Merge 461cd045c6 into da59379612

--disable-accelerator-aware-mpi

.github/ISSUE_TEMPLATE/bug-report.yml

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

.gitignore

@@ -1,3 +1,7 @@
+# Doxygen stuff
+html/*
+latex/*
+
 # Compiled Object files #
 #########################
 *.slo

Grid/DisableWarnings.h

@@ -44,14 +44,22 @@ directory
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
+
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+ //disables nvcc specific warning in json.hpp
+#pragma nv_diag_suppress unsigned_compare_with_zero
+#pragma nv_diag_suppress cast_to_qualified_type
+ //disables nvcc specific warning in many files
+#pragma nv_diag_suppress esa_on_defaulted_function_ignored
+#pragma nv_diag_suppress extra_semicolon
+#else
+ //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
-
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon
-
-//Eigen only
+#endif
 #endif
 
 // Disable vectorisation in Eigen on the Power8/9 and PowerPC

Grid/GridCore.h

@@ -44,9 +44,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridStd.h>
 #include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
-#include <Grid/perfmon/PerfCount.h>
+//#include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
+#include <Grid/perfmon/Tracing.h>
 #include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
 #include <Grid/threads/ThreadReduction.h>

Grid/GridQCDcore.h

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

Grid/GridStd.h

@@ -16,6 +16,7 @@
 #include <functional>
 #include <stdio.h>
 #include <stdlib.h>
+#include <strings.h>
 #include <stdio.h>
 #include <signal.h>
 #include <ctime>

Grid/Grid_Eigen_Dense.h

@@ -14,7 +14,11 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+#pragma nv_diag_suppress code_is_unreachable
+#else
 #pragma diag_suppress code_is_unreachable
+#endif
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")
@@ -30,7 +34,7 @@
 #pragma push_macro("__SYCL_DEVICE_ONLY__")
 #undef __SYCL_DEVICE_ONLY__
 #define EIGEN_DONT_VECTORIZE
-//#undef EIGEN_USE_SYCL
+#undef EIGEN_USE_SYCL
 #define __SYCL__REDEFINE__
 #endif
 

Grid/Makefile.am

@@ -66,6 +66,10 @@ if BUILD_FERMION_REPS
   extra_sources+=$(ADJ_FERMION_FILES)
   extra_sources+=$(TWOIND_FERMION_FILES)
 endif
+if BUILD_SP
+    extra_sources+=$(SP_FERMION_FILES)
+    extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
 
 lib_LIBRARIES = libGrid.a
 

Grid/algorithms/Algorithms.h

@@ -54,6 +54,8 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h>
 #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>

Grid/algorithms/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
         });
@@ -324,9 +324,9 @@ public:
   GridBase*        _cbgrid;
   int hermitian;
 
-  CartesianStencil<siteVector,siteVector,int> Stencil; 
-  CartesianStencil<siteVector,siteVector,int> StencilEven;
-  CartesianStencil<siteVector,siteVector,int> StencilOdd;
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> Stencil; 
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilEven;
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilOdd;
 
   std::vector<CoarseMatrix> A;
   std::vector<CoarseMatrix> Aeven;
@@ -358,7 +358,7 @@ public:
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
     autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
     Vector<Aview> AcceleratorViewContainer;
@@ -380,7 +380,7 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom_v.npoint;point++){
+      for(int point=0;point<npoint;point++){
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
@@ -424,7 +424,7 @@ public:
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
     autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
     Vector<Aview> AcceleratorViewContainer;
@@ -454,7 +454,7 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int p=0;p<geom_v.npoint;p++){
+      for(int p=0;p<npoint;p++){
         int point = points_p[p];
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);
@@ -631,7 +631,7 @@ public:
     assert(Aself != nullptr);
   }
 
-  void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
+  void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
                        const CoarseVector &in, CoarseVector &out, int dag) {
     int point = geom.npoint-1;
     autoView( out_v, out, AcceleratorWrite);
@@ -694,7 +694,7 @@ public:
     }
   }
 
-  void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
+  void DhopInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, std::vector<CoarseMatrix> &a,
                     const CoarseVector &in, CoarseVector &out, int dag) {
     SimpleCompressor<siteVector> compressor;
 
@@ -784,9 +784,9 @@ public:
     _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements),
     A(geom.npoint,&CoarseGrid),
     Aeven(geom.npoint,_cbgrid),
     Aodd(geom.npoint,_cbgrid),
@@ -804,9 +804,9 @@ public:
     _cbgrid(&CoarseRBGrid),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements),
     A(geom.npoint,&CoarseGrid),
     Aeven(geom.npoint,&CoarseRBGrid),
     Aodd(geom.npoint,&CoarseRBGrid),

Grid/algorithms/FFT.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #define _GRID_FFT_H_
 
 #ifdef HAVE_FFTW
-#ifdef USE_MKL
+#if defined(USE_MKL) || defined(GRID_SYCL)
 #include <fftw/fftw3.h>
 #else
 #include <fftw3.h>

Grid/algorithms/LinearOperator.h

@@ -52,6 +52,7 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
+  virtual ~LinearOperatorBase(){};
 };
 
 
@@ -507,7 +508,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   virtual  void MpcDag   (const Field &in, Field &out){
     Mpc(in,out);
   }
-  virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+  virtual void MpcDagMpc(const Field &in, Field &out) {
     assert(0);// Never need with staggered
   }
 };
@@ -525,6 +526,7 @@ public:
       (*this)(Linop,in[k],out[k]);
     }
   };
+  virtual ~OperatorFunction(){};
 };
 
 template<class Field> class LinearFunction {
@@ -540,6 +542,7 @@ public:
       (*this)(in[i], out[i]);
     }
   }
+  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
@@ -585,6 +588,7 @@ class HermOpOperatorFunction : public OperatorFunction<Field> {
 template<typename Field>
 class PlainHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   LinearOperatorBase<Field> &_Linop;
       
   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
@@ -598,6 +602,7 @@ public:
 template<typename Field>
 class FunctionHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator(); 
   OperatorFunction<Field>   & _poly;
   LinearOperatorBase<Field> &_Linop;
       

Grid/algorithms/Preconditioner.h

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

Grid/algorithms/SparseMatrix.h

@@ -48,6 +48,7 @@ public:
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;
+  virtual ~SparseMatrixBase() {};
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -72,7 +73,7 @@ public:
   virtual  void MeooeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeInvDag (const Field &in, Field &out)=0;
-
+  virtual ~CheckerBoardedSparseMatrixBase() {};
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/approx/Chebyshev.h

@@ -258,26 +258,12 @@ public:
     for(int n=2;n<order;n++){
 
       Linop.HermOp(*Tn,y);
-#if 0
-      auto y_v = y.View();
-      auto Tn_v = Tn->View();
-      auto Tnp_v = Tnp->View();
-      auto Tnm_v = Tnm->View();
-      constexpr int Nsimd = vector_type::Nsimd();
-      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
-	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
-	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
-      });
-      if ( Coeffs[n] != 0.0) {
-	axpy(out,Coeffs[n],*Tnp,out);
-      }
-#else
       axpby(y,xscale,mscale,y,(*Tn));
       axpby(*Tnp,2.0,-1.0,y,(*Tnm));
       if ( Coeffs[n] != 0.0) {
 	axpy(out,Coeffs[n],*Tnp,out);
       }
-#endif
+
       // Cycle pointers to avoid copies
       Field *swizzle = Tnm;
       Tnm    =Tn;

Grid/algorithms/approx/Zolotarev.cc

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

Grid/algorithms/approx/Zolotarev.h

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

Grid/algorithms/blas/BatchedBlas.cc

@@ -0,0 +1,34 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/GridCore.h>
+#include <Grid/algorithms/blas/BatchedBlas.h>
+NAMESPACE_BEGIN(Grid);
+gridblasHandle_t GridBLAS::gridblasHandle;
+int              GridBLAS::gridblasInit;
+NAMESPACE_END(Grid);
+

Grid/algorithms/blas/BatchedBlas.h

@@ -0,0 +1,727 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#ifdef GRID_HIP
+#include <hipblas/hipblas.h>
+#endif
+#ifdef GRID_CUDA
+#include <cublas_v2.h>
+#endif
+#ifdef GRID_SYCL
+#include <oneapi/mkl.hpp>
+#endif
+#if 0
+#define GRID_ONE_MKL
+#endif
+#ifdef GRID_ONE_MKL
+#include <oneapi/mkl.hpp>
+#endif
+///////////////////////////////////////////////////////////////////////	  
+// Need to rearrange lattice data to be in the right format for a
+// batched multiply. Might as well make these static, dense packed
+///////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+#ifdef GRID_HIP
+  typedef hipblasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_CUDA
+  typedef cublasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_SYCL
+  typedef cl::sycl::queue *gridblasHandle_t;
+#endif
+#ifdef GRID_ONE_MKL
+  typedef cl::sycl::queue *gridblasHandle_t;
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
+  typedef int32_t gridblasHandle_t;
+#endif
+
+enum GridBLASOperation_t { GridBLAS_OP_N, GridBLAS_OP_T, GridBLAS_OP_C } ;
+
+class GridBLAS {
+public:
+
+  
+  static gridblasHandle_t gridblasHandle;
+  static int            gridblasInit;
+  
+  static void Init(void)
+  {
+    if ( ! gridblasInit ) {
+#ifdef GRID_CUDA
+      std::cout << "cublasCreate"<<std::endl;
+      cublasCreate(&gridblasHandle);
+      cublasSetPointerMode(gridblasHandle, CUBLAS_POINTER_MODE_DEVICE);
+#endif
+#ifdef GRID_HIP
+      std::cout << "hipblasCreate"<<std::endl;
+      hipblasCreate(&gridblasHandle);
+#endif
+#ifdef GRID_SYCL
+      gridblasHandle = theGridAccelerator;
+#endif
+#ifdef GRID_ONE_MKL
+      cl::sycl::cpu_selector selector;
+      cl::sycl::device selectedDevice { selector };
+      gridblasHandle =new sycl::queue (selectedDevice);
+#endif
+      gridblasInit=1;
+    }
+  }
+  
+  // Force construct once
+  GridBLAS() { Init(); };
+  ~GridBLAS() { };
+  
+  /////////////////////////////////////////////////////////////////////////////////////
+  // BLAS GEMM conventions:
+  /////////////////////////////////////////////////////////////////////////////////////
+  // - C = alpha A * B + beta C
+  // Dimensions:
+  // - C_m.n
+  // - A_m.k
+  // - B_k.n
+  // - Flops = 8 M N K
+  // - Bytes = 2*sizeof(word) * (MN+MK+KN)
+  // M=60, N=12
+  // Flop/Byte = 8 . 60.60.12 / (60.12+60.60+60.12)/16 = 4 so expect about 4 TF/s on a GCD
+  /////////////////////////////////////////////////////////////////////////////////////
+  void synchronise(void)
+  {
+#ifdef GRID_HIP
+    auto err = hipDeviceSynchronize();
+    assert(err==hipSuccess);
+#endif
+#ifdef GRID_CUDA
+    auto err = cudaDeviceSynchronize();
+    assert(err==cudaSuccess);
+#endif
+#ifdef GRID_SYCL
+    accelerator_barrier();
+#endif
+#ifdef GRID_ONE_MKL
+    gridblasHandle->wait();
+#endif
+  }
+  
+  void gemmBatched(int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<ComplexD> alpha_p(1);
+    static deviceVector<ComplexD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+    RealD t0=usecond();
+    //    std::cout << "ZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasZgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipblasDoubleComplex *) &alpha_p[0],
+				   (hipblasDoubleComplex **)&Amk[0], lda,
+				   (hipblasDoubleComplex **)&Bkn[0], ldb,
+				   (hipblasDoubleComplex *) &beta_p[0],
+				   (hipblasDoubleComplex **)&Cmn[0], ldc,
+				   batchCount);
+    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasZgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (cuDoubleComplex *) &alpha_p[0],
+				  (cuDoubleComplex **)&Amk[0], lda,
+				  (cuDoubleComplex **)&Bkn[0], ldb,
+				  (cuDoubleComplex *) &beta_p[0],
+				  (cuDoubleComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+    //    synchronise();
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogMessage<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<ComplexF> alpha_p(1);
+    static deviceVector<ComplexF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasCgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipblasComplex *) &alpha_p[0],
+				   (hipblasComplex **)&Amk[0], lda,
+				   (hipblasComplex **)&Bkn[0], ldb,
+				   (hipblasComplex *) &beta_p[0],
+				   (hipblasComplex **)&Cmn[0], ldc,
+				   batchCount);
+
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasCgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (cuComplex *) &alpha_p[0],
+				  (cuComplex **)&Amk[0], lda,
+				  (cuComplex **)&Bkn[0], ldb,
+				  (cuComplex *) &beta_p[0],
+				  (cuComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    ComplexF alphaf(real(alpha),imag(alpha));
+    ComplexF betaf(real(beta),imag(beta));
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexF c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alphaf)*c_mn + (betaf)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Single precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<RealF> alpha_p(1);
+    static deviceVector<RealF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasSgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (float *) &alpha_p[0],
+				   (float **)&Amk[0], lda,
+				   (float **)&Bkn[0], ldb,
+				   (float *) &beta_p[0],
+				   (float **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasSgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (float *) &alpha_p[0],
+				  (float **)&Amk[0], lda,
+				  (float **)&Bkn[0], ldb,
+				  (float *) &beta_p[0],
+				  (float **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Double precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<RealD> alpha_p(1);
+    static deviceVector<RealD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasDgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (double *) &alpha_p[0],
+				   (double **)&Amk[0], lda,
+				   (double **)&Bkn[0], ldb,
+				   (double *) &beta_p[0],
+				   (double **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasDgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (double *) &alpha_p[0],
+				  (double **)&Amk[0], lda,
+				  (double **)&Bkn[0], ldb,
+				  (double *) &beta_p[0],
+				  (double **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    /*
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t batchCount64=batchCount;
+      oneapi::mkl::blas::column_major::gemm_batch(*theGridAccelerator,
+      onemkl::transpose::N,
+      onemkl::transpose::N,
+      &m64,&n64,&k64,
+      (double *) &alpha_p[0],
+      (double **)&Amk[0], lda,
+      (double **)&Bkn[0], ldb,
+      (double *) &beta_p[0],
+      (double **)&Cmn[0], ldc,
+      1,&batchCount64);
+     */
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // Strided case used by benchmark, but generally unused in Grid
+  // Keep a code example in double complex, but don't generate the single and real variants for now
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  
+  void gemmStridedBatched(int m,int n, int k,
+			  ComplexD alpha,
+			  ComplexD* Amk,  // pointer list to matrices
+			  ComplexD* Bkn,
+			  ComplexD beta,
+			  ComplexD* Cmn,
+			  int batchCount)
+  {
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    int sda = m*k;
+    int sdb = k*n;
+    int sdc = m*n;
+    deviceVector<ComplexD> alpha_p(1);
+    deviceVector<ComplexD> beta_p(1);
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+
+    //    std::cout << "blasZgemmStridedBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    //    std::cout << "blasZgemmStridedBatched ld   "<<lda<<","<<ldb<<","<<ldc<<std::endl;
+    //    std::cout << "blasZgemmStridedBatched sd   "<<sda<<","<<sdb<<","<<sdc<<std::endl;
+#ifdef GRID_HIP
+    auto err = hipblasZgemmStridedBatched(gridblasHandle,
+					  HIPBLAS_OP_N,
+					  HIPBLAS_OP_N,
+					  m,n,k,
+					  (hipblasDoubleComplex *) &alpha_p[0],
+					  (hipblasDoubleComplex *) Amk, lda, sda,
+					  (hipblasDoubleComplex *) Bkn, ldb, sdb,
+					  (hipblasDoubleComplex *) &beta_p[0],
+					  (hipblasDoubleComplex *) Cmn, ldc, sdc,
+					  batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasZgemmStridedBatched(gridblasHandle,
+			      CUBLAS_OP_N,
+			      CUBLAS_OP_N,
+			      m,n,k,
+			      (cuDoubleComplex *) &alpha_p[0],
+			      (cuDoubleComplex *) Amk, lda, sda,
+			      (cuDoubleComplex *) Bkn, ldb, sdb,
+			      (cuDoubleComplex *) &beta_p[0],
+			      (cuDoubleComplex *) Cmn, ldc, sdc,
+			      batchCount);
+#endif
+#if defined(GRID_SYCL) || defined(GRID_ONE_MKL)
+    oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						oneapi::mkl::transpose::N,
+						oneapi::mkl::transpose::N,
+						m,n,k,
+						alpha,
+						(const ComplexD *)Amk,lda,sda,
+						(const ComplexD *)Bkn,ldb,sdb,
+						beta,
+						(ComplexD *)Cmn,ldc,sdc,
+						batchCount);
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
+     // Need a default/reference implementation
+     for (int p = 0; p < batchCount; ++p) {
+       for (int mm = 0; mm < m; ++mm) {
+	 for (int nn = 0; nn < n; ++nn) {
+	   ComplexD c_mn(0.0);
+	   for (int kk = 0; kk < k; ++kk)
+	     c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	   Cmn[mm + nn*ldc + p*sdc] =  (alpha)*c_mn + (beta)*Cmn[mm + nn*ldc + p*sdc];
+	 }
+       }
+     }
+#endif
+  }
+
+  double benchmark(int M, int N, int K, int BATCH)
+  {
+    int32_t N_A = M*K*BATCH;
+    int32_t N_B = K*N*BATCH;
+    int32_t N_C = M*N*BATCH;
+    deviceVector<ComplexD> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(ComplexD));
+    deviceVector<ComplexD> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(ComplexD));
+    deviceVector<ComplexD> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(ComplexD));
+    ComplexD alpha(1.0);
+    ComplexD beta (1.0);
+    RealD flops = 8.0*M*N*K*BATCH;
+    int ncall=10;
+    RealD t0 = usecond();
+    for(int i=0;i<ncall;i++){
+      gemmStridedBatched(M,N,K,
+			 alpha,
+			 &A[0], // m x k 
+			 &B[0], // k x n
+			 beta, 
+			 &C[0], // m x n
+			 BATCH);
+    }
+    synchronise();
+    RealD t1 = usecond();
+    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K)*BATCH;
+    flops = 8.0*M*N*K*BATCH*ncall;
+    flops = flops/(t1-t0)/1.e3;
+    return flops; // Returns gigaflops
+  }
+
+
+
+
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/iterative/BiCGSTABMixedPrec.h

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

Grid/algorithms/iterative/ConjugateGradient.h

@@ -58,6 +58,7 @@ public:
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
+    GRID_TRACE("ConjugateGradient");
     psi.Checkerboard() = src.Checkerboard();
 
     conformable(psi, src);
@@ -117,9 +118,13 @@ public:
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
 
+    RealD usecs = -usecond();
     SolverTimer.Start();
     int k;
     for (k = 1; k <= MaxIterations; k++) {
+
+      GridStopWatch IterationTimer;
+      IterationTimer.Start();
       c = cp;
 
       MatrixTimer.Start();
@@ -152,31 +157,41 @@ public:
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
-      std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+      IterationTimer.Stop();
+      if ( (k % 500) == 0 ) {
+	std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
                 << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
+      } else { 
+	std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+		  << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
+      }
 
       // Stopping condition
       if (cp <= rsq) {
+	usecs +=usecond();
         SolverTimer.Stop();
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
-
+	GridBase *grid = src.Grid();
+	RealD DwfFlops = (1452. )*grid->gSites()*4*k
+   	               + (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra
         RealD srcnorm = std::sqrt(norm2(src));
         RealD resnorm = std::sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
-
         std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k 
 		  << "\tComputed residual " << std::sqrt(cp / ssq)
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogIterative << "Time breakdown "<<std::endl;
-	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
+	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+
+	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -35,7 +35,8 @@ NAMESPACE_BEGIN(Grid);
     typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
-  public:                                                
+  public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;
@@ -48,6 +49,7 @@ NAMESPACE_BEGIN(Grid);
     Integer TotalInnerIterations; //Number of inner CG iterations
     Integer TotalOuterIterations; //Number of restarts
     Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
+    RealD TrueResidual;
 
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
@@ -67,6 +69,7 @@ NAMESPACE_BEGIN(Grid);
     }
   
   void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::cout << GridLogMessage << "MixedPrecisionConjugateGradient: Starting mixed precision CG with outer tolerance " << Tolerance << " and inner tolerance " << InnerTolerance << std::endl;
     TotalInnerIterations = 0;
 	
     GridStopWatch TotalTimer;
@@ -96,6 +99,7 @@ NAMESPACE_BEGIN(Grid);
     FieldF sol_f(SinglePrecGrid);
     sol_f.Checkerboard() = cb;
     
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting initial inner CG with tolerance " << inner_tol << std::endl;
     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
     CG_f.ErrorOnNoConverge = false;
 
@@ -104,7 +108,10 @@ NAMESPACE_BEGIN(Grid);
     GridStopWatch PrecChangeTimer;
     
     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
-      
+
+    precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
+    
     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
@@ -119,7 +126,7 @@ NAMESPACE_BEGIN(Grid);
       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d);
+      precisionChange(src_f, src_d, pc_wk_dp_to_sp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -129,6 +136,7 @@ NAMESPACE_BEGIN(Grid);
 	(*guesser)(src_f, sol_f);
 
       //Inner CG
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " << outer_iter << " starting inner CG with tolerance " << inner_tol << std::endl;
       CG_f.Tolerance = inner_tol;
       InnerCGtimer.Start();
       CG_f(Linop_f, src_f, sol_f);
@@ -137,7 +145,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f);
+      precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);
@@ -149,6 +157,7 @@ NAMESPACE_BEGIN(Grid);
     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
     CG_d(Linop_d, src_d_in, sol_d);
     TotalFinalStepIterations = CG_d.IterationsToComplete;
+    TrueResidual = CG_d.TrueResidual;
 
     TotalTimer.Stop();
     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -0,0 +1,213 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
+
+    Copyright (C) 2015
+
+    Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+
+NAMESPACE_BEGIN(Grid);
+
+//Mixed precision restarted defect correction CG
+template<class FieldD,class FieldF, 
+  typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+  typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
+public:
+  using LinearFunction<FieldD>::operator();
+  RealD   Tolerance;
+  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
+  Integer MaxInnerIterations;
+  Integer MaxOuterIterations;
+  Integer MaxPatchupIterations;
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+  LinearOperatorBase<FieldF> &Linop_f;
+  LinearOperatorBase<FieldD> &Linop_d;
+
+  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+  LinearFunction<FieldF> *guesser;
+  bool updateResidual;
+  
+  MixedPrecisionConjugateGradientBatched(RealD tol, 
+          Integer maxinnerit, 
+          Integer maxouterit, 
+          Integer maxpatchit,
+          GridBase* _sp_grid, 
+          LinearOperatorBase<FieldF> &_Linop_f, 
+          LinearOperatorBase<FieldD> &_Linop_d,
+          bool _updateResidual=true) :
+    Linop_f(_Linop_f), Linop_d(_Linop_d),
+    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
+    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
+
+  void useGuesser(LinearFunction<FieldF> &g){
+    guesser = &g;
+  }
+  
+  void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::vector<FieldD> srcs_d_in{src_d_in};
+    std::vector<FieldD> sols_d{sol_d};
+
+    (*this)(srcs_d_in,sols_d);
+
+    sol_d = sols_d[0];
+  }
+
+  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
+    assert(src_d_in.size() == sol_d.size());
+    int NBatch = src_d_in.size();
+
+    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
+
+    Integer TotalOuterIterations = 0; //Number of restarts
+    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
+    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
+  
+    GridStopWatch TotalTimer;
+    TotalTimer.Start();
+
+    GridStopWatch InnerCGtimer;
+    GridStopWatch PrecChangeTimer;
+    
+    int cb = src_d_in[0].Checkerboard();
+    
+    std::vector<RealD> src_norm;
+    std::vector<RealD> norm;
+    std::vector<RealD> stop;
+    
+    GridBase* DoublePrecGrid = src_d_in[0].Grid();
+    FieldD tmp_d(DoublePrecGrid);
+    tmp_d.Checkerboard() = cb;
+    
+    FieldD tmp2_d(DoublePrecGrid);
+    tmp2_d.Checkerboard() = cb;
+
+    std::vector<FieldD> src_d;
+    std::vector<FieldF> src_f;
+    std::vector<FieldF> sol_f;
+
+    for (int i=0; i<NBatch; i++) {
+      sol_d[i].Checkerboard() = cb;
+
+      src_norm.push_back(norm2(src_d_in[i]));
+      norm.push_back(0.);
+      stop.push_back(src_norm[i] * Tolerance*Tolerance);
+
+      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
+
+      src_f.push_back(SinglePrecGrid);
+      src_f[i].Checkerboard() = cb;
+
+      sol_f.push_back(SinglePrecGrid);
+      sol_f[i].Checkerboard() = cb;
+    }
+    
+    RealD inner_tol = InnerTolerance;
+    
+    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+    CG_f.ErrorOnNoConverge = false;
+    
+    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      std::cout << GridLogMessage << std::endl;
+      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      
+      bool allConverged = true;
+      
+      for (int i=0; i<NBatch; i++) {
+        //Compute double precision rsd and also new RHS vector.
+        Linop_d.HermOp(sol_d[i], tmp_d);
+        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        
+        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
+
+        PrecChangeTimer.Start();
+        precisionChange(src_f[i], src_d[i]);
+        PrecChangeTimer.Stop();
+        
+        sol_f[i] = Zero();
+      
+        if(norm[i] > OuterLoopNormMult * stop[i]) {
+          allConverged = false;
+        }
+      }
+      if (allConverged) break;
+
+      if (updateResidual) {
+        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
+        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
+        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+        CG_f.Tolerance = inner_tol;
+      }
+
+      //Optionally improve inner solver guess (eg using known eigenvectors)
+      if(guesser != NULL) {
+        (*guesser)(src_f, sol_f);
+      }
+
+      for (int i=0; i<NBatch; i++) {
+        //Inner CG
+        InnerCGtimer.Start();
+        CG_f(Linop_f, src_f[i], sol_f[i]);
+        InnerCGtimer.Stop();
+        TotalInnerIterations[i] += CG_f.IterationsToComplete;
+        
+        //Convert sol back to double and add to double prec solution
+        PrecChangeTimer.Start();
+        precisionChange(tmp_d, sol_f[i]);
+        PrecChangeTimer.Stop();
+        
+        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
+      }
+
+    }
+    
+    //Final trial CG
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
+    
+    for (int i=0; i<NBatch; i++) {
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
+      CG_d(Linop_d, src_d_in[i], sol_d[i]);
+      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
+    }
+
+    TotalTimer.Stop();
+
+    std::cout << GridLogMessage << std::endl;
+    for (int i=0; i<NBatch; i++) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
+    }
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -44,7 +44,7 @@ public:
 
   using OperatorFunction<Field>::operator();
 
-  RealD   Tolerance;
+  //  RealD   Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -52,7 +52,7 @@ public:
   MultiShiftFunction shifts;
   std::vector<RealD> TrueResidualShift;
 
-  ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
+  ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) : 
     MaxIterations(maxit),
     shifts(_shifts)
   { 
@@ -84,6 +84,7 @@ public:
 
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
+    GRID_TRACE("ConjugateGradientMultiShift");
   
     GridBase *grid = src.Grid();
   
@@ -182,6 +183,9 @@ public:
     for(int s=0;s<nshift;s++) {
       axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
     }
+
+    std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
+    
   
   ///////////////////////////////////////
   // Timers
@@ -321,8 +325,8 @@ public:
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -0,0 +1,373 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+//For safety the a final regular CG is applied to clean up if necessary
+
+//PB Pure single, then double fixup
+
+template<class FieldD, class FieldF,
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class ConjugateGradientMultiShiftMixedPrecCleanup : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterationsMshift;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrecCleanup(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+    MaxIterations(20000)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GRID_TRACE("ConjugateGradientMultiShiftMixedPrecCleanup");
+    GridBase *DoublePrecGrid = src_d.Grid();
+
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldF> ps_f (nshift, SinglePrecGrid);// Search directions (single precision)
+    std::vector<FieldF> psi_f(nshift, SinglePrecGrid);// solutions (single precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldF r_f(SinglePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    assert(psi_d.size()==nshift);
+    assert(mass.size()==nshift);
+    assert(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  rsqf[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF p_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      assert( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	psi_f[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      rsqf[s] =rsq[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      //      ps_d[s] = src_d;
+      precisionChange(ps_f[s],src_d);
+    }
+    // r and p for primary
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
+    
+    //MdagM+m[0]
+    precisionChange(p_f,p_d);
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    precisionChange(tmp_d,mmp_f);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    //    assert(norm2(tmp_d)< 1.0e-4);
+
+    axpy(mmp_d,mass[0],p_d,mmp_d);
+    RealD rn = norm2(p_d);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_d,b,mmp_d,r_d);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+      precisionChange(psi_f[s],psi_d[s]);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterationsMshift;k++){    
+
+      a = c /cp;
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChange(r_f, r_d);
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_f[s],a,ps_f[s],r_f);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_f[s],z[s][iz],as,r_f,ps_f[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      cp=c;
+      PrecChangeTimer.Start();
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f);
+      MatrixTimer.Stop();  
+      PrecChangeTimer.Start();
+      precisionChange(mmp_d, mmp_f); // From Float to Double
+      PrecChangeTimer.Stop();
+
+      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_d,mass[0],p_d,mmp_d);
+      RealD rn = norm2(p_d);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update single precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_f[ss],-bs[s]*alpha[s],ps_f[s],psi_f[ss]);
+	}
+      }
+      c = axpy_norm(r_d,b,mmp_d,r_d);
+      AXPYTimer.Stop();
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsqf[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged || k == MaxIterationsMshift-1){
+
+	SolverTimer.Stop();
+
+	for(int s=0;s<nshift;s++){
+	  precisionChange(psi_d[s],psi_f[s]);
+	}
+
+	
+	if ( all_converged ){
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: All shifts have converged iteration "<<k<<std::endl;
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Not all shifts have converged iteration "<<k<<std::endl;
+	}
+	
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrecCleanup: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+   
+    }
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    assert(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -0,0 +1,416 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+    
+//For safety the a final regular CG is applied to clean up if necessary
+
+//Linop to add shift to input linop, used in cleanup CG
+namespace ConjugateGradientMultiShiftMixedPrecSupport{
+template<typename Field>
+class ShiftedLinop: public LinearOperatorBase<Field>{
+public:
+  LinearOperatorBase<Field> &linop_base;
+  RealD shift;
+
+  ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
+
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
+  
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+
+  void HermOp(const Field &in, Field &out){
+    linop_base.HermOp(in, out);
+    axpy(out, shift, in, out);
+  }    
+
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+};
+};
+
+
+template<class FieldD, class FieldF,
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterationsMshift;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+    MaxIterations(20000)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
+    GridBase *DoublePrecGrid = src_d.Grid();
+
+    precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid);
+    
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    assert(psi_d.size()==nshift);
+    assert(mass.size()==nshift);
+    assert(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  rsqf[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF p_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      assert( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      rsqf[s] =rsq[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      ps_d[s] = src_d;
+    }
+    // r and p for primary
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
+    
+    //MdagM+m[0]
+    precisionChange(p_f, p_d, pc_wk_d_to_s);
+
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    precisionChange(tmp_d, mmp_f, pc_wk_s_to_d);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    assert(norm2(tmp_d)< 1.0);
+
+    axpy(mmp_d,mass[0],p_d,mmp_d);
+    RealD rn = norm2(p_d);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_d,b,mmp_d,r_d);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterationsMshift;k++){    
+
+      a = c /cp;
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_d[s],a,ps_d[s],r_d);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+
+      cp=c;
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f);
+      MatrixTimer.Stop();  
+
+      PrecChangeTimer.Start();
+      precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_d,mass[0],p_d,mmp_d);
+      AXPYTimer.Stop();
+      RealD rn = norm2(p_d);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update double precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
+	}
+      }
+
+      //Perform reliable update if necessary; otherwise update residual from single-prec mmp
+      c = axpy_norm(r_d,b,mmp_d,r_d);
+
+      AXPYTimer.Stop();
+
+      if(k % ReliableUpdateFreq == 0){
+	RealD c_old = c;
+	//Replace r with true residual
+	MatrixTimer.Start();  
+	Linop_d.HermOp(psi_d[0],mmp_d); 
+	MatrixTimer.Stop();  
+
+	AXPYTimer.Start();
+	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
+
+	c = axpy_norm(r_d, -1.0, mmp_d, src_d);
+	AXPYTimer.Stop();
+
+	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_old <<" with |r|^2 = "<<c<<std::endl;
+      }
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsqf[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged || k == MaxIterationsMshift-1){
+
+	SolverTimer.Stop();
+
+	if ( all_converged ){
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
+	}
+	
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+   
+    }
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    assert(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

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

Grid/algorithms/iterative/Deflation.h

@@ -33,16 +33,19 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
 };
 template<class Field>
 class DoNothingGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) {  };
 };
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) { guess = src; };
 };
 
@@ -57,6 +60,7 @@ private:
   const unsigned int       N;
 
 public:
+  using LinearFunction<Field>::operator();
 
   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
   : DeflatedGuesser(_evec, _eval, _evec.size())
@@ -87,6 +91,7 @@ private:
   const std::vector<RealD>       &eval_coarse;
 public:
   
+  using LinearFunction<FineField>::operator();
   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
 				const std::vector<CoarseField> &_evec_coarse,
 				const std::vector<RealD>       &_eval_coarse)
@@ -108,7 +113,43 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
-};
+
+  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
+    int Nevec = (int)evec_coarse.size();
+    int Nsrc = (int)src.size();
+    // make temp variables
+    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
+    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
+    //Preporcessing
+    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    guess_coarse[j] = Zero();
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockProject(src_coarse[j],src[j],subspace);
+    }
+    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
+    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
+    for (int i=0;i<Nevec;i++)
+    {
+      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
+      const CoarseField & tmp = evec_coarse[i];
+      for (int j=0;j<Nsrc;j++)
+      {
+        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
+      }
+    }
+    //postprocessing
+    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockPromote(guess_coarse[j],guess[j],subspace);
+    guess[j].Checkerboard() = src[j].Checkerboard();
+    }
+  };
+
+  };
 
 
 

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -419,14 +419,15 @@ until convergence
 	}
       }
 
-      if ( Nconv < Nstop )
+      if ( Nconv < Nstop ) {
 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-
+	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
+      }
       eval=eval2;
       
       //Keep only converged
-      eval.resize(Nconv);// Nstop?
-      evec.resize(Nconv,grid);// Nstop?
+      eval.resize(Nstop);// was Nconv
+      evec.resize(Nstop,grid);// was Nconv
       basisSortInPlace(evec,eval,reverse);
       
     }

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -44,6 +44,7 @@ public:
 				  int, MinRes);    // Must restart
 };
 
+//This class is the input parameter class for some testing programs
 struct LocalCoherenceLanczosParams : Serializable {
 public:
   GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
@@ -67,6 +68,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -97,6 +99,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -143,16 +146,24 @@ public:
   LinearOperatorBase<FineField> &_Linop;
   RealD                             _coarse_relax_tol;
   std::vector<FineField>        &_subspace;
+
+  int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
+                                //As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
+                                //To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
+                                //out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
+                                //NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
   
   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
 					   OperatorFunction<FineField>   &smoother,
 					   LinearOperatorBase<FineField> &Linop,
 					   std::vector<FineField>        &subspace,
-					   RealD coarse_relax_tol=5.0e3) 
+					   RealD coarse_relax_tol=5.0e3,
+					   int largestEvalIdxForReport=-1) 
     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
-      _coarse_relax_tol(coarse_relax_tol)  
+      _coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
   {    };
 
+  //evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
   int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
   {
     CoarseField v(B);
@@ -175,12 +186,26 @@ public:
 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
 	     <<std::endl;
 
+    if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
+      std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
+      RealD tmp_eval;
+      ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
+    }
+    
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
     return conv;
   }
-  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+
+  //This function is called at the end of the coarse grid Lanczos. It promotes the coarse eigenvector 'B' to the fine grid,
+  //applies a smoother to the result then computes the computes the *fine grid* eigenvalue (output as 'eval').
+
+  //evalMaxApprox should be the approximation of the largest eval of the fine Hermop. However when this function is called by IRL it actually passes the largest eval of the *Chebyshev* operator (as this is the max approx used for the TestConvergence above)
+  //As the largest eval of the Chebyshev is typically several orders of magnitude larger this makes the convergence test pass even when it should not.
+  //We therefore ignore evalMaxApprox here and use a value of 1.0 (note this value is already used by TestCoarse)
+  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)  
   {
+    evalMaxApprox = 1.0; //cf above
     GridBase *FineGrid = _subspace[0].Grid();    
     int checkerboard   = _subspace[0].Checkerboard();
     FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
@@ -199,13 +224,13 @@ public:
     eval   = vnum/vden;
     fv -= eval*fB;
     RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
-
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
+    
     std::cout.precision(13);
     std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
 	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv << " target " << eresid*eresid
 	     <<std::endl;
-    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
     if( (vv<eresid*eresid) ) return 1;
     return 0;
   }
@@ -283,6 +308,10 @@ public:
     evals_coarse.resize(0);
   };
 
+  //The block inner product is the inner product on the fine grid locally summed over the blocks
+  //to give a Lattice<Scalar> on the coarse grid. This function orthnormalizes the fine-grid subspace
+  //vectors under the block inner product. This step must be performed after computing the fine grid
+  //eigenvectors and before computing the coarse grid eigenvectors.    
   void Orthogonalise(void ) {
     CoarseScalar InnerProd(_CoarseGrid);
     std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
@@ -326,6 +355,8 @@ public:
     }
   }
 
+  //While this method serves to check the coarse eigenvectors, it also recomputes the eigenvalues from the smoothed reconstructed eigenvectors
+  //hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
   {
     assert(evals_fine.size() == nbasis);
@@ -374,25 +405,31 @@ public:
     evals_fine.resize(nbasis);
     subspace.resize(nbasis,_FineGrid);
   }
+
+
+  //cheby_op: Parameters of the fine grid Chebyshev polynomial used for the Lanczos acceleration
+  //cheby_smooth: Parameters of a separate Chebyshev polynomial used after the Lanczos has completed to smooth out high frequency noise in the reconstructed fine grid eigenvectors prior to computing the eigenvalue
+  //relax: Reconstructed eigenvectors (post smoothing) are naturally not as precise as true eigenvectors. This factor acts as a multiplier on the stopping condition when determining whether the results satisfy the user provided stopping condition
   void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
 		  int Nstop, int Nk, int Nm,RealD resid, 
 		  RealD MaxIt, RealD betastp, int MinRes)
   {
-    Chebyshev<FineField>                          Cheby(cheby_op);
-    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace);
-    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
+    Chebyshev<FineField>                          Cheby(cheby_op); //Chebyshev of fine operator on fine grid
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace); //Fine operator on coarse grid with intermediate fine grid conversion
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); //Chebyshev of fine operator on coarse grid with intermediate fine grid conversion
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
 
-    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1); 
 
     evals_coarse.resize(Nm);
     evec_coarse.resize(Nm,_CoarseGrid);
 
     CoarseField src(_CoarseGrid);     src=1.0; 
 
+    //Note the "tester" here is also responsible for generating the fine grid eigenvalues which are output into the "evals_coarse" array
     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
     int Nconv=0;
     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
@@ -403,6 +440,14 @@ public:
       std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl;
     }
   }
+
+  //Get the fine eigenvector 'i' by reconstruction
+  void getFineEvecEval(FineField &evec, RealD &eval, const int i) const{
+    blockPromote(evec_coarse[i],evec,subspace);  
+    eval = evals_coarse[i];
+  }
+    
+    
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/PowerMethod.h

@@ -29,6 +29,8 @@ template<class Field> class PowerMethod
       RealD vnum = real(innerProduct(src_n,tmp)); // HermOp. 
       RealD vden = norm2(src_n); 
       RealD na = vnum/vden; 
+
+      std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
       
       if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) { 
  	evalMaxApprox = na; 

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
 template<class Field>
 class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
 public:                                                
-
+  using LinearFunction<Field>::operator();
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
 template<class Field>
 class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
 public:                                                
-
+  using LinearFunction<Field>::operator();
   RealD   Tolerance;
   Integer MaxIterations;
   int verbose;
@@ -119,7 +119,8 @@ public:
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
 
     RealD cp;
-    ComplexD a, b, zAz;
+    ComplexD a, b;
+    //    ComplexD zAz;
     RealD zAAz;
     ComplexD rq;
 
@@ -146,7 +147,7 @@ public:
     //////////////////////////////////
     MatTimer.Start();
     Linop.Op(psi,Az);
-    zAz = innerProduct(Az,psi);
+    //    zAz = innerProduct(Az,psi);
     zAAz= norm2(Az);
     MatTimer.Stop();
     
@@ -170,7 +171,7 @@ public:
 
     LinalgTimer.Start();
 
-    zAz = innerProduct(Az,psi);
+    //    zAz = innerProduct(Az,psi);
     zAAz= norm2(Az);
 
     //p[0],q[0],qq[0] 
@@ -212,7 +213,7 @@ public:
       MatTimer.Start();
       Linop.Op(z,Az);
       MatTimer.Stop();
-      zAz = innerProduct(Az,psi);
+      //      zAz = innerProduct(Az,psi);
       zAAz= norm2(Az);
 
       LinalgTimer.Start();

Grid/allocator/AlignedAllocator.h

@@ -176,6 +176,7 @@ template<class T> using cshiftAllocator = std::allocator<T>;
 template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
 template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
 template<class T> using commVector = std::vector<T,devAllocator<T> >;
+template<class T> using deviceVector  = std::vector<T,devAllocator<T> >;
 template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);

Grid/allocator/MemoryManager.cc

@@ -4,11 +4,14 @@ 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)
+#define CpuHuge  (1)
+#define CpuSmall (2)
+#define Acc      (3)
+#define AccHuge  (4)
+#define AccSmall (5)
+#define Shared   (6)
+#define SharedHuge  (7)
+#define SharedSmall (8)
 #undef GRID_MM_VERBOSE 
 uint64_t total_shared;
 uint64_t total_device;
@@ -35,12 +38,15 @@ void MemoryManager::PrintBytes(void)
   
 }
 
+uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
+uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
+
 //////////////////////////////////////////////////////////////////////
 // Data tables for recently freed pooiniter caches
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils
@@ -159,7 +165,6 @@ void MemoryManager::Init(void)
 
   char * str;
   int Nc;
-  int NcS;
   
   str= getenv("GRID_ALLOC_NCACHE_LARGE");
   if ( str ) {
@@ -171,6 +176,16 @@ void MemoryManager::Init(void)
     }
   }
 
+  str= getenv("GRID_ALLOC_NCACHE_HUGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuHuge]=Nc;
+      Ncache[AccHuge]=Nc;
+      Ncache[SharedHuge]=Nc;
+    }
+  }
+
   str= getenv("GRID_ALLOC_NCACHE_SMALL");
   if ( str ) {
     Nc = atoi(str);
@@ -191,7 +206,9 @@ void MemoryManager::InitMessage(void) {
   
   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;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
 #endif
   
 #ifdef GRID_UVM
@@ -223,8 +240,11 @@ void MemoryManager::InitMessage(void) {
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  int cache = type + small;
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
@@ -233,11 +253,12 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 
 void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
 
+  if (ncache == 0) return ptr;
+
   void * ret = NULL;
   int v = -1;
 
@@ -272,8 +293,11 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 void *MemoryManager::Lookup(size_t bytes,int type)
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  int cache = type+small;
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
   return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
@@ -282,7 +306,6 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 
 void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 

Grid/allocator/MemoryManager.h

@@ -35,6 +35,12 @@ NAMESPACE_BEGIN(Grid);
 // Move control to configure.ac and Config.h?
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
+#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
 
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
@@ -65,6 +71,21 @@ enum ViewMode {
   CpuWriteDiscard = 0x10 // same for now
 };
 
+struct MemoryStatus {
+  uint64_t     DeviceBytes;
+  uint64_t     DeviceLRUBytes;
+  uint64_t     DeviceMaxBytes;
+  uint64_t     HostToDeviceBytes;
+  uint64_t     DeviceToHostBytes;
+  uint64_t     HostToDeviceXfer;
+  uint64_t     DeviceToHostXfer;
+  uint64_t     DeviceEvictions;
+  uint64_t     DeviceDestroy;
+  uint64_t     DeviceAllocCacheBytes;
+  uint64_t     HostAllocCacheBytes;
+};
+
+
 class MemoryManager {
 private:
 
@@ -78,7 +99,7 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=6;
+  static const int NallocType=9;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
@@ -92,8 +113,9 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
+  static void Audit(std::string s);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -113,7 +135,28 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
- 
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
+  
+  static uint64_t     DeviceCacheBytes();
+  static uint64_t     HostCacheBytes();
+
+  static MemoryStatus GetFootprint(void) {
+    MemoryStatus stat;
+    stat.DeviceBytes       = DeviceBytes;
+    stat.DeviceLRUBytes    = DeviceLRUBytes;
+    stat.DeviceMaxBytes    = DeviceMaxBytes;
+    stat.HostToDeviceBytes = HostToDeviceBytes;
+    stat.DeviceToHostBytes = DeviceToHostBytes;
+    stat.HostToDeviceXfer  = HostToDeviceXfer;
+    stat.DeviceToHostXfer  = DeviceToHostXfer;
+    stat.DeviceEvictions   = DeviceEvictions;
+    stat.DeviceDestroy     = DeviceDestroy;
+    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
+    stat.HostAllocCacheBytes   = HostCacheBytes();
+    return stat;
+  };
+  
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////
@@ -170,6 +213,8 @@ private:
 
  public:
   static void Print(void);
+  static void PrintAll(void);
+  static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);
   static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);

Grid/allocator/MemoryManagerCache.cc

@@ -3,8 +3,13 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define dprintf(...)
+
+#define MAXLINE 512
+static char print_buffer [ MAXLINE ];
+
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+//#define dprintf(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -23,6 +28,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -104,15 +111,17 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    AccCache.AccPtr=(uint64_t) NULL;
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -121,26 +130,36 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove entry
-  // Cannot be locked. If allocated must be in LRU pool.
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-  assert(AccCache.accLock==0);
-  assert(AccCache.cpuLock==0);
+  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
+	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  if (AccCache.accLock!=0) return;
+  if (AccCache.cpuLock!=0) return;
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
-  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);  
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
+    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  uint64_t CpuPtr = AccCache.CpuPtr;
-  EntryErase(CpuPtr);
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -150,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   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);
+  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -165,7 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     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);
+  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -191,6 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -202,6 +222,7 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -212,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back();
+      uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
+    } else {
+      return;
     }
   }
 }
@@ -241,11 +265,12 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
+    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-		    (uint64_t)bytes);
+	            (uint64_t)bytes,
+		    (uint64_t)AccCache.accLock);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -280,6 +305,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
+    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -292,28 +318,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
+    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);
+    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  // If view is opened on device remove from LRU
+  assert(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
+    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -334,10 +362,12 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
-
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
+  } else {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -374,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  if (!AccCache.AccPtr) {
-     EvictVictims(bytes);
-  }
+  // CPU doesn't need to free space
+  //  if (!AccCache.AccPtr) {
+  //    EvictVictims(bytes);
+  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -430,20 +461,28 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  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;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -453,13 +492,13 @@ void  MemoryManager::Print(void)
     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
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -473,6 +512,89 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
+void MemoryManager::Audit(std::string s)
+{
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    assert(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    assert(AccCache.LRU_valid==1);
+    assert(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      assert(AccCache.LRU_valid==0);
+
+      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+		<< "\t cpuLock  " << AccCache.cpuLock
+		<< "\t accLock  " << AccCache.accLock
+		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
+    }
+
+    assert( AccCache.cpuLock== 0 ) ;
+    assert( AccCache.accLock== 0 ) ;
+  }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  assert(LruBytes1==LruBytes2);
+  assert(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  assert(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  assert(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+
+}
+
+void MemoryManager::PrintState(void* _CpuPtr)
+{
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+    if ( AccCache.state==EvictNext) str = std::string("EvictNext");
+
+    std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\t" << AccCache.cpuLock
+    << "\t" << AccCache.accLock
+    << "\t" << AccCache.LRU_valid<<std::endl;
+
+  } else {
+    std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl; 
+  }
+}
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryManagerShared.cc

@@ -12,11 +12,19 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
+void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
+void  MemoryManager::PrintState(void* CpuPtr)
+{
+std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
+};
 void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_base.h

@@ -53,10 +53,11 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
+  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processor_coor;  // linear processor coordinate
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -97,14 +98,16 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                      ProcessorCount(void)    ;
+  int                ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
   static void BroadcastWorld(int root,void* data, int bytes);
+  static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
   // Reduction
@@ -128,7 +131,7 @@ public:
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
-    scalar_type * ptr = (scalar_type *)& o;
+    scalar_type * ptr = (scalar_type *)& o; // Safe alias 
     GlobalSumVector(ptr,words);
   }
   
@@ -142,17 +145,17 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,
+			       int xmit_to_rank,int do_xmit,
 			       void *recv,
-			       int recv_from_rank,
+			       int recv_from_rank,int do_recv,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,
+				    int xmit_to_rank,int do_xmit,
 				    void *recv,
-				    int recv_from_rank,
-				    int bytes,int dir);
+				    int recv_from_rank,int do_recv,
+				    int xbytes,int rbytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);

Grid/communicator/Communicator_mpi3.cc

@@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
   // Remap using the shared memory optimising routine
   // The remap creates a comm which must be freed
   ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,12 +124,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
-
+  Coordinate shm_processors       (_ndimension,1);
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
   for(int d=0;d<parent_ndimension;d++){
     parent_processor_coor[pad+d]=parent._processor_coor[d];
     parent_processors    [pad+d]=parent._processors[d];
+    shm_processors       [pad+d]=parent._shm_processors[d];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -154,6 +155,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     ccoor[d] = parent_processor_coor[d] % processors[d];
     scoor[d] = parent_processor_coor[d] / processors[d];
     ssize[d] = parent_processors[d]     / processors[d];
+    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@@ -335,23 +337,24 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest,
+						     int dest, int dox,
 						     void *recv,
-						     int from,
+						     int from, int dor,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
+#undef NVLINK_GET // Define to use get instead of put DMA
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,
+							 int dest,int dox,
 							 void *recv,
-							 int from,
-							 int bytes,int dir)
+							 int from,int dor,
+							 int xbytes,int rbytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -370,30 +373,37 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    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 ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      list.push_back(rrq);
+      off_node_bytes+=rbytes;
+    }
+#ifdef NVLINK_GET
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+#endif
   }
-
-  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    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;
-  } else {
-    // TODO : make a OMP loop on CPU, call threaded bcopy
-    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-    assert(shm!=NULL);
-    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
-    acceleratorCopySynchronise(); // MPI prob slower
-  }
-
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-    this->StencilSendToRecvFromComplete(list,dir);
+  
+  if (dox) {
+    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+_processor*32;
+      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      list.push_back(xrq);
+      off_node_bytes+=xbytes;
+    } else {
+#ifndef NVLINK_GET
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+#endif
+      
+    }
   }
 
   return off_node_bytes;
@@ -402,6 +412,8 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
 {
   int nreq=list.size();
 
+  acceleratorCopySynchronise();
+
   if (nreq==0) return;
 
   std::vector<MPI_Status> status(nreq);
@@ -435,6 +447,10 @@ int CartesianCommunicator::RankWorld(void){
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
+void CartesianCommunicator::BarrierWorld(void){
+  int ierr = MPI_Barrier(communicator_world);
+  assert(ierr==0);
+}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,

Grid/communicator/Communicator_none.cc

@@ -45,12 +45,14 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
+  _shm_processors = Coordinate(processors.size(),1);
   srank=0;
   SetCommunicator(communicator_world);
 }
 
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
+  _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
   _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
@@ -102,6 +104,7 @@ int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
+void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
@@ -111,21 +114,21 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,
+						     int xmit_to_rank,int dox,
 						     void *recv,
-						     int recv_from_rank,
+						     int recv_from_rank,int dor,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,
+							 int xmit_to_rank,int dox,
 							 void *recv,
-							 int recv_from_rank,
-							 int bytes, int dir)
+							 int recv_from_rank,int dor,
+							 int xbytes,int rbytes, int dir)
 {
-  return 2.0*bytes;
+  return xbytes+rbytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {

Grid/communicator/SharedMemory.cc

@@ -40,6 +40,9 @@ int                 GlobalSharedMemory::_ShmAlloc;
 uint64_t            GlobalSharedMemory::_ShmAllocBytes;
 
 std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
+#ifndef ACCELERATOR_AWARE_MPI
+void * GlobalSharedMemory::HostCommBuf;
+#endif
 
 Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm;
 int                 GlobalSharedMemory::WorldShmRank;
@@ -66,6 +69,26 @@ void GlobalSharedMemory::SharedMemoryFree(void)
 /////////////////////////////////
 // Alloc, free shmem region
 /////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+void *SharedMemory::HostBufferMalloc(size_t bytes){
+  void *ptr = (void *)host_heap_top;
+  host_heap_top  += bytes;
+  host_heap_bytes+= bytes;
+  if (host_heap_bytes >= host_heap_size) {
+    std::cout<< " HostBufferMalloc exceeded heap size -- try increasing with --shm <MB> flag" <<std::endl;
+    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
+    std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
+    std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
+    std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
+    assert(host_heap_bytes<host_heap_size);
+  }
+  return ptr;
+}
+void SharedMemory::HostBufferFreeAll(void) { 
+  host_heap_top  =(size_t)HostCommBuf;
+  host_heap_bytes=0;
+}
+#endif
 void *SharedMemory::ShmBufferMalloc(size_t bytes){
   //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
   void *ptr = (void *)heap_top;
@@ -91,6 +114,59 @@ void *SharedMemory::ShmBufferSelf(void)
   //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
   return ShmCommBufs[ShmRank];
 }
+static inline int divides(int a,int b)
+{
+  return ( b == ( (b/a)*a ) );
+}
+void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
+{
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
+  ////////////////////////////////////////////////////////////////
+  // Powers of 2,3,5 only in prime decomposition for now
+  ////////////////////////////////////////////////////////////////
+  int ndimension = WorldDims.size();
+  ShmDims=Coordinate(ndimension,1);
+
+  std::vector<int> primes({2,3,5});
+
+  int dim = 0;
+  int last_dim = ndimension - 1;
+  int AutoShmSize = 1;
+  while(AutoShmSize != WorldShmSize) {
+    int p;
+    for(p=0;p<primes.size();p++) {
+      int prime=primes[p];
+      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
+        && divides(prime,WorldShmSize/AutoShmSize)  ) {
+  AutoShmSize*=prime;
+  ShmDims[dim]*=prime;
+  last_dim = dim;
+  break;
+      }
+    }
+    if (p == primes.size() && last_dim == dim) {
+      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
+      exit(EXIT_FAILURE);
+    }
+    dim=(dim+1) %ndimension;
+  }
+}
 
 NAMESPACE_END(Grid); 
 

Grid/communicator/SharedMemory.h

@@ -75,7 +75,9 @@ public:
   static int           Hugepages;
 
   static std::vector<void *> WorldShmCommBufs;
-
+#ifndef ACCELERATOR_AWARE_MPI
+  static void *HostCommBuf;
+#endif
   static Grid_MPI_Comm WorldComm;
   static int           WorldRank;
   static int           WorldSize;
@@ -93,9 +95,10 @@ public:
   // Create an optimal reordered communicator that makes MPI_Cart_create get it right
   //////////////////////////////////////////////////////////////////////////////////////
   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
   static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
   ///////////////////////////////////////////////////
   // Provide shared memory facilities off comm world
@@ -119,6 +122,13 @@ private:
   size_t heap_bytes;
   size_t heap_size;
 
+#ifndef ACCELERATOR_AWARE_MPI
+  size_t host_heap_top;  // set in free all
+  size_t host_heap_bytes;// set in free all
+  void *HostCommBuf;     // set in SetCommunicator
+  size_t host_heap_size; // set in SetCommunicator
+#endif
+  
 protected:
 
   Grid_MPI_Comm    ShmComm; // for barriers
@@ -150,7 +160,10 @@ public:
   void *ShmBufferTranslate(int rank,void * local_p);
   void *ShmBufferMalloc(size_t bytes);
   void  ShmBufferFreeAll(void) ;
-  
+#ifndef ACCELERATOR_AWARE_MPI
+  void *HostBufferMalloc(size_t bytes);
+  void HostBufferFreeAll(void);
+#endif  
   //////////////////////////////////////////////////////////////////////////
   // Make info on Nodes & ranks and Shared memory available
   //////////////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryMPI.cc

@@ -27,6 +27,8 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
+#define Mheader "SharedMemoryMpi: "
+
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
@@ -36,12 +38,122 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
 #endif
-#ifdef GRID_SYCl
-
+#ifdef GRID_SYCL
+#ifdef ACCELERATOR_AWARE_MPI
+#define GRID_SYCL_LEVEL_ZERO_IPC
+#define SHM_SOCKETS
+#endif 
+#include <syscall.h>
 #endif
 
+#include <sys/socket.h>
+#include <sys/un.h>
+
 NAMESPACE_BEGIN(Grid); 
-#define header "SharedMemoryMpi: "
+
+#ifdef SHM_SOCKETS
+
+/*
+ * Barbaric extra intranode communication route in case we need sockets to pass FDs
+ * Forced by level_zero not being nicely designed
+ */
+static int sock;
+static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
+static char sock_path[256];
+class UnixSockets {
+public:
+  static void Open(int rank)
+  {
+    int errnum;
+
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
+    unlink(sa_un.sun_path);
+    if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
+      perror("bind failure");
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  static int RecvFileDescriptor(void)
+  {
+    int n;
+    int fd;
+    char buf[1];
+    struct iovec iov;
+    struct msghdr msg;
+    struct cmsghdr *cmsg;
+    char cms[CMSG_SPACE(sizeof(int))];
+
+    iov.iov_base = buf;
+    iov.iov_len = 1;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_name = 0;
+    msg.msg_namelen = 0;
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+
+    msg.msg_control = (caddr_t)cms;
+    msg.msg_controllen = sizeof cms;
+
+    if((n=recvmsg(sock, &msg, 0)) < 0) {
+      perror("recvmsg failed");
+      return -1;
+    }
+    if(n == 0){
+      perror("recvmsg returned 0");
+      return -1;
+    }
+    cmsg = CMSG_FIRSTHDR(&msg);
+
+    memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
+
+    return fd;
+  }
+
+  static void SendFileDescriptor(int fildes,int xmit_to_rank)
+  {
+    struct msghdr msg;
+    struct iovec iov;
+    struct cmsghdr *cmsg = NULL;
+    char ctrl[CMSG_SPACE(sizeof(int))];
+    char data = ' ';
+
+    memset(&msg, 0, sizeof(struct msghdr));
+    memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
+    iov.iov_base = &data;
+    iov.iov_len = sizeof(data);
+    
+    sprintf(sock_path,sock_path_fmt,xmit_to_rank);
+    
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
+
+    msg.msg_name = (void *)&sa_un;
+    msg.msg_namelen = sizeof(sa_un);
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+    msg.msg_controllen =  CMSG_SPACE(sizeof(int));
+    msg.msg_control = ctrl;
+
+    cmsg = CMSG_FIRSTHDR(&msg);
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int));
+
+    *((int *) CMSG_DATA(cmsg)) = fildes;
+
+    sendmsg(sock, &msg, 0);
+  };
+};
+#endif
+
+
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -64,8 +176,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
   if ( WorldRank == 0) {
-    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
-    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+    std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl;
   }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
@@ -152,7 +264,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   //////////////////////////////////////////////////////////////////////////////
   // Look and see if it looks like an HPE 8600 based on hostname conventions
@@ -165,63 +277,11 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
 
-  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
-static inline int divides(int a,int b)
-{
-  return ( b == ( (b/a)*a ) );
-}
-void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
-{
-  ////////////////////////////////////////////////////////////////
-  // Allow user to configure through environment variable
-  ////////////////////////////////////////////////////////////////
-  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
-  if ( str ) {
-    std::vector<int> IntShmDims;
-    GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
-    long ShmSize = 1;
-    for (int dim=0;dim<WorldDims.size();dim++) {
-      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
-    }
-    assert(ShmSize == WorldShmSize);
-    return;
-  }
-  
-  ////////////////////////////////////////////////////////////////
-  // Powers of 2,3,5 only in prime decomposition for now
-  ////////////////////////////////////////////////////////////////
-  int ndimension = WorldDims.size();
-  ShmDims=Coordinate(ndimension,1);
 
-  std::vector<int> primes({2,3,5});
-
-  int dim = 0;
-  int last_dim = ndimension - 1;
-  int AutoShmSize = 1;
-  while(AutoShmSize != WorldShmSize) {
-    int p;
-    for(p=0;p<primes.size();p++) {
-      int prime=primes[p];
-      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
-        && divides(prime,WorldShmSize/AutoShmSize)  ) {
-	AutoShmSize*=prime;
-	ShmDims[dim]*=prime;
-	last_dim = dim;
-	break;
-      }
-    }
-    if (p == primes.size() && last_dim == dim) {
-      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
-      exit(EXIT_FAILURE);
-    }
-    dim=(dim+1) %ndimension;
-  }
-}
-void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -294,7 +354,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-
+  SHM = ShmDims;
+  
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -341,7 +402,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
 }
-void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -353,6 +414,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
+  SHM=ShmDims;
+
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -391,7 +454,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -451,46 +514,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
 #if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
-
-//if defined(GRID_SYCL)
-#if 0
-void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
-{
-  void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // allocate the pointer array for shared windows for our group
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  MPI_Barrier(WorldShmComm);
-  WorldShmCommBufs.resize(WorldShmSize);
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Each MPI rank should allocate our own buffer
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
-
-  if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
-    exit(EXIT_FAILURE);  
-  }
-
-  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
-
-  SharedMemoryZero(ShmCommBuf,bytes);
-
-  assert(WorldShmSize == 1);
-  for(int r=0;r<WorldShmSize;r++){
-    WorldShmCommBufs[r] = ShmCommBuf;
-  }
-  _ShmAllocBytes=bytes;
-  _ShmAlloc=1;
-}
-#endif
-
-#if defined(GRID_CUDA) ||defined(GRID_HIP) ||defined(GRID_SYCL)  
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -513,22 +536,30 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+  HostCommBuf= malloc(bytes);
+#endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
+    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
   std::cout<< "Setting up IPC"<<std::endl;
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef SHM_SOCKETS
+  UnixSockets::Open(WorldShmRank);
+#endif
   for(int r=0;r<WorldShmSize;r++){
 
+    MPI_Barrier(WorldShmComm);
+
 #ifndef GRID_MPI3_SHM_NONE
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
@@ -536,24 +567,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void * thisBuf = ShmCommBuf;
     if(!Stencil_force_mpi) {
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+    typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
-
+    
     if ( r==WorldShmRank ) { 
       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
+      memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
+#ifdef SHM_SOCKETS
+      for(int rr=0;rr<WorldShmSize;rr++){
+	if(rr!=r){
+	  UnixSockets::SendFileDescriptor(handle.fd,rr);
+	}
+      }
+#endif
     }
 #endif
 #ifdef GRID_CUDA
@@ -581,6 +620,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
     { 
+      MPI_Barrier(WorldShmComm);
       int ierr=MPI_Bcast(&handle,
 			 sizeof(handle),
 			 MPI_BYTE,
@@ -596,6 +636,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     if ( r!=WorldShmRank ) {
       thisBuf = nullptr;
+      int myfd;
+#ifdef SHM_SOCKETS
+      myfd=UnixSockets::RecvFileDescriptor();
+#else
       std::cout<<"mapping seeking remote pid/fd "
 	       <<handle.pid<<"/"
 	       <<handle.fd<<std::endl;
@@ -603,16 +647,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
       std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
-      int myfd  = syscall(438,pidfd,handle.fd,0);
-
-      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-      
+      myfd  = syscall(438,pidfd,handle.fd,0);
+      int err_t = errno;
+      if (myfd < 0) {
+        fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr);
+	perror("pidfd_getfd failed ");
+	assert(0);
+      }
+#endif
+      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
       auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
@@ -647,18 +697,18 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
+    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
-#endif
 
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -695,7 +745,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -705,7 +755,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -752,7 +802,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
@@ -876,6 +926,12 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
   ShmBufferFreeAll();
 
+#ifndef ACCELERATOR_AWARE_MPI
+  host_heap_size = heap_size;
+  HostCommBuf= GlobalSharedMemory::HostCommBuf;
+  HostBufferFreeAll();
+#endif  
+
   /////////////////////////////////////////////////////////////////////
   // find comm ranks in our SHM group (i.e. which ranks are on our node)
   /////////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryNone.cc

@@ -48,9 +48,10 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   _ShmSetup=1;
 }
 
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   optimal_comm = WorldComm;
+  SHM = Coordinate(processors.size(),1);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////

Grid/cshift/Cshift_common.h

@@ -29,8 +29,27 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-extern Vector<std::pair<int,int> > Cshift_table; 
+extern std::vector<std::pair<int,int> > Cshift_table; 
+extern commVector<std::pair<int,int> > Cshift_table_device; 
 
+inline std::pair<int,int> *MapCshiftTable(void)
+{
+  // GPU version
+#ifdef ACCELERATOR_CSHIFT    
+  uint64_t sz=Cshift_table.size();
+  if (Cshift_table_device.size()!=sz )    {
+    Cshift_table_device.resize(sz);
+  }
+  acceleratorCopyToDevice((void *)&Cshift_table[0],
+			  (void *)&Cshift_table_device[0],
+			  sizeof(Cshift_table[0])*sz);
+
+  return &Cshift_table_device[0];
+#else 
+  return &Cshift_table[0];
+#endif
+  // CPU version use identify map
+}
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
@@ -74,8 +93,8 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
   }
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
+    auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
@@ -225,7 +244,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
@@ -340,7 +359,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
@@ -392,7 +411,7 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);

Grid/cshift/Cshift_mpi.h

@@ -52,7 +52,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
   int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
 
-
+  RealD t1,t0;
+  t0=usecond();
   if ( !comm_dim ) {
     //std::cout << "CSHIFT: Cshift_local" <<std::endl;
     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
@@ -63,6 +64,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
     //std::cout << "CSHIFT: Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift);
   }
+  t1=usecond();
+  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
 
@@ -127,16 +130,20 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
   for(int x=0;x<rd;x++){       
 
     int sx        =  (x+sshift)%rd;
     int comm_proc = ((x+sshift)/rd)%pd;
     
     if (comm_proc==0) {
-
+      tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-
+      tcopy+=usecond();
     } else {
 
       int words = buffer_size;
@@ -144,26 +151,39 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
+      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-
-      grid->Barrier();
+      
+      tcomms-=usecond();
+      //      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      xbytes+=bytes;
+      //      grid->Barrier();
+      tcomms+=usecond();
 
-      grid->Barrier();
-
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  /*
+  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  */
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -190,6 +210,12 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   assert(shift>=0);
   assert(shift<fd);
 
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
+  
   int permute_type=grid->PermuteType(dimension);
 
   ///////////////////////////////////////////////
@@ -227,7 +253,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       pointers[i] = &send_buf_extract[i][0];
     }
     int sx   = (x+sshift)%rd;
+    tgather-=usecond();
     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+    tgather+=usecond();
 
     for(int i=0;i<Nsimd;i++){
       
@@ -252,7 +280,8 @@ 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->Barrier();
+	tcomms-=usecond();
+	//	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
@@ -262,7 +291,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 
-	grid->Barrier();
+	xbytes+=bytes;
+	//	grid->Barrier();
+	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -270,9 +301,17 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       }
 
     }
+    tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
+    tscatter+=usecond();
   }
-
+  /*
+  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  */
 }
 #else 
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -292,6 +331,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(comm_dim==1);
   assert(shift>=0);
   assert(shift<fd);
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
   static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
@@ -315,7 +359,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     
     if (comm_proc==0) {
 
+      tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
+      tcopy+=usecond();
 
     } else {
 
@@ -324,7 +370,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
+      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
@@ -332,7 +380,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
 
-      grid->Barrier();
+      tcomms-=usecond();
+      //      grid->Barrier();
 
       acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
       grid->SendToRecvFrom((void *)&send_buf[0],
@@ -340,13 +389,24 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      xbytes+=bytes;
       acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
 
-      grid->Barrier();
+      //      grid->Barrier();
+      tcomms+=usecond();
 
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  /*
+  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  */
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -372,6 +432,11 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   assert(simd_layout==2);
   assert(shift>=0);
   assert(shift<fd);
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
 
   int permute_type=grid->PermuteType(dimension);
 
@@ -414,8 +479,10 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     for(int i=0;i<Nsimd;i++){       
       pointers[i] = &send_buf_extract[i][0];
     }
+    tgather-=usecond();
     int sx   = (x+sshift)%rd;
     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+    tgather+=usecond();
 
     for(int i=0;i<Nsimd;i++){
       
@@ -440,7 +507,8 @@ 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->Barrier();
+	tcomms-=usecond();
+	//	grid->Barrier();
 
 	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
@@ -449,17 +517,28 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
+	xbytes+=bytes;
 
-	grid->Barrier();
+	//	grid->Barrier();
+	tcomms+=usecond();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
 	rpointers[i] = &send_buf_extract[nbr_lane][0];
       }
 
     }
+    tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-  }
+    tscatter+=usecond();
 
+  }
+  /*
+  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
+  */
 }
 #endif
 NAMESPACE_END(Grid); 

Grid/cshift/Cshift_table.cc

@@ -1,4 +1,5 @@
 #include <Grid/GridCore.h>       
 NAMESPACE_BEGIN(Grid);
-Vector<std::pair<int,int> > Cshift_table; 
+std::vector<std::pair<int,int> > Cshift_table; 
+commVector<std::pair<int,int> > Cshift_table_device; 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice.h

@@ -35,6 +35,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_transpose.h>
 #include <Grid/lattice/Lattice_local.h>
 #include <Grid/lattice/Lattice_reduction.h>
+#include <Grid/lattice/Lattice_crc.h>
 #include <Grid/lattice/Lattice_peekpoke.h>
 #include <Grid/lattice/Lattice_reality.h>
 #include <Grid/lattice/Lattice_real_imag.h>
@@ -46,3 +47,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_unary.h>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
+#include <Grid/lattice/PaddedCell.h>

Grid/lattice/Lattice_ET.h

@@ -63,7 +63,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
+  //  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   const int Nsimd = vobj::vector_type::Nsimd();
@@ -345,7 +345,9 @@ GridUnopClass(UnaryNot, Not(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
+GridUnopClass(UnarySpTa, SpTa(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
+GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
@@ -456,7 +458,9 @@ GRID_DEF_UNOP(operator!, UnaryNot);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
+GRID_DEF_UNOP(SpTa, UnarySpTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
+GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup);
 GRID_DEF_UNOP(timesI, UnaryTimesI);
 GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
 GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the

Grid/lattice/Lattice_arith.h

@@ -36,6 +36,7 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("mult");
   ret.Checkerboard() = lhs.Checkerboard();
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
@@ -53,6 +54,7 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("mac");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -70,6 +72,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("sub");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -86,6 +89,7 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("add");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -106,6 +110,7 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
+  GRID_TRACE("mult");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -119,6 +124,7 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
+  GRID_TRACE("mac");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -133,6 +139,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
+  GRID_TRACE("sub");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -146,6 +153,7 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
+  GRID_TRACE("add");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -163,6 +171,7 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("mult");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -177,6 +186,7 @@ void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("mac");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -191,6 +201,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("sub");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -204,6 +215,7 @@ void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
+  GRID_TRACE("add");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -218,6 +230,7 @@ void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class sobj,class vobj> inline
 void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
+  GRID_TRACE("axpy");
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
@@ -231,6 +244,7 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
 }
 template<class sobj,class vobj> inline
 void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
+  GRID_TRACE("axpby");
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
@@ -246,13 +260,52 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
+  GRID_TRACE("axpy_norm");
     return axpy_norm_fast(ret,a,x,y);
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
+  GRID_TRACE("axpby_norm");
     return axpby_norm_fast(ret,a,b,x,y);
 }
 
+/// Trace product
+template<class obj> auto traceProduct(const Lattice<obj> &rhs_1,const Lattice<obj> &rhs_2)
+  -> Lattice<decltype(trace(obj()))>
+{
+  typedef decltype(trace(obj())) robj;
+  Lattice<robj> ret_i(rhs_1.Grid());
+  autoView( rhs1 , rhs_1, AcceleratorRead);
+  autoView( rhs2 , rhs_2, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
+  ret.Checkerboard() = rhs_1.Checkerboard();
+  accelerator_for(ss,rhs1.size(),obj::Nsimd(),{
+      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2(ss)));
+  });
+  return ret_i;
+}
+
+template<class obj1,class obj2> auto traceProduct(const Lattice<obj1> &rhs_1,const obj2 &rhs2)
+  -> Lattice<decltype(trace(obj1()))>
+{
+  typedef decltype(trace(obj1())) robj;
+  Lattice<robj> ret_i(rhs_1.Grid());
+  autoView( rhs1 , rhs_1, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
+  ret.Checkerboard() = rhs_1.Checkerboard();
+  accelerator_for(ss,rhs1.size(),obj1::Nsimd(),{
+      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2));
+  });
+  return ret_i;
+}
+template<class obj1,class obj2> auto traceProduct(const obj2 &rhs_2,const Lattice<obj1> &rhs_1)
+  -> Lattice<decltype(trace(obj1()))>
+{
+  return traceProduct(rhs_1,rhs_2);
+}
+
+
+
 NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_base.h

@@ -88,6 +88,13 @@ public:
     LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
     accessor.ViewClose();
   }
+
+  // Helper function to print the state of this object in the AccCache
+  void PrintCacheState(void)
+  {
+    MemoryManager::PrintState(this->_odata);
+  }
+
   /////////////////////////////////////////////////////////////////////////////////
   // Return a view object that may be dereferenced in site loops.
   // The view is trivially copy constructible and may be copied to an accelerator device
@@ -110,6 +117,7 @@ public:
   ////////////////////////////////////////////////////////////////////////////////
   template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
   {
+    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -133,6 +141,7 @@ public:
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
   {
+    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -156,6 +165,7 @@ public:
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
   {
+    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -281,8 +291,8 @@ 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(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
@@ -296,8 +306,8 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });

Grid/lattice/Lattice_basis.h

@@ -62,7 +62,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
     basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
-#if ( (!defined(GRID_CUDA)) )
+#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
   int max_threads = thread_max();
   Vector < vobj > Bt(Nm * max_threads);
   thread_region

Grid/lattice/Lattice_crc.h

@@ -0,0 +1,55 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/Lattice_crc.h
+
+    Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
+{
+  auto ff = localNorm2(f);
+  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
+  typedef typename vobj::tensor_reduced normtype;
+  typedef typename normtype::scalar_object scalar;
+  std::vector<scalar> sff;
+  sliceSum(ff,sff,mu);
+  for(int t=0;t<sff.size();t++){
+    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
+  }
+}
+
+template<class vobj> uint32_t crc(const Lattice<vobj> & buf)
+{
+  autoView( buf_v , buf, CpuRead);
+  return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
+}
+
+#define CRC(U) std::cerr << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
+
+NAMESPACE_END(Grid);
+
+

Grid/lattice/Lattice_matrix_reduction.h

@@ -32,7 +32,6 @@ template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -82,7 +81,6 @@ template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -130,7 +128,6 @@ template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs.Grid();

Grid/lattice/Lattice_peekpoke.h

@@ -96,9 +96,6 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   GridBase *grid=l.Grid();
 
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
   int Nsimd = grid->Nsimd();
 
   assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
@@ -125,14 +122,17 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+template<class vobj>
+typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
+  typename vobj::scalar_object s;
+  peekSite(s,l,site);
+  return s;
+}        
 template<class vobj,class sobj>
 void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         
   GridBase *grid=l.Grid();
 
-  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));
@@ -173,11 +173,11 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  scalar_type * vp = (scalar_type *)&l[odx];
+  const vector_type *vp = (const vector_type *) &l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
-    pt[w] = vp[idx+w*Nsimd];
+    pt[w] = getlane(vp[w],idx);
   }
       
   return;
@@ -210,10 +210,10 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  scalar_type * vp = (scalar_type *)&l[odx];
+  vector_type * vp = (vector_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
   for(int w=0;w<words;w++){
-    vp[idx+w*Nsimd] = pt[w];
+    putlane(vp[w],pt[w],idx);
   }
   return;
 };

Grid/lattice/Lattice_reduction.h

@@ -28,6 +28,10 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
+#if defined(GRID_SYCL)
+#include <Grid/lattice/Lattice_reduction_sycl.h>
+#endif
+#include <Grid/lattice/Lattice_slicesum_core.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -91,10 +95,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
   for(int i=0;i<nthread;i++){
     ssum = ssum+sumarray[i];
   } 
-  
-  typedef typename vobj::scalar_object ssobj;
-  ssobj ret = ssum;
-  return ret;
+  return ssum;
 }
 /*
 Threaded max, don't use for now
@@ -127,7 +128,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -136,25 +137,61 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
 #endif  
 }
+template<class vobj>
+inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  return sumD_gpu_large(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
+
+template<class vobj>
+inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
+{
+  Integer osites = arg.Grid()->oSites();
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  autoView( arg_v, arg, AcceleratorRead);
+  return sum_gpu(&arg_v[0],osites);
+#else
+  autoView(arg_v, arg, CpuRead);
+  return sum_cpu(&arg_v[0],osites);
+#endif  
+}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+  auto ssum = rankSum(arg);
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
+template<class vobj>
+inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu(&arg_v[0],osites);
+  return sum_gpu_large(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
-#endif  
+  return sum_cpu(&arg_v[0],osites);
+#endif
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSumLarge(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -197,7 +234,6 @@ template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
 template<class vobj>
 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;
   
@@ -207,33 +243,67 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   const uint64_t sites = grid->oSites();
   
   // Might make all code paths go this way.
+#if 0
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
+  {
+    autoView( left_v , left, AcceleratorRead);
+    autoView( right_v,right, AcceleratorRead);
+    // This code could read coalesce
+    // GPU - SIMT lane compliance...
+    accelerator_for( ss, sites, nsimd,{
+	auto x_l = left_v(ss);
+	auto y_l = right_v(ss);
+	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
+    });
+  }
+#else
+  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
     
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
 
     // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, 1,{
-	auto x_l = left_v[ss];
-	auto y_l = right_v[ss];
-	inner_tmp_v[ss]=innerProductD(x_l,y_l);
+    accelerator_for( ss, sites, nsimd,{
+	auto x_l = left_v(ss);
+	auto y_l = right_v(ss);
+	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
     });
   }
-
+#endif
   // This is in single precision and fails some tests
-  auto anrm = sum(inner_tmp_v,sites);  
+  auto anrm = sumD(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
 
+
 template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
+
+#ifdef GRID_SYCL
+  uint64_t csum=0;
+  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  {
+    // Hack
+    // Fast integer xor checksum. Can also be used in comms now.
+    autoView(l_v,left,AcceleratorRead);
+    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
+    uint64_t *base= (uint64_t *)&l_v[0];
+    csum=svm_xor(base,words);
+  }
+  FlightRecorder::CsumLog(csum);
+#endif
   ComplexD nrm = rankInnerProduct(left,right);
+  RealD local = real(nrm);
+  FlightRecorder::NormLog(real(nrm)); 
   grid->GlobalSum(nrm);
+  FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
 
@@ -257,8 +327,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   conformable(z,x);
   conformable(x,y);
 
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_typeD vector_type;
+  //  typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
   GridBase *grid = x.Grid();
@@ -270,17 +339,29 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-
+#if 0
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
 
-  accelerator_for( ss, sites, 1,{
-      auto tmp = a*x_v[ss]+b*y_v[ss];
-      inner_tmp_v[ss]=innerProductD(tmp,tmp);
-      z_v[ss]=tmp;
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
   });
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
+#else
+  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
+  });
+  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
+#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -290,7 +371,6 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
 {
   conformable(left,right);
 
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
   Vector<ComplexD> tmp(2);
 
@@ -387,19 +467,10 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int e1=    grid->_slice_nblock[orthogdim];
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
-
-  // sum over reduced dimension planes, breaking out orthog dir
-  // Parallel over orthog direction
-  autoView( Data_v, Data, CpuRead);
-  thread_for( r,rd, {
-    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int ss= so+n*stride+b;
-	lvSum[r]=lvSum[r]+Data_v[ss];
-      }
-    }
-  });
+  int ostride=grid->_ostride[orthogdim];
+  
+  //Reduce Data down to lvSum
+  sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
   Coordinate icoor(Nd);
@@ -434,6 +505,15 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
 }
+template<class vobj> inline
+std::vector<typename vobj::scalar_object> 
+sliceSum(const Lattice<vobj> &Data,int orthogdim)
+{
+  std::vector<typename vobj::scalar_object> result;
+  sliceSum(Data,result,orthogdim);
+  return result;
+}
+
 
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
@@ -538,7 +618,8 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho
 template<class vobj>
 static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
 			    int orthogdim,RealD scale=1.0) 
-{    
+{
+  // perhaps easier to just promote A to a field and use regular madd
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
@@ -569,8 +650,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
     for(int l=0;l<Nsimd;l++){
       grid->iCoorFromIindex(icoor,l);
       int ldx =r+icoor[orthogdim]*rd;
-      scalar_type *as =(scalar_type *)&av;
-      as[l] = scalar_type(a[ldx])*zscale;
+      av.putlane(scalar_type(a[ldx])*zscale,l);
     }
 
     tensor_reduced at; at=av;
@@ -610,7 +690,6 @@ template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -664,7 +743,6 @@ template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -718,7 +796,6 @@ template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs.Grid();

Grid/lattice/Lattice_reduction_gpu.h

@@ -23,28 +23,27 @@ unsigned int nextPow2(Iterator x) {
 }
 
 template <class Iterator>
-void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+int 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);
+  auto r=hipGetDevice(&device);
 #endif
   
   Iterator warpSize            = gpu_props[device].warpSize;
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  
+  /*  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
-  std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  
+  */  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -52,10 +51,14 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   
   // let the number of threads in a block be a multiple of 2, starting from warpSize
   threads = warpSize;
+  if ( threads*sizeofsobj > sharedMemPerBlock ) {
+    std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
+    return 0;
+  }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  
+  return 1;
 }
 
 template <class sobj, class Iterator>
@@ -195,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -204,17 +207,77 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   Integer size = osites*nsimd;
 
   Integer numThreads, numBlocks;
-  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  Integer smemSize = numThreads * sizeof(sobj);
+  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  assert(ok);
 
+  Integer smemSize = numThreads * sizeof(sobj);
+  // Move out of UVM
+  // Turns out I had messed up the synchronise after move to compute stream
+  // as running this on the default stream fools the synchronise
+#undef UVM_BLOCK_BUFFER  
+#ifndef UVM_BLOCK_BUFFER  
+  commVector<sobj> buffer(numBlocks);
+  sobj *buffer_v = &buffer[0];
+  sobj result;
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
+  accelerator_barrier();
+  acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
+#else
   Vector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
-  
-  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  sobj result;
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
-  auto result = buffer_v[0];
+  result = *buffer_v;
+#endif
   return result;
 }
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobj;
+  sobj ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<vector> buffer(osites);
+  vector *dat = (vector *)lat;
+  vector *buf = &buffer[0];
+  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,1,{
+	buf[ss] = dat[ss*words+w];
+      });
+      
+    ret_p[w] = sumD_gpu_small(tbuf,osites);
+  }
+  return ret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_objectD sobj;
+  sobj ret;
+  
+  Integer nsimd= vobj::Nsimd();
+  Integer size = osites*nsimd;
+  Integer numThreads, numBlocks;
+  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  
+  if ( ok ) {
+    ret = sumD_gpu_small(lat,osites);
+  } else {
+    ret = sumD_gpu_large(lat,osites);
+  }
+  return ret;
+}
+
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Return as same precision as input performing reduction in double precision though
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -227,6 +290,13 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
-
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu_large(lat,osites);
+  return result;
+}
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_sycl.h

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

Grid/lattice/Lattice_rng.h

@@ -152,6 +152,7 @@ public:
 #ifdef RNG_FAST_DISCARD
   static void Skip(RngEngine &eng,uint64_t site)
   {
+#if 0
     /////////////////////////////////////////////////////////////////////////////////////
     // Skip by 2^40 elements between successive lattice sites
     // This goes by 10^12.
@@ -162,9 +163,9 @@ public:
     // tens of seconds per trajectory so this is clean in all reasonable cases,
     // and margin of safety is orders of magnitude.
     // We could hack Sitmo to skip in the higher order words of state if necessary
-      //
-      // Replace with 2^30 ; avoid problem on large volumes
-      //
+    //
+    // Replace with 2^30 ; avoid problem on large volumes
+    //
     /////////////////////////////////////////////////////////////////////////////////////
     //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
     const int shift = 30;
@@ -179,6 +180,9 @@ public:
     assert((skip >> shift)==site); // check for overflow
 
     eng.discard(skip);
+#else
+    eng.discardhi(site);
+#endif
     //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
   } 
 #endif
@@ -407,7 +411,7 @@ public:
       std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
       SeedFixedIntegers(seeds);
     }
-  void SeedFixedIntegers(const std::vector<int> &seeds){
+  void SeedFixedIntegers(const std::vector<int> &seeds, int britney=0){
 
     // Everyone generates the same seed_seq based on input seeds
     CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
@@ -424,22 +428,29 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
+    thread_for( lidx, _grid->lSites(), {
 
-    // Everybody loops over global volume.
-    thread_for( gidx, _grid->_gsites, {
-	// Where is it?
-	int rank;
+	int gidx;
 	int o_idx;
 	int i_idx;
-
+	int rank;
+	Coordinate pcoor;
+	Coordinate lcoor;
 	Coordinate gcoor;
-	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
+	_grid->LocalIndexToLocalCoor(lidx,lcoor);
+	pcoor=_grid->ThisProcessorCoor();
+	_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
+	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
+
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
+
+	assert(rank == _grid->ThisRank() );
 	
-	// If this is one of mine we take it
-	if( rank == _grid->ThisRank() ){
-	  int l_idx=generator_idx(o_idx,i_idx);
-	  _generators[l_idx] = master_engine;
+	int l_idx=generator_idx(o_idx,i_idx);
+	_generators[l_idx] = master_engine;
+	if ( britney ) { 
+	  Skip(_generators[l_idx],l_idx); // Skip to next RNG sequence
+	} else { 	
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
     });

Grid/lattice/Lattice_slicesum_core.h

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

Grid/lattice/Lattice_trace.h

@@ -66,6 +66,65 @@ inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<
   return ret;
 };
 
+template<int N, class Vec>
+Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
+  typedef typename Vec::scalar_type scalar;
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<scalar, N> > > Us;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	scalar tmp= Us()()(i,j);
+	ComplexD ztmp(real(tmp),imag(tmp));
+	EigenU(i,j)=ztmp;
+      }}
+    ComplexD detD  = EigenU.determinant();
+    typename Vec::scalar_type det(detD.real(),detD.imag());
+    pokeLocalSite(det,ret_v,lcoor);
+  });
+  return ret;
+}
+
+template<int N>
+Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
+  
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us()()(i,j);
+      }}
+    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	Ui()()(i,j) = EigenUinv(i,j);
+      }}
+    pokeLocalSite(Ui,ret_v,lcoor);
+  });
+  return ret;
+}
+
+
 NAMESPACE_END(Grid);
 #endif
 

Grid/lattice/Lattice_transfer.h

@@ -85,6 +85,76 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
   });
 }
 
+template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
+{
+  half.Checkerboard() = cb;
+  autoView(half_v, half, AcceleratorWrite);
+  autoView(full_v, full, AcceleratorRead);
+  Coordinate rdim_full             = full.Grid()->_rdimensions;
+  Coordinate rdim_half             = half.Grid()->_rdimensions;
+  unsigned long ndim_half          = half.Grid()->_ndimension;
+  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
+  Coordinate ostride_half          = half.Grid()->_ostride;
+  accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{
+    
+    Coordinate coor;
+    int cbos;
+    int linear=0;
+
+    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
+    assert(coor.size()==ndim_half);
+
+    for(int d=0;d<ndim_half;d++){ 
+      if(checker_dim_mask_half[d]) linear += coor[d];
+    }
+    cbos = (linear&0x1);
+
+    if (cbos==cb) {
+      int ssh=0;
+      for(int d=0;d<ndim_half;d++) {
+        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
+        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
+      }
+      coalescedWrite(half_v[ssh],full_v(ss));
+    }
+  });
+}
+template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
+{
+  int cb = half.Checkerboard();
+  autoView(half_v , half, AcceleratorRead);
+  autoView(full_v , full, AcceleratorWrite);
+  Coordinate rdim_full             = full.Grid()->_rdimensions;
+  Coordinate rdim_half             = half.Grid()->_rdimensions;
+  unsigned long ndim_half          = half.Grid()->_ndimension;
+  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
+  Coordinate ostride_half          = half.Grid()->_ostride;
+  accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{
+
+    Coordinate coor;
+    int cbos;
+    int linear=0;
+  
+    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
+    assert(coor.size()==ndim_half);
+
+    for(int d=0;d<ndim_half;d++){ 
+      if(checker_dim_mask_half[d]) linear += coor[d];
+    }
+    cbos = (linear&0x1);
+
+    if (cbos==cb) {
+      int ssh=0;
+      for(int d=0;d<ndim_half;d++){
+        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
+        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
+      }
+      coalescedWrite(full_v[ss],half_v(ssh));
+    }
+
+  });
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Flexible Type Conversion for internal promotion to double as well as graceful
 // treatment of scalar-compatible types
@@ -124,11 +194,11 @@ accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
 #endif
 
 accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) {
-  out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v);
+  precisionChange(out,in);
 }
 
 accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
-  Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
+  precisionChange(out,in);
 }
 
 template<typename T1,typename T2>
@@ -218,7 +288,36 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
   }
 }
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               const std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = fineData.size();
+  assert(coarseData.size() == NBatch);
 
+  GridBase * fine  = fineData[0].Grid();
+  GridBase * coarse= coarseData[0].Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  std::vector<Lattice<vobj>> fineDataCopy = fineData;
+
+  autoView(ip_, ip, AcceleratorWrite);
+  for(int v=0;v<nbasis;v++) {
+    for (int k=0; k<NBatch; k++) {
+      autoView( coarseData_ , coarseData[k], AcceleratorWrite);
+      blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // 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(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]); 
+    }
+  }
+}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -370,15 +469,13 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   Coordinate fine_rdimensions = fine->_rdimensions;
   Coordinate coarse_rdimensions = coarse->_rdimensions;
 
-  vobj zz = Zero();
-  
   accelerator_for(sc,coarse->oSites(),1,{
 
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
 
-      vobj cd = zz;
+      vobj cd = Zero();
       
       for(int sb=0;sb<blockVol;sb++){
 
@@ -520,6 +617,26 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 }
 #endif
 
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = coarseData.size();
+  assert(fineData.size() == NBatch);
+
+  GridBase * fine   = fineData[0].Grid();
+  GridBase * coarse = coarseData[0].Grid();
+  for (int k=0; k<NBatch; k++)
+    fineData[k]=Zero();
+  for (int i=0;i<nbasis;i++) {
+    for (int k=0; k<NBatch; k++) {
+      Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
+      blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
+    }
+  }
+}
+
 // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
 // Simd layouts need not match since we use peek/poke Local
 template<class vobj,class vvobj>
@@ -578,8 +695,68 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
-
   // the above should guarantee that the operations are local
+  
+#if 1
+
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
+  
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+ 
+  thread_for(idx, nsite, {
+      Coordinate from_coor, to_coor;
+      size_t rem = idx;
+      for(int i=0;i<nd;i++){
+	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i];
+	from_coor[i] = base_i + FromLowerLeft[i];
+	to_coor[i] = base_i + ToLowerLeft[i];
+      }
+      
+      int foidx = Fg->oIndex(from_coor);
+      int fiidx = Fg->iIndex(from_coor);
+      int toidx = Tg->oIndex(to_coor);
+      int tiidx = Tg->iIndex(to_coor);
+      int* tt = table + 4*idx;
+      tt[0] = foidx;
+      tt[1] = fiidx;
+      tt[2] = toidx;
+      tt[3] = tiidx;
+    });
+  
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+
+#else  
   Coordinate ldf = Fg->_ldimensions;
   Coordinate rdf = Fg->_rdimensions;
   Coordinate isf = Fg->_istride;
@@ -588,9 +765,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
 
-  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, AcceleratorRead);
-  accelerator_for(idx,Fg->lSites(),1,{
+  autoView( t_v , To, CpuWrite);
+  autoView( f_v , From, CpuRead);
+  thread_for(idx,Fg->lSites(),{
     sobj s;
     Coordinate Fcoor(nd);
     Coordinate Tcoor(nd);
@@ -603,17 +780,24 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
     }
     if (in_region) {
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
+#if 0      
+      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
+      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
+      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
+      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
       scalar_type * fp = (scalar_type *)&f_v[odx_f];
       scalar_type * tp = (scalar_type *)&t_v[odx_t];
       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
+	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
       }
+#else
+    peekLocalSite(s,f_v,Fcoor);
+    pokeLocalSite(s,t_v,Tcoor);
+#endif
     }
   });
+
+#endif
 }
 
 
@@ -706,6 +890,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 }
 
 
+//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
+//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -722,11 +908,70 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-  }
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
+#if 1
+  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+  
+  thread_for(idx,nsite,{
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
+    lcoor[orthog] = slice_lo;
+    hcoor[orthog] = slice_hi;
+    size_t rem = idx;
+    for(int mu=0;mu<nl;mu++){
+      if(mu != orthog){
+	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
+	lcoor[mu] = hcoor[mu] = xmu;
+      }
+    }
+    int loidx = lg->oIndex(lcoor);
+    int liidx = lg->iIndex(lcoor);
+    int hoidx = hg->oIndex(hcoor);
+    int hiidx = hg->iIndex(hcoor);
+    int* tt = table + 4*idx;
+    tt[0] = loidx;
+    tt[1] = liidx;
+    tt[2] = hoidx;
+    tt[3] = hiidx;
+    });
+   
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(lowDim_v,lowDim,AcceleratorRead);
+  autoView(higherDim_v,higherDim,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
+      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+#else
   // the above should guarantee that the operations are local
   autoView(lowDimv,lowDim,CpuRead);
   autoView(higherDimv,higherDim,CpuWrite);
@@ -742,6 +987,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       pokeLocalSite(s,higherDimv,hcoor);
     }
   });
+#endif
 }
 
 
@@ -785,7 +1031,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
 
 
 template<class vobj>
-void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
+void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 {
   typedef typename vobj::scalar_object sobj;
 
@@ -1010,9 +1256,27 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//Convert a Lattice from one precision to another
+//Very fast precision change. Requires in/out objects to reside on same Grid (e.g. by using double2 for the double-precision field)
 template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+{
+  typedef typename VobjOut::vector_type Vout;
+  typedef typename VobjIn::vector_type Vin;
+  const int N = sizeof(VobjOut)/sizeof(Vout);
+  conformable(out.Grid(),in.Grid());
+  out.Checkerboard() = in.Checkerboard();
+  int nsimd = out.Grid()->Nsimd();
+  autoView( out_v  , out, AcceleratorWrite);
+  autoView(  in_v ,   in, AcceleratorRead);
+  accelerator_for(idx,out.Grid()->oSites(),1,{
+      Vout *vout = (Vout *)&out_v[idx];
+      Vin  *vin  = (Vin  *)&in_v[idx];
+      precisionChange(vout,vin,N);
+  });
+}
+//Convert a Lattice from one precision to another (original, slow implementation)
+template<class VobjOut, class VobjIn>
+void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 {
   assert(out.Grid()->Nd() == in.Grid()->Nd());
   for(int d=0;d<out.Grid()->Nd();d++){
@@ -1027,7 +1291,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 
   int ndim = out.Grid()->Nd();
   int out_nsimd = out_grid->Nsimd();
-    
+  int in_nsimd = in_grid->Nsimd();
   std::vector<Coordinate > out_icoor(out_nsimd);
       
   for(int lane=0; lane < out_nsimd; lane++){
@@ -1058,6 +1322,128 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   });
 }
 
+//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
+class precisionChangeWorkspace{
+  std::pair<Integer,Integer>* fmap_device; //device pointer
+  //maintain grids for checking
+  GridBase* _out_grid;
+  GridBase* _in_grid;
+public:
+  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
+    //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
+    assert(out_grid->Nd() == in_grid->Nd());
+    for(int d=0;d<out_grid->Nd();d++){
+      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
+    }
+    int Nsimd_out = out_grid->Nsimd();
+
+    std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
+    for(int lane=0; lane < out_grid->Nsimd(); lane++)
+      out_grid->iCoorFromIindex(out_icorrs[lane], lane);
+  
+    std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
+    thread_for(out_oidx,out_grid->oSites(),{
+	Coordinate out_ocorr; 
+	out_grid->oCoorFromOindex(out_ocorr, out_oidx);
+      
+	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
+	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
+	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
+	
+	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
+	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
+	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
+	  int in_oidx = 0, in_lane = 0;
+	  for(int d=0;d<in_grid->_ndimension;d++){
+	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
+	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
+	  }
+	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
+	}
+      });
+
+    //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
+    size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
+    fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
+    acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes); 
+  }
+
+  //Prevent moving or copying
+  precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
+  precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
+  
+  std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
+
+  void checkGrids(GridBase* out, GridBase* in) const{
+    conformable(out, _out_grid);
+    conformable(in, _in_grid);
+  }
+  
+  ~precisionChangeWorkspace(){
+    acceleratorFreeDevice(fmap_device);
+  }
+};
+
+
+//We would like to use precisionChangeFast when possible. However usage of this requires the Grids to be the same (runtime check)
+//*and* the precisionChange(VobjOut::vector_type, VobjIn, int) function to be defined for the types; this requires an extra compile-time check which we do using some SFINAE trickery
+template<class VobjOut, class VobjIn>
+auto _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, int dummy)->decltype( precisionChange( ((typename VobjOut::vector_type*)0), ((typename VobjIn::vector_type*)0), 1), int()){
+  if(out.Grid() == in.Grid()){
+    precisionChangeFast(out,in);
+    return 1;
+  }else{
+    return 0;
+  }
+}
+template<class VobjOut, class VobjIn>
+int _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, long dummy){ //note long here is intentional; it means the above is preferred if available
+  return 0;
+}
+
+
+//Convert a lattice of one precision to another. Much faster than original implementation but requires a pregenerated workspace
+//which contains the mapping data.
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
+  if(_precisionChangeFastWrap(out,in,0)) return;
+  
+  static_assert( std::is_same<typename VobjOut::scalar_typeD, typename VobjIn::scalar_typeD>::value == 1, "precisionChange: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+
+  out.Checkerboard() = in.Checkerboard();
+  constexpr int Nsimd_out = VobjOut::Nsimd();
+
+  workspace.checkGrids(out.Grid(),in.Grid());
+  std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
+
+  //Do the copy/precision change
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( in_v , in, AcceleratorRead);
+
+  accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
+      std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
+      for(int out_lane=0; out_lane < Nsimd_out; out_lane++){      
+	int in_oidx = fmap_osite[out_lane].first;
+	int in_lane = fmap_osite[out_lane].second;
+	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
+      }
+    });
+}
+
+//Convert a Lattice from one precision to another. Much faster than original implementation but slower than precisionChangeFast
+//or precisionChange called with pregenerated workspace, as it needs to internally generate the workspace on the host and copy to device
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
+  if(_precisionChangeFastWrap(out,in,0)) return;   
+  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
+  precisionChange(out, in, workspace);
+}
+
+
+
+
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////

Grid/lattice/Lattice_view.h

@@ -45,6 +45,7 @@ public:
   };
   // Host only
   GridBase * getGrid(void) const { return _grid; };
+  vobj* getHostPointer(void) const { return _odata; };
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////

Grid/lattice/PaddedCell.h

@@ -0,0 +1,174 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/PaddedCell.h
+
+    Copyright (C) 2019
+
+Author: Peter Boyle pboyle@bnl.gov
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include<Grid/cshift/Cshift.h>
+
+NAMESPACE_BEGIN(Grid);
+
+//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
+template<typename vobj>
+struct CshiftImplBase{
+  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
+  virtual ~CshiftImplBase(){}
+};
+template<typename vobj>
+struct CshiftImplDefault: public CshiftImplBase<vobj>{
+  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
+};
+template<typename Gimpl>
+struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
+  typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
+};  
+
+class PaddedCell {
+public:
+  GridCartesian * unpadded_grid;
+  int dims;
+  int depth;
+  std::vector<GridCartesian *> grids;
+
+  ~PaddedCell()
+  {
+    DeleteGrids();
+  }
+  PaddedCell(int _depth,GridCartesian *_grid)
+  {
+    unpadded_grid = _grid;
+    depth=_depth;
+    dims=_grid->Nd();
+    AllocateGrids();
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    for(int d=0;d<dims;d++){
+      assert(local[d]>=depth);
+    }
+  }
+  void DeleteGrids(void)
+  {
+    for(int d=0;d<grids.size();d++){
+      delete grids[d];
+    }
+    grids.resize(0);
+  };
+  void AllocateGrids(void)
+  {
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate simd      =unpadded_grid->_simd_layout;
+    Coordinate processors=unpadded_grid->_processors;
+    Coordinate plocal    =unpadded_grid->LocalDimensions();
+    Coordinate global(dims);
+
+    // expand up one dim at a time
+    for(int d=0;d<dims;d++){
+
+      plocal[d] += 2*depth; 
+
+      for(int d=0;d<dims;d++){
+	global[d] = plocal[d]*processors[d];
+      }
+
+      grids.push_back(new GridCartesian(global,simd,processors));
+    }
+  };
+  template<class vobj>
+  inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
+  {
+    Lattice<vobj> out(unpadded_grid);
+
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate fll(dims,depth); // depends on the MPI spread
+    Coordinate tll(dims,0); // depends on the MPI spread
+    localCopyRegion(in,out,fll,tll,local);
+    return out;
+  }
+  template<class vobj>
+  inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = Expand(d,tmp,cshift); // rvalue && assignment
+    }
+    return tmp;
+  }
+  // expand up one dim at a time
+  template<class vobj>
+  inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    GridCartesian *new_grid = grids[dim];//These are new grids
+    Lattice<vobj>  padded(new_grid);
+    Lattice<vobj> shifted(old_grid);    
+    Coordinate local     =old_grid->LocalDimensions();
+    Coordinate plocal    =new_grid->LocalDimensions();
+    if(dim==0) conformable(old_grid,unpadded_grid);
+    else       conformable(old_grid,grids[dim-1]);
+
+    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
+
+    double tins=0, tshift=0;
+    
+    // Middle bit
+    double t = usecond();
+    for(int x=0;x<local[dim];x++){
+      InsertSliceLocal(in,padded,x,depth+x,dim);
+    }
+    tins += usecond() - t;
+    
+    // High bit
+    t = usecond();
+    shifted = cshift.Cshift(in,dim,depth);
+    tshift += usecond() - t;
+
+    t=usecond();
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
+    }
+    tins += usecond() - t;
+    
+    // Low bit
+    t = usecond();
+    shifted = cshift.Cshift(in,dim,-depth);
+    tshift += usecond() - t;
+    
+    t = usecond();
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,x,x,dim);
+    }
+    tins += usecond() - t;
+
+    std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
+    
+    return padded;
+  }
+
+};
+ 
+
+NAMESPACE_END(Grid);
+

Grid/log/Log.cc

@@ -65,29 +65,40 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
+GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
+GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(0);
+  GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
+  GridLogMemory.Active(0); 
+  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
+  GridLogDslash.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
+  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
+    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
+    if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,7 +138,8 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      stream << log.colour() <<  std::left;
+      //      stream << log.colour() <<  std::left;
+      stream <<  std::left;
       if (log.chanWidth > 0)
       {
         stream << std::setw(log.chanWidth);
@@ -153,9 +154,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      stream << log.colour();
+      //      stream << log.colour();
+      stream <<  std::right;
       stream.flags(f);
-
       return stream;
     } else { 
       return devnull;
@@ -178,14 +179,53 @@ extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
-extern GridLogger GridLogDebug  ;
+extern GridLogger GridLogDebug;
 extern GridLogger GridLogPerformance;
-extern GridLogger GridLogIterative  ;
-extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogDslash;
+extern GridLogger GridLogIterative;
+extern GridLogger GridLogIntegrator;
+extern GridLogger GridLogHMC;
+extern GridLogger GridLogMemory;
+extern GridLogger GridLogTracing;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;
 
+template<typename... Args>
+inline std::string sjoin(Args&&... args) noexcept {
+    std::ostringstream msg;
+    (msg << ... << args);
+    return msg.str();
+}
+
+/*!  @brief make log messages work like python print */
+template <typename... Args>
+inline void Grid_log(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << GridLogMessage << msg << std::endl;
+}
+
+/*!  @brief make warning messages work like python print */
+template <typename... Args>
+inline void Grid_warn(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[33m" << GridLogWarning << msg << "\033[0m" << std::endl;
+}
+
+/*!  @brief make error messages work like python print */
+template <typename... Args>
+inline void Grid_error(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[31m" << GridLogError << msg << "\033[0m" << std::endl;
+}
+
+/*!  @brief make pass messages work like python print */
+template <typename... Args>
+inline void Grid_pass(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[32m" << GridLogMessage << msg << "\033[0m" << std::endl;
+}
+
 #define _NBACKTRACE (256)
 extern void * Grid_backtrace_buffer[_NBACKTRACE];
 

Grid/parallelIO/IldgIO.h

@@ -31,6 +31,7 @@ directory
 #include <fstream>
 #include <iomanip>
 #include <iostream>
+#include <string>
 #include <map>
 
 #include <pwd.h>
@@ -576,6 +577,8 @@ class ScidacReader : public GridLimeReader {
     std::string rec_name(ILDG_BINARY_DATA);
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
       if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+  // in principle should do the line below, but that breaks backard compatibility with old data
+  // skipPastObjectRecord(std::string(GRID_FIELD_NORM));
 	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
 	return;
       }
@@ -652,7 +655,8 @@ class IldgWriter : public ScidacWriter {
     // Fill ILDG header data struct
     //////////////////////////////////////////////////////
     ildgFormat ildgfmt ;
-    ildgfmt.field     = std::string("su3gauge");
+    const std::string stNC = std::to_string( Nc ) ;
+    ildgfmt.field          = std::string("su"+stNC+"gauge");
 
     if ( format == std::string("IEEE32BIG") ) { 
       ildgfmt.precision = 32;
@@ -869,7 +873,8 @@ class IldgReader : public GridLimeReader {
     } else { 
 
       assert(found_ildgFormat);
-      assert ( ildgFormat_.field == std::string("su3gauge") );
+      const std::string stNC = std::to_string( Nc ) ;
+      assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
 
       ///////////////////////////////////////////////////////////////////////////////////////
       // Populate our Grid metadata as best we can
@@ -877,7 +882,7 @@ class IldgReader : public GridLimeReader {
 
       std::ostringstream vers; vers << ildgFormat_.version;
       FieldMetaData_.hdr_version = vers.str();
-      FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
+      FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
 
       FieldMetaData_.nd=4;
       FieldMetaData_.dimension.resize(4);

Grid/parallelIO/MetaData.h

@@ -6,8 +6,8 @@
 
     Copyright (C) 2015
 
-
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -182,8 +182,8 @@ class GaugeStatistics
 public:
   void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
   {
-    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
-    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
+    header.link_trace = WilsonLoops<Impl>::linkTrace(data);
+    header.plaquette  = WilsonLoops<Impl>::avgPlaquette(data);
   }
 };
 typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
@@ -203,20 +203,24 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
 //////////////////////////////////////////////////////////////////////
 inline void reconstruct3(LorentzColourMatrix & cm)
 {
-  const int x=0;
-  const int y=1;
-  const int z=2;
+  assert( Nc < 4 && Nc > 1 ) ;
   for(int mu=0;mu<Nd;mu++){
-    cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
-    cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
-    cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
+    #if Nc == 2
+      cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ;
+      cm(mu)()(1,1) =  adj(cm(mu)()(0,x)) ;
+    #else
+      const int x=0 , y=1 , z=2 ; // a little disinenuous labelling
+      cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
+      cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
+      cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
+    #endif
   }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Some data types for intermediate storage
 ////////////////////////////////////////////////////////////////////////////////
-template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
+template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, Nc-1>, Nd >;
 
 typedef iLorentzColour2x3<Complex>  LorentzColour2x3;
 typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
@@ -278,7 +282,6 @@ struct GaugeSimpleMunger{
 
 template <class fobj, class sobj>
 struct GaugeSimpleUnmunger {
-
   void operator()(sobj &in, fobj &out) {
     for (int mu = 0; mu < Nd; mu++) {
       for (int i = 0; i < Nc; i++) {
@@ -317,8 +320,8 @@ template<class fobj,class sobj>
 struct Gauge3x2munger{
   void operator() (fobj &in,sobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<2;i++){
-	for(int j=0;j<3;j++){
+      for(int i=0;i<Nc-1;i++){
+	for(int j=0;j<Nc;j++){
 	  out(mu)()(i,j) = in(mu)(i)(j);
 	}}
     }
@@ -330,8 +333,8 @@ template<class fobj,class sobj>
 struct Gauge3x2unmunger{
   void operator() (sobj &in,fobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<2;i++){
-	for(int j=0;j<3;j++){
+      for(int i=0;i<Nc-1;i++){
+	for(int j=0;j<Nc;j++){
 	  out(mu)(i)(j) = in(mu)()(i,j);
 	}}
     }

Grid/parallelIO/NerscIO.h

@@ -9,6 +9,7 @@
     Author: Matt Spraggs <matthew.spraggs@gmail.com>
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -30,6 +31,8 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
+#include <string>
+
 NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
@@ -39,9 +42,11 @@ using namespace Grid;
 ////////////////////////////////////////////////////////////////////////////////
 class NerscIO : public BinaryIO { 
 public:
-
   typedef Lattice<vLorentzColourMatrixD> GaugeField;
 
+  // Enable/disable exiting if the plaquette in the header does not match the value computed (default true)
+  static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; }
+
   static inline void truncate(std::string file){
     std::ofstream fout(file,std::ios::out);
   }
@@ -145,15 +150,17 @@ public:
 
     std::string format(header.floating_point);
 
-    int ieee32big = (format == std::string("IEEE32BIG"));
-    int ieee32    = (format == std::string("IEEE32"));
-    int ieee64big = (format == std::string("IEEE64BIG"));
-    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
+    const int ieee32big = (format == std::string("IEEE32BIG"));
+    const int ieee32    = (format == std::string("IEEE32"));
+    const int ieee64big = (format == std::string("IEEE64BIG"));
+    const 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") ) {
+    const std::string stNC = std::to_string( Nc ) ;
+    if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@@ -164,7 +171,7 @@ public:
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
-    } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
+    } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@@ -198,7 +205,7 @@ public:
       std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
       exit(0);
     }
-    assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
+    if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
     assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
     assert(nersc_csum == header.checksum );
       
@@ -209,27 +216,29 @@ public:
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
-					std::string ens_label = std::string("DWF"))
+					std::string ens_label = std::string("DWF"),
+					std::string ens_id = std::string("UKQCD"),
+					unsigned int sequence_number = 1)
   {
-    writeConfiguration(Umu,file,0,1,ens_label);
+    writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
   }
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32,
-					std::string ens_label = std::string("DWF"))
+					std::string ens_label = std::string("DWF"),
+					std::string ens_id = std::string("UKQCD"),
+					unsigned int sequence_number = 1)
   {
     typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
-    ///////////////////////////////////////////
-    // Following should become arguments
-    ///////////////////////////////////////////
-    header.sequence_number = 1;
-    header.ensemble_id     = std::string("UKQCD");
+    header.sequence_number = sequence_number;
+    header.ensemble_id     = ens_id;
     header.ensemble_label  = ens_label;
+    header.hdr_version     = "1.0" ;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -243,10 +252,14 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write 3x3 double
-    header.floating_point = std::string("IEEE64BIG");
-    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
-    GaugeSimpleUnmunger<fobj3D,sobj> munge;
+    // Sod it -- always write NcxNc double
+    header.floating_point  = std::string("IEEE64BIG");
+    const std::string stNC = std::to_string( Nc ) ;
+    if( two_row ) {
+      header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
+    } else {
+      header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
+    }
     if ( grid->IsBoss() ) { 
       truncate(file);
       offset = writeHeader(header,file);
@@ -254,8 +267,15 @@ public:
     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);
+    if( two_row ) {
+      Gauge3x2unmunger<fobj2D,sobj> munge;
+      BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
+						nersc_csum,scidac_csuma,scidac_csumb);
+    } else {
+      GaugeSimpleUnmunger<fobj3D,sobj> munge;
+      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);
@@ -287,8 +307,7 @@ public:
     header.plaquette=0.0;
     MachineCharacteristics(header);
 
-	uint64_t offset;
-  
+    uint64_t offset;
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");
@@ -328,7 +347,7 @@ public:
 
     GridBase *grid = parallel.Grid();
 
-	uint64_t offset = readHeader(file,grid,header);
+    uint64_t offset = readHeader(file,grid,header);
 
     FieldMetaData clone(header);
 

Grid/perfmon/PerfCount.cc

@@ -27,10 +27,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
-#include <Grid/perfmon/PerfCount.h>
 
+#include <Grid/perfmon/Timer.h>
+#include <Grid/perfmon/PerfCount.h>
 NAMESPACE_BEGIN(Grid);
 
+GridTimePoint theProgramStart = GridClock::now();
+
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
 #define RawConfig(A,B) (A<<8|B)
 const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::PerformanceCounterConfigs [] = {

Grid/perfmon/PerfCount.h

@@ -30,6 +30,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
+
+#ifndef __SSC_START
+#define __SSC_START
+#define __SSC_STOP
+#endif
+
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>
@@ -72,17 +78,9 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
-#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
-#define __SSC_STOP  __SSC_MARK(0x110)
-#define __SSC_START __SSC_MARK(0x111)
-
 
 #else
 
-#define __SSC_MARK(mark) 
-#define __SSC_STOP  
-#define __SSC_START 
-
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

@@ -35,17 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid)
 
-// Dress the output; use std::chrono
-// C++11 time facilities better?
-inline double usecond(void) {
-  struct timeval tv;
-#ifdef TIMERS_ON
-  gettimeofday(&tv,NULL);
-#endif
-  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
-}
-
-typedef  std::chrono::system_clock          GridClock;
+//typedef  std::chrono::system_clock          GridClock;
+typedef  std::chrono::high_resolution_clock   GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
 
 typedef  std::chrono::seconds               GridSecs;
@@ -53,6 +44,15 @@ typedef  std::chrono::milliseconds          GridMillisecs;
 typedef  std::chrono::microseconds          GridUsecs;
 typedef  std::chrono::microseconds          GridTime;
 
+extern GridTimePoint theProgramStart;
+// Dress the output; use std::chrono
+// C++11 time facilities better?
+inline double usecond(void) {
+  auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart); 
+  return 1.0*usecs.count();
+}
+
+
 inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
   stream << time.count()<<" s";

Grid/perfmon/Tracing.h

@@ -0,0 +1,70 @@
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+#ifdef GRID_TRACING_NVTX
+#include <nvToolsExt.h>
+class GridTracer {
+public:
+  GridTracer(const char* name) {
+    nvtxRangePushA(name);
+  }
+  ~GridTracer() {
+    nvtxRangePop();
+  }
+};
+inline void tracePush(const char *name) { nvtxRangePushA(name); }
+inline void tracePop(const char *name) { nvtxRangePop(); }
+inline int  traceStart(const char *name) {  }
+inline void traceStop(int ID) {  }
+#endif
+
+#ifdef GRID_TRACING_ROCTX
+#include <roctracer/roctx.h>
+class GridTracer {
+ public:
+  GridTracer(const char* name) {
+    roctxRangePushA(name);
+    std::cout << "roctxRangePush "<<name<<std::endl;
+  }
+  ~GridTracer() {
+    roctxRangePop();
+    std::cout << "roctxRangePop "<<std::endl;
+  }
+};
+inline void tracePush(const char *name) { roctxRangePushA(name); }
+inline void tracePop(const char *name) { roctxRangePop(); }
+inline int  traceStart(const char *name) { return roctxRangeStart(name); }
+inline void traceStop(int ID) { roctxRangeStop(ID); }
+#endif
+
+#ifdef GRID_TRACING_TIMER
+class GridTracer {
+ public:
+  const char *name;
+  double elapsed;
+  GridTracer(const char* _name) {
+    name = _name;
+    elapsed=-usecond();
+  }
+  ~GridTracer() {
+    elapsed+=usecond();
+    std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl;
+  }
+};
+inline void tracePush(const char *name) {  }
+inline void tracePop(const char *name) {  }
+inline int  traceStart(const char *name) { return 0; }
+inline void traceStop(int ID) {  }
+#endif
+
+#ifdef GRID_TRACING_NONE
+#define GRID_TRACE(name) 
+inline void tracePush(const char *name) {  }
+inline void tracePop(const char *name) {  }
+inline int  traceStart(const char *name) { return 0;  }
+inline void traceStop(int ID) {  }
+#else
+#define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
+#endif
+NAMESPACE_END(Grid);

Grid/pugixml/pugixml.cc

@@ -16,8 +16,12 @@
 
 #ifdef __NVCC__
 #pragma push
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
+#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
+#else
 #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
 #endif
+#endif
 
 #include "pugixml.h"
 

Grid/qcd/QCD.h

@@ -63,6 +63,7 @@ static constexpr int Ngp=2; // gparity index range
 #define ColourIndex  (2)
 #define SpinIndex    (1)
 #define LorentzIndex (0)
+#define GparityFlavourIndex (0)
 
 // Also should make these a named enum type
 static constexpr int DaggerNo=0;
@@ -87,6 +88,8 @@ template<typename T> struct isCoarsened {
 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> > ;
 
+const int GparityFlavourTensorIndex = 3; //TensorLevel counts from the bottom!
+
 // ChrisK very keen to add extra space for Gparity doubling.
 //
 // Also add domain wall index, in a way where Wilson operator 
@@ -101,6 +104,7 @@ template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iSca
 template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
 template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
 template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
+template<typename vtype> using iLorentzComplex            = iVector<iScalar<iScalar<vtype> >, Nd > ;
 template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
 template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >;
 template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >;
@@ -110,8 +114,10 @@ template<typename vtype> using iHalfSpinColourVector      = iScalar<iVector<iVec
     template<typename vtype> using iSpinColourSpinColourMatrix  = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
 
 
+template<typename vtype> using iGparityFlavourVector                = iVector<iScalar<iScalar<vtype> >, Ngp>;
 template<typename vtype> using iGparitySpinColourVector       = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
 template<typename vtype> using iGparityHalfSpinColourVector   = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
+template<typename vtype> using iGparityFlavourMatrix = iMatrix<iScalar<iScalar<vtype> >, Ngp>;
 
 // Spin matrix
 typedef iSpinMatrix<Complex  >          SpinMatrix;
@@ -121,6 +127,7 @@ typedef iSpinMatrix<ComplexD >          SpinMatrixD;
 typedef iSpinMatrix<vComplex >          vSpinMatrix;
 typedef iSpinMatrix<vComplexF>          vSpinMatrixF;
 typedef iSpinMatrix<vComplexD>          vSpinMatrixD;
+typedef iSpinMatrix<vComplexD2>         vSpinMatrixD2;
 
 // Colour Matrix
 typedef iColourMatrix<Complex  >        ColourMatrix;
@@ -130,6 +137,7 @@ typedef iColourMatrix<ComplexD >        ColourMatrixD;
 typedef iColourMatrix<vComplex >        vColourMatrix;
 typedef iColourMatrix<vComplexF>        vColourMatrixF;
 typedef iColourMatrix<vComplexD>        vColourMatrixD;
+typedef iColourMatrix<vComplexD2>       vColourMatrixD2;
 
 // SpinColour matrix
 typedef iSpinColourMatrix<Complex  >    SpinColourMatrix;
@@ -139,6 +147,7 @@ typedef iSpinColourMatrix<ComplexD >    SpinColourMatrixD;
 typedef iSpinColourMatrix<vComplex >    vSpinColourMatrix;
 typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF;
 typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD;
+typedef iSpinColourMatrix<vComplexD2>   vSpinColourMatrixD2;
 
 // SpinColourSpinColour matrix
 typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
@@ -148,6 +157,7 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2;
 
 // SpinColourSpinColour matrix
 typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
@@ -157,24 +167,47 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2;
 
 // LorentzColour
 typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix;
 typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
 typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
 
-typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix;
-typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF;
-typedef iLorentzColourMatrix<vComplexD> vLorentzColourMatrixD;
+typedef iLorentzColourMatrix<vComplex >  vLorentzColourMatrix;
+typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF;
+typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD;
+typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2;
+
+// LorentzComplex
+typedef iLorentzComplex<Complex  > LorentzComplex;
+typedef iLorentzComplex<ComplexF > LorentzComplexF;
+typedef iLorentzComplex<ComplexD > LorentzComplexD;
+
+typedef iLorentzComplex<vComplex > vLorentzComplex;
+typedef iLorentzComplex<vComplexF> vLorentzComplexF;
+typedef iLorentzComplex<vComplexD> vLorentzComplexD;
 
 // DoubleStored gauge field
 typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
 typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD;
 
-typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
-typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
-typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
+typedef iDoubleStoredColourMatrix<vComplex >  vDoubleStoredColourMatrix;
+typedef iDoubleStoredColourMatrix<vComplexF>  vDoubleStoredColourMatrixF;
+typedef iDoubleStoredColourMatrix<vComplexD>  vDoubleStoredColourMatrixD;
+typedef iDoubleStoredColourMatrix<vComplexD2> vDoubleStoredColourMatrixD2;
+
+//G-parity flavour matrix
+typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
+typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
+
+typedef iGparityFlavourMatrix<vComplex>   vGparityFlavourMatrix;
+typedef iGparityFlavourMatrix<vComplexF>  vGparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<vComplexD>  vGparityFlavourMatrixD;
+typedef iGparityFlavourMatrix<vComplexD2> vGparityFlavourMatrixD2;
+
 
 // Spin vector
 typedef iSpinVector<Complex >           SpinVector;
@@ -184,6 +217,7 @@ typedef iSpinVector<ComplexD>           SpinVectorD;
 typedef iSpinVector<vComplex >           vSpinVector;
 typedef iSpinVector<vComplexF>           vSpinVectorF;
 typedef iSpinVector<vComplexD>           vSpinVectorD;
+typedef iSpinVector<vComplexD2>          vSpinVectorD2;
 
 // Colour vector
 typedef iColourVector<Complex >         ColourVector;
@@ -193,6 +227,7 @@ typedef iColourVector<ComplexD>         ColourVectorD;
 typedef iColourVector<vComplex >         vColourVector;
 typedef iColourVector<vComplexF>         vColourVectorF;
 typedef iColourVector<vComplexD>         vColourVectorD;
+typedef iColourVector<vComplexD2>        vColourVectorD2;
 
 // SpinColourVector
 typedef iSpinColourVector<Complex >     SpinColourVector;
@@ -202,6 +237,7 @@ typedef iSpinColourVector<ComplexD>     SpinColourVectorD;
 typedef iSpinColourVector<vComplex >     vSpinColourVector;
 typedef iSpinColourVector<vComplexF>     vSpinColourVectorF;
 typedef iSpinColourVector<vComplexD>     vSpinColourVectorD;
+typedef iSpinColourVector<vComplexD2>    vSpinColourVectorD2;
 
 // HalfSpin vector
 typedef iHalfSpinVector<Complex >       HalfSpinVector;
@@ -211,15 +247,27 @@ typedef iHalfSpinVector<ComplexD>       HalfSpinVectorD;
 typedef iHalfSpinVector<vComplex >       vHalfSpinVector;
 typedef iHalfSpinVector<vComplexF>       vHalfSpinVectorF;
 typedef iHalfSpinVector<vComplexD>       vHalfSpinVectorD;
+typedef iHalfSpinVector<vComplexD2>      vHalfSpinVectorD2;
 
 // HalfSpinColour vector
 typedef iHalfSpinColourVector<Complex > HalfSpinColourVector;
 typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF;
 typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
     
-typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
-typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
-typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
+typedef iHalfSpinColourVector<vComplex >  vHalfSpinColourVector;
+typedef iHalfSpinColourVector<vComplexF>  vHalfSpinColourVectorF;
+typedef iHalfSpinColourVector<vComplexD>  vHalfSpinColourVectorD;
+typedef iHalfSpinColourVector<vComplexD2> vHalfSpinColourVectorD2;
+
+//G-parity flavour vector
+typedef iGparityFlavourVector<Complex >         GparityFlavourVector;
+typedef iGparityFlavourVector<ComplexF>         GparityFlavourVectorF;
+typedef iGparityFlavourVector<ComplexD>         GparityFlavourVectorD;
+
+typedef iGparityFlavourVector<vComplex >         vGparityFlavourVector;
+typedef iGparityFlavourVector<vComplexF>         vGparityFlavourVectorF;
+typedef iGparityFlavourVector<vComplexD>         vGparityFlavourVectorD;
+typedef iGparityFlavourVector<vComplexD2>        vGparityFlavourVectorD2;
     
 // singlets
 typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type.
@@ -229,6 +277,7 @@ typedef iSinglet<ComplexD>         TComplexD;    // FIXME This is painful. Tenso
 typedef iSinglet<vComplex >        vTComplex ;   // what if we don't know the tensor structure
 typedef iSinglet<vComplexF>        vTComplexF;   // what if we don't know the tensor structure
 typedef iSinglet<vComplexD>        vTComplexD;   // what if we don't know the tensor structure
+typedef iSinglet<vComplexD2>       vTComplexD2;   // what if we don't know the tensor structure
 
 typedef iSinglet<Real >            TReal;        // Shouldn't need these; can I make it work without?
 typedef iSinglet<RealF>            TRealF;       // Shouldn't need these; can I make it work without?
@@ -246,47 +295,62 @@ typedef iSinglet<Integer >         TInteger;
 typedef Lattice<vColourMatrix>          LatticeColourMatrix;
 typedef Lattice<vColourMatrixF>         LatticeColourMatrixF;
 typedef Lattice<vColourMatrixD>         LatticeColourMatrixD;
+typedef Lattice<vColourMatrixD2>        LatticeColourMatrixD2;
 
 typedef Lattice<vSpinMatrix>            LatticeSpinMatrix;
 typedef Lattice<vSpinMatrixF>           LatticeSpinMatrixF;
 typedef Lattice<vSpinMatrixD>           LatticeSpinMatrixD;
+typedef Lattice<vSpinMatrixD2>          LatticeSpinMatrixD2;
 
 typedef Lattice<vSpinColourMatrix>      LatticeSpinColourMatrix;
 typedef Lattice<vSpinColourMatrixF>     LatticeSpinColourMatrixF;
 typedef Lattice<vSpinColourMatrixD>     LatticeSpinColourMatrixD;
+typedef Lattice<vSpinColourMatrixD2>    LatticeSpinColourMatrixD2;
 
 typedef Lattice<vSpinColourSpinColourMatrix>      LatticeSpinColourSpinColourMatrix;
 typedef Lattice<vSpinColourSpinColourMatrixF>     LatticeSpinColourSpinColourMatrixF;
 typedef Lattice<vSpinColourSpinColourMatrixD>     LatticeSpinColourSpinColourMatrixD;
+typedef Lattice<vSpinColourSpinColourMatrixD2>    LatticeSpinColourSpinColourMatrixD2;
 
-typedef Lattice<vLorentzColourMatrix>  LatticeLorentzColourMatrix;
-typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
-typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD;
+typedef Lattice<vLorentzColourMatrix>   LatticeLorentzColourMatrix;
+typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF;
+typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD;
+typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2;
+
+typedef Lattice<vLorentzComplex>  LatticeLorentzComplex;
+typedef Lattice<vLorentzComplexF> LatticeLorentzComplexF;
+typedef Lattice<vLorentzComplexD> LatticeLorentzComplexD;
 
 // DoubleStored gauge field
-typedef Lattice<vDoubleStoredColourMatrix>  LatticeDoubleStoredColourMatrix;
-typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;
-typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD;
+typedef Lattice<vDoubleStoredColourMatrix>   LatticeDoubleStoredColourMatrix;
+typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF;
+typedef Lattice<vDoubleStoredColourMatrixD>  LatticeDoubleStoredColourMatrixD;
+typedef Lattice<vDoubleStoredColourMatrixD2> LatticeDoubleStoredColourMatrixD2;
 
 typedef Lattice<vSpinVector>            LatticeSpinVector;
 typedef Lattice<vSpinVectorF>           LatticeSpinVectorF;
 typedef Lattice<vSpinVectorD>           LatticeSpinVectorD;
+typedef Lattice<vSpinVectorD2>          LatticeSpinVectorD2;
 
 typedef Lattice<vColourVector>          LatticeColourVector;
 typedef Lattice<vColourVectorF>         LatticeColourVectorF;
 typedef Lattice<vColourVectorD>         LatticeColourVectorD;
+typedef Lattice<vColourVectorD2>        LatticeColourVectorD2;
 
 typedef Lattice<vSpinColourVector>      LatticeSpinColourVector;
 typedef Lattice<vSpinColourVectorF>     LatticeSpinColourVectorF;
 typedef Lattice<vSpinColourVectorD>     LatticeSpinColourVectorD;
+typedef Lattice<vSpinColourVectorD2>    LatticeSpinColourVectorD2;
 
 typedef Lattice<vHalfSpinVector>        LatticeHalfSpinVector;
 typedef Lattice<vHalfSpinVectorF>       LatticeHalfSpinVectorF;
 typedef Lattice<vHalfSpinVectorD>       LatticeHalfSpinVectorD;
+typedef Lattice<vHalfSpinVectorD2>      LatticeHalfSpinVectorD2;
 
-typedef Lattice<vHalfSpinColourVector>  LatticeHalfSpinColourVector;
-typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF;
-typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD;
+typedef Lattice<vHalfSpinColourVector>   LatticeHalfSpinColourVector;
+typedef Lattice<vHalfSpinColourVectorF>  LatticeHalfSpinColourVectorF;
+typedef Lattice<vHalfSpinColourVectorD>  LatticeHalfSpinColourVectorD;
+typedef Lattice<vHalfSpinColourVectorD2> LatticeHalfSpinColourVectorD2;
 
 typedef Lattice<vTReal>            LatticeReal;
 typedef Lattice<vTRealF>           LatticeRealF;
@@ -295,6 +359,7 @@ typedef Lattice<vTRealD>           LatticeRealD;
 typedef Lattice<vTComplex>         LatticeComplex;
 typedef Lattice<vTComplexF>        LatticeComplexF;
 typedef Lattice<vTComplexD>        LatticeComplexD;
+typedef Lattice<vTComplexD2>       LatticeComplexD2;
 
 typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where"
 
@@ -302,37 +367,42 @@ typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where"
 ///////////////////////////////////////////
 // Physical names for things
 ///////////////////////////////////////////
-typedef LatticeHalfSpinColourVector  LatticeHalfFermion;
-typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF;
-typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD;
+typedef LatticeHalfSpinColourVector   LatticeHalfFermion;
+typedef LatticeHalfSpinColourVectorF  LatticeHalfFermionF;
+typedef LatticeHalfSpinColourVectorD  LatticeHalfFermionD;
+typedef LatticeHalfSpinColourVectorD2 LatticeHalfFermionD2;
 
 typedef LatticeSpinColourVector      LatticeFermion;
 typedef LatticeSpinColourVectorF     LatticeFermionF;
 typedef LatticeSpinColourVectorD     LatticeFermionD;
+typedef LatticeSpinColourVectorD2    LatticeFermionD2;
 
 typedef LatticeSpinColourMatrix                LatticePropagator;
 typedef LatticeSpinColourMatrixF               LatticePropagatorF;
 typedef LatticeSpinColourMatrixD               LatticePropagatorD;
+typedef LatticeSpinColourMatrixD2              LatticePropagatorD2;
 
 typedef LatticeLorentzColourMatrix             LatticeGaugeField;
 typedef LatticeLorentzColourMatrixF            LatticeGaugeFieldF;
 typedef LatticeLorentzColourMatrixD            LatticeGaugeFieldD;
+typedef LatticeLorentzColourMatrixD2           LatticeGaugeFieldD2;
 
 typedef LatticeDoubleStoredColourMatrix        LatticeDoubledGaugeField;
 typedef LatticeDoubleStoredColourMatrixF       LatticeDoubledGaugeFieldF;
 typedef LatticeDoubleStoredColourMatrixD       LatticeDoubledGaugeFieldD;
+typedef LatticeDoubleStoredColourMatrixD2      LatticeDoubledGaugeFieldD2;
 
 template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >;
 
-// Uhgg... typing this hurt  ;)
-// (my keyboard got burning hot when I typed this, must be the anti-Fermion)
 typedef Lattice<vColourVector>          LatticeStaggeredFermion;    
 typedef Lattice<vColourVectorF>         LatticeStaggeredFermionF;    
 typedef Lattice<vColourVectorD>         LatticeStaggeredFermionD;    
+typedef Lattice<vColourVectorD2>        LatticeStaggeredFermionD2;    
 
 typedef Lattice<vColourMatrix>          LatticeStaggeredPropagator; 
 typedef Lattice<vColourMatrixF>         LatticeStaggeredPropagatorF; 
 typedef Lattice<vColourMatrixD>         LatticeStaggeredPropagatorD; 
+typedef Lattice<vColourMatrixD2>        LatticeStaggeredPropagatorD2; 
 
 //////////////////////////////////////////////////////////////////////////////
 // Peek and Poke named after physics attributes
@@ -451,9 +521,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
+#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuWrite);
+  autoView(f_v,f,CpuRead);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -465,12 +546,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
+#endif
 }
     
 //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)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuRead);
+  autoView(f_v,f,CpuWrite);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -482,6 +574,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/ActionBase.h

@@ -34,21 +34,117 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+///////////////////////////////////
+// Smart configuration base class
+///////////////////////////////////
+template< class Field >
+class ConfigurationBase
+{
+public:
+  ConfigurationBase() {}
+  virtual ~ConfigurationBase() {}
+  virtual void set_Field(Field& U) =0;
+  virtual void smeared_force(Field&) = 0;
+  virtual Field& get_SmearedU() =0;
+  virtual Field &get_U(bool smeared = false) = 0;
+};
+
 template <class GaugeField >
 class Action 
 {
-
 public:
   bool is_smeared = false;
+  RealD deriv_norm_sum;
+  RealD deriv_max_sum;
+  RealD Fdt_norm_sum;
+  RealD Fdt_max_sum;
+  int   deriv_num;
+  RealD deriv_us;
+  RealD S_us;
+  RealD refresh_us;
+  void  reset_timer(void)        {
+    deriv_us = S_us = refresh_us = 0.0;
+    deriv_norm_sum = deriv_max_sum=0.0;
+    Fdt_max_sum =  Fdt_norm_sum = 0.0;
+    deriv_num=0;
+  }
+  void  deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) {
+    if ( max > deriv_max_sum ) {
+      deriv_max_sum=max;
+    }
+    deriv_norm_sum+=nrm;
+    if ( Fdt_max > Fdt_max_sum ) {
+      Fdt_max_sum=Fdt_max;
+    }
+    Fdt_norm_sum+=Fdt_nrm; deriv_num++;
+  }
+  RealD deriv_max_average(void)       { return deriv_max_sum; };
+  RealD deriv_norm_average(void)      { return deriv_norm_sum/deriv_num; };
+  RealD Fdt_max_average(void)         { return Fdt_max_sum; };
+  RealD Fdt_norm_average(void)        { return Fdt_norm_sum/deriv_num; };
+  RealD deriv_timer(void)        { return deriv_us; };
+  RealD S_timer(void)            { return S_us; };
+  RealD refresh_timer(void)      { return refresh_us; };
+  void deriv_timer_start(void)   { deriv_us-=usecond(); }
+  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
+  void refresh_timer_start(void) { refresh_us-=usecond(); }
+  void refresh_timer_stop(void)  { refresh_us+=usecond(); }
+  void S_timer_start(void)       { S_us-=usecond(); }
+  void S_timer_stop(void)        { S_us+=usecond(); }
+  /////////////////////////////
   // Heatbath?
+  /////////////////////////////
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
+  virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
+
+  /////////////////////////////////////////////////////////////
+  // virtual smeared interface through configuration container
+  /////////////////////////////////////////////////////////////
+  virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
+  {
+    refresh(U.get_U(is_smeared),sRNG,pRNG);
+  }
+  virtual RealD S(ConfigurationBase<GaugeField>& U)
+  {
+    return S(U.get_U(is_smeared));
+  }
+  virtual RealD Sinitial(ConfigurationBase<GaugeField>& U) 
+  {
+    return Sinitial(U.get_U(is_smeared));
+  }
+  virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU)
+  {
+    deriv(U.get_U(is_smeared),dSdU); 
+    if ( is_smeared ) {
+      U.smeared_force(dSdU);
+    }
+  }
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
   virtual std::string action_name()    = 0;                             // return the action name
   virtual std::string LogParameters()  = 0;                             // prints action parameters
   virtual ~Action(){}
 };
 
+template <class GaugeField >
+class EmptyAction : public Action <GaugeField>
+{
+  virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) { assert(0);}; // refresh pseudofermions
+  virtual RealD S(const GaugeField& U) { return 0.0;};                             // evaluate the action
+  virtual void deriv(const GaugeField& U, GaugeField& dSdU) { assert(0); };        // evaluate the action derivative
+
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
+  virtual std::string action_name()    { return std::string("Level Force Log"); };
+  virtual std::string LogParameters()  { return std::string("No parameters");};
+};
+
+
+
 NAMESPACE_END(Grid);
 
 #endif // ACTION_BASE_H

Grid/qcd/action/ActionCore.h

@@ -30,6 +30,8 @@ directory
 #ifndef QCD_ACTION_CORE
 #define QCD_ACTION_CORE
 
+#include <Grid/qcd/action/gauge/GaugeImplementations.h>
+
 #include <Grid/qcd/action/ActionBase.h>
 NAMESPACE_CHECK(ActionBase);
 #include <Grid/qcd/action/ActionSet.h>
@@ -37,6 +39,10 @@ NAMESPACE_CHECK(ActionSet);
 #include <Grid/qcd/action/ActionParams.h>
 NAMESPACE_CHECK(ActionParams);
 
+#include <Grid/qcd/action/filters/MomentumFilter.h>
+#include <Grid/qcd/action/filters/DirichletFilter.h>
+#include <Grid/qcd/action/filters/DDHMCFilter.h>
+
 ////////////////////////////////////////////
 // Gauge Actions
 ////////////////////////////////////////////

Grid/qcd/action/ActionParams.h

@@ -34,27 +34,45 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
-// These can move into a params header and be given MacroMagic serialisation
+
 struct GparityWilsonImplParams {
   Coordinate twists;
-  GparityWilsonImplParams() : twists(Nd, 0) {};
+                     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  int  partialDirichlet;
+  GparityWilsonImplParams() : twists(Nd, 0) {
+    dirichlet.resize(0);
+    partialDirichlet=0;
+  };
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  int  partialDirichlet;
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
+    dirichlet.resize(0);
+    partialDirichlet=0;
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
+    partialDirichlet=0;
+    dirichlet.resize(0);
   }
 };
 
 struct StaggeredImplParams {
-  StaggeredImplParams()  {};
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  int  partialDirichlet;
+  StaggeredImplParams()
+  {
+    partialDirichlet=0;
+    dirichlet.resize(0);
+  };
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -63,9 +81,11 @@ struct StaggeredImplParams {
 				    RealD, hi, 
 				    int,   MaxIter, 
 				    RealD, tolerance, 
+				    RealD, mdtolerance, 
 				    int,   degree, 
 				    int,   precision,
-				    int,   BoundsCheckFreq);
+				    int,   BoundsCheckFreq,
+				    RealD, BoundsCheckTol);
     
   // MaxIter and tolerance, vectors??
     
@@ -76,16 +96,62 @@ struct StaggeredImplParams {
 				RealD tol      = 1.0e-8, 
                            	int _degree    = 10,
 				int _precision = 64,
-				int _BoundsCheckFreq=20)
+				int _BoundsCheckFreq=20,
+				RealD mdtol    = 1.0e-6,
+				double _BoundsCheckTol=1e-6)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
+        mdtolerance(mdtol),
 	degree(_degree),
         precision(_precision),
-        BoundsCheckFreq(_BoundsCheckFreq){};
+        BoundsCheckFreq(_BoundsCheckFreq),
+        BoundsCheckTol(_BoundsCheckTol){};
   };
   
+  /*Action parameters for the generalized rational action
+    The approximation is for (M^dag M)^{1/inv_pow}
+    where inv_pow is the denominator of the fractional power.
+    Default inv_pow=2 for square root, making this equivalent to 
+    the OneFlavourRational action
+  */
+    struct RationalActionParams : Serializable {
+    GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams, 
+				    int, inv_pow, 
+				    RealD, lo, //low eigenvalue bound of rational approx
+				    RealD, hi, //high eigenvalue bound of rational approx
+				    int,   MaxIter,  //maximum iterations in msCG
+				    RealD, action_tolerance,  //msCG tolerance in action evaluation
+				    int,   action_degree, //rational approx tolerance in action evaluation
+				    RealD, md_tolerance,  //msCG tolerance in MD integration
+				    int,   md_degree, //rational approx tolerance in MD integration
+				    int,   precision, //precision of floating point arithmetic
+				    int,   BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
+  // constructor 
+  RationalActionParams(int _inv_pow = 2,
+		       RealD _lo      = 0.0, 
+		       RealD _hi      = 1.0, 
+		       int _maxit     = 1000,
+		       RealD _action_tolerance      = 1.0e-8, 
+		       int _action_degree    = 10,
+		       RealD _md_tolerance      = 1.0e-8, 
+		       int _md_degree    = 10,
+		       int _precision = 64,
+		       int _BoundsCheckFreq=20)
+    : inv_pow(_inv_pow), 
+      lo(_lo),
+      hi(_hi),
+      MaxIter(_maxit),
+      action_tolerance(_action_tolerance),
+      action_degree(_action_degree),
+      md_tolerance(_md_tolerance),
+      md_degree(_md_degree),
+      precision(_precision),
+      BoundsCheckFreq(_BoundsCheckFreq){};
+  };
+
+
 NAMESPACE_END(Grid);
 
 #endif

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -68,9 +68,17 @@ public:
   ///////////////////////////////////////////////////////////////
   // Support for MADWF tricks
   ///////////////////////////////////////////////////////////////
-  RealD Mass(void) { return mass; };
+  RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
+  RealD MassPlus(void) { return mass_plus; };
+  RealD MassMinus(void) { return mass_minus; };
+
   void  SetMass(RealD _mass) { 
-    mass=_mass; 
+    mass_plus=mass_minus=_mass; 
+    SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
+  } ;
+  void  SetMass(RealD _mass_plus, RealD _mass_minus) { 
+    mass_plus=_mass_plus;
+    mass_minus=_mass_minus;
     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
   } ;
   void  P(const FermionField &psi, FermionField &chi);
@@ -108,7 +116,7 @@ public:
   void   MeooeDag5D    (const FermionField &in, FermionField &out);
 
   //    protected:
-  RealD mass;
+  RealD mass_plus, mass_minus;
 
   // Save arguments to SetCoefficientsInternal
   Vector<Coeff_t> _gamma;
@@ -175,16 +183,6 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  RealD _mass,RealD _M5,const ImplParams &p= ImplParams());
 
-  void CayleyReport(void);
-  void CayleyZeroCounters(void);
-
-  double M5Dflops;
-  double M5Dcalls;
-  double M5Dtime;
-
-  double MooeeInvFlops;
-  double MooeeInvCalls;
-  double MooeeInvTime;
 
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);

Grid/qcd/action/fermion/CloverHelpers.h

@@ -0,0 +1,334 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
+
+    Copyright (C) 2017 - 2022
+
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Mattia Bruno <mattia.bruno@cern.ch>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#pragma once
+
+#include <Grid/Grid.h>
+#include <Grid/qcd/spin/Dirac.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+
+////////////////////////////////////////////
+// Standard Clover
+//   (4+m0) + csw * clover_term
+// Exp Clover
+//   (4+m0) * exp(csw/(4+m0) clover_term)
+//   = (4+m0) + csw * clover_term + ...
+////////////////////////////////////////////
+
+NAMESPACE_BEGIN(Grid);
+
+
+//////////////////////////////////
+// Generic Standard Clover
+//////////////////////////////////
+
+template<class Impl>
+class CloverHelpers: public WilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+
+  typedef WilsonCloverHelpers<Impl> Helpers;
+
+  static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
+    GridBase *grid = CloverTerm.Grid();
+    CloverTerm += diag_mass;
+
+    int lvol = grid->lSites();
+    int DimRep = Impl::Dimension;
+    {
+      autoView(CTv,CloverTerm,CpuRead);
+      autoView(CTIv,CloverTermInv,CpuWrite);
+      thread_for(site, lvol, {
+        Coordinate lcoor;
+        grid->LocalIndexToLocalCoor(site, lcoor);
+        Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+        Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+        typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
+        peekLocalSite(Qx, CTv, lcoor);
+
+        for (int j = 0; j < Ns; j++)
+          for (int k = 0; k < Ns; k++)
+            for (int a = 0; a < DimRep; a++)
+              for (int b = 0; b < DimRep; b++){
+                auto zz =  Qx()(j, k)(a, b);
+                EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+              }
+
+        EigenInvCloverOp = EigenCloverOp.inverse();
+        for (int j = 0; j < Ns; j++)
+          for (int k = 0; k < Ns; k++)
+            for (int a = 0; a < DimRep; a++)
+              for (int b = 0; b < DimRep; b++)
+                Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+               pokeLocalSite(Qxinv, CTIv, lcoor);
+      });
+    }
+  }
+
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    return Helpers::Cmunu(U, lambda, mu, nu);
+  }
+
+};
+
+
+//////////////////////////////////
+// Generic Exp Clover
+//////////////////////////////////
+
+template<class Impl>
+class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+
+  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+  typedef WilsonCloverHelpers<Impl> Helpers;
+
+  // Can this be avoided?
+  static void IdentityTimesC(const CloverField& in, RealD c) {
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
+  }
+
+  static int getNMAX(RealD prec, RealD R) {
+    /* compute stop condition for exponential */
+    int NMAX=1;
+    RealD cond=R*R/2.;
+
+    while (cond*std::exp(R)>prec) {
+      NMAX++;
+      cond*=R/(double)(NMAX+1);
+    }
+    return NMAX;
+  }
+
+  static int getNMAX(Lattice<iImplClover<vComplexD2>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+
+  static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
+
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
+
+    Clover *= (1.0/diag_mass);
+
+    // Taylor expansion, slow but generic
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+    // qN = cN
+    // qn = cn + qn+1 X
+    std::vector<RealD> cn(NMAX+1);
+    cn[0] = 1.0;
+    for (int i=1; i<=NMAX; i++)
+      cn[i] = cn[i-1] / RealD(i);
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
+
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
+
+    Clover = ExpClover * diag_mass;
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
+
+    CloverInv = ExpClover * (1.0/diag_mass);
+
+  }
+
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    assert(0);
+    return lambda;
+  }
+
+};
+
+
+//////////////////////////////////
+// Compact Standard Clover
+//////////////////////////////////
+
+
+template<class Impl>
+class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
+                            public WilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  typedef WilsonCloverHelpers<Impl> Helpers;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+    Clover += diag_mass;
+  }
+
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
+
+    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+  }
+
+  // TODO: implement Cmunu for better performances with compact layout, but don't do it
+  // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    return Helpers::Cmunu(U, lambda, mu, nu);
+  }
+};
+
+//////////////////////////////////
+// Compact Exp Clover
+//////////////////////////////////
+
+template<class Impl>
+class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  // Can this be avoided?
+  static void IdentityTimesC(const CloverField& in, RealD c) {
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
+  }
+
+  static int getNMAX(RealD prec, RealD R) {
+    /* compute stop condition for exponential */
+    int NMAX=1;
+    RealD cond=R*R/2.;
+
+    while (cond*std::exp(R)>prec) {
+      NMAX++;
+      cond*=R/(double)(NMAX+1);
+    }
+    return NMAX;
+  }
+
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
+
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
+
+    Clover *= (1.0/diag_mass);
+
+    // Taylor expansion, slow but generic
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+    // qN = cN
+    // qn = cn + qn+1 X
+    std::vector<RealD> cn(NMAX+1);
+    cn[0] = 1.0;
+    for (int i=1; i<=NMAX; i++)
+      cn[i] = cn[i-1] / RealD(i);
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
+
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
+
+    Clover = ExpClover * diag_mass;
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
+
+    CloverInv = ExpClover * (1.0/diag_mass);
+
+  }
+
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
+
+    if (fixedBoundaries)
+    {
+      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+    }
+    else
+    {
+      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
+    }
+  }
+
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
+    assert(0);
+    return lambda;
+  }
+
+};
+
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -0,0 +1,241 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h
+
+    Copyright (C) 2020 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Nils Meyer <nils.meyer@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 <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
+
+NAMESPACE_BEGIN(Grid);
+
+// see Grid/qcd/action/fermion/WilsonCloverFermion.h for description
+//
+// Modifications done here:
+//
+// Original: clover term = 12x12 matrix per site
+//
+// But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running)
+// Sufficient to store/transfer only the real parts of the diagonal and one triangular part
+// 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered
+//
+// Here: Above but diagonal as complex numbers, i.e., need to store/transfer
+// 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words
+//
+// Words per site and improvement compared to original (combined with the input and output spinors):
+//
+// - Original: 2*12 + 12*12 = 168 words -> 1.00 x less
+// - Minimal:  2*12 + 36    =  60 words -> 2.80 x less
+// - Here:     2*12 + 42    =  66 words -> 2.55 x less
+//
+// These improvements directly translate to wall-clock time
+//
+// Data layout:
+//
+// - diagonal and triangle part as separate lattice fields,
+//   this was faster than as 1 combined field on all tested machines
+// - diagonal: as expected
+// - triangle: store upper right triangle in row major order
+// - graphical:
+//        0  1  2  3  4
+//           5  6  7  8
+//              9 10 11 = upper right triangle indices
+//                12 13
+//                   14
+//     0
+//        1
+//           2
+//              3       = diagonal indices
+//                 4
+//                    5
+//     0
+//     1  5
+//     2  6  9          = lower left triangle indices
+//     3  7 10 12
+//     4  8 11 13 14
+//
+// Impact on total memory consumption:
+// - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site
+// - Here:     (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts =  24 complex words per site
+//           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
+//                                                                 =  84 complex words per site
+
+template<class Impl, class CloverHelpers>
+class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
+                                   public WilsonCloverHelpers<Impl>,
+                                   public CompactWilsonCloverHelpers<Impl> {
+  /////////////////////////////////////////////
+  // Sizes
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_COMPACT_CLOVER_SIZES(Impl);
+
+  /////////////////////////////////////////////
+  // Type definitions
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  typedef WilsonFermion<Impl>              WilsonBase;
+  typedef WilsonCloverHelpers<Impl>        Helpers;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  /////////////////////////////////////////////
+  // Constructors
+  /////////////////////////////////////////////
+
+public:
+
+  CompactWilsonCloverFermion(GaugeField& _Umu,
+			    GridCartesian& Fgrid,
+			    GridRedBlackCartesian& Hgrid,
+			    const RealD _mass,
+			    const RealD _csw_r = 0.0,
+			    const RealD _csw_t = 0.0,
+			    const RealD _cF = 1.0,
+			    const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
+			    const ImplParams& impl_p = ImplParams());
+
+  /////////////////////////////////////////////
+  // Member functions (implementing interface)
+  /////////////////////////////////////////////
+
+public:
+
+  virtual void Instantiatable() {};
+  int          ConstEE()     override { return 0; };
+  int          isTrivialEE() override { return 0; };
+
+  void Dhop(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopOE(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopEO(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
+
+  void M(const FermionField& in, FermionField& out) override;
+
+  void Mdag(const FermionField& in, FermionField& out) override;
+
+  void Meooe(const FermionField& in, FermionField& out) override;
+
+  void MeooeDag(const FermionField& in, FermionField& out) override;
+
+  void Mooee(const FermionField& in, FermionField& out) override;
+
+  void MooeeDag(const FermionField& in, FermionField& out) override;
+
+  void MooeeInv(const FermionField& in, FermionField& out) override;
+
+  void MooeeInvDag(const FermionField& in, FermionField& out) override;
+
+  void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
+
+  void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
+
+  void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  /////////////////////////////////////////////
+  // Member functions (internals)
+  /////////////////////////////////////////////
+
+  void MooeeInternal(const FermionField&        in,
+                     FermionField&              out,
+                     const CloverDiagonalField& diagonal,
+                     const CloverTriangleField& triangle);
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+  void ImportGauge(const GaugeField& _Umu) override;
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+private:
+
+  template<class Field>
+  const MaskField* getCorrectMaskField(const Field &in) const {
+    if(in.Grid()->_isCheckerBoarded) {
+      if(in.Checkerboard() == Odd) {
+        return &this->BoundaryMaskOdd;
+      } else {
+        return &this->BoundaryMaskEven;
+      }
+    } else {
+      return &this->BoundaryMask;
+    }
+  }
+
+  template<class Field>
+  void ApplyBoundaryMask(Field& f) {
+    const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
+    assert(m != nullptr);
+    CompactHelpers::ApplyBoundaryMask(f, *m);
+  }
+
+  /////////////////////////////////////////////
+  // Member Data
+  /////////////////////////////////////////////
+
+public:
+
+  RealD csw_r;
+  RealD csw_t;
+  RealD cF;
+
+  bool fixedBoundaries;
+
+  CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
+  CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
+
+  CloverTriangleField Triangle,    TriangleEven,    TriangleOdd;
+  CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
+
+  FermionField Tmp;
+
+  MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/DWFSlow.h

@@ -0,0 +1,291 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DWFSlow.h
+
+Copyright (C) 2022
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template <class Impl>
+class DWFSlowFermion : public FermionOperator<Impl>
+{
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+  ///////////////////////////////////////////////////////////////
+  // Implement the abstract base
+  ///////////////////////////////////////////////////////////////
+  GridBase *GaugeGrid(void) { return _grid4; }
+  GridBase *GaugeRedBlackGrid(void) { return _cbgrid4; }
+  GridBase *FermionGrid(void) { return _grid; }
+  GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
+
+  FermionField _tmp;
+  FermionField &tmp(void) { return _tmp; }
+
+  //////////////////////////////////////////////////////////////////
+  // override multiply; cut number routines if pass dagger argument
+  // and also make interface more uniformly consistent
+  //////////////////////////////////////////////////////////////////
+  virtual void  M(const FermionField &in, FermionField &out)
+  {
+    FermionField tmp(_grid);
+    out = (5.0 - M5) * in;
+    Dhop(in,tmp,DaggerNo);
+    out = out + tmp;
+  }
+  virtual void  Mdag(const FermionField &in, FermionField &out)
+  {
+    FermionField tmp(_grid);
+    out = (5.0 - M5) * in;
+    Dhop(in,tmp,DaggerYes);
+    out = out + tmp;
+  };
+
+  /////////////////////////////////////////////////////////
+  // half checkerboard operations 5D redblack so just site identiy
+  /////////////////////////////////////////////////////////
+  void Meooe(const FermionField &in, FermionField &out)
+  {
+    if ( in.Checkerboard() == Odd ) {
+      this->DhopEO(in,out,DaggerNo);
+    } else {
+      this->DhopOE(in,out,DaggerNo);
+    }
+  }
+  void MeooeDag(const FermionField &in, FermionField &out)
+  {
+    if ( in.Checkerboard() == Odd ) {
+      this->DhopEO(in,out,DaggerYes);
+    } else {
+      this->DhopOE(in,out,DaggerYes);
+    }
+  };
+
+  // allow override for twisted mass and clover
+  virtual void Mooee(const FermionField &in, FermionField &out)
+  {
+    out = (5.0 - M5) * in;
+  }
+  virtual void MooeeDag(const FermionField &in, FermionField &out)
+  {
+    out = (5.0 - M5) * in;
+  }
+  virtual void MooeeInv(const FermionField &in, FermionField &out)
+  {
+    out = (1.0/(5.0 - M5)) * in;
+  };
+  virtual void MooeeInvDag(const FermionField &in, FermionField &out)
+  {
+    out = (1.0/(5.0 - M5)) * in;
+  };
+
+  virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _mass,std::vector<double> twist) {} ;
+
+  ////////////////////////
+  // Derivative interface
+  ////////////////////////
+  // Interface calls an internal routine
+  void DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)  { assert(0);};
+  void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
+  void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
+
+  ///////////////////////////////////////////////////////////////
+  // non-hermitian hopping term; half cb or both
+  ///////////////////////////////////////////////////////////////
+  void Dhop(const FermionField &in, FermionField &out, int dag)
+  {
+    FermionField tmp(in.Grid());
+    Dhop5(in,out,MassField,MassField,dag );
+    for(int mu=0;mu<4;mu++){
+      DhopDirU(in,Umu[mu],Umu[mu],tmp,mu,dag );    out = out + tmp;
+    }
+  };
+  void DhopOE(const FermionField &in, FermionField &out, int dag)
+  {
+    FermionField tmp(in.Grid());
+    assert(in.Checkerboard()==Even);
+    Dhop5(in,out,MassFieldOdd,MassFieldEven,dag);
+    for(int mu=0;mu<4;mu++){
+      DhopDirU(in,UmuOdd[mu],UmuEven[mu],tmp,mu,dag );    out = out + tmp;
+    }
+  };
+  void DhopEO(const FermionField &in, FermionField &out, int dag)
+  {
+    FermionField tmp(in.Grid());
+    assert(in.Checkerboard()==Odd);
+    Dhop5(in,out, MassFieldEven,MassFieldOdd ,dag );  
+    for(int mu=0;mu<4;mu++){
+      DhopDirU(in,UmuEven[mu],UmuOdd[mu],tmp,mu,dag );    out = out + tmp;
+    }
+  };
+
+  ///////////////////////////////////////////////////////////////
+  // Multigrid assistance; force term uses too
+  ///////////////////////////////////////////////////////////////
+  void Mdir(const FermionField &in, FermionField &out, int dir, int disp){ assert(0);};
+  void MdirAll(const FermionField &in, std::vector<FermionField> &out)   { assert(0);};
+  void DhopDir(const FermionField &in, FermionField &out, int dir, int disp) { assert(0);};
+  void DhopDirAll(const FermionField &in, std::vector<FermionField> &out)    { assert(0);};
+  void DhopDirCalc(const FermionField &in, FermionField &out, int dirdisp,int gamma, int dag) { assert(0);};
+
+  void DhopDirU(const FermionField &in, const GaugeLinkField &U5e, const GaugeLinkField &U5o, FermionField &out, int mu, int dag)
+  {
+    RealD     sgn= 1.0;
+    if (dag ) sgn=-1.0;
+
+    Gamma::Algebra Gmu [] = {
+			 Gamma::Algebra::GammaX,
+			 Gamma::Algebra::GammaY,
+			 Gamma::Algebra::GammaZ,
+			 Gamma::Algebra::GammaT
+    };
+
+    //    mass is  1,1,1,1,-m has to multiply the round the world term
+    FermionField tmp (in.Grid());
+    tmp = U5e * Cshift(in,mu+1,1);
+    out = tmp - Gamma(Gmu[mu])*tmp*sgn;
+    
+    tmp = Cshift(adj(U5o)*in,mu+1,-1);
+    out = out + tmp + Gamma(Gmu[mu])*tmp*sgn;
+
+    out = -0.5*out;
+  };
+
+  void Dhop5(const FermionField &in, FermionField &out, ComplexField &massE, ComplexField &massO, int dag)
+  {
+    // Mass term.... must multiple the round world with mass = 1,1,1,1, -m
+    RealD     sgn= 1.0;
+    if (dag ) sgn=-1.0;
+
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    FermionField tmp (in.Grid());
+    tmp = massE*Cshift(in,0,1);
+    out = tmp - G5*tmp*sgn;
+    
+    tmp = Cshift(massO*in,0,-1);
+    out = out + tmp + G5*tmp*sgn;
+    out = -0.5*out;
+  };
+
+  // Constructor
+  DWFSlowFermion(GaugeField &_Umu, GridCartesian &Fgrid,
+		 GridRedBlackCartesian &Hgrid, RealD _mass, RealD _M5)
+    :
+    _grid(&Fgrid),
+    _cbgrid(&Hgrid),
+    _grid4(_Umu.Grid()),
+    Umu(Nd,&Fgrid),
+    UmuEven(Nd,&Hgrid),
+    UmuOdd(Nd,&Hgrid),
+    MassField(&Fgrid),
+    MassFieldEven(&Hgrid),
+    MassFieldOdd(&Hgrid),
+    M5(_M5),
+    mass(_mass),
+    _tmp(&Hgrid)
+    {
+      Ls=Fgrid._fdimensions[0];
+      ImportGauge(_Umu);
+
+      typedef typename FermionField::scalar_type scalar;
+
+      Lattice<iScalar<vInteger> > coor(&Fgrid);
+      LatticeCoordinate(coor, 0); // Scoor
+      ComplexField one(&Fgrid);
+      MassField =scalar(-mass);
+      one       =scalar(1.0);
+      MassField =where(coor==Integer(Ls-1),MassField,one);
+      for(int mu=0;mu<Nd;mu++){
+	pickCheckerboard(Even,UmuEven[mu],Umu[mu]);
+	pickCheckerboard(Odd ,UmuOdd[mu],Umu[mu]);
+      }
+      pickCheckerboard(Even,MassFieldEven,MassField);
+      pickCheckerboard(Odd ,MassFieldOdd,MassField);
+      
+    }
+  
+  // DoubleStore impl dependent
+  void ImportGauge(const GaugeField &_Umu4)
+  {
+    GaugeLinkField U4(_grid4);
+    for(int mu=0;mu<Nd;mu++){
+      U4 = PeekIndex<LorentzIndex>(_Umu4, mu);
+      for(int s=0;s<this->Ls;s++){
+	InsertSlice(U4,Umu[mu],s,0);
+      }
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////
+  // Data members require to support the functionality
+  ///////////////////////////////////////////////////////////////
+
+public:
+  virtual RealD Mass(void) { return mass; }
+  virtual int   isTrivialEE(void) { return 1; };
+  RealD mass;
+  RealD M5;
+  int Ls;
+
+  GridBase *_grid4;
+  GridBase *_grid;
+  GridBase *_cbgrid4;
+  GridBase *_cbgrid;
+
+  // Copy of the gauge field , with even and odd subsets
+  std::vector<GaugeLinkField> Umu;
+  std::vector<GaugeLinkField> UmuEven;
+  std::vector<GaugeLinkField> UmuOdd;
+  ComplexField MassField;
+  ComplexField MassFieldEven;
+  ComplexField MassFieldOdd;
+
+  ///////////////////////////////////////////////////////////////
+  // 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 tmin,
+			   unsigned int tmax,
+			   ComplexField &lattice_cmplx){}
+};
+
+typedef DWFSlowFermion<WilsonImplF> DWFSlowFermionF;
+typedef DWFSlowFermion<WilsonImplD> DWFSlowFermionD;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -47,12 +47,14 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 ////////////////////////////////////////////
 // Fermion operators / actions
 ////////////////////////////////////////////
+#include <Grid/qcd/action/fermion/DWFSlow.h>       // Slow DWF
 
 #include <Grid/qcd/action/fermion/WilsonFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(Wilson);
 #include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(WilsonTM);
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
 NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
@@ -111,183 +113,161 @@ NAMESPACE_CHECK(DWFutils);
 // Cayley 5d
 NAMESPACE_BEGIN(Grid);
 
-typedef WilsonFermion<WilsonImplR> WilsonFermionR;
+typedef WilsonFermion<WilsonImplD2> WilsonFermionD2;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
-//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
-//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
-//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
-
-typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
 typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD;
 
-typedef WilsonFermion<WilsonTwoIndexSymmetricImplR> WilsonTwoIndexSymmetricFermionR;
 typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD;
 
-typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonTwoIndexAntiSymmetricFermionR;
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD;
 
+// Sp(2n)
+typedef WilsonFermion<SpWilsonImplF> SpWilsonFermionF;
+typedef WilsonFermion<SpWilsonImplD> SpWilsonFermionD;
+
+typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplF> SpWilsonTwoIndexAntiSymmetricFermionF;
+typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplD> SpWilsonTwoIndexAntiSymmetricFermionD;
+
+typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplF> SpWilsonTwoIndexSymmetricFermionF;
+typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplD> SpWilsonTwoIndexSymmetricFermionD;
+
 // Twisted mass fermion
-typedef WilsonTMFermion<WilsonImplR> WilsonTMFermionR;
+typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
 // Clover fermions
-typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
-typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
-typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
+template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
+template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
 
-typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
-typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
-typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
+typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2;
+typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
+typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2;
+typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
+typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
 
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
+
+typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
+typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
+
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+
+// Compact Clover fermions
+template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
+template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
+
+typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
+typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
+
+typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
+typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
+typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
+
+typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
+
+typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
+typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
+
+typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Domain Wall fermions
-typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
+typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2;
 
-//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
-//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
-//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
-
-typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
+typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
-//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
-//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
-
-typedef MobiusFermion<WilsonImplR> MobiusFermionR;
+typedef MobiusFermion<WilsonImplD2> MobiusFermionD2;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
-//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
-//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
-
-typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
+typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
-//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
-//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
-
-typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
+typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
-//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
-//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
-
-// Ls vectorised
-typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
+typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2;
 typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
 typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD;
 
-typedef MobiusZolotarevFermion<WilsonImplR> MobiusZolotarevFermionR;
+typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2;
 typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF;
 typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD;
-typedef ShamirZolotarevFermion<WilsonImplR> ShamirZolotarevFermionR;
+typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2;
 typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF;
 typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD;
 
-typedef OverlapWilsonCayleyTanhFermion<WilsonImplR> OverlapWilsonCayleyTanhFermionR;
+typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD;
-typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplR> OverlapWilsonCayleyZolotarevFermionR;
+typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD;
 
 // Continued fraction
-typedef OverlapWilsonContFracTanhFermion<WilsonImplR> OverlapWilsonContFracTanhFermionR;
+typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD;
-typedef OverlapWilsonContFracZolotarevFermion<WilsonImplR> OverlapWilsonContFracZolotarevFermionR;
+typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD;
 
 // Partial fraction
-typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplR> OverlapWilsonPartialFractionTanhFermionR;
+typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD;
 
-typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplR> OverlapWilsonPartialFractionZolotarevFermionR;
+typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD;
 
 // Gparity cases; partial list until tested
-typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 
-//typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
-//typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
-//typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
-
-typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
-//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
-//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
-
-typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
+typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
-//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
-//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
-
-typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
+typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
-//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
-//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
-
-typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
+typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
-//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
-//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
-
-typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
+typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 
-//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
-//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
-//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
-
-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;
 

Grid/qcd/action/fermion/FermionOperator.h

@@ -49,6 +49,8 @@ public:
 
   virtual FermionField &tmp(void) = 0;
 
+  virtual void DirichletBlock(const Coordinate & _Block) { assert(0); };
+  
   GridBase * Grid(void)   { return FermionGrid(); };   // this is all the linalg routines need to know
   GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };
 

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -30,6 +30,18 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+/*
+  Policy implementation for G-parity boundary conditions
+
+  Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular
+  field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first
+  with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field
+  to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to
+  either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor.
+
+  Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
+  mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
+ */
 template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal>
 class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > {
 public:
@@ -113,7 +125,7 @@ public:
     || ((distance== 1)&&(icoor[direction]==1))
     || ((distance==-1)&&(icoor[direction]==0));
 
-    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
+    permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction
 
     //Apply the links
     int f_upper = permute_lane ? 1 : 0;
@@ -139,10 +151,10 @@ public:
     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code
     assert((sl == 1) || (sl == 2));
 
-    if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
-
+    //If this site is an global boundary site, perform the G-parity flavor twist
+    if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) {
       if ( sl == 2 ) {
-       
+	//Only do the twist for lanes on the edge of the physical node
 	ExtractBuffer<sobj> vals(Nsimd);
 
 	extract(chi,vals);
@@ -197,6 +209,19 @@ public:
     reg = memory;
   }
 
+
+  //Poke 'poke_f0' onto flavor 0 and 'poke_f1' onto flavor 1 in direction mu of the doubled gauge field Uds
+  inline void pokeGparityDoubledGaugeField(DoubledGaugeField &Uds, const GaugeLinkField &poke_f0, const GaugeLinkField &poke_f1, const int mu){
+    autoView(poke_f0_v, poke_f0, CpuRead);
+    autoView(poke_f1_v, poke_f1, CpuRead);
+    autoView(Uds_v, Uds, CpuWrite);
+    thread_foreach(ss,poke_f0_v,{
+	Uds_v[ss](0)(mu) = poke_f0_v[ss]();
+	Uds_v[ss](1)(mu) = poke_f1_v[ss]();
+      });
+  }
+    
+
   inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
   {
     conformable(Uds.Grid(),GaugeGrid);
@@ -207,14 +232,19 @@ public:
     GaugeLinkField Uconj(GaugeGrid);
    
     Lattice<iScalar<vInteger> > coor(GaugeGrid);
-        
-    for(int mu=0;mu<Nd;mu++){
-          
-      LatticeCoordinate(coor,mu);
+
+    //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
+    //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs        
+    for(int mu=0;mu<Nd-1;mu++){
+
+      if( Params.twists[mu] ){
+	LatticeCoordinate(coor,mu);
+      }
           
       U     = PeekIndex<LorentzIndex>(Umu,mu);
       Uconj = conjugate(U);
      
+      // Implement the isospin rotation sign on the boundary between f=1 and f=0
       // This phase could come from a simple bc 1,1,-1,1 ..
       int neglink = GaugeGrid->GlobalDimensions()[mu]-1;
       if ( Params.twists[mu] ) { 
@@ -229,7 +259,7 @@ public:
 	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
@@ -260,6 +290,38 @@ public:
         });
       }
     }
+
+    { //periodic / antiperiodic temporal BCs
+      int mu = Nd-1;
+      int L   = GaugeGrid->GlobalDimensions()[mu];
+      int Lmu = L - 1;
+
+      LatticeCoordinate(coor, mu);
+
+      U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links
+      
+      GaugeLinkField *Upoke = &U;
+
+      if(Params.twists[mu]){ //antiperiodic
+	Utmp =  where(coor == Lmu, -U, U);
+	Upoke = &Utmp;
+      }
+    
+      Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links      
+      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu);
+
+      //Get the barrel-shifted field
+      Utmp = adj(Cshift(U, mu, -1)); //is a forward shift!
+      Upoke = &Utmp;
+
+      if(Params.twists[mu]){
+	U = where(coor == 0, -Utmp, Utmp);  //boundary phase
+	Upoke = &U;
+      }
+      
+      Uconj = conjugate(*Upoke);
+      pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4);
+    }
   }
       
   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
@@ -298,28 +360,48 @@ public:
   inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
     assert(0);
   }
-  
+ 
   inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
-
-    int Ls = Btilde.Grid()->_fdimensions[0];
-        
-    GaugeLinkField tmp(mat.Grid());
-    tmp = Zero();
+    int Ls=Btilde.Grid()->_fdimensions[0];
+    
     {
-      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));
-	  }
-	});
+      GridBase *GaugeGrid = mat.Grid();
+      Lattice<iScalar<vInteger> > coor(GaugeGrid);
+
+      if( Params.twists[mu] ){
+	LatticeCoordinate(coor,mu);
+      }
+
+      autoView( mat_v , mat, AcceleratorWrite);
+      autoView( Btilde_v , Btilde, AcceleratorRead);
+      autoView( Atilde_v , Atilde, AcceleratorRead);
+      accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{	  
+  	  int sU=sss;
+  	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
+  	  ColorMatrixType sum;
+  	  zeroit(sum);
+  	  for(int s=0;s<Ls;s++){
+  	    int sF = s+Ls*sU;
+  	    for(int spn=0;spn<Ns;spn++){ //sum over spin
+	      //Flavor 0
+  	      auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector
+  	      auto aa = coalescedRead(Atilde_v[sF](0)(spn) );
+  	      sum = sum + outerProduct(bb,aa);
+
+  	      //Flavor 1
+  	      bb = coalescedRead(Btilde_v[sF](1)(spn) );
+  	      aa = coalescedRead(Atilde_v[sF](1)(spn) );
+  	      sum = sum + conjugate(outerProduct(bb,aa));
+  	    }
+  	  }	    
+  	  coalescedWrite(mat_v[sU](mu)(), sum);
+  	});
     }
-    PokeIndex<LorentzIndex>(mat, tmp, mu);
-    return;
   }
+
+
+  
+
   
 };
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -47,18 +47,6 @@ public:
   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
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -52,18 +52,6 @@ public:
   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
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -47,18 +47,6 @@ public:
   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
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -4,10 +4,11 @@
 
     Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
 
-    Copyright (C) 2017
+    Copyright (C) 2017 - 2022
 
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: David Preti <>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
 
     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
@@ -29,7 +30,9 @@
 
 #pragma once
 
-#include <Grid/Grid.h>
+#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -49,19 +52,16 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template <class Impl>
-class WilsonCloverFermion : public WilsonFermion<Impl>
+template<class Impl, class CloverHelpers>
+class WilsonCloverFermion : public WilsonFermion<Impl>,
+                            public WilsonCloverHelpers<Impl>
 {
 public:
-  // Types definitions
   INHERIT_IMPL_TYPES(Impl);
-  template <typename vtype>
-  using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef iImplClover<Simd> SiteCloverType;
-  typedef Lattice<SiteCloverType> CloverFieldType;
+  INHERIT_CLOVER_TYPES(Impl);
 
-public:
-  typedef WilsonFermion<Impl> WilsonBase;
+  typedef WilsonFermion<Impl>       WilsonBase;
+  typedef WilsonCloverHelpers<Impl> Helpers;
 
   virtual int    ConstEE(void)     { return 0; };
   virtual void Instantiatable(void){};
@@ -72,42 +72,7 @@ public:
                       const RealD _csw_r = 0.0,
                       const RealD _csw_t = 0.0,
                       const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
-                      const ImplParams &impl_p = ImplParams()) : WilsonFermion<Impl>(_Umu,
-                                                                                     Fgrid,
-                                                                                     Hgrid,
-                                                                                     _mass, impl_p, clover_anisotropy),
-                                                                 CloverTerm(&Fgrid),
-                                                                 CloverTermInv(&Fgrid),
-                                                                 CloverTermEven(&Hgrid),
-                                                                 CloverTermOdd(&Hgrid),
-                                                                 CloverTermInvEven(&Hgrid),
-                                                                 CloverTermInvOdd(&Hgrid),
-                                                                 CloverTermDagEven(&Hgrid),
-                                                                 CloverTermDagOdd(&Hgrid),
-                                                                 CloverTermInvDagEven(&Hgrid),
-                                                                 CloverTermInvDagOdd(&Hgrid)
-  {
-    assert(Nd == 4); // require 4 dimensions
-
-    if (clover_anisotropy.isAnisotropic)
-    {
-      csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
-      diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
-    }
-    else
-    {
-      csw_r = _csw_r * 0.5;
-      diag_mass = 4.0 + _mass;
-    }
-    csw_t = _csw_t * 0.5;
-
-    if (csw_r == 0)
-      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
-    if (csw_t == 0)
-      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
-
-    ImportGauge(_Umu);
-  }
+                      const ImplParams &impl_p = ImplParams());
 
   virtual void M(const FermionField &in, FermionField &out);
   virtual void Mdag(const FermionField &in, FermionField &out);
@@ -124,250 +89,21 @@ public:
   void ImportGauge(const GaugeField &_Umu);
 
   // Derivative parts unpreconditioned pseudofermions
-  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
-  {
-    conformable(X.Grid(), Y.Grid());
-    conformable(X.Grid(), force.Grid());
-    GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
-    GaugeField clover_force(force.Grid());
-    PropagatorField Lambda(force.Grid());
+  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
 
-    // Guido: Here we are hitting some performance issues:
-    // need to extract the components of the DoubledGaugeField
-    // for each call
-    // Possible solution
-    // Create a vector object to store them? (cons: wasting space)
-    std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
-
-    Impl::extractLinkField(U, this->Umu);
-
-    force = Zero();
-    // Derivative of the Wilson hopping term
-    this->DhopDeriv(force, X, Y, dag);
-
-    ///////////////////////////////////////////////////////////
-    // Clover term derivative
-    ///////////////////////////////////////////////////////////
-    Impl::outerProductImpl(Lambda, X, Y);
-    //std::cout << "Lambda:" << Lambda << std::endl;
-
-    Gamma::Algebra sigma[] = {
-        Gamma::Algebra::SigmaXY,
-        Gamma::Algebra::SigmaXZ,
-        Gamma::Algebra::SigmaXT,
-        Gamma::Algebra::MinusSigmaXY,
-        Gamma::Algebra::SigmaYZ,
-        Gamma::Algebra::SigmaYT,
-        Gamma::Algebra::MinusSigmaXZ,
-        Gamma::Algebra::MinusSigmaYZ,
-        Gamma::Algebra::SigmaZT,
-        Gamma::Algebra::MinusSigmaXT,
-        Gamma::Algebra::MinusSigmaYT,
-        Gamma::Algebra::MinusSigmaZT};
-
-    /*
-      sigma_{\mu \nu}=
-      | 0         sigma[0]  sigma[1]  sigma[2] |
-      | sigma[3]    0       sigma[4]  sigma[5] |
-      | sigma[6]  sigma[7]     0      sigma[8] |
-      | sigma[9]  sigma[10] sigma[11]   0      |
-    */
-
-    int count = 0;
-    clover_force = Zero();
-    for (int mu = 0; mu < 4; mu++)
-    {
-      force_mu = Zero();
-      for (int nu = 0; nu < 4; nu++)
-      {
-        if (mu == nu)
-        continue;
-        
-        RealD factor;
-        if (nu == 4 || mu == 4)
-        {
-          factor = 2.0 * csw_t;
-        }
-        else
-        {
-          factor = 2.0 * csw_r;
-        }
-        PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
-        Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-        force_mu -= factor*Cmunu(U, lambda, mu, nu);                   // checked
-        count++;
-      }
-
-      pokeLorentz(clover_force, U[mu] * force_mu, mu);
-    }
-    //clover_force *= csw;
-    force += clover_force;
-  }
-
-  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
-  GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
-  {
-    conformable(lambda.Grid(), U[0].Grid());
-    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
-    // insertion in upper staple
-    // please check redundancy of shift operations
-
-    // C1+
-    tmp = lambda * U[nu];
-    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C2+
-    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C3+
-    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
-
-    // C4+
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
-
-    // insertion in lower staple
-    // C1-
-    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C2-
-    tmp = adj(lambda) * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C3-
-    tmp = lambda * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
-
-    // C4-
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
-
-    return out;
-  }
-
-protected:
+public:
   // here fixing the 4 dimensions, make it more general?
 
   RealD csw_r;                                               // Clover coefficient - spatial
   RealD csw_t;                                               // Clover coefficient - temporal
   RealD diag_mass;                                           // Mass term
-  CloverFieldType CloverTerm, CloverTermInv;                 // Clover term
-  CloverFieldType CloverTermEven, CloverTermOdd;             // Clover term EO
-  CloverFieldType CloverTermInvEven, CloverTermInvOdd;       // Clover term Inv EO
-  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();
-    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]()());
-      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverXZ(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-    
-    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]()();
-      T_v[i]()(2, 3) = -F_v[i]()();
-      T_v[i]()(3, 2) = F_v[i]()();
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverXY(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-
-    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]()());
-      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 3) = timesI(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverXT(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-
-    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]()());
-      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverYT(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-    
-    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]()());
-      T_v[i]()(2, 3) = (F_v[i]()());
-      T_v[i]()(3, 2) = -(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverZT(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-
-    T = Zero();
-
-    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]()());
-      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 3) = timesI(F_v[i]()());
-    });
-
-    return T;
-  }
+  CloverField CloverTerm, CloverTermInv;                     // Clover term
+  CloverField CloverTermEven, CloverTermOdd;                 // Clover term EO
+  CloverField CloverTermInvEven, CloverTermInvOdd;           // Clover term Inv EO
+  CloverField CloverTermDagEven, CloverTermDagOdd;           // Clover term Dag EO
+  CloverField CloverTermInvDagEven, CloverTermInvDagOdd;     // Clover term Inv Dag EO
 };
+
 NAMESPACE_END(Grid);
 
 

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -0,0 +1,763 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonCloverHelpers.h
+
+    Copyright (C) 2021 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+
+    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
+
+// Helper routines that implement common clover functionality
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Impl> class WilsonCloverHelpers {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+
+  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
+  {
+    conformable(lambda.Grid(), U[0].Grid());
+    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
+    // insertion in upper staple
+    // please check redundancy of shift operations
+
+    // C1+
+    tmp = lambda * U[nu];
+    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C2+
+    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C3+
+    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
+
+    // C4+
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
+
+    // insertion in lower staple
+    // C1-
+    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C2-
+    tmp = adj(lambda) * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C3-
+    tmp = lambda * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
+
+    // C4-
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
+
+    return out;
+  }
+
+  static CloverField fillCloverYZ(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverXZ(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+    
+    autoView(T_v, T,AcceleratorWrite);
+    autoView(F_v, F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-F_v[i]()()));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead(F_v[i]()()));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead(-F_v[i]()()));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(F_v[i]()()));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverXY(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverXT(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+
+    autoView( T_v , T, AcceleratorWrite);
+    autoView( F_v , F, AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverYT(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+    
+    autoView( T_v ,T,AcceleratorWrite);
+    autoView( F_v ,F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead((F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead((F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(-(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverZT(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+
+    T = Zero();
+
+    autoView( T_v , T,AcceleratorWrite);
+    autoView( F_v , F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  template<class _Spinor>
+  static accelerator_inline void multClover(_Spinor& phi, const SiteClover& C, const _Spinor& chi) {
+    auto CC = coalescedRead(C);
+    mult(&phi, &CC, &chi);
+  }
+
+  template<class _SpinorField>
+  inline void multCloverField(_SpinorField& out, const CloverField& C, const _SpinorField& phi) {
+    const int Nsimd = SiteSpinor::Nsimd();
+    autoView(out_v, out, AcceleratorWrite);
+    autoView(phi_v, phi, AcceleratorRead);
+    autoView(C_v,   C,   AcceleratorRead);
+    typedef decltype(coalescedRead(out_v[0])) calcSpinor;
+    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
+      calcSpinor tmp;
+      multClover(tmp,C_v[sss],phi_v(sss));
+      coalescedWrite(out_v[sss],tmp);
+    });
+  }
+};
+
+
+////////////////////////////////////////////////////////
+
+template<class Impl> class CompactWilsonCloverHelpers {
+public:
+
+  INHERIT_COMPACT_CLOVER_SIZES(Impl);
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  #if 0
+  static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
+    assert(i != j);
+    if(i < j) {
+      return triangle()(block)(triangle_index(i, j));
+    } else { // i > j
+      return conjugate(triangle()(block)(triangle_index(i, j)));
+    }
+  }
+  #else
+  template<typename vobj>
+  static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
+    assert(i != j);
+    if(i < j) {
+      return triangle()(block)(triangle_index(i, j));
+    } else { // i > j
+      return conjugate(triangle()(block)(triangle_index(i, j)));
+    }
+  }
+  #endif
+
+  static accelerator_inline int triangle_index(int i, int j) {
+    if(i == j)
+      return 0;
+    else if(i < j)
+      return Nred * (Nred - 1) / 2 - (Nred - i) * (Nred - i - 1) / 2 + j - i - 1;
+    else // i > j
+      return Nred * (Nred - 1) / 2 - (Nred - j) * (Nred - j - 1) / 2 + i - j - 1;
+  }
+
+  static void MooeeKernel_gpu(int                        Nsite,
+                              int                        Ls,
+                              const FermionField&        in,
+                              FermionField&              out,
+                              const CloverDiagonalField& diagonal,
+                              const CloverTriangleField& triangle) {
+    autoView(diagonal_v, diagonal, AcceleratorRead);
+    autoView(triangle_v, triangle, AcceleratorRead);
+    autoView(in_v,       in,       AcceleratorRead);
+    autoView(out_v,      out,      AcceleratorWrite);
+
+    typedef decltype(coalescedRead(out_v[0])) CalcSpinor;
+
+    const uint64_t NN = Nsite * Ls;
+
+    accelerator_for(ss, NN, Simd::Nsimd(), {
+      int sF = ss;
+      int sU = ss/Ls;
+      CalcSpinor res;
+      CalcSpinor in_t = in_v(sF);
+      auto diagonal_t = diagonal_v(sU);
+      auto triangle_t = triangle_v(sU);
+      for(int block=0; block<Nhs; block++) {
+        int s_start = block*Nhs;
+        for(int i=0; i<Nred; i++) {
+          int si = s_start + i/Nc, ci = i%Nc;
+          res()(si)(ci) = diagonal_t()(block)(i) * in_t()(si)(ci);
+          for(int j=0; j<Nred; j++) {
+            if (j == i) continue;
+            int sj = s_start + j/Nc, cj = j%Nc;
+            res()(si)(ci) = res()(si)(ci) + triangle_elem(triangle_t, block, i, j) * in_t()(sj)(cj);
+          };
+        };
+      };
+      coalescedWrite(out_v[sF], res);
+    });
+  }
+
+  static void MooeeKernel_cpu(int                        Nsite,
+                              int                        Ls,
+                              const FermionField&        in,
+                              FermionField&              out,
+                              const CloverDiagonalField& diagonal,
+                              const CloverTriangleField& triangle) {
+    autoView(diagonal_v, diagonal, CpuRead);
+    autoView(triangle_v, triangle, CpuRead);
+    autoView(in_v,       in,       CpuRead);
+    autoView(out_v,      out,      CpuWrite);
+
+    typedef SiteSpinor CalcSpinor;
+
+#if defined(A64FX) || defined(A64FXFIXEDSIZE)
+#define PREFETCH_CLOVER(BASE) {                                     \
+    uint64_t base;                                                  \
+    int pf_dist_L1 = 1;                                             \
+    int pf_dist_L2 = -5; /* -> penalty -> disable */                \
+                                                                    \
+    if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {           \
+      base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);               \
+      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL1STRM); \
+    }                                                               \
+                                                                    \
+    if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {           \
+      base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);               \
+      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL2STRM); \
+    }                                                               \
+  }
+// TODO: Implement/generalize this for other architectures
+// I played around a bit on KNL (see below) but didn't bring anything
+// #elif defined(AVX512)
+// #define PREFETCH_CLOVER(BASE) {                              \
+//     uint64_t base;                                           \
+//     int pf_dist_L1 = 1;                                      \
+//     int pf_dist_L2 = +4;                                     \
+//                                                              \
+//     if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {    \
+//       base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);        \
+//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T0); \
+//     }                                                        \
+//                                                              \
+//     if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {    \
+//       base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);        \
+//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T1); \
+//     }                                                        \
+//   }
+#else
+#define PREFETCH_CLOVER(BASE)
+#endif
+
+    const uint64_t NN = Nsite * Ls;
+
+    thread_for(ss, NN, {
+      int sF = ss;
+      int sU = ss/Ls;
+      CalcSpinor res;
+      CalcSpinor in_t = in_v[sF];
+      auto diag_t     = diagonal_v[sU]; // "diag" instead of "diagonal" here to make code below easier to read
+      auto triangle_t = triangle_v[sU];
+
+      // upper half
+      PREFETCH_CLOVER(0);
+
+      auto in_cc_0_0 = conjugate(in_t()(0)(0)); // Nils: reduces number
+      auto in_cc_0_1 = conjugate(in_t()(0)(1)); // of conjugates from
+      auto in_cc_0_2 = conjugate(in_t()(0)(2)); // 30 to 20
+      auto in_cc_1_0 = conjugate(in_t()(1)(0));
+      auto in_cc_1_1 = conjugate(in_t()(1)(1));
+
+      res()(0)(0) =               diag_t()(0)( 0) * in_t()(0)(0)
+                  +           triangle_t()(0)( 0) * in_t()(0)(1)
+                  +           triangle_t()(0)( 1) * in_t()(0)(2)
+                  +           triangle_t()(0)( 2) * in_t()(1)(0)
+                  +           triangle_t()(0)( 3) * in_t()(1)(1)
+                  +           triangle_t()(0)( 4) * in_t()(1)(2);
+
+      res()(0)(1) =           triangle_t()(0)( 0) * in_cc_0_0;
+      res()(0)(1) =               diag_t()(0)( 1) * in_t()(0)(1)
+                  +           triangle_t()(0)( 5) * in_t()(0)(2)
+                  +           triangle_t()(0)( 6) * in_t()(1)(0)
+                  +           triangle_t()(0)( 7) * in_t()(1)(1)
+                  +           triangle_t()(0)( 8) * in_t()(1)(2)
+                  + conjugate(       res()(0)( 1));
+
+      res()(0)(2) =           triangle_t()(0)( 1) * in_cc_0_0
+                  +           triangle_t()(0)( 5) * in_cc_0_1;
+      res()(0)(2) =               diag_t()(0)( 2) * in_t()(0)(2)
+                  +           triangle_t()(0)( 9) * in_t()(1)(0)
+                  +           triangle_t()(0)(10) * in_t()(1)(1)
+                  +           triangle_t()(0)(11) * in_t()(1)(2)
+                  + conjugate(       res()(0)( 2));
+
+      res()(1)(0) =           triangle_t()(0)( 2) * in_cc_0_0
+                  +           triangle_t()(0)( 6) * in_cc_0_1
+                  +           triangle_t()(0)( 9) * in_cc_0_2;
+      res()(1)(0) =               diag_t()(0)( 3) * in_t()(1)(0)
+                  +           triangle_t()(0)(12) * in_t()(1)(1)
+                  +           triangle_t()(0)(13) * in_t()(1)(2)
+                  + conjugate(       res()(1)( 0));
+
+      res()(1)(1) =           triangle_t()(0)( 3) * in_cc_0_0
+                  +           triangle_t()(0)( 7) * in_cc_0_1
+                  +           triangle_t()(0)(10) * in_cc_0_2
+                  +           triangle_t()(0)(12) * in_cc_1_0;
+      res()(1)(1) =               diag_t()(0)( 4) * in_t()(1)(1)
+                  +           triangle_t()(0)(14) * in_t()(1)(2)
+                  + conjugate(       res()(1)( 1));
+
+      res()(1)(2) =           triangle_t()(0)( 4) * in_cc_0_0
+                  +           triangle_t()(0)( 8) * in_cc_0_1
+                  +           triangle_t()(0)(11) * in_cc_0_2
+                  +           triangle_t()(0)(13) * in_cc_1_0
+                  +           triangle_t()(0)(14) * in_cc_1_1;
+      res()(1)(2) =               diag_t()(0)( 5) * in_t()(1)(2)
+                  + conjugate(       res()(1)( 2));
+
+      vstream(out_v[sF]()(0)(0), res()(0)(0));
+      vstream(out_v[sF]()(0)(1), res()(0)(1));
+      vstream(out_v[sF]()(0)(2), res()(0)(2));
+      vstream(out_v[sF]()(1)(0), res()(1)(0));
+      vstream(out_v[sF]()(1)(1), res()(1)(1));
+      vstream(out_v[sF]()(1)(2), res()(1)(2));
+
+      // lower half
+      PREFETCH_CLOVER(1);
+
+      auto in_cc_2_0 = conjugate(in_t()(2)(0));
+      auto in_cc_2_1 = conjugate(in_t()(2)(1));
+      auto in_cc_2_2 = conjugate(in_t()(2)(2));
+      auto in_cc_3_0 = conjugate(in_t()(3)(0));
+      auto in_cc_3_1 = conjugate(in_t()(3)(1));
+
+      res()(2)(0) =               diag_t()(1)( 0) * in_t()(2)(0)
+                  +           triangle_t()(1)( 0) * in_t()(2)(1)
+                  +           triangle_t()(1)( 1) * in_t()(2)(2)
+                  +           triangle_t()(1)( 2) * in_t()(3)(0)
+                  +           triangle_t()(1)( 3) * in_t()(3)(1)
+                  +           triangle_t()(1)( 4) * in_t()(3)(2);
+
+      res()(2)(1) =           triangle_t()(1)( 0) * in_cc_2_0;
+      res()(2)(1) =               diag_t()(1)( 1) * in_t()(2)(1)
+                  +           triangle_t()(1)( 5) * in_t()(2)(2)
+                  +           triangle_t()(1)( 6) * in_t()(3)(0)
+                  +           triangle_t()(1)( 7) * in_t()(3)(1)
+                  +           triangle_t()(1)( 8) * in_t()(3)(2)
+                  + conjugate(       res()(2)( 1));
+
+      res()(2)(2) =           triangle_t()(1)( 1) * in_cc_2_0
+                  +           triangle_t()(1)( 5) * in_cc_2_1;
+      res()(2)(2) =               diag_t()(1)( 2) * in_t()(2)(2)
+                  +           triangle_t()(1)( 9) * in_t()(3)(0)
+                  +           triangle_t()(1)(10) * in_t()(3)(1)
+                  +           triangle_t()(1)(11) * in_t()(3)(2)
+                  + conjugate(       res()(2)( 2));
+
+      res()(3)(0) =           triangle_t()(1)( 2) * in_cc_2_0
+                  +           triangle_t()(1)( 6) * in_cc_2_1
+                  +           triangle_t()(1)( 9) * in_cc_2_2;
+      res()(3)(0) =               diag_t()(1)( 3) * in_t()(3)(0)
+                  +           triangle_t()(1)(12) * in_t()(3)(1)
+                  +           triangle_t()(1)(13) * in_t()(3)(2)
+                  + conjugate(       res()(3)( 0));
+
+      res()(3)(1) =           triangle_t()(1)( 3) * in_cc_2_0
+                  +           triangle_t()(1)( 7) * in_cc_2_1
+                  +           triangle_t()(1)(10) * in_cc_2_2
+                  +           triangle_t()(1)(12) * in_cc_3_0;
+      res()(3)(1) =               diag_t()(1)( 4) * in_t()(3)(1)
+                  +           triangle_t()(1)(14) * in_t()(3)(2)
+                  + conjugate(       res()(3)( 1));
+
+      res()(3)(2) =           triangle_t()(1)( 4) * in_cc_2_0
+                  +           triangle_t()(1)( 8) * in_cc_2_1
+                  +           triangle_t()(1)(11) * in_cc_2_2
+                  +           triangle_t()(1)(13) * in_cc_3_0
+                  +           triangle_t()(1)(14) * in_cc_3_1;
+      res()(3)(2) =               diag_t()(1)( 5) * in_t()(3)(2)
+                  + conjugate(       res()(3)( 2));
+
+      vstream(out_v[sF]()(2)(0), res()(2)(0));
+      vstream(out_v[sF]()(2)(1), res()(2)(1));
+      vstream(out_v[sF]()(2)(2), res()(2)(2));
+      vstream(out_v[sF]()(3)(0), res()(3)(0));
+      vstream(out_v[sF]()(3)(1), res()(3)(1));
+      vstream(out_v[sF]()(3)(2), res()(3)(2));
+    });
+  }
+
+  static void MooeeKernel(int                        Nsite,
+                          int                        Ls,
+                          const FermionField&        in,
+                          FermionField&              out,
+                          const CloverDiagonalField& diagonal,
+                          const CloverTriangleField& triangle) {
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+    MooeeKernel_gpu(Nsite, Ls, in, out, diagonal, triangle);
+#else
+    MooeeKernel_cpu(Nsite, Ls, in, out, diagonal, triangle);
+#endif
+  }
+
+  static void Invert(const CloverDiagonalField& diagonal,
+                     const CloverTriangleField& triangle,
+                     CloverDiagonalField&       diagonalInv,
+                     CloverTriangleField&       triangleInv) {
+    conformable(diagonal, diagonalInv);
+    conformable(triangle, triangleInv);
+    conformable(diagonal, triangle);
+
+    diagonalInv.Checkerboard() = diagonal.Checkerboard();
+    triangleInv.Checkerboard() = triangle.Checkerboard();
+
+    GridBase* grid = diagonal.Grid();
+
+    long lsites = grid->lSites();
+
+    typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
+    typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
+
+    autoView(diagonal_v,  diagonal,  CpuRead);
+    autoView(triangle_v,  triangle,  CpuRead);
+    autoView(diagonalInv_v, diagonalInv, CpuWrite);
+    autoView(triangleInv_v, triangleInv, CpuWrite);
+
+    thread_for(site, lsites, { // NOTE: Not on GPU because of Eigen & (peek/poke)LocalSite
+      Eigen::MatrixXcd clover_inv_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
+      Eigen::MatrixXcd clover_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
+
+      scalar_object_diagonal diagonal_tmp     = Zero();
+      scalar_object_diagonal diagonal_inv_tmp = Zero();
+      scalar_object_triangle triangle_tmp     = Zero();
+      scalar_object_triangle triangle_inv_tmp = Zero();
+
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(site, lcoor);
+
+      peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
+      peekLocalSite(triangle_tmp, triangle_v, lcoor);
+
+      // TODO: can we save time here by inverting the two 6x6 hermitian matrices separately?
+      for (long s_row=0;s_row<Ns;s_row++) {
+        for (long s_col=0;s_col<Ns;s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for (long c_row=0;c_row<Nc;c_row++) {
+            for (long c_col=0;c_col<Nc;c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
+              else
+                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(triangle_elem(triangle_tmp, block, i, j)));
+            }
+          }
+        }
+      }
+
+      clover_inv_eigen = clover_eigen.inverse();
+
+      for (long s_row=0;s_row<Ns;s_row++) {
+        for (long s_col=0;s_col<Ns;s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for (long c_row=0;c_row<Nc;c_row++) {
+            for (long c_col=0;c_col<Nc;c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                diagonal_inv_tmp()(block)(i) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
+              else if(i < j)
+                triangle_inv_tmp()(block)(triangle_index(i, j)) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
+              else
+                continue;
+            }
+          }
+        }
+      }
+
+      pokeLocalSite(diagonal_inv_tmp, diagonalInv_v, lcoor);
+      pokeLocalSite(triangle_inv_tmp, triangleInv_v, lcoor);
+    });
+  }
+
+  static void ConvertLayout(const CloverField&   full,
+                            CloverDiagonalField& diagonal,
+                            CloverTriangleField& triangle) {
+    conformable(full, diagonal);
+    conformable(full, triangle);
+
+    diagonal.Checkerboard() = full.Checkerboard();
+    triangle.Checkerboard() = full.Checkerboard();
+
+    autoView(full_v,     full,     AcceleratorRead);
+    autoView(diagonal_v, diagonal, AcceleratorWrite);
+    autoView(triangle_v, triangle, AcceleratorWrite);
+
+    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
+    accelerator_for(ss, full.Grid()->oSites(), 1, {
+      for(int s_row = 0; s_row < Ns; s_row++) {
+        for(int s_col = 0; s_col < Ns; s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for(int c_row = 0; c_row < Nc; c_row++) {
+            for(int c_col = 0; c_col < Nc; c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                diagonal_v[ss]()(block)(i) = full_v[ss]()(s_row, s_col)(c_row, c_col);
+              else if(i < j)
+                triangle_v[ss]()(block)(triangle_index(i, j)) = full_v[ss]()(s_row, s_col)(c_row, c_col);
+              else
+                continue;
+            }
+          }
+        }
+      }
+    });
+  }
+
+
+  static void ConvertLayout(const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverField&               full) {
+    conformable(full, diagonal);
+    conformable(full, triangle);
+
+    full.Checkerboard() = diagonal.Checkerboard();
+
+    full = Zero();
+
+    autoView(diagonal_v, diagonal, AcceleratorRead);
+    autoView(triangle_v, triangle, AcceleratorRead);
+    autoView(full_v,     full,     AcceleratorWrite);
+
+    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
+    accelerator_for(ss, full.Grid()->oSites(), 1, {
+      for(int s_row = 0; s_row < Ns; s_row++) {
+        for(int s_col = 0; s_col < Ns; s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for(int c_row = 0; c_row < Nc; c_row++) {
+            for(int c_col = 0; c_col < Nc; c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                full_v[ss]()(s_row, s_col)(c_row, c_col) = diagonal_v[ss]()(block)(i);
+              else
+                full_v[ss]()(s_row, s_col)(c_row, c_col) = triangle_elem(triangle_v[ss], block, i, j);
+            }
+          }
+        }
+      }
+    });
+  }
+
+  static void ModifyBoundaries(CloverDiagonalField& diagonal, CloverTriangleField& triangle, RealD csw_t, RealD cF, RealD diag_mass) {
+    // Checks/grid
+    double t0 = usecond();
+    conformable(diagonal, triangle);
+    GridBase* grid = diagonal.Grid();
+
+    // Determine the boundary coordinates/sites
+    double t1 = usecond();
+    int t_dir = Nd - 1;
+    Lattice<iScalar<vInteger>> t_coor(grid);
+    LatticeCoordinate(t_coor, t_dir);
+    int T = grid->GlobalDimensions()[t_dir];
+
+    // Set off-diagonal parts at boundary to zero -- OK
+    double t2 = usecond();
+    CloverTriangleField zeroTriangle(grid);
+    zeroTriangle.Checkerboard() = triangle.Checkerboard();
+    zeroTriangle = Zero();
+    triangle = where(t_coor == 0,   zeroTriangle, triangle);
+    triangle = where(t_coor == T-1, zeroTriangle, triangle);
+
+    // Set diagonal to unity (scaled correctly) -- OK
+    double t3 = usecond();
+    CloverDiagonalField tmp(grid);
+    tmp.Checkerboard() = diagonal.Checkerboard();
+    tmp                = -1.0 * csw_t + diag_mass;
+    diagonal           = where(t_coor == 0,   tmp, diagonal);
+    diagonal           = where(t_coor == T-1, tmp, diagonal);
+
+    // Correct values next to boundary
+    double t4 = usecond();
+    if(cF != 1.0) {
+      tmp = cF - 1.0;
+      tmp += diagonal;
+      diagonal = where(t_coor == 1,   tmp, diagonal);
+      diagonal = where(t_coor == T-2, tmp, diagonal);
+    }
+
+    // Report timings
+    double t5 = usecond();
+#if 0
+    std::cout << GridLogMessage << "CompactWilsonCloverHelpers::ModifyBoundaries timings:"
+              << " checks = "          << (t1 - t0) / 1e6
+              << ", coordinate = "     << (t2 - t1) / 1e6
+              << ", off-diag zero = "  << (t3 - t2) / 1e6
+              << ", diagonal unity = " << (t4 - t3) / 1e6
+              << ", near-boundary = "  << (t5 - t4) / 1e6
+              << ", total = "          << (t5 - t0) / 1e6
+              << std::endl;
+#endif
+  }
+
+  template<class Field, class Mask>
+  static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
+    conformable(f, m);
+    auto grid  = f.Grid();
+    const uint32_t Nsite = grid->oSites();
+    const uint32_t Nsimd = grid->Nsimd();
+    autoView(f_v, f, AcceleratorWrite);
+    autoView(m_v, m, AcceleratorRead);
+    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
+    accelerator_for(ss, Nsite, Nsimd, {
+      coalescedWrite(f_v[ss], m_v(ss) * f_v(ss));
+    });
+  }
+
+  template<class MaskField>
+  static void SetupMasks(MaskField& full, MaskField& even, MaskField& odd) {
+    assert(even.Grid()->_isCheckerBoarded && even.Checkerboard() == Even);
+    assert(odd.Grid()->_isCheckerBoarded  && odd.Checkerboard()  == Odd);
+    assert(!full.Grid()->_isCheckerBoarded);
+
+    GridBase* grid = full.Grid();
+    int t_dir = Nd-1;
+    Lattice<iScalar<vInteger>> t_coor(grid);
+    LatticeCoordinate(t_coor, t_dir);
+    int T = grid->GlobalDimensions()[t_dir];
+
+    MaskField zeroMask(grid); zeroMask = Zero();
+    full = 1.0;
+    full = where(t_coor == 0,   zeroMask, full);
+    full = where(t_coor == T-1, zeroMask, full);
+
+    pickCheckerboard(Even, even, full);
+    pickCheckerboard(Odd,  odd,  full);
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -0,0 +1,90 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonCloverTypes.h
+
+    Copyright (C) 2021 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+
+    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 Impl>
+class WilsonCloverTypes {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+
+  typedef iImplClover<Simd> SiteClover;
+
+  typedef Lattice<SiteClover> CloverField;
+};
+
+template<class Impl>
+class CompactWilsonCloverTypes {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+  static constexpr int Nred      = Nc * Nhs;        // 6
+  static constexpr int Nblock    = Nhs;             // 2
+  static constexpr int Ndiagonal = Nred;            // 6
+  static constexpr int Ntriangle = (Nred - 1) * Nc; // 15
+
+  template<typename vtype> using iImplCloverDiagonal = iScalar<iVector<iVector<vtype, Ndiagonal>, Nblock>>;
+  template<typename vtype> using iImplCloverTriangle = iScalar<iVector<iVector<vtype, Ntriangle>, Nblock>>;
+
+  typedef iImplCloverDiagonal<Simd> SiteCloverDiagonal;
+  typedef iImplCloverTriangle<Simd> SiteCloverTriangle;
+  typedef iSinglet<Simd>            SiteMask;
+
+  typedef Lattice<SiteCloverDiagonal> CloverDiagonalField;
+  typedef Lattice<SiteCloverTriangle> CloverTriangleField;
+  typedef Lattice<SiteMask>           MaskField;
+};
+
+#define INHERIT_CLOVER_TYPES(Impl)                                 \
+  typedef typename WilsonCloverTypes<Impl>::SiteClover SiteClover; \
+  typedef typename WilsonCloverTypes<Impl>::CloverField CloverField;
+
+#define INHERIT_COMPACT_CLOVER_TYPES(Impl) \
+  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverDiagonal  SiteCloverDiagonal; \
+  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverTriangle  SiteCloverTriangle; \
+  typedef typename CompactWilsonCloverTypes<Impl>::SiteMask            SiteMask; \
+  typedef typename CompactWilsonCloverTypes<Impl>::CloverDiagonalField CloverDiagonalField; \
+  typedef typename CompactWilsonCloverTypes<Impl>::CloverTriangleField CloverTriangleField; \
+  typedef typename CompactWilsonCloverTypes<Impl>::MaskField           MaskField; \
+  /* ugly duplication but needed inside functionality classes */ \
+  template<typename vtype> using iImplCloverDiagonal = \
+    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ndiagonal>, CompactWilsonCloverTypes<Impl>::Nblock>>; \
+  template<typename vtype> using iImplCloverTriangle = \
+    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ntriangle>, CompactWilsonCloverTypes<Impl>::Nblock>>;
+
+#define INHERIT_COMPACT_CLOVER_SIZES(Impl)                                    \
+  static constexpr int Nred      = CompactWilsonCloverTypes<Impl>::Nred;      \
+  static constexpr int Nblock    = CompactWilsonCloverTypes<Impl>::Nblock;    \
+  static constexpr int Ndiagonal = CompactWilsonCloverTypes<Impl>::Ndiagonal; \
+  static constexpr int Ntriangle = CompactWilsonCloverTypes<Impl>::Ntriangle;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -32,17 +32,218 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+///////////////////////////////////////////////////////////////
+// Wilson compressor will need FaceGather policies for:
+// Periodic, Dirichlet, and partial Dirichlet for DWF
+///////////////////////////////////////////////////////////////
+const int dwf_compressor_depth=2;
+#define DWF_COMPRESS
+class FaceGatherPartialDWF
+{
+public:
+#ifdef DWF_COMPRESS
+  static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
+#else
+  static int PartialCompressionFactor(GridBase *grid) { return 1;}
+#endif
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
+				   const Lattice<vobj> &rhs,
+				   cobj *buffer,
+				   compressor &compress,
+				   int off,int so,int partial)
+  {
+    //DWF only hack: If a direction that is OFF node we use Partial Dirichlet
+    //  Shrinks local and remote comms buffers
+    GridBase *Grid = rhs.Grid();
+    int Ls = Grid->_rdimensions[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else 
+    int depth=Ls/2;
+#endif
+    std::pair<int,int> *table_v = & table[0];
+    auto rhs_v = rhs.View(AcceleratorRead);
+    int vol=table.size()/Ls;
+    accelerator_forNB( idx,table.size(), vobj::Nsimd(), {
+	Integer i=idx/Ls;
+	Integer s=idx%Ls;
+	Integer sc=depth+s-(Ls-depth);
+	if(s<depth)     compress.Compress(buffer[off+i+s*vol],rhs_v[so+table_v[idx].second]);
+	if(s>=Ls-depth) compress.Compress(buffer[off+i+sc*vol],rhs_v[so+table_v[idx].second]);
+    });
+    rhs_v.ViewClose();
+  }
+  template<class decompressor,class Decompression>
+  static void DecompressFace(decompressor decompress,Decompression &dd)
+  {
+    auto Ls = dd.dims[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else
+    int depth=Ls/2;
+#endif    
+    // Just pass in the Grid
+    auto kp = dd.kernel_p;
+    auto mp = dd.mpi_p;
+    int size= dd.buffer_size;
+    int vol= size/Ls;
+    accelerator_forNB(o,size,1,{
+	int idx=o/Ls;
+	int   s=o%Ls;
+	if ( s < depth ) {
+	  int oo=s*vol+idx;
+	  kp[o]=mp[oo];
+	} else if ( s >= Ls-depth ) {
+	  int sc = depth + s - (Ls-depth);
+	  int oo=sc*vol+idx; 
+	  kp[o]=mp[oo];
+	} else {
+	  kp[o] = Zero();//fill rest with zero if partial dirichlet
+	}
+    });
+  }
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Need to gather *interior portions* for ALL s-slices in simd directions
+  // Do the gather as need to treat SIMD lanes differently, and insert zeroes on receive side
+  // Reorder the fifth dim to be s=Ls-1 , s=0, s=1,...,Ls-2.
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    compressor &compress,int type,int partial)
+  {
+    GridBase *Grid = rhs.Grid();
+    int Ls = Grid->_rdimensions[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else
+    int depth = Ls/2;
+#endif
+    
+    // insertion of zeroes...
+    assert( (table.size()&0x1)==0);
+    int num=table.size()/2;
+    int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
+    
+    auto rhs_v = rhs.View(AcceleratorRead);
+    auto p0=&pointers[0][0];
+    auto p1=&pointers[1][0];
+    auto tp=&table[0];
+    int nnum=num/Ls;
+    accelerator_forNB(j, num, vobj::Nsimd(), {
+	//  Reorders both local and remote comms buffers
+	//  
+	int s  = j % Ls;
+	int sp1 = (s+depth)%Ls;  // peri incremented s slice
+	
+	int hxyz= j/Ls;
+
+	int xyz0= hxyz*2; // xyzt part of coor
+	int xyz1= hxyz*2+1;
+	
+	int jj= hxyz + sp1*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
+	
+	int kk0= xyz0*Ls + s ; // s=0 goes to s=1
+	int kk1= xyz1*Ls + s ; // s=Ls-1 -> s=0
+	compress.CompressExchange(p0[jj],p1[jj],
+				  rhs_v[so+tp[kk0 ].second], // Same s, consecutive xyz sites
+				  rhs_v[so+tp[kk1 ].second], 
+				  type);
+    });
+    rhs_v.ViewClose();
+  }
+  // Merge routine is for SIMD faces
+  template<class decompressor,class Merger>
+  static void MergeFace(decompressor decompress,Merger &mm)
+  {
+    auto Ls = mm.dims[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else
+    int depth = Ls/2;
+#endif
+    int  num= mm.buffer_size/2; // relate vol and Ls to buffer size
+    auto mp = &mm.mpointer[0];
+    auto vp0= &mm.vpointers[0][0]; // First arg is exchange first
+    auto vp1= &mm.vpointers[1][0];
+    auto type= mm.type;
+    int nnum = num/Ls;
+    accelerator_forNB(o,num,Merger::Nsimd,{
+
+	int  s=o%Ls;
+	int hxyz=o/Ls; // xyzt related component
+	int xyz0=hxyz*2;
+	int xyz1=hxyz*2+1;
+
+	int sp = (s+depth)%Ls; 
+	int jj= hxyz + sp*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
+
+	int oo0= s+xyz0*Ls;
+	int oo1= s+xyz1*Ls;
+
+	// same ss0, ss1 pair goes to new layout
+	decompress.Exchange(mp[oo0],mp[oo1],vp0[jj],vp1[jj],type);
+      });
+  }
+};
+class FaceGatherDWFMixedBCs
+{
+public:
+#ifdef DWF_COMPRESS
+  static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
+#else 
+  static int PartialCompressionFactor(GridBase *grid) {return 1;}
+#endif
+  
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
+					 const Lattice<vobj> &rhs,
+					 cobj *buffer,
+					 compressor &compress,
+					 int off,int so,int partial)
+  {
+    //    std::cout << " face gather simple DWF partial "<<partial <<std::endl;
+    if(partial) FaceGatherPartialDWF::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
+    else        FaceGatherSimple::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
+  }
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    compressor &compress,int type,int partial)
+  {
+    //    std::cout << " face gather exch DWF partial "<<partial <<std::endl;
+    if(partial) FaceGatherPartialDWF::Gather_plane_exchange(table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
+    else        FaceGatherSimple::Gather_plane_exchange    (table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
+  }
+  template<class decompressor,class Merger>
+  static void MergeFace(decompressor decompress,Merger &mm)
+  {
+    int partial = mm.partial;
+    //    std::cout << " merge DWF partial "<<partial <<std::endl;
+    if ( partial ) FaceGatherPartialDWF::MergeFace(decompress,mm);
+    else           FaceGatherSimple::MergeFace(decompress,mm);
+  }
+
+  template<class decompressor,class Decompression>
+  static void DecompressFace(decompressor decompress,Decompression &dd)
+  {
+    int partial = dd.partial;
+    //    std::cout << " decompress DWF partial "<<partial <<std::endl;
+    if ( partial ) FaceGatherPartialDWF::DecompressFace(decompress,dd);
+    else           FaceGatherSimple::DecompressFace(decompress,dd);
+  }
+};
+
 /////////////////////////////////////////////////////////////////////////////////////////////
-// optimised versions supporting half precision too
+// optimised versions supporting half precision too??? Deprecate
 /////////////////////////////////////////////////////////////////////////////////////////////
 
-template<class _HCspinor,class _Hspinor,class _Spinor, class projector,typename SFINAE = void >
-class WilsonCompressorTemplate;
-
 
+//Could make FaceGather a template param, but then behaviour is runtime not compile time
 template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
-class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
-				typename std::enable_if<std::is_same<_HCspinor,_Hspinor>::value>::type >
+class WilsonCompressorTemplate  : public FaceGatherDWFMixedBCs
+//  : public FaceGatherSimple
 {
 public:
   
@@ -79,172 +280,81 @@ public:
   /*****************************************************/
   /* Exchange includes precision change if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Exchange(SiteHalfSpinor *mp,
-				   const SiteHalfSpinor * __restrict__ vp0,
-				   const SiteHalfSpinor * __restrict__ vp1,
-				   Integer type,Integer o) const {
+  accelerator_inline void Exchange(SiteHalfSpinor &mp0,
+				   SiteHalfSpinor &mp1,
+				   const SiteHalfSpinor & vp0,
+				   const SiteHalfSpinor & vp1,
+				   Integer type) const {
 #ifdef GRID_SIMT
-    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+    exchangeSIMT(mp0,mp1,vp0,vp1,type);
 #else
     SiteHalfSpinor tmp1;
     SiteHalfSpinor tmp2;
-    exchange(tmp1,tmp2,vp0[o],vp1[o],type);
-    vstream(mp[2*o  ],tmp1);
-    vstream(mp[2*o+1],tmp2);
+    exchange(tmp1,tmp2,vp0,vp1,type);
+    vstream(mp0,tmp1);
+    vstream(mp1,tmp2);
 #endif
   }
-
+  
 
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
-				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
-    assert(0);
+  accelerator_inline void Decompress(SiteHalfSpinor &out,
+				     SiteHalfSpinor &in) const {    
+    out = in;
   }
 
   /*****************************************************/
   /* Compress Exchange                                 */
   /*****************************************************/
-  accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
-					   SiteHalfSpinor * __restrict__ out1,
-					   const SiteSpinor * __restrict__ in,
-					   Integer j,Integer k, Integer m,Integer type) const
+  accelerator_inline void CompressExchange(SiteHalfSpinor &out0,
+					   SiteHalfSpinor &out1,
+					   const SiteSpinor &in0,
+					   const SiteSpinor &in1,
+					   Integer type) const
   {
 #ifdef GRID_SIMT
     typedef SiteSpinor vobj;
     typedef SiteHalfSpinor hvobj;
-    typedef decltype(coalescedRead(*in))    sobj;
-    typedef decltype(coalescedRead(*out0)) hsobj;
+    typedef decltype(coalescedRead(in0))    sobj;
+    typedef decltype(coalescedRead(out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
-    const vobj *vp1 = &in[m];
+    const vobj *vp0 = &in0;
+    const vobj *vp1 = &in1;
     const vobj *vp = (lane&mask) ? vp1:vp0;
     auto sa = coalescedRead(*vp,j0);
     auto sb = coalescedRead(*vp,j1);
     hsobj psa, psb;
     projector::Proj(psa,sa,mu,dag);
     projector::Proj(psb,sb,mu,dag);
-    coalescedWrite(out0[j],psa);
-    coalescedWrite(out1[j],psb);
+    coalescedWrite(out0,psa);
+    coalescedWrite(out1,psb);
 #else
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
-    projector::Proj(temp1,in[k],mu,dag);
-    projector::Proj(temp2,in[m],mu,dag);
+    projector::Proj(temp1,in0,mu,dag);
+    projector::Proj(temp2,in1,mu,dag);
     exchange(temp3,temp4,temp1,temp2,type);
-    vstream(out0[j],temp3);
-    vstream(out1[j],temp4);
+    vstream(out0,temp3);
+    vstream(out1,temp4);
 #endif
   }
 
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return false; }
+  accelerator_inline bool DecompressionStep(void) const {
+    return false;
+  }
 
 };
 
-#if 0
-template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
-class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
-				typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
-{
-public:
-  
-  int mu,dag;  
-
-  void Point(int p) { mu=p; };
-
-  WilsonCompressorTemplate(int _dag=0){
-    dag = _dag;
-  }
-
-  typedef _Spinor         SiteSpinor;
-  typedef _Hspinor     SiteHalfSpinor;
-  typedef _HCspinor SiteHalfCommSpinor;
-  typedef typename SiteHalfCommSpinor::vector_type vComplexLow;
-  typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
-  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
-
-  accelerator_inline int CommDatumSize(void) const {
-    return sizeof(SiteHalfCommSpinor);
-  }
-
-  /*****************************************************/
-  /* Compress includes precision change if mpi data is not same */
-  /*****************************************************/
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-    SiteHalfSpinor hsp;
-    SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
-    projector::Proj(hsp,in,mu,dag);
-    precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
-  }
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-#ifdef GRID_SIMT
-    typedef decltype(coalescedRead(buf)) sobj;
-    sobj sp;
-    auto sin = coalescedRead(in);
-    projector::Proj(sp,sin,mu,dag);
-    coalescedWrite(buf,sp);
-#else
-    projector::Proj(buf,in,mu,dag);
-#endif
-  }
-
-  /*****************************************************/
-  /* Exchange includes precision change if mpi data is not same */
-  /*****************************************************/
-  accelerator_inline void Exchange(SiteHalfSpinor *mp,
-                       SiteHalfSpinor *vp0,
-                       SiteHalfSpinor *vp1,
-		       Integer type,Integer o) const {
-    SiteHalfSpinor vt0,vt1;
-    SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
-    SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
-    precisionChange((vComplexHigh *)&vt0,(vComplexLow *)&vpp0[o],Nw);
-    precisionChange((vComplexHigh *)&vt1,(vComplexLow *)&vpp1[o],Nw);
-    exchange(mp[2*o],mp[2*o+1],vt0,vt1,type);
-  }
-
-  /*****************************************************/
-  /* Have a decompression step if mpi data is not same */
-  /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
-    SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
-    precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
-  }
-
-  /*****************************************************/
-  /* Compress Exchange                                 */
-  /*****************************************************/
-  accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
-			       SiteHalfSpinor *out1,
-			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type) const {
-    SiteHalfSpinor temp1, temp2,temp3,temp4;
-    SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
-    SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
-    projector::Proj(temp1,in[k],mu,dag);
-    projector::Proj(temp2,in[m],mu,dag);
-    exchange(temp3,temp4,temp1,temp2,type);
-    precisionChange((vComplexLow *)&hout0[j],(vComplexHigh *)&temp3,Nw);
-    precisionChange((vComplexLow *)&hout1[j],(vComplexHigh *)&temp4,Nw);
-  }
-
-  /*****************************************************/
-  /* Pass the info to the stencil */
-  /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return true; }
-
-};
-#endif
-
 #define DECLARE_PROJ(Projector,Compressor,spProj)			\
   class Projector {							\
   public:								\
@@ -294,11 +404,7 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  void ZeroCountersi(void)  {  }
-  void Reporti(int calls)  {  }
-
-  std::vector<int> surface_list;
-
+  //  Vector<int> surface_list;
   WilsonStencil(GridBase *grid,
 		int npoints,
 		int checkerboard,
@@ -306,11 +412,11 @@ public:
 		const std::vector<int> &distances,Parameters p)  
     : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
   { 
-    ZeroCountersi();
-    surface_list.resize(0);
+    //    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
 
+  /*
   void BuildSurfaceList(int Ls,int vol4){
 
     // find same node for SHM
@@ -331,7 +437,8 @@ public:
       }
     }
   }
-
+  */
+  
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
   {
@@ -377,24 +484,26 @@ public:
 
     int dag = compress.dag;
     int face_idx=0;
+#define vet_same_node(a,b) \
+      { auto tmp = b;  }
     if ( dag ) { 
-      assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
-      assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
-      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
-      assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
-      assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
-      assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
-      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
-      assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
+      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
+      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
+      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
+      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
+      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
+      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
+      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
+      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
     } else {
-      assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
-      assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
-      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
-      assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
-      assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
-      assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
-      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
-      assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
+      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
+      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
+      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
+      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
+      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
+      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
+      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
+      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);

Grid/qcd/action/fermion/WilsonFermion.h

@@ -74,20 +74,6 @@ 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

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,19 +75,8 @@ 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;
+  int Dirichlet;
+  Coordinate Block; 
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
@@ -173,7 +162,10 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-    
+
+  virtual void DirichletBlock(const Coordinate & block)
+  {
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

Grid/qcd/action/fermion/WilsonImpl.h

@@ -37,7 +37,7 @@ NAMESPACE_BEGIN(Grid);
 template <class S, class Representation = FundamentalRepresentation,class Options = CoeffReal >
 class WilsonImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation::Dimension > > {
 public:
-
+  
   static const int Dimension = Representation::Dimension;
   static const bool isFundamental = Representation::isFundamental;
   static const bool LsVectorised=false;
@@ -242,19 +242,13 @@ public:
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR;  // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
-
-//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
-//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
-//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
+typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffReal > WilsonImplD2;  // Double
 
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
+typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffComplex > ZWilsonImplD2; // Double
 
-//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
-//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
-//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
- 
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float
 typedef WilsonImpl<vComplexD, AdjointRepresentation, CoeffReal > WilsonAdjImplD;  // Double
@@ -267,6 +261,22 @@ typedef WilsonImpl<vComplex,  TwoIndexAntiSymmetricRepresentation, CoeffReal > W
 typedef WilsonImpl<vComplexF, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplF;  // Float
 typedef WilsonImpl<vComplexD, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplD;  // Double
 
+//sp 2n
+
+typedef WilsonImpl<vComplex,  SpFundamentalRepresentation, CoeffReal > SpWilsonImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpFundamentalRepresentation, CoeffReal > SpWilsonImplF;  // Float
+typedef WilsonImpl<vComplexD, SpFundamentalRepresentation, CoeffReal > SpWilsonImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplF;  // Float
+typedef WilsonImpl<vComplexD, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplF;  // Float
+typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplR;  // Real.. whichever prec    // adj = 2indx symmetric for Sp(2N)
+typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplF;  // Float     // adj = 2indx symmetric for Sp(2N)
+typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplD;  // Double    // adj = 2indx symmetric for Sp(2N)
 
 NAMESPACE_END(Grid);
-

Grid/qcd/action/fermion/WilsonKernels.h

@@ -52,13 +52,6 @@ public:
   typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;   
 public:
 
-#ifdef GRID_SYCL
-#define SYCL_HACK
-#endif  
-#ifdef SYCL_HACK
-  static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
-			       int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
-#endif
   
   static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,

Grid/qcd/action/fermion/WilsonTMFermion.h

@@ -63,7 +63,9 @@ public:
   virtual void MooeeDag(const FermionField &in, FermionField &out) ;
   virtual void MooeeInv(const FermionField &in, FermionField &out) ;
   virtual void MooeeInvDag(const FermionField &in, FermionField &out) ;
-
+  virtual void M(const FermionField &in, FermionField &out) ;
+  virtual void Mdag(const FermionField &in, FermionField &out) ;
+  
 private:
   RealD mu; // TwistedMass parameter