F> 
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D> 
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3D> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
     } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
       if ( ieee64 || ieee64big ) {
-	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
+	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixD>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
@@ -177,7 +179,7 @@ public:
       assert(0);
     }
 
-    GaugeStatistics(Umu,clone);
+    GaugeStats Stats; Stats(Umu,clone);
 
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl;
@@ -203,15 +205,13 @@ public:
     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
   }
 
-  template<class vsimd>
-  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
+  template<class GaugeStats=PeriodicGaugeStatistics>
+  static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32)
   {
-    typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
-
-    typedef iLorentzColourMatrix<vsimd> vobj;
+    typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
@@ -229,7 +229,7 @@ public:
 
     GridMetaData(grid,header);
     assert(header.nd==4);
-    GaugeStatistics(Umu,header);
+    GaugeStats Stats; Stats(Umu,header);
     MachineCharacteristics(header);
 
 	uint64_t offset;
@@ -238,19 +238,19 @@ public:
     header.floating_point = std::string("IEEE64BIG");
     header.data_type      = std::string("4D_SU3_GAUGE_3x3");
     GaugeSimpleUnmunger<fobj3D,sobj> munge;
-	if ( grid->IsBoss() ) { 
-	  truncate(file);
-    offset = writeHeader(header,file);
-	}
-	grid->Broadcast(0,(void *)&offset,sizeof(offset));
+    if ( grid->IsBoss() ) { 
+      truncate(file);
+      offset = writeHeader(header,file);
+    }
+    grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
 					      nersc_csum,scidac_csuma,scidac_csumb);
     header.checksum = nersc_csum;
-	if ( grid->IsBoss() ) { 
-    writeHeader(header,file);
-	}
+    if ( grid->IsBoss() ) { 
+      writeHeader(header,file);
+    }
 
     std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
 	     <<std::hex<<header.checksum

Grid/parallelIO/OpenQcdIO.h

@@ -0,0 +1,224 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/parallelIO/OpenQcdIO.h
+
+Copyright (C) 2015 - 2020
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+struct OpenQcdHeader : Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(OpenQcdHeader,
+                                  int,    Nt,
+                                  int,    Nx,
+                                  int,    Ny,
+                                  int,    Nz,
+                                  double, plaq);
+};
+
+class OpenQcdIO : public BinaryIO {
+public:
+  static constexpr double normalisationFactor = Nc; // normalisation difference: grid 18, openqcd 6
+
+  static inline int readHeader(std::string file, GridBase* grid, FieldMetaData& field) {
+    OpenQcdHeader header;
+
+    {
+      std::ifstream fin(file, std::ios::in | std::ios::binary);
+      fin.read(reinterpret_cast<char*>(&header), sizeof(OpenQcdHeader));
+      assert(!fin.fail());
+      field.data_start = fin.tellg();
+      fin.close();
+    }
+
+    header.plaq /= normalisationFactor;
+
+    // sanity check (should trigger on endian issues)
+    assert(0 < header.Nt && header.Nt <= 1024);
+    assert(0 < header.Nx && header.Nx <= 1024);
+    assert(0 < header.Ny && header.Ny <= 1024);
+    assert(0 < header.Nz && header.Nz <= 1024);
+
+    field.dimension[0] = header.Nx;
+    field.dimension[1] = header.Ny;
+    field.dimension[2] = header.Nz;
+    field.dimension[3] = header.Nt;
+
+    std::cout << GridLogDebug << "header: " << header << std::endl;
+    std::cout << GridLogDebug << "grid dimensions: " << grid->_fdimensions << std::endl;
+    std::cout << GridLogDebug << "file dimensions: " << field.dimension << std::endl;
+
+    assert(grid->_ndimension == Nd);
+    for(int d = 0; d < Nd; d++)
+      assert(grid->_fdimensions[d] == field.dimension[d]);
+
+    field.plaquette = header.plaq;
+
+    return field.data_start;
+  }
+
+  template<class vsimd>
+  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
+                                       FieldMetaData&                        header,
+                                       std::string                           file) {
+    typedef Lattice<iDoubleStoredColourMatrix<vsimd>> DoubleStoredGaugeField;
+
+    assert(Ns == 4 and Nd == 4 and Nc == 3);
+
+    auto grid = dynamic_cast<GridCartesian*>(Umu.Grid());
+    assert(grid != nullptr); assert(grid->_ndimension == Nd);
+
+    uint64_t offset = readHeader(file, Umu.Grid(), header);
+
+    FieldMetaData clone(header);
+
+    std::string format("IEEE64"); // they always store little endian double precsision
+    uint32_t    nersc_csum, scidac_csuma, scidac_csumb;
+
+    GridCartesian*         grid_openqcd = createOpenQcdGrid(grid);
+    GridRedBlackCartesian* grid_rb      = SpaceTimeGrid::makeFourDimRedBlackGrid(grid);
+
+    typedef DoubleStoredColourMatrixD                                              fobj;
+    typedef typename DoubleStoredGaugeField::vector_object::scalar_object          sobj;
+    typedef typename DoubleStoredGaugeField::vector_object::Realified::scalar_type word;
+
+    word w = 0;
+
+    std::vector<fobj> iodata(grid_openqcd->lSites()); // Munge, checksum, byte order in here
+    std::vector<sobj> scalardata(grid->lSites());
+
+    IOobject(w, grid_openqcd, iodata, file, offset, format, BINARYIO_READ | BINARYIO_LEXICOGRAPHIC,
+             nersc_csum, scidac_csuma, scidac_csumb);
+
+    GridStopWatch timer;
+    timer.Start();
+
+    DoubleStoredGaugeField Umu_ds(grid);
+
+    auto munge = GaugeDoubleStoredMunger<DoubleStoredColourMatrixD, DoubleStoredColourMatrix>();
+
+    Coordinate ldim = grid->LocalDimensions();
+    thread_for(idx_g, grid->lSites(), {
+        Coordinate coor;
+        grid->LocalIndexToLocalCoor(idx_g, coor);
+
+        bool isOdd = grid_rb->CheckerBoard(coor) == Odd;
+
+        if(!isOdd) continue;
+
+        int idx_o = (coor[Tdir] * ldim[Xdir] * ldim[Ydir] * ldim[Zdir]
+                  +  coor[Xdir] * ldim[Ydir] * ldim[Zdir]
+                  +  coor[Ydir] * ldim[Zdir]
+                  +  coor[Zdir])/2;
+
+        munge(iodata[idx_o], scalardata[idx_g]);
+    });
+
+    grid->Barrier(); timer.Stop();
+    std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: munge overhead " << timer.Elapsed() << std::endl;
+
+    timer.Reset(); timer.Start();
+
+    vectorizeFromLexOrdArray(scalardata, Umu_ds);
+
+    grid->Barrier(); timer.Stop();
+    std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: vectorize overhead " << timer.Elapsed() << std::endl;
+
+    timer.Reset(); timer.Start();
+
+    undoDoubleStore(Umu, Umu_ds);
+
+    grid->Barrier(); timer.Stop();
+    std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: redistribute overhead " << timer.Elapsed() << std::endl;
+
+    PeriodicGaugeStatistics Stats; Stats(Umu, clone);
+
+    RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
+
+    // clang-format off
+    std::cout << GridLogMessage << "OpenQcd Configuration " << file
+              << " plaquette " << clone.plaquette
+              << " header " << header.plaquette
+              << " difference " << plaq_diff
+              << std::endl;
+    // clang-format on
+
+    RealD precTol = (getPrecision<vsimd>::value == 1) ? 2e-7 : 2e-15;
+    RealD tol     = precTol * std::sqrt(grid->_Nprocessors); // taken from RQCD chroma code
+
+    if(plaq_diff >= tol)
+      std::cout << " Plaquette mismatch (diff = " << plaq_diff << ", tol = " << tol << ")" << std::endl;
+    assert(plaq_diff < tol);
+
+    std::cout << GridLogMessage << "OpenQcd Configuration " << file << " and plaquette agree" << std::endl;
+  }
+
+  template<class vsimd>
+  static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
+                                        std::string                           file) {
+    std::cout << GridLogError << "Writing to openQCD file format is not implemented" << std::endl;
+    exit(EXIT_FAILURE);
+  }
+
+private:
+  static inline GridCartesian* createOpenQcdGrid(GridCartesian* grid) {
+    // exploit GridCartesian to be able to still use IOobject
+    Coordinate gdim  = grid->GlobalDimensions();
+    Coordinate ldim  = grid->LocalDimensions();
+    Coordinate pcoor = grid->ThisProcessorCoor();
+
+    // openqcd does rb on the z direction
+    gdim[Zdir] /= 2;
+    ldim[Zdir] /= 2;
+
+    // and has the order T X Y Z (from slowest to fastest)
+    std::swap(gdim[Xdir], gdim[Zdir]);
+    std::swap(ldim[Xdir], ldim[Zdir]);
+    std::swap(pcoor[Xdir], pcoor[Zdir]);
+
+    GridCartesian* ret   = SpaceTimeGrid::makeFourDimGrid(gdim, grid->_simd_layout, grid->ProcessorGrid());
+    ret->_ldimensions    = ldim;
+    ret->_processor_coor = pcoor;
+    return ret;
+  }
+
+  template<class vsimd>
+  static inline void undoDoubleStore(Lattice<iLorentzColourMatrix<vsimd>>&            Umu,
+                                     Lattice<iDoubleStoredColourMatrix<vsimd>> const& Umu_ds) {
+    conformable(Umu.Grid(), Umu_ds.Grid());
+    Lattice<iColourMatrix<vsimd>> U(Umu.Grid());
+
+    // they store T+, T-, X+, X-, Y+, Y-, Z+, Z-
+    for(int mu_g = 0; mu_g < Nd; ++mu_g) {
+      int mu_o = (mu_g + 1) % Nd;
+      U        = PeekIndex<LorentzIndex>(Umu_ds, 2 * mu_o)
+               + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 2 * mu_o + 1), mu_g, +1);
+      PokeIndex<LorentzIndex>(Umu, U, mu_g);
+    }
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/parallelIO/OpenQcdIOChromaReference.h

@@ -0,0 +1,281 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/parallelIO/OpenQcdIOChromaReference.h
+
+Copyright (C) 2015 - 2020
+
+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 */
+#pragma once
+
+#include <ios>
+#include <iostream>
+#include <limits>
+#include <iomanip>
+#include <mpi.h>
+#include <ostream>
+#include <string>
+
+#define CHECK {std::cerr << __FILE__ << " @l " << __LINE__ << ": CHECK" << grid->ThisRank() << std::endl;}
+#define CHECK_VAR(a)   { std::cerr << __FILE__ << "@l" << __LINE__ << " on "<< grid->ThisRank() << ": " << __func__ << " " << #a << "=" << (a) << std::endl; }
+// #undef CHECK
+// #define CHECK
+
+NAMESPACE_BEGIN(Grid);
+
+class ParRdr {
+private:
+  bool const swap;
+
+  MPI_Status status;
+  MPI_File   fp;
+
+  int err;
+
+  MPI_Datatype oddSiteType;
+  MPI_Datatype fileViewType;
+
+  GridBase* grid;
+
+public:
+  ParRdr(MPI_Comm comm, std::string const& filename, GridBase* gridPtr)
+    : swap(false)
+    , grid(gridPtr) {
+    err = MPI_File_open(comm, const_cast<char*>(filename.c_str()), MPI_MODE_RDONLY, MPI_INFO_NULL, &fp);
+    assert(err == MPI_SUCCESS);
+  }
+
+  virtual ~ParRdr() { MPI_File_close(&fp); }
+
+  inline void errInfo(int const err, std::string const& func) {
+    static char estring[MPI_MAX_ERROR_STRING];
+    int         eclass = -1, len = 0;
+    MPI_Error_class(err, &eclass);
+    MPI_Error_string(err, estring, &len);
+    std::cerr << func << " - Error " << eclass << ": " << estring << std::endl;
+  }
+
+  int readHeader(FieldMetaData& field) {
+    assert((grid->_ndimension == Nd) && (Nd == 4));
+    assert(Nc == 3);
+
+    OpenQcdHeader header;
+
+    readBlock(reinterpret_cast<char*>(&header), 0, sizeof(OpenQcdHeader), MPI_CHAR);
+
+    header.plaq /= 3.; // TODO change this into normalizationfactor
+
+    // sanity check (should trigger on endian issues) TODO remove?
+    assert(0 < header.Nt && header.Nt <= 1024);
+    assert(0 < header.Nx && header.Nx <= 1024);
+    assert(0 < header.Ny && header.Ny <= 1024);
+    assert(0 < header.Nz && header.Nz <= 1024);
+
+    field.dimension[0] = header.Nx;
+    field.dimension[1] = header.Ny;
+    field.dimension[2] = header.Nz;
+    field.dimension[3] = header.Nt;
+
+    for(int d = 0; d < Nd; d++)
+      assert(grid->FullDimensions()[d] == field.dimension[d]);
+
+    field.plaquette = header.plaq;
+
+    field.data_start = sizeof(OpenQcdHeader);
+
+    return field.data_start;
+  }
+
+  void readBlock(void* const dest, uint64_t const pos, uint64_t const nbytes, MPI_Datatype const datatype) {
+    err = MPI_File_read_at_all(fp, pos, dest, nbytes, datatype, &status);
+    errInfo(err, "MPI_File_read_at_all");
+    // CHECK_VAR(err)
+
+    int read = -1;
+    MPI_Get_count(&status, datatype, &read);
+    // CHECK_VAR(read)
+    assert(nbytes == (uint64_t)read);
+    assert(err == MPI_SUCCESS);
+  }
+
+  void createTypes() {
+    constexpr int elem_size = Nd * 2 * 2 * Nc * Nc * sizeof(double); // 2_complex 2_fwdbwd
+
+    err = MPI_Type_contiguous(elem_size, MPI_BYTE, &oddSiteType); assert(err == MPI_SUCCESS);
+    err = MPI_Type_commit(&oddSiteType); assert(err == MPI_SUCCESS);
+
+    Coordinate const L = grid->GlobalDimensions();
+    Coordinate const l = grid->LocalDimensions();
+    Coordinate const i = grid->ThisProcessorCoor();
+
+    Coordinate sizes({L[2] / 2, L[1], L[0], L[3]});
+    Coordinate subsizes({l[2] / 2, l[1], l[0], l[3]});
+    Coordinate starts({i[2] * l[2] / 2, i[1] * l[1], i[0] * l[0], i[3] * l[3]});
+
+    err = MPI_Type_create_subarray(grid->_ndimension, &sizes[0], &subsizes[0], &starts[0], MPI_ORDER_FORTRAN, oddSiteType, &fileViewType); assert(err == MPI_SUCCESS);
+    err = MPI_Type_commit(&fileViewType); assert(err == MPI_SUCCESS);
+  }
+
+  void freeTypes() {
+    err = MPI_Type_free(&fileViewType); assert(err == MPI_SUCCESS);
+    err = MPI_Type_free(&oddSiteType); assert(err == MPI_SUCCESS);
+  }
+
+  bool readGauge(std::vector<ColourMatrixD>& domain_buff, FieldMetaData& meta) {
+    auto hdr_offset = readHeader(meta);
+    CHECK
+    createTypes();
+    err = MPI_File_set_view(fp, hdr_offset, oddSiteType, fileViewType, "native", MPI_INFO_NULL); errInfo(err, "MPI_File_set_view0"); assert(err == MPI_SUCCESS);
+    CHECK
+    int const domainSites = grid->lSites();
+    domain_buff.resize(Nd * domainSites); // 2_fwdbwd * 4_Nd * domainSites / 2_onlyodd
+
+    // the actual READ
+    constexpr uint64_t cm_size   = 2 * Nc * Nc * sizeof(double);    // 2_complex
+    constexpr uint64_t os_size   = Nd * 2 * cm_size;                // 2_fwdbwd
+    constexpr uint64_t max_elems = std::numeric_limits<int>::max(); // int adressable elems: floor is fine
+    uint64_t const     n_os      = domainSites / 2;
+
+    for(uint64_t os_idx = 0; os_idx < n_os;) {
+      uint64_t const read_os = os_idx + max_elems <= n_os ? max_elems : n_os - os_idx;
+      uint64_t const cm      = os_idx * Nd * 2;
+      readBlock(&(domain_buff[cm]), os_idx, read_os, oddSiteType);
+      os_idx += read_os;
+    }
+
+    CHECK
+    err = MPI_File_set_view(fp, 0, MPI_BYTE, MPI_BYTE, "native", MPI_INFO_NULL);
+  errInfo(err, "MPI_File_set_view1");
+    assert(err == MPI_SUCCESS);
+    freeTypes();
+
+    std::cout << GridLogMessage << "read sum: " << n_os * os_size << " bytes" << std::endl;
+    return true;
+  }
+};
+
+class OpenQcdIOChromaReference : public BinaryIO {
+public:
+  template<class vsimd>
+  static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
+                                       Grid::FieldMetaData&                  header,
+                                       std::string                           file) {
+    typedef Lattice<iDoubleStoredColourMatrix<vsimd>> DoubledGaugeField;
+
+    assert(Ns == 4 and Nd == 4 and Nc == 3);
+
+    auto grid = Umu.Grid();
+
+    typedef ColourMatrixD fobj;
+
+    std::vector<fobj> iodata(
+      Nd * grid->lSites()); // actual size = 2*Nd*lsites but have only lsites/2 sites in file
+
+    {
+      ParRdr rdr(MPI_COMM_WORLD, file, grid);
+      rdr.readGauge(iodata, header);
+    } // equivalent to using binaryio
+
+    std::vector<iDoubleStoredColourMatrix<typename vsimd::scalar_type>> Umu_ds_scalar(grid->lSites());
+
+    copyToLatticeObject(Umu_ds_scalar, iodata, grid); // equivalent to munging
+
+    DoubledGaugeField Umu_ds(grid);
+
+    vectorizeFromLexOrdArray(Umu_ds_scalar, Umu_ds);
+
+    redistribute(Umu, Umu_ds); // equivalent to undoDoublestore
+
+    FieldMetaData clone(header);
+
+    PeriodicGaugeStatistics Stats; Stats(Umu, clone);
+
+    RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
+
+    // clang-format off
+    std::cout << GridLogMessage << "OpenQcd Configuration " << file
+              << " plaquette " << clone.plaquette
+              << " header " << header.plaquette
+              << " difference " << plaq_diff
+              << std::endl;
+    // clang-format on
+
+    RealD precTol = (getPrecision<vsimd>::value == 1) ? 2e-7 : 2e-15;
+    RealD tol     = precTol * std::sqrt(grid->_Nprocessors); // taken from RQCD chroma code
+
+    if(plaq_diff >= tol)
+      std::cout << " Plaquette mismatch (diff = " << plaq_diff << ", tol = " << tol << ")" << std::endl;
+    assert(plaq_diff < tol);
+
+    std::cout << GridLogMessage << "OpenQcd Configuration " << file << " and plaquette agree" << std::endl;
+  }
+
+private:
+  template<class vsimd>
+  static inline void redistribute(Lattice<iLorentzColourMatrix<vsimd>>&            Umu,
+                                  Lattice<iDoubleStoredColourMatrix<vsimd>> const& Umu_ds) {
+    Grid::conformable(Umu.Grid(), Umu_ds.Grid());
+    Lattice<iColourMatrix<vsimd>> U(Umu.Grid());
+
+    U = PeekIndex<LorentzIndex>(Umu_ds, 2) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 3), 0, +1); PokeIndex<LorentzIndex>(Umu, U, 0);
+    U = PeekIndex<LorentzIndex>(Umu_ds, 4) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 5), 1, +1); PokeIndex<LorentzIndex>(Umu, U, 1);
+    U = PeekIndex<LorentzIndex>(Umu_ds, 6) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 7), 2, +1); PokeIndex<LorentzIndex>(Umu, U, 2);
+    U = PeekIndex<LorentzIndex>(Umu_ds, 0) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 1), 3, +1); PokeIndex<LorentzIndex>(Umu, U, 3);
+  }
+
+  static inline void copyToLatticeObject(std::vector<DoubleStoredColourMatrix>& u_fb,
+                                         std::vector<ColourMatrixD> const&      node_buff,
+                                         GridBase*                              grid) {
+    assert(node_buff.size() == Nd * grid->lSites());
+
+    Coordinate const& l = grid->LocalDimensions();
+
+    Coordinate coord(Nd);
+    int&       x = coord[0];
+    int&       y = coord[1];
+    int&       z = coord[2];
+    int&       t = coord[3];
+
+    int buff_idx = 0;
+    for(t = 0; t < l[3]; ++t) // IMPORTANT: openQCD file ordering
+      for(x = 0; x < l[0]; ++x)
+        for(y = 0; y < l[1]; ++y)
+          for(z = 0; z < l[2]; ++z) {
+            if((t + z + y + x) % 2 == 0) continue;
+
+            int local_idx;
+            Lexicographic::IndexFromCoor(coord, local_idx, grid->LocalDimensions());
+            for(int mu = 0; mu < 2 * Nd; ++mu)
+              for(int c1 = 0; c1 < Nc; ++c1) {
+                for(int c2 = 0; c2 < Nc; ++c2) {
+                  u_fb[local_idx](mu)()(c1,c2) = node_buff[mu+buff_idx]()()(c1,c2);
+                }
+              }
+            buff_idx += 2 * Nd;
+          }
+
+    assert(node_buff.size() == buff_idx);
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/perfmon/PerfCount.h

@@ -44,7 +44,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <sys/syscall.h>
 #endif
 #ifdef __x86_64__
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 accelerator_inline uint64_t __rdtsc(void) {  return 0; }
 accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
 #else
@@ -95,7 +95,8 @@ inline uint64_t cyclecount(void){
 }
 #elif defined __x86_64__
 inline uint64_t cyclecount(void){ 
-  return __rdtsc();
+  uint64_t ret = __rdtsc();
+  return (uint64_t)ret;
 }
 #else
 
@@ -111,7 +112,6 @@ class PerformanceCounter {
 private:
 
   typedef struct { 
-  public:
     uint32_t type;
     uint64_t config;
     const char *name;

Grid/perfmon/Timer.h

@@ -110,15 +110,15 @@ public:
 #endif
     accumulator = std::chrono::duration_cast<GridUsecs>(start-start); 
   }
-  GridTime Elapsed(void) {
+  GridTime Elapsed(void) const {
     assert(running == false);
     return std::chrono::duration_cast<GridTime>( accumulator );
   }
-  uint64_t useconds(void){
+  uint64_t useconds(void) const {
     assert(running == false);
     return (uint64_t) accumulator.count();
   }
-  bool isRunning(void){
+  bool isRunning(void) const {
     return running;
   }
 };

Grid/pugixml/pugixml.cc

@@ -12773,7 +12773,7 @@ namespace pugi
 #undef PUGI__THROW_ERROR
 #undef PUGI__CHECK_ERROR
 
-#ifdef GRID_NVCC
+#ifdef GRID_CUDA
 #pragma pop
 #endif
 

Grid/qcd/QCD.h

@@ -47,7 +47,7 @@ static constexpr int Ym = 5;
 static constexpr int Zm = 6;
 static constexpr int Tm = 7;
 
-static constexpr int Nc=3;
+static constexpr int Nc=Config_Nc;
 static constexpr int Ns=4;
 static constexpr int Nd=4;
 static constexpr int Nhs=2; // half spinor
@@ -80,6 +80,13 @@ template<typename T> struct isSpinor {
 template <typename T> using IfSpinor    = Invoke<std::enable_if< isSpinor<T>::value,int> > ;
 template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ;
 
+const int CoarseIndex = 4;
+template<typename T> struct isCoarsened {
+   static constexpr bool value = (CoarseIndex<=T::TensorLevel);
+};
+template <typename T> using IfCoarsened    = Invoke<std::enable_if< isCoarsened<T>::value,int> > ;
+template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ;
+
 // ChrisK very keen to add extra space for Gparity doubling.
 //
 // Also add domain wall index, in a way where Wilson operator 
@@ -133,23 +140,23 @@ typedef iSpinColourMatrix<vComplex >    vSpinColourMatrix;
 typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF;
 typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD;
 
-    // SpinColourSpinColour matrix
-    typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
-    typedef iSpinColourSpinColourMatrix<ComplexF >    SpinColourSpinColourMatrixF;
-    typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
+// SpinColourSpinColour matrix
+typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
+typedef iSpinColourSpinColourMatrix<ComplexF >    SpinColourSpinColourMatrixF;
+typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 
-    typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
-    typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
-    typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
+typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
+typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
 
-    // SpinColourSpinColour matrix
-    typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
-    typedef iSpinColourSpinColourMatrix<ComplexF >    SpinColourSpinColourMatrixF;
-    typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
+// SpinColourSpinColour matrix
+typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
+typedef iSpinColourSpinColourMatrix<ComplexF >    SpinColourSpinColourMatrixF;
+typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 
-    typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
-    typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
-    typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
+typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
+typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
 
 // LorentzColour
 typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix;
@@ -443,16 +450,16 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 //////////////////////////////////////////////
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
-    //template <class Prop, class Ferm>
-    template <class Fimpl>
-      void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
+//template <class Prop, class Ferm>
+template <class Fimpl>
+void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
       auto fj  = peekSpin(f, j);
             
-            for(int i = 0; i < Fimpl::Dimension; ++i)
+      for(int i = 0; i < Fimpl::Dimension; ++i)
 	{
 	  pokeColour(pjs, peekColour(fj, i), i, c);
 	}
@@ -460,16 +467,16 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
     }
 }
     
-    //template <class Prop, class Ferm>
-    template <class Fimpl>
-      void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
+//template <class Prop, class Ferm>
+template <class Fimpl>
+void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
       auto fj  = peekSpin(f, j);
             
-            for(int i = 0; i < Fimpl::Dimension; ++i)
+      for(int i = 0; i < Fimpl::Dimension; ++i)
 	{
 	  pokeColour(fj, peekColour(pjs, i, c), i);
 	}

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -40,8 +40,8 @@ public:
 public:
 
   // override multiply
-  virtual RealD  M    (const FermionField &in, FermionField &out);
-  virtual RealD  Mdag (const FermionField &in, FermionField &out);
+  virtual void   M    (const FermionField &in, FermionField &out);
+  virtual void   Mdag (const FermionField &in, FermionField &out);
 
   // half checkerboard operations
   virtual void   Meooe       (const FermionField &in, FermionField &out);
@@ -141,7 +141,33 @@ public:
   Vector<iSinglet<Simd> >  MatpInvDag;
   Vector<iSinglet<Simd> >  MatmInvDag;
 
+  ///////////////////////////////////////////////////////////////
+  // Conserved current utilities
+  ///////////////////////////////////////////////////////////////
+
+  // Virtual can't template
+  void ContractConservedCurrent(PropagatorField &q_in_1,
+				PropagatorField &q_in_2,
+				PropagatorField &q_out,
+				PropagatorField &phys_src,
+				Current curr_type, 
+				unsigned int mu);
+
+   void SeqConservedCurrent(PropagatorField &q_in,
+			   PropagatorField &q_out,
+			   PropagatorField &phys_src,
+			   Current curr_type,
+			   unsigned int mu,
+			   unsigned int tmin,
+			   unsigned int tmax,
+			   ComplexField &lattice_cmplx);
+
+  void ContractJ5q(PropagatorField &q_in,ComplexField &J5q);
+  void ContractJ5q(FermionField &q_in,ComplexField &J5q);
+
+  ///////////////////////////////////////////////////////////////
   // Constructors
+  ///////////////////////////////////////////////////////////////
   CayleyFermion5D(GaugeField &_Umu,
 		  GridCartesian         &FiveDimGrid,
 		  GridRedBlackCartesian &FiveDimRedBlackGrid,

Grid/qcd/action/fermion/ContinuedFractionFermion5D.h

@@ -41,8 +41,8 @@ public:
 public:
 
   // override multiply
-  virtual RealD  M    (const FermionField &in, FermionField &out);
-  virtual RealD  Mdag (const FermionField &in, FermionField &out);
+  virtual void   M    (const FermionField &in, FermionField &out);
+  virtual void   Mdag (const FermionField &in, FermionField &out);
 
   // half checkerboard operaions
   virtual void   Meooe       (const FermionField &in, FermionField &out);

Grid/qcd/action/fermion/DomainWallEOFAFermion.h

@@ -53,8 +53,8 @@ public:
   virtual void  DtildeInv  (const FermionField& in, FermionField& out);
 
   // override multiply
-  virtual RealD M          (const FermionField& in, FermionField& out);
-  virtual RealD Mdag       (const FermionField& in, FermionField& out);
+  virtual void  M          (const FermionField& in, FermionField& out);
+  virtual void  Mdag       (const FermionField& in, FermionField& out);
 
   // half checkerboard operations
   virtual void  Mooee      (const FermionField& in, FermionField& out);

Grid/qcd/action/fermion/DomainWallVec5dImpl.h

@@ -114,19 +114,22 @@ public:
       U = adj(Cshift(U, mu, -1));
       PokeIndex<LorentzIndex>(Uadj, U, mu);
     }
-    
-    for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
+
+    autoView(Umu_v,Umu,CpuRead);
+    autoView(Uadj_v,Uadj,CpuRead);
+    autoView(Uds_v,Uds,CpuWrite);
+    thread_for( lidx, GaugeGrid->lSites(), {
       Coordinate lcoor;
       GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
       
-      peekLocalSite(ScalarUmu, Umu, lcoor);
+      peekLocalSite(ScalarUmu, Umu_v, lcoor);
       for (int mu = 0; mu < 4; mu++) ScalarUds(mu) = ScalarUmu(mu);
       
-      peekLocalSite(ScalarUmu, Uadj, lcoor);
+      peekLocalSite(ScalarUmu, Uadj_v, lcoor);
       for (int mu = 0; mu < 4; mu++) ScalarUds(mu + 4) = ScalarUmu(mu);
       
-      pokeLocalSite(ScalarUds, Uds, lcoor);
-    }
+      pokeLocalSite(ScalarUds, Uds_v, lcoor);
+    });
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu) 

Grid/qcd/action/fermion/Fermion.h

@@ -57,6 +57,7 @@ NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
 
+#include <Grid/qcd/action/fermion/NaiveStaggeredFermion.h>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion.h>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
 NAMESPACE_CHECK(Staggered);
@@ -282,11 +283,15 @@ typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
 typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
 
+typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR;
+typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
+typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
+
 typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
 typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
 typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
 
-#ifndef GRID_NVCC
+#ifndef GRID_CUDA
 typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR;
 typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF;
 typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD;

Grid/qcd/action/fermion/FermionOperator.h

@@ -58,8 +58,8 @@ public:
   virtual GridBase *GaugeRedBlackGrid(void)   =0;
 
   // override multiply
-  virtual RealD  M    (const FermionField &in, FermionField &out)=0;
-  virtual RealD  Mdag (const FermionField &in, FermionField &out)=0;
+  virtual void  M    (const FermionField &in, FermionField &out)=0;
+  virtual void  Mdag (const FermionField &in, FermionField &out)=0;
 
   // half checkerboard operaions
   virtual void   Meooe       (const FermionField &in, FermionField &out)=0;
@@ -86,15 +86,14 @@ public:
   virtual void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)=0;
   virtual void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)=0;
 
-
   virtual void  Mdiag  (const FermionField &in, FermionField &out) { Mooee(in,out);};   // Same as Mooee applied to both CB's
   virtual void  Mdir   (const FermionField &in, FermionField &out,int dir,int disp)=0;   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
   virtual void  MdirAll(const FermionField &in, std::vector<FermionField> &out)=0;   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
 
 
-      virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
+  virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
 
-      virtual void  FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary,std::vector<double> twist) 
+  virtual void  FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary,std::vector<double> twist) 
       {
 	FFT theFFT((GridCartesian *) in.Grid());
 
@@ -148,15 +147,19 @@ public:
   virtual void ContractConservedCurrent(PropagatorField &q_in_1,
 					PropagatorField &q_in_2,
 					PropagatorField &q_out,
+					PropagatorField &phys_src,
 					Current curr_type,
-					unsigned int mu)=0;
+					unsigned int mu)
+  {assert(0);};
   virtual void SeqConservedCurrent(PropagatorField &q_in, 
 				   PropagatorField &q_out,
+				   PropagatorField &phys_src,
 				   Current curr_type,
 				   unsigned int mu,
 				   unsigned int tmin, 
 				   unsigned int tmax,
-				   ComplexField &lattice_cmplx)=0;
+				   ComplexField &lattice_cmplx)
+  {assert(0);};
 
       // Only reimplemented in Wilson5D 
       // Default to just a zero correlation function

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -38,6 +38,7 @@ public:
  static const bool isFundamental = Representation::isFundamental;
  static const int Nhcs = Options::Nhcs;
  static const bool LsVectorised=false;
+ static const bool isGparity=true;
 
  typedef ConjugateGaugeImpl< GaugeImplTypes<S,Dimension> > Gimpl;
  INHERIT_GIMPL_TYPES(Gimpl);
@@ -46,7 +47,7 @@ public:
  typedef typename Options::template PrecisionMapper<Simd>::LowerPrecVector SimdL;
       
  template <typename vtype> using iImplSpinor            = iVector<iVector<iVector<vtype, Dimension>, Ns>,   Ngp>;
- template <typename vtype> using iImplPropagator        = iVector<iMatrix<iMatrix<vtype, Dimension>, Ns>,   Ngp>;
+ template <typename vtype> using iImplPropagator        = iMatrix<iMatrix<iMatrix<vtype, Dimension>, Ns>,   Ngp>;
  template <typename vtype> using iImplHalfSpinor        = iVector<iVector<iVector<vtype, Dimension>, Nhs>,  Ngp>;
  template <typename vtype> using iImplHalfCommSpinor    = iVector<iVector<iVector<vtype, Dimension>, Nhcs>, Ngp>;
  template <typename vtype> using iImplDoubledGaugeField = iVector<iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>, Ngp>;
@@ -80,6 +81,7 @@ public:
   {
     assert(0);
   } 
+
   template<class _Spinor>
   static accelerator_inline void multLink(_Spinor &phi, 
 					  const SiteDoubledGaugeField &U,
@@ -94,43 +96,31 @@ public:
     int sl        = St._simd_layout[direction];
     Coordinate icoor;
 
-#ifdef __CUDA_ARCH__
-    _Spinor tmp;
-
+#ifdef GRID_SIMT
     const int Nsimd =SiteDoubledGaugeField::Nsimd();
-    int s = SIMTlane(Nsimd);
+    int s = acceleratorSIMTlane(Nsimd);
     St.iCoorFromIindex(icoor,s);
 
     int mmu = mu % Nd;
-    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
-      
-      int permute_lane = (sl==1) 
-    	|| ((distance== 1)&&(icoor[direction]==1))
-	|| ((distance==-1)&&(icoor[direction]==0));
 
-      if ( permute_lane ) { 
-	tmp(0) = chi(1);
-	tmp(1) = chi(0);
-      } else {
-	tmp(0) = chi(0);
-	tmp(1) = chi(1);
-      }
+    auto UU0=coalescedRead(U(0)(mu));
+    auto UU1=coalescedRead(U(1)(mu));
+    
+    //Decide whether we do a G-parity flavor twist
+    //Note: this assumes (but does not check) that sl==1 || sl==2 i.e. max 2 SIMD lanes in G-parity dir
+    //It also assumes (but does not check) that abs(distance) == 1
+    int permute_lane = (sl==1) 
+    || ((distance== 1)&&(icoor[direction]==1))
+    || ((distance==-1)&&(icoor[direction]==0));
 
-      auto UU0=coalescedRead(U(0)(mu));
-      auto UU1=coalescedRead(U(1)(mu));
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
 
-      mult(&phi(0),&UU0,&tmp(0));
-      mult(&phi(1),&UU1,&tmp(1));
+    //Apply the links
+    int f_upper = permute_lane ? 1 : 0;
+    int f_lower = !f_upper;
 
-    } else {
-
-      auto UU0=coalescedRead(U(0)(mu));
-      auto UU1=coalescedRead(U(1)(mu));
-
-      mult(&phi(0),&UU0,&chi(0));
-      mult(&phi(1),&UU1,&chi(1));
-
-    }
+    mult(&phi(0),&UU0,&chi(f_upper));
+    mult(&phi(1),&UU1,&chi(f_lower));
 
 #else
     typedef _Spinor vobj;
@@ -191,6 +181,16 @@ public:
 #endif   
   }
 
+
+  template<class _SpinorField>
+  inline void multLinkField(_SpinorField & out,
+			    const DoubledGaugeField &Umu,
+			    const _SpinorField & phi,
+			    int mu)
+  {
+    assert(0);
+  }
+
   template <class ref>
   static accelerator_inline void loadLinkElement(Simd &reg, ref &memory) 
   {
@@ -220,15 +220,17 @@ public:
       if ( Params.twists[mu] ) { 
 	Uconj = where(coor==neglink,-Uconj,Uconj);
       }
-	  
-      auto U_v = U.View();
-      auto Uds_v = Uds.View();
-      auto Uconj_v = Uconj.View();
-      auto Utmp_v= Utmp.View();
-      thread_foreach(ss,U_v,{
-	Uds_v[ss](0)(mu) = U_v[ss]();
-	Uds_v[ss](1)(mu) = Uconj_v[ss]();
-      });
+
+      {
+	autoView( U_v , U, CpuRead);
+	autoView( Uconj_v , Uconj, CpuRead);
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,U_v,{
+	    Uds_v[ss](0)(mu) = U_v[ss]();
+	    Uds_v[ss](1)(mu) = Uconj_v[ss]();
+	  });
+      }
           
       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary
       Uconj = adj(Cshift(Uconj,mu,-1));
@@ -238,19 +240,25 @@ public:
 	Utmp = where(coor==0,Uconj,Utmp);
       }
 
-      thread_foreach(ss,Utmp_v,{
-	Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
-      });
-          
+      {
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,Utmp_v,{
+	    Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
+	  });
+      }
       Utmp = Uconj;
       if ( Params.twists[mu] ) { 
 	Utmp = where(coor==0,U,Utmp);
       }
-	  
-      thread_foreach(ss,Utmp_v,{
-        Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
-      });
-          
+
+      {	  
+	autoView( Uds_v , Uds, CpuWrite);
+	autoView( Utmp_v, Utmp, CpuWrite);
+	thread_foreach(ss,Utmp_v,{
+	    Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
+        });
+      }
     }
   }
       
@@ -260,11 +268,14 @@ public:
     GaugeLinkField link(mat.Grid());
     // use lorentz for flavour as hack.
     auto tmp = TraceIndex<SpinIndex>(outerProduct(Btilde, A));
-    auto link_v = link.View();
-    auto tmp_v = tmp.View();
-    thread_foreach(ss,tmp_v,{
-      link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
-    });
+
+    {
+      autoView( link_v , link, CpuWrite);
+      autoView( tmp_v , tmp, CpuRead);
+      thread_foreach(ss,tmp_v,{
+	  link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
+	});
+    }
     PokeIndex<LorentzIndex>(mat, link, mu);
     return;
   }
@@ -294,16 +305,18 @@ public:
         
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    auto tmp_v = tmp.View();
-    auto Atilde_v = Atilde.View();
-    auto Btilde_v = Btilde.View();
-    thread_for(ss,tmp.Grid()->oSites(),{
-      for (int s = 0; s < Ls; s++) {
-	int sF = s + Ls * ss;
-	auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
-	tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
-      }
-    });
+    {
+      autoView( tmp_v , tmp, CpuWrite);
+      autoView( Atilde_v , Atilde, CpuRead);
+      autoView( Btilde_v , Btilde, CpuRead);
+      thread_for(ss,tmp.Grid()->oSites(),{
+	  for (int s = 0; s < Ls; s++) {
+	    int sF = s + Ls * ss;
+	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
+	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
+	  }
+	});
+    }
     PokeIndex<LorentzIndex>(mat, tmp, mu);
     return;
   }

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -71,8 +71,8 @@ public:
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent
   //////////////////////////////////////////////////////////////////
-  RealD M(const FermionField &in, FermionField &out);
-  RealD Mdag(const FermionField &in, FermionField &out);
+  void M(const FermionField &in, FermionField &out);
+  void Mdag(const FermionField &in, FermionField &out);
 
   /////////////////////////////////////////////////////////
   // half checkerboard operations
@@ -185,10 +185,12 @@ public:
   void ContractConservedCurrent(PropagatorField &q_in_1,
                                 PropagatorField &q_in_2,
                                 PropagatorField &q_out,
+                                PropagatorField &src,
                                 Current curr_type,
                                 unsigned int mu);
   void SeqConservedCurrent(PropagatorField &q_in,
                            PropagatorField &q_out,
+                           PropagatorField &srct,
                            Current curr_type,
                            unsigned int mu, 
                            unsigned int tmin,

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -62,8 +61,8 @@ public:
   double DhopCalls;
   double DhopCommTime;
   double DhopComputeTime;
-      double DhopComputeTime2;
-      double DhopFaceTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
@@ -74,8 +73,8 @@ public:
   GridBase *FermionRedBlackGrid(void)    { return _FiveDimRedBlackGrid;}
 
   // full checkerboard operations; leave unimplemented as abstract for now
-  RealD  M    (const FermionField &in, FermionField &out);
-  RealD  Mdag (const FermionField &in, FermionField &out);
+  void  M    (const FermionField &in, FermionField &out);
+  void  Mdag (const FermionField &in, FermionField &out);
 
   // half checkerboard operations
   void   Meooe       (const FermionField &in, FermionField &out);
@@ -209,7 +208,7 @@ public:
   LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
-  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
+  //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
     
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
@@ -217,15 +216,17 @@ public:
   void ContractConservedCurrent(PropagatorField &q_in_1,
 				PropagatorField &q_in_2,
 				PropagatorField &q_out,
+				PropagatorField &src,
 				Current curr_type,
 				unsigned int mu);
   void SeqConservedCurrent(PropagatorField &q_in,
 			   PropagatorField &q_out,
+			   PropagatorField &src,
 			   Current curr_type,
 			   unsigned int mu, 
 			   unsigned int tmin,
-                             unsigned int tmax,
-                 	     ComplexField &lattice_cmplx);
+			   unsigned int tmax,
+			   ComplexField &lattice_cmplx);
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/MADWF.h

@@ -40,6 +40,11 @@ inline void convert(const Fieldi &from,Fieldo &to)
   to=from;
 }
 
+struct MADWFinnerIterCallbackBase{
+  virtual void operator()(const RealD current_resid){}
+  virtual ~MADWFinnerIterCallbackBase(){}
+};
+
 template<class Matrixo,class Matrixi,class PVinverter,class SchurSolver, class Guesser> 
 class MADWF 
 {
@@ -56,24 +61,30 @@ class MADWF
 
   RealD target_resid;
   int   maxiter;
- public:
 
+  //operator() is called on "callback" at the end of every inner iteration. This allows for example the adjustment of the inner
+  //tolerance to speed up subsequent iteration
+  MADWFinnerIterCallbackBase* callback;
+  
+ public:
   MADWF(Matrixo &_Mato,
-	Matrixi &_Mati, 
-	PVinverter &_PauliVillarsSolvero, 
+	Matrixi &_Mati,
+	PVinverter &_PauliVillarsSolvero,
 	SchurSolver &_SchurSolveri,
 	Guesser & _Guesseri,
 	RealD resid,
-	int _maxiter) :
+	int _maxiter,
+	MADWFinnerIterCallbackBase* _callback = NULL) :
 
   Mato(_Mato),Mati(_Mati),
     SchurSolveri(_SchurSolveri),
-    PauliVillarsSolvero(_PauliVillarsSolvero),Guesseri(_Guesseri)
-  {   
-    target_resid=resid;
-    maxiter     =_maxiter; 
-  };
-
+    PauliVillarsSolvero(_PauliVillarsSolvero),Guesseri(_Guesseri),
+    callback(_callback)
+    {
+      target_resid=resid;
+      maxiter     =_maxiter;
+    };
+   
   void operator() (const FermionFieldo &src4,FermionFieldo &sol5)
   {
     std::cout << GridLogMessage<< " ************************************************" << std::endl;
@@ -177,6 +188,8 @@ class MADWF
        std::cout << GridLogMessage << "Residual " << i << ": " << resid  << std::endl;
        std::cout << GridLogMessage << "***************************************" <<std::endl;
 
+       if(callback != NULL) (*callback)(resid);       
+       
        if (resid < target_resid) {
 	 return;
        }

Grid/qcd/action/fermion/MobiusEOFAFermion.h

@@ -56,8 +56,8 @@ public:
   virtual void  DtildeInv        (const FermionField& in, FermionField& out);
 
   // override multiply
-  virtual RealD M                (const FermionField& in, FermionField& out);
-  virtual RealD Mdag             (const FermionField& in, FermionField& out);
+  virtual void  M                (const FermionField& in, FermionField& out);
+  virtual void  Mdag             (const FermionField& in, FermionField& out);
 
   // half checkerboard operations
   virtual void  Mooee            (const FermionField& in, FermionField& out);

Grid/qcd/action/fermion/MobiusFermion.h

@@ -59,7 +59,7 @@ public:
   {
     RealD eps = 1.0;
 
-    std::cout<<GridLogMessage << "MobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" Tanh approx"<<std::endl;
+    //    std::cout<<GridLogMessage << "MobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" Tanh approx"<<std::endl;
     Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);// eps is ignored for higham
     assert(zdata->n==this->Ls);
 	

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -0,0 +1,194 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/ImprovedStaggered.h
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi, Peter Boyle
+
+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_QCD_NAIVE_STAG_FERMION_H
+#define GRID_QCD_NAIVE_STAG_FERMION_H
+
+NAMESPACE_BEGIN(Grid);
+
+class NaiveStaggeredFermionStatic {
+public:
+  static const std::vector<int> directions;
+  static const std::vector<int> displacements;
+  static const int npoint = 8;
+};
+
+template <class Impl>
+class NaiveStaggeredFermion : public StaggeredKernels<Impl>, public NaiveStaggeredFermionStatic {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+  typedef StaggeredKernels<Impl> Kernels;
+
+  FermionField _tmp;
+  FermionField &tmp(void) { return _tmp; }
+
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
+  ///////////////////////////////////////////////////////////////
+  // Implement the abstract base
+  ///////////////////////////////////////////////////////////////
+  GridBase *GaugeGrid(void) { return _grid; }
+  GridBase *GaugeRedBlackGrid(void) { return _cbgrid; }
+  GridBase *FermionGrid(void) { return _grid; }
+  GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
+
+  //////////////////////////////////////////////////////////////////
+  // override multiply; cut number routines if pass dagger argument
+  // and also make interface more uniformly consistent
+  //////////////////////////////////////////////////////////////////
+  void M(const FermionField &in, FermionField &out);
+  void Mdag(const FermionField &in, FermionField &out);
+
+  /////////////////////////////////////////////////////////
+  // half checkerboard operations
+  /////////////////////////////////////////////////////////
+  void Meooe(const FermionField &in, FermionField &out);
+  void MeooeDag(const FermionField &in, FermionField &out);
+  void Mooee(const FermionField &in, FermionField &out);
+  void MooeeDag(const FermionField &in, FermionField &out);
+  void MooeeInv(const FermionField &in, FermionField &out);
+  void MooeeInvDag(const FermionField &in, FermionField &out);
+
+  ////////////////////////
+  // Derivative interface
+  ////////////////////////
+  // Interface calls an internal routine
+  void DhopDeriv  (GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+  void DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+  void DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
+
+  ///////////////////////////////////////////////////////////////
+  // non-hermitian hopping term; half cb or both
+  ///////////////////////////////////////////////////////////////
+  void Dhop  (const FermionField &in, FermionField &out, int dag);
+  void DhopOE(const FermionField &in, FermionField &out, int dag);
+  void DhopEO(const FermionField &in, FermionField &out, int dag);
+
+  ///////////////////////////////////////////////////////////////
+  // Multigrid assistance; force term uses too
+  ///////////////////////////////////////////////////////////////
+  void Mdir(const FermionField &in, FermionField &out, int dir, int disp);
+  void MdirAll(const FermionField &in, std::vector<FermionField> &out);
+  void DhopDir(const FermionField &in, FermionField &out, int dir, int disp);
+
+  ///////////////////////////////////////////////////////////////
+  // Extra methods added by derived
+  ///////////////////////////////////////////////////////////////
+  void DerivInternal(StencilImpl &st, 
+		     DoubledGaugeField &U,
+		     GaugeField &mat, 
+		     const FermionField &A, const FermionField &B, int dag);
+
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
+  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+			       const FermionField &in, FermionField &out, int dag);
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+				   const FermionField &in, FermionField &out, int dag);
+
+  //////////////////////////////////////////////////////////////////////////
+  // Grid own interface Constructor
+  //////////////////////////////////////////////////////////////////////////
+  NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
+			GridRedBlackCartesian &Hgrid, RealD _mass,
+			RealD _c1, RealD _u0,
+			const ImplParams &p = ImplParams());
+  NaiveStaggeredFermion(GridCartesian &Fgrid,
+			GridRedBlackCartesian &Hgrid, RealD _mass,
+			RealD _c1, RealD _u0,
+			const ImplParams &p = ImplParams());
+
+  // DoubleStore impl dependent
+  void ImportGauge      (const GaugeField &_U );
+  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  void CopyGaugeCheckerboards(void);
+
+  ///////////////////////////////////////////////////////////////
+  // Data members require to support the functionality
+  ///////////////////////////////////////////////////////////////
+
+  //    protected:
+public:
+  // any other parameters of action ???
+  virtual int   isTrivialEE(void) { return 1; };
+  virtual RealD Mass(void) { return mass; }
+  RealD mass;
+  RealD u0;
+  RealD c1;
+
+  GridBase *_grid;
+  GridBase *_cbgrid;
+
+  // Defines the stencils for even and odd
+  StencilImpl Stencil;
+  StencilImpl StencilEven;
+  StencilImpl StencilOdd;
+
+  // Copy of the gauge field , with even and odd subsets
+  DoubledGaugeField Umu;
+  DoubledGaugeField UmuEven;
+  DoubledGaugeField UmuOdd;
+
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+  
+  ///////////////////////////////////////////////////////////////
+  // Conserved current utilities
+  ///////////////////////////////////////////////////////////////
+  void ContractConservedCurrent(PropagatorField &q_in_1,
+                                PropagatorField &q_in_2,
+                                PropagatorField &q_out,
+                                PropagatorField &src,
+                                Current curr_type,
+                                unsigned int mu);
+  void SeqConservedCurrent(PropagatorField &q_in,
+                           PropagatorField &q_out,
+                           PropagatorField &srct,
+                           Current curr_type,
+                           unsigned int mu, 
+                           unsigned int tmin,
+                           unsigned int tmax,
+			   ComplexField &lattice_cmplx);
+};
+
+typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
+typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/qcd/action/fermion/PartialFractionFermion5D.h

@@ -47,8 +47,8 @@ public:
   void   M_internal(const FermionField &in, FermionField &out,int dag);
 
   // override multiply
-  virtual RealD  M    (const FermionField &in, FermionField &out);
-  virtual RealD  Mdag (const FermionField &in, FermionField &out);
+  virtual void   M    (const FermionField &in, FermionField &out);
+  virtual void   Mdag (const FermionField &in, FermionField &out);
 
   // half checkerboard operaions
   virtual void   Meooe       (const FermionField &in, FermionField &out);

Grid/qcd/action/fermion/StaggeredKernels.h

@@ -47,23 +47,37 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
   INHERIT_IMPL_TYPES(Impl);
   typedef FermionOperator<Impl> Base;
    
-public:
-    
-   void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
-		      int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+ public:
+
+  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
+		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
+		    const FermionField &in, FermionField &out, int dag, int interior,int exterior);
+  void DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
+		 DoubledGaugeField &U,
+		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
+  
+  void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
+		     int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
+ protected:    
 
    ///////////////////////////////////////////////////////////////////////////////////////
    // Generic Nc kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+   template<int Naik> 
+   static accelerator_inline
+   void DhopSiteGeneric(StencilView &st, 
 			DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteGenericInt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteGenericExt(StencilView &st, 
 			   DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
 			   SiteSpinor * buf, int LLs, int sU, 
 			   const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -71,15 +85,21 @@ public:
    ///////////////////////////////////////////////////////////////////////////////////////
    // Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteHand(StencilView &st, 
 		     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		     SiteSpinor * buf, int LLs, int sU, 
 		     const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteHandInt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
-   void DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+   
+   template<int Naik> static accelerator_inline
+   void DhopSiteHandExt(StencilView &st, 
 			DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 			SiteSpinor * buf, int LLs, int sU, 
 			const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -87,27 +107,11 @@ public:
    ///////////////////////////////////////////////////////////////////////////////////////
    // Asm Nc=3 specific kernels
    ///////////////////////////////////////////////////////////////////////////////////////
-   void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+   
+   void DhopSiteAsm(StencilView &st, 
 		    DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, 
 		    SiteSpinor * buf, int LLs, int sU, 
 		    const FermionFieldView &in, FermionFieldView &out,int dag);
-   ///////////////////////////////////////////////////////////////////////////////////////////////////
-   // Generic interface; fan out to right routine
-   ///////////////////////////////////////////////////////////////////////////////////////////////////
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		 SiteSpinor * buf, int LLs, int sU,
-		 const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
-
-   void DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, 
-		    DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		    SiteSpinor * buf, int LLs, int sU,
-		    const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
-
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
-		 DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, 
-		 SiteSpinor * buf, int LLs, int sU,
-		 const FermionFieldView &in, FermionFieldView &out, int dag, int interior,int exterior);
   
 public:
 

Grid/qcd/action/fermion/StaggeredVec5dImpl.h

@@ -113,20 +113,7 @@ public:
       
   inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
   {
-    GridBase *GaugeGrid = U_ds.Grid();
-    thread_for(lidx, GaugeGrid->lSites(),{
-
-	SiteScalarGaugeLink   ScalarU;
-	SiteDoubledGaugeField ScalarUds;
-	
-	Coordinate lcoor;
-	GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
-	peekLocalSite(ScalarUds, U_ds, lcoor);
-	
-	peekLocalSite(ScalarU, U, lcoor);
-	ScalarUds(mu) = ScalarU();
-	
-    });
+    assert(0);
   }
   inline void DoubleStore(GridBase *GaugeGrid,
 			  DoubledGaugeField &UUUds, // for Naik term

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -109,9 +109,8 @@ public:
     ImportGauge(_Umu);
   }
 
-  virtual RealD M(const FermionField &in, FermionField &out);
-  virtual RealD Mdag(const FermionField &in, FermionField &out);
-
+  virtual void M(const FermionField &in, FermionField &out);
+  virtual void Mdag(const FermionField &in, FermionField &out);
   virtual void Mooee(const FermionField &in, FermionField &out);
   virtual void MooeeDag(const FermionField &in, FermionField &out);
   virtual void MooeeInv(const FermionField &in, FermionField &out);
@@ -258,15 +257,16 @@ private:
   CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term Dag EO
   CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
 
+ public:
   // eventually these can be compressed into 6x6 blocks instead of the 12x12
   // using the DeGrand-Rossi basis for the gamma matrices
   CloverFieldType fillCloverYZ(const GaugeLinkField &F)
   {
     CloverFieldType T(F.Grid());
     T = Zero();
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
       T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
@@ -282,9 +282,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
     
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v, T,AcceleratorWrite);
+    autoView(F_v, F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = -F_v[i]()();
       T_v[i]()(1, 0) = F_v[i]()();
@@ -300,9 +300,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
       T_v[i]()(1, 1) = timesI(F_v[i]()());
@@ -318,9 +318,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v , T, AcceleratorWrite);
+    autoView( F_v , F, AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = timesI(F_v[i]()());
       T_v[i]()(1, 0) = timesI(F_v[i]()());
@@ -336,9 +336,9 @@ private:
     CloverFieldType T(F.Grid());
     T = Zero();
     
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v ,T,AcceleratorWrite);
+    autoView( F_v ,F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 1) = -(F_v[i]()());
       T_v[i]()(1, 0) = (F_v[i]()());
@@ -355,9 +355,9 @@ private:
 
     T = Zero();
 
-    auto T_v = T.View();
-    auto F_v = F.View();
-    thread_for(i, CloverTerm.Grid()->oSites(),
+    autoView( T_v , T,AcceleratorWrite);
+    autoView( F_v , F,AcceleratorRead);
+    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
     {
       T_v[i]()(0, 0) = timesI(F_v[i]()());
       T_v[i]()(1, 1) = timesMinusI(F_v[i]()());

Grid/qcd/action/fermion/WilsonFermion.h

@@ -50,14 +50,14 @@ public:
   double, nu);
 
   WilsonAnisotropyCoefficients():
-    isAnisotropic(false), 
-    t_direction(Nd-1), 
-    xi_0(1.0), 
+    isAnisotropic(false),
+    t_direction(Nd-1),
+    xi_0(1.0),
     nu(1.0){}
 };
 
 template <class Impl>
-class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic 
+class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic
 {
 public:
   INHERIT_IMPL_TYPES(Impl);
@@ -74,12 +74,26 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
+
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent
   //////////////////////////////////////////////////////////////////
-  virtual RealD M(const FermionField &in, FermionField &out);
-  virtual RealD Mdag(const FermionField &in, FermionField &out);
+  virtual void  M(const FermionField &in, FermionField &out);
+  virtual void  Mdag(const FermionField &in, FermionField &out);
 
   /////////////////////////////////////////////////////////
   // half checkerboard operations
@@ -138,7 +152,7 @@ public:
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
                 GridRedBlackCartesian &Hgrid, RealD _mass,
-                const ImplParams &p = ImplParams(), 
+                const ImplParams &p = ImplParams(),
                 const WilsonAnisotropyCoefficients &anis = WilsonAnisotropyCoefficients() );
 
   // DoubleStore impl dependent
@@ -170,29 +184,29 @@ public:
 
   LebesgueOrder Lebesgue;
   LebesgueOrder LebesgueEvenOdd;
-  
+
   WilsonAnisotropyCoefficients anisotropyCoeff;
-  
+
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
   ///////////////////////////////////////////////////////////////
   void ContractConservedCurrent(PropagatorField &q_in_1,
                                 PropagatorField &q_in_2,
                                 PropagatorField &q_out,
+                                PropagatorField &phys_src,
                                 Current curr_type,
                                 unsigned int mu);
   void SeqConservedCurrent(PropagatorField &q_in,
                            PropagatorField &q_out,
+                           PropagatorField &phys_src,
                            Current curr_type,
-                           unsigned int mu, 
+                           unsigned int mu,
                            unsigned int tmin,
-                             unsigned int tmax,
-			     ComplexField &lattice_cmplx);
+			   unsigned int tmax,
+			   ComplexField &lattice_cmplx);
 };
 
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
 NAMESPACE_END(Grid);
-
-

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -99,8 +98,8 @@ public:
   GridBase *FermionRedBlackGrid(void)    { return _FiveDimRedBlackGrid;}
 
   // full checkerboard operations; leave unimplemented as abstract for now
-  virtual RealD  M    (const FermionField &in, FermionField &out){assert(0); return 0.0;};
-  virtual RealD  Mdag (const FermionField &in, FermionField &out){assert(0); return 0.0;};
+  virtual void   M    (const FermionField &in, FermionField &out){assert(0);};
+  virtual void   Mdag (const FermionField &in, FermionField &out){assert(0);};
 
   // half checkerboard operations; leave unimplemented as abstract for now
   virtual void   Meooe       (const FermionField &in, FermionField &out){assert(0);};
@@ -216,26 +215,8 @@ public:
   LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
-  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
-    
-  ///////////////////////////////////////////////////////////////
-  // Conserved current utilities
-  ///////////////////////////////////////////////////////////////
-  void ContractConservedCurrent(PropagatorField &q_in_1,
-				PropagatorField &q_in_2,
-				PropagatorField &q_out,
-				Current curr_type, 
-				unsigned int mu);
-  void SeqConservedCurrent(PropagatorField &q_in,
-			   PropagatorField &q_out,
-			   Current curr_type,
-			   unsigned int mu,
-			   unsigned int tmin,
-			   unsigned int tmax,
-			   ComplexField &lattice_cmplx);
+  //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
 
-  void ContractJ5q(PropagatorField &q_in,ComplexField &J5q);
-  void ContractJ5q(FermionField &q_in,ComplexField &J5q);
 
 };
 

Grid/qcd/action/fermion/WilsonImpl.h

@@ -41,6 +41,7 @@ public:
   static const int Dimension = Representation::Dimension;
   static const bool isFundamental = Representation::isFundamental;
   static const bool LsVectorised=false;
+  static const bool isGparity=false;
   static const int Nhcs = Options::Nhcs;
 
   typedef PeriodicGaugeImpl<GaugeImplTypes<S, Dimension > > Gimpl;
@@ -98,8 +99,21 @@ public:
   {
     multLink(phi,U,chi,mu);
   }
-    
-      
+
+  template<class _SpinorField> 
+  inline void multLinkField(_SpinorField & out,
+			    const DoubledGaugeField &Umu,
+			    const _SpinorField & phi,
+			    int mu)
+  {
+    autoView( out_v, out, AcceleratorWrite);
+    autoView( phi_v, phi, AcceleratorRead);
+    autoView( Umu_v, Umu, AcceleratorRead);
+    accelerator_for(sss,out.Grid()->oSites(),1,{
+	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
+    });
+  }
+					   
   template <class ref>
   static accelerator_inline void loadLinkElement(Simd &reg, ref &memory) 
   {
@@ -177,18 +191,19 @@ public:
     int Ls=Btilde.Grid()->_fdimensions[0];
     GaugeLinkField tmp(mat.Grid());
     tmp = Zero();
-    auto tmp_v = tmp.View();
-    auto Btilde_v = Btilde.View();
-    auto Atilde_v = Atilde.View();
-    thread_for(sss,tmp.Grid()->oSites(),{
-      int sU=sss;
-      for(int s=0;s<Ls;s++){
-	int sF = s+Ls*sU;
-	tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
-      }
-    });
+    {
+      autoView( tmp_v , tmp, AcceleratorWrite);
+      autoView( Btilde_v , Btilde, AcceleratorRead);
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,tmp.Grid()->oSites(),1,{
+	  int sU=sss;
+	  for(int s=0;s<Ls;s++){
+	    int sF = s+Ls*sU;
+	    tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
+	  }
+	});
+    }
     PokeIndex<LorentzIndex>(mat,tmp,mu);
-      
   }
 };
 

Grid/qcd/action/fermion/WilsonKernels.h

@@ -66,41 +66,6 @@ public:
   static void DhopDirKernel(StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor * buf,
 			    int Ls, int Nsite, const FermionField &in, FermionField &out, int dirdisp, int gamma);
 
-  //////////////////////////////////////////////////////////////////////////////
-  // Utilities for inserting Wilson conserved current.
-  //////////////////////////////////////////////////////////////////////////////
-  static void ContractConservedCurrentSiteFwd(const SitePropagator &q_in_1,
-                                       const SitePropagator &q_in_2,
-                                       SitePropagator &q_out,
-                                       DoubledGaugeFieldView &U,
-                                       unsigned int sU,
-                                       unsigned int mu,
-                                       bool switch_sign = false);
-
-  static void ContractConservedCurrentSiteBwd(const SitePropagator &q_in_1,
-                                       const SitePropagator &q_in_2,
-                                       SitePropagator &q_out,
-                                       DoubledGaugeFieldView &U,
-                                       unsigned int sU,
-                                       unsigned int mu,
-                                       bool switch_sign = false);
-
-  static void SeqConservedCurrentSiteFwd(const SitePropagator &q_in, 
-                                  SitePropagator &q_out,
-                                  DoubledGaugeFieldView &U,
-                                  unsigned int sU,
-                                  unsigned int mu,
-                                  vPredicate t_mask,
-                                  bool switch_sign = false);
-
-  static void SeqConservedCurrentSiteBwd(const SitePropagator &q_in,
-                                  SitePropagator &q_out,
-                                  DoubledGaugeFieldView &U,
-                                  unsigned int sU,
-                                  unsigned int mu,
-                                  vPredicate t_mask,
-                                  bool switch_sign = false);
-
 private:
 
   static accelerator_inline void DhopDirK(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor * buf,