Merge pull request #409 from giltirn/feature/dirichlet-gparity-stage

Import round 5

Grid/DisableWarnings.h

@@ -34,9 +34,6 @@ directory
 
 #if defined __GNUC__ && __GNUC__>=6
 #pragma GCC diagnostic ignored "-Wignored-attributes"
-#endif
-#if defined __GNUC__ 
-#pragma GCC diagnostic ignored "-Wpsabi"
 #endif
 
  //disables and intel compiler specific warning (in json.hpp)
@@ -47,14 +44,22 @@ directory
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
+
+#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+ //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/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
+#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#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__")

Grid/algorithms/approx/Chebyshev.h

@@ -292,7 +292,6 @@ public:
 template<class Field>
 class ChebyshevLanczos : public Chebyshev<Field> {
 private:
-
   std::vector<RealD> Coeffs;
   int order;
   RealD alpha;

Grid/algorithms/iterative/ConjugateGradient.h

@@ -120,6 +120,9 @@ public:
     SolverTimer.Start();
     int k;
     for (k = 1; k <= MaxIterations; k++) {
+
+      GridStopWatch IterationTimer;
+      IterationTimer.Start();
       c = cp;
 
       MatrixTimer.Start();
@@ -152,8 +155,14 @@ 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) {
@@ -170,13 +179,13 @@ public:
 		  << "\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;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -82,11 +82,6 @@ NAMESPACE_BEGIN(Grid);
     RealD stop = src_norm * Tolerance*Tolerance;
 
     GridBase* DoublePrecGrid = src_d_in.Grid();
-
-    //Generate precision change workspaces
-    precisionChangeWorkspace wk_dp_from_sp(DoublePrecGrid, SinglePrecGrid);
-    precisionChangeWorkspace wk_sp_from_dp(SinglePrecGrid, DoublePrecGrid);
-
     FieldD tmp_d(DoublePrecGrid);
     tmp_d.Checkerboard() = cb;
     
@@ -128,7 +123,7 @@ 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, wk_sp_from_dp);
+      precisionChange(src_f, src_d);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -147,7 +142,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f, wk_dp_from_sp);
+      precisionChange(tmp_d, sol_f);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);

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
@@ -325,7 +325,7 @@ 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 << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -128,8 +128,6 @@ public:
   void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
   { 
     GridBase *DoublePrecGrid = src_d.Grid();
-    precisionChangeWorkspace wk_f_from_d(SinglePrecGrid, DoublePrecGrid);
-    precisionChangeWorkspace wk_d_from_f(DoublePrecGrid, SinglePrecGrid);
 
     ////////////////////////////////////////////////////////////////////////
     // Convenience references to the info stored in "MultiShiftFunction"
@@ -170,7 +168,7 @@ public:
     FieldF tmp_f(SinglePrecGrid);
     FieldF mmp_f(SinglePrecGrid);
     FieldF src_f(SinglePrecGrid);
-    precisionChange(src_f, src_d, wk_f_from_d);
+    precisionChange(src_f, src_d);
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
@@ -245,7 +243,7 @@ public:
 
       //Update double precision search direction by residual
       PrecChangeTimer.Start();
-      precisionChange(r_d, r_f, wk_d_from_f);
+      precisionChange(r_d, r_f);
       PrecChangeTimer.Stop();
 
       AXPYTimer.Start();
@@ -264,7 +262,7 @@ public:
       AXPYTimer.Stop();
 
       PrecChangeTimer.Start();
-      precisionChange(p_f, p_d, wk_f_from_d); //get back single prec search direction for linop
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
       PrecChangeTimer.Stop();
 
       cp=c;
@@ -327,7 +325,7 @@ public:
 	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
 	
 	PrecChangeTimer.Start();
-	precisionChange(r_f, r_d, wk_f_from_d);
+	precisionChange(r_f, r_d);
 	PrecChangeTimer.Stop();
 	c = c_d;
       }

Grid/algorithms/iterative/Deflation.h

@@ -113,6 +113,42 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
+
+  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
+    int Nevec = (int)evec_coarse.size();
+    int Nsrc = (int)src.size();
+    // make temp variables
+    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
+    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
+    //Preporcessing
+    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    guess_coarse[j] = Zero();
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockProject(src_coarse[j],src[j],subspace);
+    }
+    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
+    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
+    for (int i=0;i<Nevec;i++)
+    {
+      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
+      const CoarseField & tmp = evec_coarse[i];
+      for (int j=0;j<Nsrc;j++)
+      {
+        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
+      }
+    }
+    //postprocessing
+    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
+    for (int j=0;j<Nsrc;j++)
+    {
+    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+    blockPromote(guess_coarse[j],guess[j],subspace);
+    guess[j].Checkerboard() = src[j].Checkerboard();
+    }
+  };
+
   };
 
 

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -147,13 +147,20 @@ public:
   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)
@@ -179,6 +186,12 @@ 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;
@@ -409,7 +422,7 @@ public:
     //////////////////////////////////////////////////////////////////////////////////////////////////
 
     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); 
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1); 
 
     evals_coarse.resize(Nm);
     evec_coarse.resize(Nm,_CoarseGrid);

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;
@@ -97,6 +98,7 @@ 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)    ;
 
@@ -142,16 +144,16 @@ 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 recv_from_rank,int do_recv,
 				    int bytes,int dir);
   
   

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
@@ -335,22 +337,22 @@ 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,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,
+							 int dest,int dox,
 							 void *recv,
-							 int from,
+							 int from,int dor,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -370,6 +372,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
+  if ( dor ) {
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
       ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
@@ -377,7 +380,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(rrq);
       off_node_bytes+=bytes;
     }
+  }
   
+  if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
       ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
@@ -385,15 +390,15 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       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);
-    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
     }
+  }
   
   if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);
+    list.resize(0);
   }
 
   return off_node_bytes;

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);
@@ -111,18 +113,18 @@ 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 recv_from_rank,int dor,
 							 int bytes, int dir)
 {
   return 2.0*bytes;

Grid/communicator/SharedMemory.h

@@ -93,9 +93,10 @@ 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

Grid/communicator/SharedMemoryMPI.cc

@@ -152,7 +152,7 @@ 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
@@ -165,8 +165,8 @@ 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)
 {
@@ -221,7 +221,7 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
     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.
@@ -294,6 +294,7 @@ 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 +342,7 @@ 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 +354,8 @@ 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
   ////////////////////////////////////////////////////////////////

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/lattice/Lattice_crc.h

@@ -29,6 +29,19 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 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(Lattice<vobj> & buf)
 {
   autoView( buf_v , buf, CpuRead);

Grid/lattice/Lattice_peekpoke.h

@@ -125,6 +125,12 @@ 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){
         

Grid/lattice/Lattice_reduction.h

@@ -142,6 +142,15 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
   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)
+  return sumD_gpu_large(arg,osites);
+#else
+  return sumD_cpu(arg,osites);
+#endif  
+}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
@@ -159,6 +168,22 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
   return ssum;
 }
 
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  autoView( arg_v, arg, AcceleratorRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_gpu_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
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Deterministic Reduction operations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -207,6 +232,7 @@ 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];
@@ -216,15 +242,31 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
     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],innerProductD(x_l,y_l));
     });
   }
+#else
+  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
     
+  {
+    autoView( left_v , left, AcceleratorRead);
+    autoView( right_v,right, AcceleratorRead);
+
+    // GPU - SIMT lane compliance...
+    accelerator_for( ss, sites, nsimd,{
+	auto x_l = left_v(ss);
+	auto y_l = right_v(ss);
+	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
+    });
+  }
+#endif
   // This is in single precision and fails some tests
-  auto anrm = sum(inner_tmp_v,sites);  
+  auto anrm = sumD(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
@@ -258,7 +300,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   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 +312,29 @@ 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; 
 }

Grid/lattice/Lattice_reduction_gpu.h

@@ -23,7 +23,7 @@ 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
@@ -37,13 +37,13 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  
+  /*  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  
+  */  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -53,12 +53,12 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   threads = warpSize;
   if ( threads*sizeofsobj > sharedMemPerBlock ) {
     std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
-    exit(EXIT_FAILURE);
+    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>
@@ -198,7 +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;
@@ -207,7 +207,9 @@ 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);
+  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  assert(ok);
+
   Integer smemSize = numThreads * sizeof(sobj);
 
   Vector<sobj> buffer(numBlocks);
@@ -218,6 +220,54 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
   auto result = buffer_v[0];
   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::vector_type  vector;
+  typedef typename vobj::scalar_typeD scalarD;
+  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
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -230,6 +280,13 @@ 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_rng.h

@@ -32,9 +32,8 @@
 #include <random>
 
 #ifdef RNG_SITMO
-#include <Grid/random/sitmo_prng_engine.hpp>
+#include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
 #endif 
-#include <Grid/random/gaussian.h>
 
 #if defined(RNG_SITMO)
 #define RNG_FAST_DISCARD
@@ -143,8 +142,8 @@ public:
 
   std::vector<RngEngine>                             _generators;
   std::vector<std::uniform_real_distribution<RealD> > _uniform;
-  std::vector<Grid::gaussian_distribution<RealD> >       _gaussian;
-  //  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
+  std::vector<std::normal_distribution<RealD> >       _gaussian;
+  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
   std::vector<std::uniform_int_distribution<uint32_t> > _uid;
 
   ///////////////////////
@@ -244,8 +243,8 @@ public:
   GridSerialRNG() : GridRNGbase() {
     _generators.resize(1);
     _uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(1,gaussian_distribution<RealD>(0.0,1.0) );
-    //    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
+    _gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
+    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(1,std::uniform_int_distribution<uint32_t>() );
   }
 
@@ -358,8 +357,8 @@ public:
 
     _generators.resize(_vol);
     _uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(_vol,gaussian_distribution<RealD>(0.0,1.0) );
-    //    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
+    _gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
+    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
   }
 
@@ -425,9 +424,32 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
+#if 1
+    thread_for( lidx, _grid->lSites(), {
 
+	int gidx;
+	int o_idx;
+	int i_idx;
+	int rank;
+	Coordinate pcoor;
+	Coordinate lcoor;
+	Coordinate gcoor;
+	_grid->LocalIndexToLocalCoor(lidx,lcoor);
+	pcoor=_grid->ThisProcessorCoor();
+	_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
+	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
+
+	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
+	assert(rank == _grid->ThisRank() );
+	
+	int l_idx=generator_idx(o_idx,i_idx);
+	_generators[l_idx] = master_engine;
+	Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
+    });
+#else
     // Everybody loops over global volume.
     thread_for( gidx, _grid->_gsites, {
+
 	// Where is it?
 	int rank;
 	int o_idx;
@@ -444,6 +466,7 @@ public:
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
     });
+#endif
 #else 
     ////////////////////////////////////////////////////////////////
     // Machine and thread decomposition dependent seeding is efficient
@@ -516,11 +539,11 @@ public:
 
 template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l)   { rng.fill(l,rng._uniform);  }
 template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
-//template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
+template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
 
 template <class sobj> inline void random(GridSerialRNG &rng,sobj &l)   { rng.fill(l,rng._uniform  ); }
 template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
-//template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
+template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
 
 NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_transfer.h

@@ -1080,95 +1080,53 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//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
-public:
-  precisionChangeWorkspace(GridBase *out_grid, GridBase *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; }
-
-  ~precisionChangeWorkspace(){
-    acceleratorFreeDevice(fmap_device);
-  }
-};
-
-
-//Convert a lattice of one precision to another. The input workspace contains the mapping data.
-template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
-  static_assert( std::is_same<typename VobjOut::DoublePrecision, typename VobjIn::DoublePrecision>::value == 1, "copyLane: 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();
-
-  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
-//Generate the workspace in place; if multiple calls with the same mapping are performed, consider pregenerating the workspace and reusing
 template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
-  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
-  precisionChange(out, in, workspace);
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+{
+  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]);
+  }
+  out.Checkerboard() = in.Checkerboard();
+  GridBase *in_grid=in.Grid();
+  GridBase *out_grid = out.Grid();
+
+  typedef typename VobjOut::scalar_object SobjOut;
+  typedef typename VobjIn::scalar_object SobjIn;
+
+  int ndim = out.Grid()->Nd();
+  int out_nsimd = out_grid->Nsimd();
+    
+  std::vector<Coordinate > out_icoor(out_nsimd);
+      
+  for(int lane=0; lane < out_nsimd; lane++){
+    out_icoor[lane].resize(ndim);
+    out_grid->iCoorFromIindex(out_icoor[lane], lane);
   }
         
+  std::vector<SobjOut> in_slex_conv(in_grid->lSites());
+  unvectorizeToLexOrdArray(in_slex_conv, in);
+    
+  autoView( out_v , out, CpuWrite);
+  thread_for(out_oidx,out_grid->oSites(),{
+    Coordinate out_ocoor(ndim);
+    out_grid->oCoorFromOindex(out_ocoor, out_oidx);
+
+    ExtractPointerArray<SobjOut> ptrs(out_nsimd);      
+
+    Coordinate lcoor(out_grid->Nd());
+      
+    for(int lane=0; lane < out_nsimd; lane++){
+      for(int mu=0;mu<ndim;mu++)
+	lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
+	
+      int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
+      ptrs[lane] = &in_slex_conv[llex];
+    }
+    merge(out_v[out_oidx], ptrs, 0);
+  });
+}
 
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids

Grid/log/Log.cc

@@ -65,6 +65,7 @@ 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 GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
@@ -72,9 +73,10 @@ 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); // at least the messages should be always on
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
@@ -83,7 +85,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   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("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);

Grid/log/Log.h

@@ -183,6 +183,7 @@ extern GridLogger GridLogPerformance;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
 extern GridLogger GridLogHMC;
+extern GridLogger GridLogMemory;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/IldgIO.h

@@ -31,6 +31,7 @@ directory
 #include <fstream>
 #include <iomanip>
 #include <iostream>
+#include <string>
 #include <map>
 
 #include <pwd.h>
@@ -654,7 +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;
@@ -871,7 +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
@@ -879,7 +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
@@ -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++){
+    #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;
@@ -147,15 +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,
@@ -166,7 +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,
@@ -211,27 +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;
@@ -245,10 +252,14 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write 3x3 double
+    // Sod it -- always write NcxNc double
     header.floating_point  = std::string("IEEE64BIG");
-    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
-    GaugeSimpleUnmunger<fobj3D,sobj> munge;
+    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);
@@ -256,8 +267,15 @@ public:
     grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t 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);
@@ -290,7 +308,6 @@ public:
     MachineCharacteristics(header);
 
     uint64_t offset;
-  
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");

Grid/perfmon/PerfCount.h

@@ -72,17 +72,9 @@ 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

@@ -39,9 +39,9 @@ NAMESPACE_BEGIN(Grid)
 // C++11 time facilities better?
 inline double usecond(void) {
   struct timeval tv;
-#ifdef TIMERS_ON
+  tv.tv_sec = 0;
+  tv.tv_usec = 0;
   gettimeofday(&tv,NULL);
-#endif
   return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
 }
 

Grid/pugixml/pugixml.cc

@@ -16,8 +16,12 @@
 
 #ifdef __NVCC__
 #pragma push
+#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#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/action/ActionBase.h

@@ -40,6 +40,29 @@ class Action
 
 public:
   bool is_smeared = false;
+  RealD deriv_norm_sum;
+  RealD deriv_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_num=0;
+    deriv_norm_sum = deriv_max_sum=0.0;
+  }
+  void  deriv_log(RealD nrm, RealD max) { deriv_max_sum+=max; deriv_norm_sum+=nrm; deriv_num++;}
+  RealD deriv_max_average(void)         { return deriv_max_sum/deriv_num; };
+  RealD deriv_norm_average(void)        { return deriv_norm_sum/deriv_num; };
+  RealD deriv_timer(void)        { return deriv_us; };
+  RealD S_timer(void)            { return deriv_us; };
+  RealD refresh_timer(void)      { return deriv_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

Grid/qcd/action/ActionCore.h

@@ -37,6 +37,10 @@ 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

@@ -36,26 +36,34 @@ NAMESPACE_BEGIN(Grid);
 
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
-  Coordinate twists; //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
+  Coordinate twists;
                      //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
-  GparityWilsonImplParams() : twists(Nd, 0) {};
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  GparityWilsonImplParams() : twists(Nd, 0), dirichlet(Nd, 0) {};
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
+    dirichlet.resize(Nd,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);
+    dirichlet.resize(Nd,0);
   }
 };
 
 struct StaggeredImplParams {
-  StaggeredImplParams()  {};
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  StaggeredImplParams()
+  {
+    dirichlet.resize(Nd,0);
+  };
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -64,6 +72,7 @@ struct StaggeredImplParams {
 				    RealD, hi, 
 				    int,   MaxIter, 
 				    RealD, tolerance, 
+				    RealD, mdtolerance, 
 				    int,   degree, 
 				    int,   precision,
 				    int,   BoundsCheckFreq,
@@ -79,18 +88,19 @@ struct StaggeredImplParams {
                            	int _degree    = 10,
 				int _precision = 64,
 				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),
         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.
@@ -133,7 +143,6 @@ struct StaggeredImplParams {
   };
 
 
-  
 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;

Grid/qcd/action/fermion/CloverHelpers.h

@@ -0,0 +1,435 @@
+/*************************************************************************************
+
+    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<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 MassTerm(CloverField& Clover, RealD diag_mass) {
+    Clover += diag_mass;
+  }
+
+  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
+                          CloverTriangleField& Triangle,
+                          RealD csw_t, RealD diag_mass) {
+
+    // Do nothing
+  }
+
+  // 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;
+
+  static void MassTerm(CloverField& Clover, RealD diag_mass) {
+    // do nothing!
+    // mass term is multiplied to exp(Clover) below
+  }
+
+  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<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+
+  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
+
+  	  typedef iMatrix<ComplexD,6> mat;
+
+  	  RealD qn[6];
+  	  RealD qnold[6];
+  	  RealD p[5];
+  	  RealD trA2, trA3, trA4;
+
+  	  mat A2, A3, A4, A5;
+  	  A2 = alpha * alpha * arg * arg;
+  	  A3 = alpha * arg * A2;
+  	  A4 = A2 * A2;
+  	  A5 = A2 * A3;
+
+  	  trA2 = toReal( trace(A2) );
+  	  trA3 = toReal( trace(A3) );
+  	  trA4 = toReal( trace(A4));
+
+  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
+  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
+  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
+  	  p[3] = trA3 / 3.0;
+  	  p[4] = 0.5 * trA2;
+
+  	  qnold[0] = cN[Niter];
+  	  qnold[1] = 0.0;
+  	  qnold[2] = 0.0;
+  	  qnold[3] = 0.0;
+  	  qnold[4] = 0.0;
+  	  qnold[5] = 0.0;
+
+  	  for(int i = Niter-1; i >= 0; i--)
+  	  {
+  	   qn[0] = p[0] * qnold[5] + cN[i];
+  	   qn[1] = p[1] * qnold[5] + qnold[0];
+  	   qn[2] = p[2] * qnold[5] + qnold[1];
+  	   qn[3] = p[3] * qnold[5] + qnold[2];
+  	   qn[4] = p[4] * qnold[5] + qnold[3];
+  	   qn[5] = qnold[4];
+
+  	   qnold[0] = qn[0];
+  	   qnold[1] = qn[1];
+  	   qnold[2] = qn[2];
+  	   qnold[3] = qn[3];
+  	   qnold[4] = qn[4];
+  	   qnold[5] = qn[5];
+  	  }
+
+  	  mat unit(1.0);
+
+  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
+
+    }
+
+  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
+
+    GridBase* grid = Diagonal.Grid();
+    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
+
+    //
+    // Implementation completely in Daniel's layout
+    //
+
+    // Taylor expansion with Cayley-Hamilton recursion
+    // underlying Horner scheme as above
+    std::vector<RealD> cn(NMAX+1);
+    cn[0] = 1.0;
+    for (int i=1; i<=NMAX; i++){
+      cn[i] = cn[i-1] / RealD(i);
+    }
+
+      // Taken over from Daniel's implementation
+      conformable(Diagonal, Triangle);
+
+      long lsites = grid->lSites();
+    {
+      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
+      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
+      typedef iMatrix<ComplexD,6> mat;
+
+      autoView(diagonal_v,  Diagonal,  CpuRead);
+      autoView(triangle_v,  Triangle,  CpuRead);
+      autoView(diagonalExp_v, Diagonal, CpuWrite);
+      autoView(triangleExp_v, Triangle, CpuWrite);
+
+      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
+
+    	  mat srcCloverOpUL(0.0); // upper left block
+    	  mat srcCloverOpLR(0.0); // lower right block
+    	  mat ExpCloverOp;
+
+        scalar_object_diagonal diagonal_tmp     = Zero();
+        scalar_object_diagonal diagonal_exp_tmp = Zero();
+        scalar_object_triangle triangle_tmp     = Zero();
+        scalar_object_triangle triangle_exp_tmp = Zero();
+
+        Coordinate lcoor;
+        grid->LocalIndexToLocalCoor(site, lcoor);
+
+        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
+        peekLocalSite(triangle_tmp, triangle_v, lcoor);
+
+        int block;
+        block = 0;
+        for(int i = 0; i < 6; i++){
+        	for(int j = 0; j < 6; j++){
+        		if (i == j){
+        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
+        		}
+        		else{
+        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
+        		}
+        	}
+        }
+        block = 1;
+        for(int i = 0; i < 6; i++){
+          	for(int j = 0; j < 6; j++){
+           		if (i == j){
+           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
+           		}
+           		else{
+           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
+           		}
+            }
+        }
+
+        // exp(Clover)
+
+        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
+
+        block = 0;
+        for(int i = 0; i < 6; i++){
+        	for(int j = 0; j < 6; j++){
+            	if (i == j){
+            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
+            	}
+            	else if(i < j){
+            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
+            	}
+           	}
+        }
+
+        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
+
+        block = 1;
+        for(int i = 0; i < 6; i++){
+        	for(int j = 0; j < 6; j++){
+              	if (i == j){
+              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
+               	}
+               	else if(i < j){
+               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
+               	}
+            }
+        }
+
+        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
+        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
+      });
+    }
+
+    Diagonal *= diag_mass;
+    Triangle *= diag_mass;
+  }
+
+
+  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

@@ -31,6 +31,7 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -85,7 +86,7 @@ NAMESPACE_BEGIN(Grid);
 //           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
 //                                                                 =  84 complex words per site
 
-template<class Impl>
+template<class Impl, class CloverHelpers>
 class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
                                    public WilsonCloverHelpers<Impl>,
                                    public CompactWilsonCloverHelpers<Impl> {

Grid/qcd/action/fermion/Fermion.h

@@ -138,38 +138,52 @@ 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<WilsonImplR> WilsonCloverFermionR;
+typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
+typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
+typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
+typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
 
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
+typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
+
+typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
+typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
+typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
+
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Compact Clover fermions
-typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
-typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
-typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
+template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
+template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
 
-typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
-typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
-typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
+typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
+typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
 
-typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
+typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
+typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
+typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
 
-typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
+typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
+
+typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
+typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
+typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
+
+typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Domain Wall fermions
 typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;

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/WilsonCloverFermion.h

@@ -32,6 +32,7 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -51,7 +52,7 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template <class Impl>
+template<class Impl, class CloverHelpers>
 class WilsonCloverFermion : public WilsonFermion<Impl>,
                             public WilsonCloverHelpers<Impl>
 {

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -209,6 +209,8 @@ public:
 };
 
 
+////////////////////////////////////////////////////////
+
 template<class Impl> class CompactWilsonCloverHelpers {
 public:
 

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -47,8 +47,6 @@ class CompactWilsonCloverTypes {
 public:
   INHERIT_IMPL_TYPES(Impl);
 
-  static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
-
   static constexpr int Nred      = Nc * Nhs;        // 6
   static constexpr int Nblock    = Nhs;             // 2
   static constexpr int Ndiagonal = Nred;            // 6

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -297,7 +297,7 @@ public:
   void ZeroCountersi(void)  {  }
   void Reporti(int calls)  {  }
 
-  std::vector<int> surface_list;
+  //  Vector<int> surface_list;
 
   WilsonStencil(GridBase *grid,
 		int npoints,
@@ -307,10 +307,11 @@ public:
     : 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
@@ -331,6 +332,7 @@ public:
       }
     }
   }
+  */
   
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,6 +75,10 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  int Dirichlet;
+  Coordinate Block; 
+
+  /********** Deprecate timers **********/
   void Report(void);
   void ZeroCounters(void);
   double DhopCalls;
@@ -174,6 +178,9 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
 
+  virtual void DirichletBlock(const Coordinate & block)
+  {
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h

@@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FiveDimRedBlackGrid,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
-  mass(_mass)
+  mass_plus(_mass), mass_minus(_mass)
 { 
 }
 
@@ -209,8 +209,8 @@ void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
   Vector<Coeff_t> diag (Ls,1.0);
-  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
-  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -220,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
   Vector<Coeff_t> diag = bs;
   Vector<Coeff_t> upper= cs;
   Vector<Coeff_t> lower= cs; 
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
 }
 // FIXME Redunant with the above routine; check this and eliminate
@@ -235,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
   }
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -250,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &
     upper[i]=-cee[i];
     lower[i]=-cee[i];
   }
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -266,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &
     // Assemble the 5d matrix
     if ( s==0 ) {
       upper[s] = -cee[s+1] ;
-      lower[s] = mass*cee[Ls-1];
+      lower[s] = mass_minus*cee[Ls-1];
     } else if ( s==(Ls-1)) { 
-      upper[s] = mass*cee[0];
+      upper[s] = mass_plus*cee[0];
       lower[s] = -cee[s-1];
     } else {
       upper[s]=-cee[s+1];
@@ -291,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
   Vector<Coeff_t> diag(Ls,1.0);
   Vector<Coeff_t> upper(Ls,-1.0);
   Vector<Coeff_t> lower(Ls,-1.0);
-  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass*lower[0];
+  upper[Ls-1]=-mass_plus*upper[Ls-1];
+  lower[0]   =-mass_minus*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
 }
 
@@ -307,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
       upper[s] = cs[s+1];
-      lower[s] =-mass*cs[Ls-1];
+      lower[s] =-mass_minus*cs[Ls-1];
     } else if ( s==(Ls-1) ) { 
-      upper[s] =-mass*cs[0];
+      upper[s] =-mass_plus*cs[0];
       lower[s] = cs[s-1];
     } else { 
       upper[s] = cs[s+1];
@@ -552,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
-      leem[i]=mass*cee[Ls-1]/bee[0];
+      leem[i]=mass_minus*cee[Ls-1]/bee[0];
       for(int j=0;j<i;j++) {
 	assert(bee[j+1]!=Coeff_t(0.0));
 	leem[i]*= aee[j]/bee[j+1];
@@ -560,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
-      ueem[i]=mass;
+      ueem[i]=mass_plus;
       for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
       ueem[i]*= aee[0]/bee[0];
       
@@ -573,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
   }
 	
   { 
-    Coeff_t delta_d=mass*cee[Ls-1];
+    Coeff_t delta_d=mass_minus*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
       assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
@@ -642,6 +642,10 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
+
+  assert(mass_plus == mass_minus);
+  RealD mass = mass_plus;
+  
 #if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
@@ -777,6 +781,8 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   assert(mu>=0);
   assert(mu<Nd);
 
+  assert(mass_plus == mass_minus);
+  RealD mass = mass_plus;
 
 #if 0
   int tshift = (mu == Nd-1) ? 1 : 0;

Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h

@@ -66,18 +66,17 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   M5Dcalls++;
   M5Dtime-=usecond();
 
-  uint64_t nloop = grid->oSites()/Ls;
+  uint64_t nloop = grid->oSites();
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t ss= sss*Ls;
+    uint64_t s = sss%Ls;
+    uint64_t ss= sss-s;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1, tmp2;
-    for(int s=0;s<Ls;s++){
     uint64_t idx_u = ss+((s+1)%Ls);
     uint64_t idx_l = ss+((s+Ls-1)%Ls);
     spProj5m(tmp1,psi(idx_u));
     spProj5p(tmp2,psi(idx_l));
     coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
-    }
   });
   M5Dtime+=usecond();
 }
@@ -108,18 +107,17 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   M5Dcalls++;
   M5Dtime-=usecond();
 
-  uint64_t nloop = grid->oSites()/Ls;
+  uint64_t nloop = grid->oSites();
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t ss=sss*Ls;
+    uint64_t s = sss%Ls;
+    uint64_t ss= sss-s;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1,tmp2;
-    for(int s=0;s<Ls;s++){
     uint64_t idx_u = ss+((s+1)%Ls);
     uint64_t idx_l = ss+((s+Ls-1)%Ls);
     spProj5p(tmp1,psi(idx_u));
     spProj5m(tmp2,psi(idx_l));
     coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
-    }
   });
   M5Dtime+=usecond();
 }

Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h

@@ -32,9 +32,10 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 
+
 NAMESPACE_BEGIN(Grid);
-template<class Impl>
-CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
+template<class Impl, class CloverHelpers>
+CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
                                                                             GridCartesian& Fgrid,
                                                                             GridRedBlackCartesian& Hgrid,
                                                                             const RealD _mass,
@@ -58,50 +59,55 @@ CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
   , BoundaryMask(&Fgrid)
   , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
 {
+  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
+
   csw_r *= 0.5;
   csw_t *= 0.5;
   if (clover_anisotropy.isAnisotropic)
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries)
+  if (open_boundaries) {
+    this->BoundaryMaskEven.Checkerboard() = Even;
+    this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
   }
+}
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
   if(this->open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
   if(this->open_boundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
@@ -109,8 +115,8 @@ void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& o
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
@@ -118,20 +124,20 @@ void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
@@ -144,13 +150,13 @@ void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionFiel
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
   Mooee(in, out); // blocks are hermitian
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
@@ -163,23 +169,23 @@ void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionF
   if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
   MooeeInv(in, out); // blocks are hermitian
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
   DhopDir(in, out, dir, disp);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
   DhopDirAll(in, out);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
   assert(!open_boundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
@@ -251,7 +257,7 @@ void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionFi
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);   // checked
+      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
       count++;
     }
 
@@ -261,18 +267,18 @@ void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionFi
   force += clover_force;
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&        in,
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
                     FermionField&              out,
                     const CloverDiagonalField& diagonal,
                     const CloverTriangleField& triangle) {
@@ -285,8 +291,8 @@ void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&
   CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
 }
 
-template<class Impl>
-void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
   // NOTE: parts copied from original implementation
 
   // Import gauge into base class
@@ -318,22 +324,27 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  TmpOriginal += this->diag_mass;
+  // Handle mass term based on clover policy
+  CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
   
   // Convert the data layout of the clover term
   double t4 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Possible modify the boundary values
+  // Exponentiate the clover (nothing happens in case of the standard clover)
   double t5 = usecond();
+  CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
+
+  // Possible modify the boundary values
+  double t6 = usecond();
   if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the clover term in the improved layout
-  double t6 = usecond();
+  // Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
+  double t7 = usecond();
   CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
 
   // Fill the remaining clover fields
-  double t7 = usecond();
+  double t8 = usecond();
   pickCheckerboard(Even, DiagonalEven,    Diagonal);
   pickCheckerboard(Even, TriangleEven,    Triangle);
   pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
@@ -344,20 +355,19 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
 
   // Report timings
-  double t8 = usecond();
-#if 0
-  std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
-            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
-            << ", allocations = "               << (t2 - t1) / 1e6
-            << ", field strength = "            << (t3 - t2) / 1e6
-            << ", fill clover = "               << (t4 - t3) / 1e6
-            << ", convert = "                   << (t5 - t4) / 1e6
-            << ", boundaries = "                << (t6 - t5) / 1e6
-            << ", inversions = "                << (t7 - t6) / 1e6
-            << ", pick cbs = "                  << (t8 - t7) / 1e6
-            << ", total = "                     << (t8 - t0) / 1e6
-            << std::endl;
-#endif
+  double t9 = usecond();
+
+  std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
+  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert =                    " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "exponentiation =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "boundaries =                 " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "inversions =                 " << (t8 - t7) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
 }
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h

@@ -34,8 +34,8 @@
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl>
-WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField&                         _Umu,
+template<class Impl, class CloverHelpers>
+WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField&                         _Umu,
                                                GridCartesian&                      Fgrid,
                                                GridRedBlackCartesian&              Hgrid,
                                                const RealD                         _mass,
@@ -74,8 +74,8 @@ WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField&
 }
 
 // *NOT* EO
-template <class Impl>
-void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -89,8 +89,8 @@ void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
   out += temp;
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -104,8 +104,8 @@ void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
   out += temp;
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
 {
   double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
@@ -131,47 +131,11 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
   CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
   CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
-  CloverTerm += diag_mass;
    
   double t4 = usecond();
-  int lvol = _Umu.Grid()->lSites();
-  int DimRep = Impl::Dimension;
+  CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
 
   double t5 = usecond();
-  {
-    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);
-      //if (csw!=0){
-      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);
-	    }
-      //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
-      
-      EigenInvCloverOp = EigenCloverOp.inverse();
-      //std::cout << EigenInvCloverOp << std::endl;
-      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);
-      //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
-      //  }
-      pokeLocalSite(Qxinv, CTIv, lcoor);
-    });
-  }
-
-  double t6 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@@ -184,48 +148,44 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 
   pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
   pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
-  double t7 = usecond();
+  double t6 = usecond();
 
-#if 0
-  std::cout << GridLogMessage << "WilsonCloverFermion::ImportGauge timings:"
-            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
-            << ", allocations = "               << (t2 - t1) / 1e6
-            << ", field strength = "            << (t3 - t2) / 1e6
-            << ", fill clover = "               << (t4 - t3) / 1e6
-            << ", misc = "                      << (t5 - t4) / 1e6
-            << ", inversions = "                << (t6 - t5) / 1e6
-            << ", pick cbs = "                  << (t7 - t6) / 1e6
-            << ", total = "                     << (t7 - t0) / 1e6
-            << std::endl;
-#endif
+  std::cout << GridLogDebug << "WilsonCloverFermion::ImportGauge timings:" << std::endl;
+  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "instantiation =              " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "pick cbs =                   " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "total =                      " << (t6 - t0) / 1e6 << std::endl;
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseNo);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseNo);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseYes);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseYes);
 }
 
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
 {
   out.Checkerboard() = in.Checkerboard();
   CloverField *Clover;
@@ -278,8 +238,8 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
 } // MooeeInternal
 
 // Derivative parts unpreconditioned pseudofermions
-template <class Impl>
-void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
 {
   conformable(X.Grid(), Y.Grid());
   conformable(X.Grid(), force.Grid());
@@ -349,7 +309,7 @@ void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X,
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);                   // checked
+      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);                   // checked
       count++;
     }
 
@@ -360,15 +320,15 @@ void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X,
 }
 
 // Derivative parts
-template <class Impl>
-void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
 {
   assert(0);
 }
 
 // Derivative parts
-template <class Impl>
-void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
+template<class Impl, class CloverHelpers>
+void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   assert(0); // not implemented yet
 }

Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h

@@ -60,7 +60,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   UmuOdd (_FourDimRedBlackGrid),
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid),
-  _tmp(&FiveDimRedBlackGrid)
+  _tmp(&FiveDimRedBlackGrid),
+  Dirichlet(0)
 {
   // some assertions
   assert(FiveDimGrid._ndimension==5);
@@ -91,6 +92,19 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
     assert(FourDimRedBlackGrid._simd_layout[d]  ==FourDimGrid._simd_layout[d]);
   }
 
+  if ( p.dirichlet.size() == Nd+1) {
+    Coordinate block = p.dirichlet;
+    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
+      Dirichlet = 1;
+      Block = block;
+    }
+  } else {
+    Coordinate block(Nd+1,0);
+    Block = block;
+  }
+
+  ZeroCounters();
+
   if (Impl::LsVectorised) { 
 
     int nsimd = Simd::Nsimd();
@@ -218,6 +232,14 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
+  if ( Dirichlet ) {
+    std::cout << GridLogMessage << " Dirichlet BCs 5d " <<Block<<std::endl;
+    Coordinate GaugeBlock(Nd);
+    for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
+    std::cout << GridLogMessage << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
+    DirichletFilter<GaugeField> Filter(GaugeBlock);
+    Filter.applyFilter(HUmu);
+  }
   Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
   pickCheckerboard(Even,UmuEven,Umu);
   pickCheckerboard(Odd ,UmuOdd,Umu);

Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h

@@ -4,12 +4,13 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
 
-Copyright (C) 2015
+Copyright (C) 2022
 
 Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Fabian Joswig <fabian.joswig@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
@@ -599,11 +600,47 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
                                                    Current curr_type,
                                                    unsigned int mu)
 {
+  if(curr_type != Current::Vector)
+  {
+    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
+    exit(1);
+  }
+
   Gamma g5(Gamma::Algebra::Gamma5);
   conformable(_grid, q_in_1.Grid());
   conformable(_grid, q_in_2.Grid());
   conformable(_grid, q_out.Grid());
-  assert(0);
+  auto UGrid= this->GaugeGrid();
+
+  PropagatorField tmp_shifted(UGrid);
+  PropagatorField g5Lg5(UGrid);
+  PropagatorField R(UGrid);
+  PropagatorField gmuR(UGrid);
+
+    Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT,
+  };
+  Gamma gmu=Gamma(Gmu[mu]);
+
+  g5Lg5=g5*q_in_1*g5;
+  tmp_shifted=Cshift(q_in_2,mu,1);
+  Impl::multLinkField(R,this->Umu,tmp_shifted,mu);
+  gmuR=gmu*R;
+
+  q_out=adj(g5Lg5)*R;
+  q_out-=adj(g5Lg5)*gmuR;
+
+  tmp_shifted=Cshift(q_in_1,mu,1);
+  Impl::multLinkField(g5Lg5,this->Umu,tmp_shifted,mu);
+  g5Lg5=g5*g5Lg5*g5;
+  R=q_in_2;
+  gmuR=gmu*R;
+
+  q_out-=adj(g5Lg5)*R;
+  q_out-=adj(g5Lg5)*gmuR;
 }
 
 
@@ -617,9 +654,51 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
                                               unsigned int tmax,
 					      ComplexField &lattice_cmplx)
 {
+  if(curr_type != Current::Vector)
+  {
+    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
+    exit(1);
+  }
+
+  int tshift = (mu == Nd-1) ? 1 : 0;
+  unsigned int LLt    = GridDefaultLatt()[Tp];
   conformable(_grid, q_in.Grid());
   conformable(_grid, q_out.Grid());
-  assert(0);
+  auto UGrid= this->GaugeGrid();
+
+  PropagatorField tmp(UGrid);
+  PropagatorField Utmp(UGrid);
+  PropagatorField L(UGrid);
+  PropagatorField zz (UGrid);
+  zz=Zero();
+  LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
+
+    Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT,
+  };
+  Gamma gmu=Gamma(Gmu[mu]);
+
+  tmp = Cshift(q_in,mu,1);
+  Impl::multLinkField(Utmp,this->Umu,tmp,mu);
+  tmp = ( Utmp*lattice_cmplx - gmu*Utmp*lattice_cmplx ); // Forward hop
+  tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
+  q_out = where((lcoor<=tmax),tmp,zz); // Position of current complicated
+
+  tmp = q_in *lattice_cmplx;
+  tmp = Cshift(tmp,mu,-1);
+  Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
+  tmp = -( Utmp + gmu*Utmp );
+  // Mask the time
+  if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
+    unsigned int t0 = 0;
+    tmp = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
+  } else {
+    tmp = where((lcoor>=tmin+tshift),tmp,zz);
+  }
+  q_out+= where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
 }
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h

@@ -440,6 +440,17 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
+#define KERNEL_CALL_EXT(A)						\
+  const uint64_t    NN = Nsite*Ls;					\
+  const uint64_t    sz = st.surface_list.size();			\
+  auto ptr = &st.surface_list[0];					\
+  accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
+      int sF = ptr[ss];							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
+    });									\
+  accelerator_barrier();
+
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
     int sU = ss;							\

Grid/qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master

@@ -9,6 +9,7 @@
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@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
@@ -32,10 +33,12 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class CompactWilsonCloverFermion<IMPLEMENTATION>; 
+template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>; 
+template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>; 
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonCloverFermionInstantiation.cc.master

@@ -8,6 +8,7 @@
 
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
+    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
@@ -31,10 +32,12 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class WilsonCloverFermion<IMPLEMENTATION>; 
+template class WilsonCloverFermion<IMPLEMENTATION, CloverHelpers<IMPLEMENTATION>>; 
+template class WilsonCloverFermion<IMPLEMENTATION, ExpCloverHelpers<IMPLEMENTATION>>; 
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

@@ -1,51 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
-
-Copyright (C) 2015, 2020
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
-
-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/qcd/action/fermion/FermionCore.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
-
-#ifndef AVX512
-#ifndef QPX
-#ifndef A64FX
-#ifndef A64FXFIXEDSIZE
-#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
-#endif
-#endif
-#endif
-#endif
-
-NAMESPACE_BEGIN(Grid);
-
-#include "impl.h"
-template class WilsonKernels<IMPLEMENTATION>;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/generate_instantiations.sh

@@ -18,6 +18,10 @@ WILSON_IMPL_LIST=" \
 	   GparityWilsonImplF \
 	   GparityWilsonImplD "
 
+COMPACT_WILSON_IMPL_LIST=" \
+	   WilsonImplF \
+	   WilsonImplD "
+
 DWF_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
@@ -40,7 +44,7 @@ EOF
 
 done
 
-CC_LIST="WilsonCloverFermionInstantiation CompactWilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
+CC_LIST="WilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
 
 for impl in $WILSON_IMPL_LIST
 do
@@ -50,6 +54,16 @@ do
 done
 done
 
+CC_LIST="CompactWilsonCloverFermionInstantiation"
+
+for impl in $COMPACT_WILSON_IMPL_LIST
+do
+for f in $CC_LIST
+do
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+done
+done
+
 CC_LIST=" \
   CayleyFermion5DInstantiation \
   ContinuedFractionFermion5DInstantiation \

Grid/qcd/action/filters/DDHMCFilter.h

@@ -0,0 +1,102 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/DirichletFilter.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+//--------------------------------------------------------------------
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+////////////////////////////////////////////////////
+// DDHMC filter with sub-block size B[mu]
+////////////////////////////////////////////////////
+
+template<typename GaugeField>
+struct DDHMCFilter: public MomentumFilterBase<GaugeField>
+{
+  Coordinate Block;
+  int Width;
+  
+  DDHMCFilter(const Coordinate &_Block,int _Width=2): Block(_Block) { Width=_Width; }
+
+  void applyFilter(GaugeField &U) const override
+  {
+    GridBase *grid = U.Grid();
+    Coordinate Global=grid->GlobalDimensions();
+    GaugeField zzz(grid); zzz = Zero();
+    LatticeInteger coor(grid); 
+    
+    auto zzz_mu = PeekIndex<LorentzIndex>(zzz,0);
+    ////////////////////////////////////////////////////
+    // Zero BDY layers
+    ////////////////////////////////////////////////////
+    std::cout<<GridLogMessage<<" DDHMC Force Filter Block "<<Block<<" width " <<Width<<std::endl;
+    for(int mu=0;mu<Nd;mu++) {
+
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+	LatticeCoordinate(coor,mu);
+
+	////////////////////////////////
+	// OmegaBar - zero all links contained in slice B-1,0 and
+	// mu links connecting to Omega
+	////////////////////////////////
+	if ( Width==1) { 
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-2),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+	if ( Width==2) { 
+	  U    = where(mod(coor,B1)==Integer(B1-2),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-3),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+	if ( Width==3) { 
+	  U    = where(mod(coor,B1)==Integer(B1-3),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-2),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(2)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-4),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+      }
+
+    }
+   
+  }
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/filters/DirichletFilter.h

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/DirichletFilter.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+//--------------------------------------------------------------------
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<typename MomentaField>
+struct DirichletFilter: public MomentumFilterBase<MomentaField>
+{
+  typedef typename MomentaField::vector_type vector_type; //SIMD-vectorized complex type
+  typedef typename MomentaField::scalar_type scalar_type; //scalar complex type
+
+  typedef iScalar<iScalar<iScalar<vector_type> > >            ScalarType; //complex phase for each site
+  
+  Coordinate Block;
+  
+  DirichletFilter(const Coordinate &_Block): Block(_Block){}
+
+  void applyFilter(MomentaField &P) const override
+  {
+    GridBase *grid = P.Grid();
+    typedef decltype(PeekIndex<LorentzIndex>(P, 0)) LatCM;
+    ////////////////////////////////////////////////////
+    // Zero strictly links crossing between domains
+    ////////////////////////////////////////////////////
+    LatticeInteger coor(grid); 
+    LatCM zz(grid); zz = Zero();
+    for(int mu=0;mu<Nd;mu++) {
+      if ( (Block[mu]) && (Block[mu] <= grid->GlobalDimensions()[mu] ) ) {
+	// If costly could provide Grid earlier and precompute masks
+	std::cout << GridLogMessage << " Dirichlet in mu="<<mu<<std::endl;
+	LatticeCoordinate(coor,mu);
+	auto P_mu = PeekIndex<LorentzIndex>(P, mu);
+	P_mu = where(mod(coor,Block[mu])==Integer(Block[mu]-1),zz,P_mu);
+	PokeIndex<LorentzIndex>(P, P_mu, mu);
+      }
+    }
+  }
+};
+
+
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/filters/MomentumFilter.h

@@ -37,7 +37,7 @@ NAMESPACE_BEGIN(Grid);
 
 template<typename MomentaField>
 struct MomentumFilterBase{
-  virtual void applyFilter(MomentaField &P) const;
+  virtual void applyFilter(MomentaField &P) const = 0;
 };
 
 //Do nothing

Grid/qcd/action/pseudofermion/Bounds.h

@@ -14,6 +14,31 @@ NAMESPACE_BEGIN(Grid);
       assert( (lambda_max < hi) && " High Bounds Check on operator failed" );
     }
 
+     template<class Field> void ChebyBoundsCheck(LinearOperatorBase<Field> &HermOp,
+						 Field &GaussNoise,
+						 RealD lo,RealD hi) 
+    {
+      int orderfilter = 1000;
+      Chebyshev<Field> Cheb(lo,hi,orderfilter);
+
+      GridBase *FermionGrid = GaussNoise.Grid();
+
+      Field X(FermionGrid);
+      Field Z(FermionGrid);
+
+      X=GaussNoise;
+      RealD Nx = norm2(X);
+      Cheb(HermOp,X,Z);
+      RealD Nz = norm2(Z);
+
+      std::cout << "************************* "<<std::endl;
+      std::cout << " noise                    = "<<Nx<<std::endl;
+      std::cout << " Cheb x noise             = "<<Nz<<std::endl;
+      std::cout << " Ratio                    = "<<Nz/Nx<<std::endl;
+      std::cout << "************************* "<<std::endl;
+      assert( ((Nz/Nx)<1.0) && " ChebyBoundsCheck ");
+    }
+      
     template<class Field> void InverseSqrtBoundsCheck(int MaxIter,double tol,
 						       LinearOperatorBase<Field> &HermOp,
 						       Field &GaussNoise,
@@ -100,6 +125,5 @@ NAMESPACE_BEGIN(Grid);
       assert( (std::sqrt(Nd/Nx)<tol) && " InversePowerBoundsCheck ");
     }
 
-
 NAMESPACE_END(Grid);
 

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion.h

@@ -0,0 +1,163 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundaryBoson.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);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class ImplD,class ImplF>
+class DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion : public Action<typename ImplD::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(ImplD);
+
+private:
+  SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
+  RealD InnerStoppingCondition;
+  RealD ActionStoppingCondition;
+  RealD DerivativeStoppingCondition;
+  FermionField Phi; // the pseudo fermion field for this trajectory
+public:
+  DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_NumOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
+    : NumOp(_NumOp), 
+      DerivativeStoppingCondition(_DerivativeTol),
+      ActionStoppingCondition(_ActionTol),
+      InnerStoppingCondition(_InnerTol),
+      Phi(_NumOp.FermionGrid()) {};
+
+  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion";}
+
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    return sstream.str();
+  }  
+  
+  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
+  {
+    // P(phi) = e^{- phi^dag P^dag P phi}
+    //
+    // NumOp == P
+    //
+    // Take phi = P^{-1} eta  ; eta = P Phi
+    //
+    // P(eta) = e^{- eta^dag eta}
+    //
+    // e^{x^2/2 sig^2} => sig^2 = 0.5.
+    // 
+    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    NumOp.tolinner=InnerStoppingCondition;
+    NumOp.tol=ActionStoppingCondition;
+    NumOp.ImportGauge(U);
+
+    FermionField eta(NumOp.FermionGrid());
+
+    gaussian(pRNG,eta);    eta=eta*scale;
+    
+    NumOp.ProjectBoundaryBar(eta);
+    //DumpSliceNorm("eta",eta);
+    NumOp.RInv(eta,Phi);
+
+    //DumpSliceNorm("Phi",Phi);
+
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag Pdag P phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    NumOp.tolinner=InnerStoppingCondition;
+    NumOp.tol=ActionStoppingCondition;
+    NumOp.ImportGauge(U);
+
+    FermionField Y(NumOp.FermionGrid());
+
+    NumOp.R(Phi,Y);
+
+    RealD action = norm2(Y);
+
+    return action;
+  };
+
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
+  {
+    NumOp.tolinner=InnerStoppingCondition;
+    NumOp.tol=DerivativeStoppingCondition;
+    NumOp.ImportGauge(U);
+
+    GridBase *fgrid = NumOp.FermionGrid();
+    GridBase *ugrid = NumOp.GaugeGrid();
+
+    FermionField  X(fgrid);
+    FermionField  Y(fgrid);
+    FermionField  tmp(fgrid);
+
+    GaugeField   force(ugrid);	
+
+    FermionField DobiDdbPhi(fgrid);      // Vector A in my notes
+    FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
+    FermionField DoidP_Phi(fgrid);    // Vector E in my notes
+    FermionField DobidDddDoidP_Phi(fgrid);    // Vector F in my notes
+    
+    FermionField P_Phi(fgrid);
+    
+    // P term
+    NumOp.dBoundaryBar(Phi,tmp);
+    NumOp.dOmegaBarInv(tmp,DobiDdbPhi);        // Vector A
+    NumOp.dBoundary(DobiDdbPhi,tmp);
+    NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi);      // Vector B
+    P_Phi  = Phi - DoiDdDobiDdbPhi;
+    NumOp.ProjectBoundaryBar(P_Phi);
+    
+    // P^dag P term
+    NumOp.dOmegaDagInv(P_Phi,DoidP_Phi); // Vector E
+    NumOp.dBoundaryDag(DoidP_Phi,tmp);
+    NumOp.dOmegaBarDagInv(tmp,DobidDddDoidP_Phi);   // Vector F
+    NumOp.dBoundaryBarDag(DobidDddDoidP_Phi,tmp);
+
+    X = DobiDdbPhi;
+    Y = DobidDddDoidP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    X = DoiDdDobiDdbPhi;
+    Y = DoidP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    dSdU *= -1.0;
+
+  };
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourPseudoFermion.h

@@ -0,0 +1,158 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.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);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class ImplD,class ImplF>
+class DomainDecomposedBoundaryTwoFlavourPseudoFermion : public Action<typename ImplD::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(ImplD);
+
+private:
+  SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
+  RealD ActionStoppingCondition;
+  RealD DerivativeStoppingCondition;
+  RealD InnerStoppingCondition;
+
+  FermionField Phi; // the pseudo fermion field for this trajectory
+
+  RealD refresh_action;
+public:
+  DomainDecomposedBoundaryTwoFlavourPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_DenOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol = 1.0e-6 )
+    : DenOp(_DenOp),
+      DerivativeStoppingCondition(_DerivativeTol),
+      ActionStoppingCondition(_ActionTol),
+      InnerStoppingCondition(_InnerTol),
+      Phi(_DenOp.FermionGrid()) {};
+      
+  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourPseudoFermion";}
+
+ 
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    return sstream.str();
+  }  
+  
+  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
+  {
+    // P(phi) = e^{- phi^dag Rdag^-1 R^-1 phi}
+    //
+    // DenOp == R
+    //
+    // Take phi = R eta  ; eta = R^-1 Phi
+    //
+    // P(eta) = e^{- eta^dag eta}
+    //
+    // e^{x^2/2 sig^2} => sig^2 = 0.5.
+    // 
+    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol     =ActionStoppingCondition;
+    DenOp.ImportGauge(U);
+
+    FermionField eta(DenOp.FermionGrid());
+
+    gaussian(pRNG,eta);    eta=eta*scale;
+    
+    DenOp.ProjectBoundaryBar(eta);
+    DenOp.R(eta,Phi);
+    //DumpSliceNorm("Phi",Phi);
+    refresh_action = norm2(eta);
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag Rdag^-1 R^-1 phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol=ActionStoppingCondition;
+    DenOp.ImportGauge(U);
+
+    FermionField X(DenOp.FermionGrid());
+
+    DenOp.RInv(Phi,X);
+
+    RealD action = norm2(X);
+
+    return action;
+  };
+
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
+  {
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol=DerivativeStoppingCondition;
+    DenOp.ImportGauge(U);
+
+    GridBase *fgrid = DenOp.FermionGrid();
+    GridBase *ugrid = DenOp.GaugeGrid();
+
+    FermionField  X(fgrid);
+    FermionField  Y(fgrid);
+    FermionField  tmp(fgrid);
+
+    GaugeField   force(ugrid);	
+
+    FermionField DiDdb_Phi(fgrid);      // Vector C in my notes
+    FermionField DidRinv_Phi(fgrid);    // Vector D in my notes
+    FermionField Rinv_Phi(fgrid);
+
+//   FermionField RinvDagRinv_Phi(fgrid);
+//   FermionField DdbdDidRinv_Phi(fgrid);
+
+    // R^-1 term
+    DenOp.dBoundaryBar(Phi,tmp);
+    DenOp.Dinverse(tmp,DiDdb_Phi);            // Vector C
+    Rinv_Phi = Phi - DiDdb_Phi;
+    DenOp.ProjectBoundaryBar(Rinv_Phi); 
+ 
+    // R^-dagger R^-1 term
+    DenOp.DinverseDag(Rinv_Phi,DidRinv_Phi); // Vector D
+/*
+    DenOp.dBoundaryBarDag(DidRinv_Phi,DdbdDidRinv_Phi);
+    RinvDagRinv_Phi = Rinv_Phi - DdbdDidRinv_Phi;
+    DenOp.ProjectBoundaryBar(RinvDagRinv_Phi);
+*/
+    X = DiDdb_Phi;
+    Y = DidRinv_Phi;
+    DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=force;
+    DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+    DumpSliceNorm("force",dSdU);
+    dSdU *= -1.0;
+  };
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/pseudofermion/DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion.h

@@ -0,0 +1,237 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.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);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class ImplD,class ImplF>
+class DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion : public Action<typename ImplD::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(ImplD);
+
+private:
+  SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
+  SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
+
+  RealD InnerStoppingCondition;
+  RealD ActionStoppingCondition;
+  RealD DerivativeStoppingCondition;
+  
+  FermionField Phi; // the pseudo fermion field for this trajectory
+
+public:
+  DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF>  &_NumOp, 
+						       SchurFactoredFermionOperator<ImplD,ImplF>  &_DenOp,
+						       RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
+    : NumOp(_NumOp), DenOp(_DenOp),
+      Phi(_NumOp.PeriodicFermOpD.FermionGrid()),
+      InnerStoppingCondition(_InnerTol),
+      DerivativeStoppingCondition(_DerivativeTol),
+      ActionStoppingCondition(_ActionTol)
+  {};
+      
+  virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion";}
+ 
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    return sstream.str();
+  }  
+  
+  virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
+  {
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField eta(NumOp.PeriodicFermOpD.FermionGrid());
+    FermionField tmp(NumOp.PeriodicFermOpD.FermionGrid());
+
+    // P(phi) = e^{- phi^dag P^dag Rdag^-1 R^-1 P phi}
+    //
+    // NumOp == P
+    // DenOp == R
+    //
+    // Take phi = P^{-1} R eta  ; eta = R^-1 P Phi
+    //
+    // P(eta) = e^{- eta^dag eta}
+    //
+    // e^{x^2/2 sig^2} => sig^2 = 0.5.
+    // 
+    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    gaussian(pRNG,eta);    eta=eta*scale;
+    
+    NumOp.ProjectBoundaryBar(eta);
+    NumOp.tolinner=InnerStoppingCondition;
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol = ActionStoppingCondition;
+    NumOp.tol = ActionStoppingCondition;
+    DenOp.R(eta,tmp);
+    NumOp.RInv(tmp,Phi);
+    DumpSliceNorm("Phi",Phi);
+
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag Pdag Rdag^-1 R^-1 P phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField X(NumOp.PeriodicFermOpD.FermionGrid());
+    FermionField Y(NumOp.PeriodicFermOpD.FermionGrid());
+
+    NumOp.tolinner=InnerStoppingCondition;
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol = ActionStoppingCondition;
+    NumOp.tol = ActionStoppingCondition;
+    NumOp.R(Phi,Y);
+    DenOp.RInv(Y,X);
+
+    RealD action = norm2(X);
+    //    std::cout << " DD boundary action is " <<action<<std::endl;
+
+    return action;
+  };
+
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU)
+  {
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    GridBase *fgrid = NumOp.PeriodicFermOpD.FermionGrid();
+    GridBase *ugrid = NumOp.PeriodicFermOpD.GaugeGrid();
+
+    FermionField  X(fgrid);
+    FermionField  Y(fgrid);
+    FermionField  tmp(fgrid);
+
+    GaugeField   force(ugrid);	
+
+    FermionField DobiDdbPhi(fgrid);      // Vector A in my notes
+    FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
+    FermionField DiDdbP_Phi(fgrid);      // Vector C in my notes
+    FermionField DidRinvP_Phi(fgrid);    // Vector D in my notes
+    FermionField DdbdDidRinvP_Phi(fgrid);
+    FermionField DoidRinvDagRinvP_Phi(fgrid);    // Vector E in my notes
+    FermionField DobidDddDoidRinvDagRinvP_Phi(fgrid);    // Vector F in my notes
+    
+    FermionField P_Phi(fgrid);
+    FermionField RinvP_Phi(fgrid);
+    FermionField RinvDagRinvP_Phi(fgrid);
+    FermionField PdagRinvDagRinvP_Phi(fgrid);
+
+    //    RealD action = S(U);
+    NumOp.tolinner=InnerStoppingCondition;
+    DenOp.tolinner=InnerStoppingCondition;
+    DenOp.tol = DerivativeStoppingCondition;
+    NumOp.tol = DerivativeStoppingCondition;
+    
+    // P term
+    NumOp.dBoundaryBar(Phi,tmp);
+    NumOp.dOmegaBarInv(tmp,DobiDdbPhi);        // Vector A
+    NumOp.dBoundary(DobiDdbPhi,tmp);
+    NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi);      // Vector B
+    P_Phi  = Phi - DoiDdDobiDdbPhi;
+    NumOp.ProjectBoundaryBar(P_Phi);
+
+    // R^-1 P term
+    DenOp.dBoundaryBar(P_Phi,tmp);
+    DenOp.Dinverse(tmp,DiDdbP_Phi);            // Vector C
+    RinvP_Phi = P_Phi - DiDdbP_Phi;
+    DenOp.ProjectBoundaryBar(RinvP_Phi); // Correct to here
+
+ 
+    // R^-dagger R^-1 P term
+    DenOp.DinverseDag(RinvP_Phi,DidRinvP_Phi); // Vector D
+    DenOp.dBoundaryBarDag(DidRinvP_Phi,DdbdDidRinvP_Phi);
+    RinvDagRinvP_Phi = RinvP_Phi - DdbdDidRinvP_Phi;
+    DenOp.ProjectBoundaryBar(RinvDagRinvP_Phi);
+
+    
+    // P^dag R^-dagger R^-1 P term
+    NumOp.dOmegaDagInv(RinvDagRinvP_Phi,DoidRinvDagRinvP_Phi); // Vector E
+    NumOp.dBoundaryDag(DoidRinvDagRinvP_Phi,tmp);
+    NumOp.dOmegaBarDagInv(tmp,DobidDddDoidRinvDagRinvP_Phi);   // Vector F
+    NumOp.dBoundaryBarDag(DobidDddDoidRinvDagRinvP_Phi,tmp);
+    PdagRinvDagRinvP_Phi = RinvDagRinvP_Phi- tmp;
+    NumOp.ProjectBoundaryBar(PdagRinvDagRinvP_Phi);
+
+    /*
+    std::cout << "S eval  "<< action << std::endl;
+    std::cout << "S - IP1 "<< innerProduct(Phi,PdagRinvDagRinvP_Phi) << std::endl;
+    std::cout << "S - IP2 "<< norm2(RinvP_Phi) << std::endl;
+
+    NumOp.R(Phi,tmp);
+    tmp = tmp - P_Phi;
+    std::cout << "diff1 "<<norm2(tmp) <<std::endl;
+    
+    
+    DenOp.RInv(P_Phi,tmp);
+    tmp = tmp - RinvP_Phi;
+    std::cout << "diff2 "<<norm2(tmp) <<std::endl;
+
+    DenOp.RDagInv(RinvP_Phi,tmp);
+    tmp  = tmp - RinvDagRinvP_Phi;
+    std::cout << "diff3 "<<norm2(tmp) <<std::endl;
+
+    DenOp.RDag(RinvDagRinvP_Phi,tmp);
+    tmp  = tmp - PdagRinvDagRinvP_Phi;
+    std::cout << "diff4 "<<norm2(tmp) <<std::endl;
+    */
+    
+    dSdU=Zero();
+
+    X = DobiDdbPhi;
+    Y = DobidDddDoidRinvDagRinvP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    X = DoiDdDobiDdbPhi;
+    Y = DoidRinvDagRinvP_Phi;
+    NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    X = DiDdbP_Phi;
+    Y = DidRinvP_Phi;
+    DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo);    dSdU=dSdU+force;
+    DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes);   dSdU=dSdU+force;
+
+    dSdU *= -1.0;
+
+  };
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h

@@ -196,7 +196,6 @@ NAMESPACE_BEGIN(Grid);
         for(int k=0; k<param.degree; ++k){
           gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
           heatbathRefreshShiftCoefficients(0, -gamma_l);
-	  //Lop.RefreshShiftCoefficients(-gamma_l);
           if(use_heatbath_forecasting){ // Forecast CG guess using solutions from previous poles
             Lop.Mdag(CG_src, Forecast_src);
             CG_soln = Forecast(Lop, Forecast_src, prev_solns);
@@ -224,7 +223,6 @@ NAMESPACE_BEGIN(Grid);
         for(int k=0; k<param.degree; ++k){
           gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
 	  heatbathRefreshShiftCoefficients(1, -gamma_l*PowerNegHalf.poles[k]);
-          //Rop.RefreshShiftCoefficients(-gamma_l*PowerNegHalf.poles[k]);
           if(use_heatbath_forecasting){
             Rop.Mdag(CG_src, Forecast_src);
             CG_soln = Forecast(Rop, Forecast_src, prev_solns);
@@ -242,8 +240,6 @@ NAMESPACE_BEGIN(Grid);
         Phi = Phi + tmp[1];
 
         // Reset shift coefficients for energy and force evals
-        //Lop.RefreshShiftCoefficients(0.0);
-        //Rop.RefreshShiftCoefficients(-1.0);
 	heatbathRefreshShiftCoefficients(0, 0.0);
 	heatbathRefreshShiftCoefficients(1, -1.0);
 

Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h

@@ -49,10 +49,12 @@ NAMESPACE_BEGIN(Grid);
       Params param;
 
       MultiShiftFunction PowerHalf   ;
-      MultiShiftFunction PowerNegHalf;
       MultiShiftFunction PowerQuarter;
+      MultiShiftFunction PowerNegHalf;
       MultiShiftFunction PowerNegQuarter;
 
+      MultiShiftFunction MDPowerQuarter;
+      MultiShiftFunction MDPowerNegHalf;
     private:
      
       FermionOperator<Impl> & NumOp;// the basic operator
@@ -73,11 +75,13 @@ NAMESPACE_BEGIN(Grid);
 	remez.generateApprox(param.degree,1,2);
 	PowerHalf.Init(remez,param.tolerance,false);
 	PowerNegHalf.Init(remez,param.tolerance,true);
+	MDPowerNegHalf.Init(remez,param.mdtolerance,true);
 
 	// MdagM^(+- 1/4)
 	std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/4)"<<std::endl;
 	remez.generateApprox(param.degree,1,4);
    	PowerQuarter.Init(remez,param.tolerance,false);
+   	MDPowerQuarter.Init(remez,param.mdtolerance,false);
 	PowerNegQuarter.Init(remez,param.tolerance,true);
       };
 
@@ -204,8 +208,8 @@ NAMESPACE_BEGIN(Grid);
 
       virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
 
-	const int n_f  = PowerNegHalf.poles.size();
-	const int n_pv = PowerQuarter.poles.size();
+	const int n_f  = MDPowerNegHalf.poles.size();
+	const int n_pv = MDPowerQuarter.poles.size();
 
 	std::vector<FermionField> MpvPhi_k     (n_pv,NumOp.FermionGrid());
 	std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionGrid());
@@ -224,8 +228,8 @@ NAMESPACE_BEGIN(Grid);
 	MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagM(DenOp);
 	MdagMLinearOperator<FermionOperator<Impl> ,FermionField> VdagV(NumOp);
 
-	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
-	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
+	ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,MDPowerQuarter);
+	ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,MDPowerNegHalf);
 
 	msCG_V(VdagV,Phi,MpvPhi_k,MpvPhi);
 	msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
@@ -244,7 +248,7 @@ NAMESPACE_BEGIN(Grid);
 
 	//(1)
 	for(int k=0;k<n_f;k++){
-	  ak = PowerNegHalf.residues[k];
+	  ak = MDPowerNegHalf.residues[k];
 	  DenOp.M(MfMpvPhi_k[k],Y);
 	  DenOp.MDeriv(tmp , MfMpvPhi_k[k], Y,DaggerYes );  dSdU=dSdU+ak*tmp;
 	  DenOp.MDeriv(tmp , Y, MfMpvPhi_k[k], DaggerNo );  dSdU=dSdU+ak*tmp;
@@ -254,7 +258,7 @@ NAMESPACE_BEGIN(Grid);
 	//(3)
 	for(int k=0;k<n_pv;k++){
 
-          ak = PowerQuarter.residues[k];
+          ak = MDPowerQuarter.residues[k];
 	  
 	  NumOp.M(MpvPhi_k[k],Y);
 	  NumOp.MDeriv(tmp,MpvMfMpvPhi_k[k],Y,DaggerYes); dSdU=dSdU+ak*tmp;

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -75,15 +75,19 @@ NAMESPACE_BEGIN(Grid);
           conformable(_NumOp.GaugeRedBlackGrid(), _DenOp.GaugeRedBlackGrid());
         };
 
-      virtual std::string action_name(){return "TwoFlavourEvenOddRatioPseudoFermionAction";}
-
-      virtual std::string LogParameters(){
+      virtual std::string action_name(){
 	std::stringstream sstream;
-	sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
+	sstream<<"TwoFlavourEvenOddRatioPseudoFermionAction det("<<DenOp.Mass()<<") / det("<<NumOp.Mass()<<")";
 	return sstream.str();
       }
 
-      //Access the fermion field
+      virtual std::string LogParameters(){
+	std::stringstream sstream;
+	sstream<< GridLogMessage << "["<<action_name()<<"] -- No further parameters "<<std::endl;
+	return sstream.str();
+      } 
+
+      
       const FermionField &getPhiOdd() const{ return PhiOdd; }
 
       virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
@@ -100,12 +104,12 @@ NAMESPACE_BEGIN(Grid);
       }
 
       void refresh(const GaugeField &U, const FermionField &eta) {
+
         // P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
         //
         // NumOp == V
         // DenOp == M
         //
-        // Take phi_o = Vpcdag^{-1} Mpcdag eta_o  ; eta_o = Mpcdag^{-1} Vpcdag Phi
         FermionField etaOdd (NumOp.FermionRedBlackGrid());
         FermionField etaEven(NumOp.FermionRedBlackGrid());
         FermionField tmp    (NumOp.FermionRedBlackGrid());
@@ -128,10 +132,6 @@ NAMESPACE_BEGIN(Grid);
         // Even det factors
         DenOp.MooeeDag(etaEven,tmp);
         NumOp.MooeeInvDag(tmp,PhiEven);
-
-        //PhiOdd =PhiOdd*scale;
-        //PhiEven=PhiEven*scale;
-        
       };
 
       //////////////////////////////////////////////////////

Grid/qcd/action/pseudofermion/TwoFlavourRatioEO4DPseudoFermion.h

@@ -0,0 +1,203 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/pseudofermion/TwoFlavourRatio.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////
+// Two flavour ratio
+///////////////////////////////////////
+template<class Impl>
+class TwoFlavourRatioEO4DPseudoFermionAction : public Action<typename Impl::GaugeField> {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+private:
+  typedef FermionOperator<Impl> FermOp;
+  FermionOperator<Impl> & NumOp;// the basic operator
+  FermionOperator<Impl> & DenOp;// the basic operator
+
+  OperatorFunction<FermionField> &DerivativeSolver;
+  OperatorFunction<FermionField> &DerivativeDagSolver;
+  OperatorFunction<FermionField> &ActionSolver;
+  OperatorFunction<FermionField> &HeatbathSolver;
+
+  FermionField phi4; // the pseudo fermion field for this trajectory
+
+public:
+  TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
+					 FermionOperator<Impl>  &_DenOp, 
+					 OperatorFunction<FermionField> & DS,
+					 OperatorFunction<FermionField> & AS ) : 
+    TwoFlavourRatioEO4DPseudoFermionAction(_NumOp,_DenOp, DS,DS,AS,AS) {};
+  TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
+					 FermionOperator<Impl>  &_DenOp, 
+					 OperatorFunction<FermionField> & DS,
+					 OperatorFunction<FermionField> & DDS,
+					 OperatorFunction<FermionField> & AS,
+					 OperatorFunction<FermionField> & HS
+				       ) : NumOp(_NumOp),
+					   DenOp(_DenOp),
+					   DerivativeSolver(DS),
+					   DerivativeDagSolver(DDS),
+					   ActionSolver(AS),
+					   HeatbathSolver(HS),
+					   phi4(_NumOp.GaugeGrid())
+  {};
+      
+  virtual std::string action_name(){return "TwoFlavourRatioEO4DPseudoFermionAction";}
+
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
+    return sstream.str();
+  }  
+      
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
+
+    // P(phi) = e^{- phi^dag (V^dag M^-dag)_11  (M^-1 V)_11 phi}
+    //
+    // NumOp == V
+    // DenOp == M
+    //
+    // Take phi = (V^{-1} M)_11 eta  ; eta = (M^{-1} V)_11 Phi
+    //
+    // P(eta) = e^{- eta^dag eta}
+    //
+    // e^{x^2/2 sig^2} => sig^2 = 0.5.
+    // 
+    // So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
+    //
+    RealD scale = std::sqrt(0.5);
+
+    FermionField eta4(NumOp.GaugeGrid());
+    FermionField eta5(NumOp.FermionGrid());
+    FermionField tmp(NumOp.FermionGrid());
+    FermionField phi5(NumOp.FermionGrid());
+
+    gaussian(pRNG,eta4);
+    NumOp.ImportFourDimPseudoFermion(eta4,eta5);
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(HeatbathSolver);
+
+    DenOp.M(eta5,tmp);               // M eta
+    PrecSolve(NumOp,tmp,phi5);  // phi = V^-1 M eta
+    phi5=phi5*scale;
+    std::cout << GridLogMessage << "4d pf refresh "<< norm2(phi5)<<"\n";
+    // Project to 4d
+    NumOp.ExportFourDimPseudoFermion(phi5,phi4);
+      
+  };
+
+  //////////////////////////////////////////////////////
+  // S = phi^dag (V^dag M^-dag)_11  (M^-1 V)_11 phi
+  //////////////////////////////////////////////////////
+  virtual RealD S(const GaugeField &U) {
+
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField Y4(NumOp.GaugeGrid());
+    FermionField X(NumOp.FermionGrid());
+    FermionField Y(NumOp.FermionGrid());
+    FermionField phi5(NumOp.FermionGrid());
+	
+    MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagMOp(DenOp);
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(ActionSolver);
+
+    NumOp.ImportFourDimPseudoFermion(phi4,phi5);
+    NumOp.M(phi5,X);              // X= V phi
+    PrecSolve(DenOp,X,Y);    // Y= (MdagM)^-1 Mdag Vdag phi = M^-1 V phi
+    NumOp.ExportFourDimPseudoFermion(Y,Y4);
+
+    RealD action = norm2(Y4);
+
+    return action;
+  };
+
+  //////////////////////////////////////////////////////
+  // dS/du = 2 Re phi^dag (V^dag M^-dag)_11  (M^-1 d V)_11  phi
+  //       - 2 Re phi^dag (dV^dag M^-dag)_11  (M^-1 dM M^-1 V)_11  phi
+  //////////////////////////////////////////////////////
+  virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
+
+    NumOp.ImportGauge(U);
+    DenOp.ImportGauge(U);
+
+    FermionField  X(NumOp.FermionGrid());
+    FermionField  Y(NumOp.FermionGrid());
+    FermionField       phi(NumOp.FermionGrid());
+    FermionField      Vphi(NumOp.FermionGrid());
+    FermionField  MinvVphi(NumOp.FermionGrid());
+    FermionField      tmp4(NumOp.GaugeGrid());
+    FermionField  MdagInvMinvVphi(NumOp.FermionGrid());
+
+    GaugeField   force(NumOp.GaugeGrid());	
+
+    //Y=V phi
+    //X = (Mdag V phi
+    //Y = (Mdag M)^-1 Mdag V phi = M^-1 V Phi
+    NumOp.ImportFourDimPseudoFermion(phi4,phi);
+    NumOp.M(phi,Vphi);               //  V phi
+    SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(DerivativeSolver);
+    PrecSolve(DenOp,Vphi,MinvVphi);// M^-1 V phi
+    std::cout << GridLogMessage << "4d deriv solve "<< norm2(MinvVphi)<<"\n";
+
+    // Projects onto the physical space and back
+    NumOp.ExportFourDimPseudoFermion(MinvVphi,tmp4);
+    NumOp.ImportFourDimPseudoFermion(tmp4,Y);
+
+    SchurRedBlackDiagMooeeDagSolve<FermionField> PrecDagSolve(DerivativeDagSolver);
+    // X = proj M^-dag V phi
+    // Need an adjoint solve
+    PrecDagSolve(DenOp,Y,MdagInvMinvVphi);
+    std::cout << GridLogMessage << "4d deriv solve dag "<< norm2(MdagInvMinvVphi)<<"\n";
+    
+    // phi^dag (Vdag Mdag^-1) (M^-1 dV)  phi
+    NumOp.MDeriv(force ,MdagInvMinvVphi , phi, DaggerNo );  dSdU=force;
+  
+    // phi^dag (dVdag Mdag^-1) (M^-1 V)  phi
+    NumOp.MDeriv(force , phi, MdagInvMinvVphi ,DaggerYes  );  dSdU=dSdU+force;
+
+    //    - 2 Re phi^dag (dV^dag M^-dag)_11  (M^-1 dM M^-1 V)_11  phi
+    DenOp.MDeriv(force,MdagInvMinvVphi,MinvVphi,DaggerNo);   dSdU=dSdU-force;
+    DenOp.MDeriv(force,MinvVphi,MdagInvMinvVphi,DaggerYes);  dSdU=dSdU-force;
+
+    dSdU *= -1.0; 
+    //dSdU = - Ta(dSdU);
+    
+  };
+};
+
+NAMESPACE_END(Grid);
+
+

Grid/qcd/hmc/GenericHMCrunner.h

@@ -186,6 +186,9 @@ private:
     typedef IntegratorType<SmearingPolicy> TheIntegrator;
     TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
 
+    // Sets the momentum filter
+    MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
+
     Smearing.set_Field(U);
 
     HybridMonteCarlo<TheIntegrator> HMC(Parameters, MDynamics,

Grid/qcd/hmc/HMC.h

@@ -34,6 +34,7 @@ directory
 			    * @brief Classes for Hybrid Monte Carlo update
 			    *
 			    * @author Guido Cossu
+			    * @author Peter Boyle
 			    */
 			   //--------------------------------------------------------------------
 #pragma once
@@ -115,22 +116,17 @@ private:
 
     random(sRNG, rn_test);
 
-    std::cout << GridLogHMC
-              << "--------------------------------------------------\n";
-    std::cout << GridLogHMC << "exp(-dH) = " << prob
-              << "  Random = " << rn_test << "\n";
-    std::cout << GridLogHMC
-              << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+    std::cout << GridLogHMC << "exp(-dH) = " << prob << "  Random = " << rn_test << "\n";
+    std::cout << GridLogHMC << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
 
     if ((prob > 1.0) || (rn_test <= prob)) {  // accepted
       std::cout << GridLogHMC << "Metropolis_test -- ACCEPTED\n";
-      std::cout << GridLogHMC
-                << "--------------------------------------------------\n";
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
       std::cout << GridLogHMC << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogHMC
-                << "--------------------------------------------------\n";
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
       return false;
     }
   }
@@ -139,18 +135,67 @@ private:
   // Evolution
   /////////////////////////////////////////////////////////
   RealD evolve_hmc_step(Field &U) {
-    TheIntegrator.refresh(U, sRNG, pRNG);  // set U and initialize P and phi's
 
-    RealD H0 = TheIntegrator.S(U);  // initial state action
+    GridBase *Grid = U.Grid();
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // Mainly for DDHMC perform a random translation of U modulo volume
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Random shifting gauge field by [";
+    for(int d=0;d<Grid->Nd();d++) {
+
+      int L = Grid->GlobalDimensions()[d];
+
+      RealD rn_uniform;  random(sRNG, rn_uniform);
+
+      int shift = (int) (rn_uniform*L);
+
+      std::cout << shift;
+      if(d<Grid->Nd()-1) std::cout <<",";
+      else               std::cout <<"]\n";
+      
+      U = Cshift(U,d,shift);
+    }
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    TheIntegrator.reset_timer();
+    
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // set U and initialize P and phi's
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Refresh momenta and pseudofermions";
+    TheIntegrator.refresh(U, sRNG, pRNG);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // initial state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Compute initial action";
+    RealD H0 = TheIntegrator.S(U);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
     std::streamsize current_precision = std::cout.precision();
     std::cout.precision(15);
     std::cout << GridLogHMC << "Total H before trajectory = " << H0 << "\n";
     std::cout.precision(current_precision);
 
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << " Molecular Dynamics evolution ";
     TheIntegrator.integrate(U);
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // updated state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Compute final action";
+    RealD H1 = TheIntegrator.S(U);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
 
-    RealD H1 = TheIntegrator.S(U);  // updated state action
     
     ///////////////////////////////////////////////////////////
     if(0){
@@ -163,18 +208,17 @@ private:
     }
     ///////////////////////////////////////////////////////////
 
-
     std::cout.precision(15);
-    std::cout << GridLogHMC << "Total H after trajectory  = " << H1
-	      << "  dH = " << H1 - H0 << "\n";
+
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+    std::cout << GridLogHMC << "Total H after trajectory  = " << H1 << "  dH = " << H1 - H0 << "\n";
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+
     std::cout.precision(current_precision);
     
     return (H1 - H0);
   }
 
-
-  
-
 public:
   /////////////////////////////////////////
   // Constructor
@@ -195,8 +239,11 @@ public:
 
     // Actual updates (evolve a copy Ucopy then copy back eventually)
     unsigned int FinalTrajectory = Params.Trajectories + Params.NoMetropolisUntil + Params.StartTrajectory;
+
     for (int traj = Params.StartTrajectory; traj < FinalTrajectory; ++traj) {
+
       std::cout << GridLogHMC << "-- # Trajectory = " << traj << "\n";
+
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
       	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
       }
@@ -216,11 +263,10 @@ public:
       if (accept)
         Ucur = Ucopy; 
       
-     
-      
       double t1=usecond();
       std::cout << GridLogHMC << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
 
+      TheIntegrator.print_timer();
 
       for (int obs = 0; obs < Observables.size(); obs++) {
       	std::cout << GridLogDebug << "Observables # " << obs << std::endl;

Grid/qcd/hmc/HMCResourceManager.h

@@ -72,6 +72,8 @@ class HMCResourceManager {
   typedef HMCModuleBase< BaseHmcCheckpointer<ImplementationPolicy> > CheckpointerBaseModule;
   typedef HMCModuleBase< HmcObservable<typename ImplementationPolicy::Field> > ObservableBaseModule;
   typedef ActionModuleBase< Action<typename ImplementationPolicy::Field>, GridModule > ActionBaseModule;
+  typedef typename ImplementationPolicy::Field MomentaField;
+  typedef typename ImplementationPolicy::Field Field;  
 
   // Named storage for grid pairs (std + red-black)
   std::unordered_map<std::string, GridModule> Grids;
@@ -80,6 +82,9 @@ class HMCResourceManager {
   // SmearingModule<ImplementationPolicy> Smearing;
   std::unique_ptr<CheckpointerBaseModule> CP;
 
+  // Momentum filter
+  std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> > Filter;
+  
   // A vector of HmcObservable modules
   std::vector<std::unique_ptr<ObservableBaseModule> > ObservablesList;
 
@@ -90,6 +95,7 @@ class HMCResourceManager {
 
   bool have_RNG;
   bool have_CheckPointer;
+  bool have_Filter;
 
   // NOTE: operator << is not overloaded for std::vector<string> 
   // so this function is necessary
@@ -101,7 +107,7 @@ class HMCResourceManager {
 
 
 public:
-  HMCResourceManager() : have_RNG(false), have_CheckPointer(false) {}
+  HMCResourceManager() : have_RNG(false), have_CheckPointer(false), have_Filter(false) {}
 
   template <class ReaderClass, class vector_type = vComplex >
   void initialize(ReaderClass &Read){
@@ -129,6 +135,7 @@ public:
     RNGModuleParameters RNGpar(Read);
     SetRNGSeeds(RNGpar);
 
+  
     // Observables
     auto &ObsFactory = HMC_ObservablesModuleFactory<observable_string, typename ImplementationPolicy::Field, ReaderClass>::getInstance(); 
     Read.push(observable_string);// here must check if existing...
@@ -208,6 +215,16 @@ public:
     AddGrid(s, Mod);
   }
 
+  void SetMomentumFilter( MomentumFilterBase<typename ImplementationPolicy::Field> * MomFilter) {
+    assert(have_Filter==false);
+    Filter = std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> >(MomFilter);
+    have_Filter = true;
+  }
+  MomentumFilterBase<typename ImplementationPolicy::Field> *GetMomentumFilter(void) {
+    if ( !have_Filter)
+      SetMomentumFilter(new MomentumFilterNone<typename ImplementationPolicy::Field>());
+    return Filter.get();
+  }
 
   GridCartesian* GetCartesian(std::string s = "") {
     if (s.empty()) s = Grids.begin()->first;

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,7 +33,6 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
-#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -67,6 +66,7 @@ public:
 template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
+
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 
@@ -119,42 +119,58 @@ protected:
     }
   } update_P_hireps{};
 
+ 
   void update_P(MomentaField& Mom, Field& U, int level, double ep) {
     // input U actually not used in the fundamental case
     // Fundamental updates, include smearing
 
     for (int a = 0; a < as[level].actions.size(); ++a) {
+
       double start_full = usecond();
       Field force(U.Grid());
       conformable(U.Grid(), Mom.Grid());
 
       Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
+      as[level].actions.at(a)->deriv_timer_start();
       as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
+      as[level].actions.at(a)->deriv_timer_stop();
 
       std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
+      auto name = as[level].actions.at(a)->action_name();
       if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
+
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
 
+      MomFilter->applyFilter(force);
+      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<< std::endl;
+      DumpSliceNorm("force ",force,Nd-1);
+      
       Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
       Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
 
-      Real max_force_abs = std::sqrt(maxLocalNorm2(force));
-      Real max_impulse_abs = max_force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
+      Real force_max   = std::sqrt(maxLocalNorm2(force));
+      Real impulse_max = force_max * ep * HMC_MOMENTUM_DENOMINATOR;    
+
+      as[level].actions.at(a)->deriv_log(force_abs,force_max);
+      
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force average: " << force_abs <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force max    : " << force_max <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt average  : " << impulse_abs <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt max      : " << impulse_max <<" "<<name<<std::endl;
 
-      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << " Max force: " << max_force_abs << " Time step: " << ep << " Impulse average: " << impulse_abs << " Max impulse: " << max_impulse_abs << std::endl;
       Mom -= force * ep* HMC_MOMENTUM_DENOMINATOR;; 
       double end_full = usecond();
       double time_full  = (end_full - start_full) / 1e3;
       double time_force = (end_force - start_force) / 1e3;
       std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)"  << std::endl;
+
     }
 
     // Force from the other representations
     as[level].apply(update_P_hireps, Representations, Mom, U, ep);
 
-    MomFilter->applyFilter(Mom);
   }
 
   void update_U(Field& U, double ep) 
@@ -168,8 +184,12 @@ protected:
   
   void update_U(MomentaField& Mom, Field& U, double ep) 
   {
+    MomentaField MomFiltered(Mom.Grid());
+    MomFiltered = Mom;
+    MomFilter->applyFilter(MomFiltered);
+
     // exponential of Mom*U in the gauge fields case
-    FieldImplementation::update_field(Mom, U, ep);
+    FieldImplementation::update_field(MomFiltered, U, ep);
 
     // Update the smeared fields, can be implemented as observer
     Smearer.set_Field(U);
@@ -212,6 +232,66 @@ public:
   const MomentaField & getMomentum() const{ return P; }
   
 
+  void reset_timer(void)
+  {
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+        as[level].actions.at(actionID)->reset_timer();
+      }
+    }
+  }
+  void print_timer(void)
+  {
+    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::" << std::endl;
+    std::cout << GridLogMessage << " Refresh cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->refresh_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Action cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->S_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Force cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->deriv_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Force average size "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] : "
+		  <<" force max " << as[level].actions.at(actionID)->deriv_max_average()
+		  <<" norm "      << as[level].actions.at(actionID)->deriv_norm_average()
+		  <<" calls "     << as[level].actions.at(actionID)->deriv_num
+		  << std::endl;
+      }
+    }
+    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
+  }
+  
   void print_parameters()
   {
     std::cout << GridLogMessage << "[Integrator] Name : "<< integrator_name() << std::endl;
@@ -230,7 +310,6 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
-
   }
 
   void reverse_momenta()
@@ -277,16 +356,19 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
-	std::cout << GridLogIntegrator << "Refreshing integrator level " << level << " index " << actionID << std::endl;
+	auto name = as[level].actions.at(actionID)->action_name();
+        std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
+
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
+	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
+	as[level].actions.at(actionID)->refresh_timer_stop();
       }
 
       // Refresh the higher representation actions
       as[level].apply(refresh_hireps, Representations, sRNG, pRNG);
     }
 
-    MomFilter->applyFilter(P);
   }
 
   // to be used by the actionlevel class to iterate
@@ -321,7 +403,9 @@ public:
         // based on the boolean is_smeared in actionID
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
+	        as[level].actions.at(actionID)->S_timer_start();
         Hterm = as[level].actions.at(actionID)->S(Us);
+   	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
       }

Grid/qcd/utils/GaugeFix.h

@@ -72,14 +72,13 @@ public:
 
   //Fix the gauge field Umu
   //0 < alpha < 1 is related to the step size, cf https://arxiv.org/pdf/1405.5812.pdf
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu, Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
-    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog);
+    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
   }
-
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
   //Fix the gauge field Umu and also return the gauge transformation from the original gauge field, xform
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform, Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
 
     GridBase *grid = Umu.Grid();
 
@@ -141,6 +140,8 @@ public:
 
       }
     }
+    std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
+    if (err_on_no_converge)
       assert(0 && "Gauge fixing did not converge within the specified number of iterations");
   };
   static Real SteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform, Real alpha, GaugeMat & dmuAmu,int orthog) {

Grid/qcd/utils/WilsonLoops.h

@@ -215,7 +215,7 @@ public:
 
     double vol = Umu.Grid()->gSites();
 
-    return p.real() / vol / 4.0 / 3.0;
+    return p.real() / vol / (4.0 * Nc ) ;
   };
 
   //////////////////////////////////////////////////

Grid/random/gaussian.h

@@ -1,200 +0,0 @@
-// -*- C++ -*-
-//===--------------------------- random -----------------------------------===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-
-// Peter Boyle: Taken from libc++ in Clang/LLVM.
-// Reason is that libstdc++ and clang differ in their return order in the normal_distribution / box mueller type step.
-// standardise on one and call it "gaussian_distribution".
-
-#pragma once
-
-#include <cstddef>
-#include <cstdint>
-#include <cmath>
-#include <type_traits>
-#include <initializer_list>
-#include <limits>
-#include <algorithm>
-#include <numeric>
-#include <vector>
-#include <string>
-#include <istream>
-#include <ostream>
-#include <random>
-
-// normal_distribution -> gaussian distribution
-namespace Grid {
-
-template<class _RealType = double>
-class  gaussian_distribution
-{
-public:
-    // types
-    typedef _RealType result_type;
-
-    class param_type
-    {
-        result_type __mean_;
-        result_type __stddev_;
-    public:
-        typedef gaussian_distribution distribution_type;
-
-        strong_inline
-        explicit param_type(result_type __mean = 0, result_type __stddev = 1)
-            : __mean_(__mean), __stddev_(__stddev) {}
-
-        strong_inline
-        result_type mean() const {return __mean_;}
-        strong_inline
-        result_type stddev() const {return __stddev_;}
-
-        friend strong_inline
-            bool operator==(const param_type& __x, const param_type& __y)
-            {return __x.__mean_ == __y.__mean_ && __x.__stddev_ == __y.__stddev_;}
-        friend strong_inline
-            bool operator!=(const param_type& __x, const param_type& __y)
-            {return !(__x == __y);}
-    };
-
-private:
-    param_type __p_;
-    result_type _V_;
-    bool _V_hot_;
-
-public:
-    // constructors and reset functions
-    strong_inline
-    explicit gaussian_distribution(result_type __mean = 0, result_type __stddev = 1)
-        : __p_(param_type(__mean, __stddev)), _V_hot_(false) {}
-    strong_inline
-    explicit gaussian_distribution(const param_type& __p)
-        : __p_(__p), _V_hot_(false) {}
-    strong_inline
-    void reset() {_V_hot_ = false;}
-
-    // generating functions
-    template<class _URNG>
-        strong_inline
-        result_type operator()(_URNG& __g)
-        {return (*this)(__g, __p_);}
-    template<class _URNG> result_type operator()(_URNG& __g, const param_type& __p);
-
-    // property functions
-    strong_inline
-    result_type mean() const {return __p_.mean();}
-    strong_inline
-    result_type stddev() const {return __p_.stddev();}
-
-    strong_inline
-    param_type param() const {return __p_;}
-    strong_inline
-    void param(const param_type& __p) {__p_ = __p;}
-
-    strong_inline
-    result_type min() const {return -std::numeric_limits<result_type>::infinity();}
-    strong_inline
-    result_type max() const {return std::numeric_limits<result_type>::infinity();}
-
-    friend strong_inline
-        bool operator==(const gaussian_distribution& __x,
-                        const gaussian_distribution& __y)
-        {return __x.__p_ == __y.__p_ && __x._V_hot_ == __y._V_hot_ &&
-                (!__x._V_hot_ || __x._V_ == __y._V_);}
-    friend strong_inline
-        bool operator!=(const gaussian_distribution& __x,
-                        const gaussian_distribution& __y)
-        {return !(__x == __y);}
-
-    template <class _CharT, class _Traits, class _RT>
-    friend
-    std::basic_ostream<_CharT, _Traits>&
-    operator<<(std::basic_ostream<_CharT, _Traits>& __os,
-               const gaussian_distribution<_RT>& __x);
-
-    template <class _CharT, class _Traits, class _RT>
-    friend
-    std::basic_istream<_CharT, _Traits>&
-    operator>>(std::basic_istream<_CharT, _Traits>& __is,
-               gaussian_distribution<_RT>& __x);
-};
-
-template <class _RealType>
-template<class _URNG>
-_RealType
-gaussian_distribution<_RealType>::operator()(_URNG& __g, const param_type& __p)
-{
-    result_type _Up;
-    if (_V_hot_)
-    {
-        _V_hot_ = false;
-        _Up = _V_;
-    }
-    else
-    {
-        std::uniform_real_distribution<result_type> _Uni(-1, 1);
-        result_type __u;
-        result_type __v;
-        result_type __s;
-        do
-        {
-            __u = _Uni(__g);
-            __v = _Uni(__g);
-            __s = __u * __u + __v * __v;
-        } while (__s > 1 || __s == 0);
-        result_type _Fp = std::sqrt(-2 * std::log(__s) / __s);
-        _V_ = __v * _Fp;
-        _V_hot_ = true;
-        _Up = __u * _Fp;
-    }
-    return _Up * __p.stddev() + __p.mean();
-}
-
-template <class _CharT, class _Traits, class _RT>
-std::basic_ostream<_CharT, _Traits>&
-operator<<(std::basic_ostream<_CharT, _Traits>& __os,
-           const gaussian_distribution<_RT>& __x)
-{
-    auto __save_flags = __os.flags();
-    __os.flags(std::ios_base::dec | std::ios_base::left | std::ios_base::fixed |
-               std::ios_base::scientific);
-    _CharT __sp = __os.widen(' ');
-    __os.fill(__sp);
-    __os << __x.mean() << __sp << __x.stddev() << __sp << __x._V_hot_;
-    if (__x._V_hot_)
-        __os << __sp << __x._V_;
-    __os.flags(__save_flags);
-    return __os;
-}
-
-template <class _CharT, class _Traits, class _RT>
-std::basic_istream<_CharT, _Traits>&
-operator>>(std::basic_istream<_CharT, _Traits>& __is,
-           gaussian_distribution<_RT>& __x)
-{
-    typedef gaussian_distribution<_RT> _Eng;
-    typedef typename _Eng::result_type result_type;
-    typedef typename _Eng::param_type param_type;
-    auto __save_flags = __is.flags();
-    __is.flags(std::ios_base::dec | std::ios_base::skipws);
-    result_type __mean;
-    result_type __stddev;
-    result_type _Vp = 0;
-    bool _V_hot = false;
-    __is >> __mean >> __stddev >> _V_hot;
-    if (_V_hot)
-        __is >> _Vp;
-    if (!__is.fail())
-    {
-        __x.param(param_type(__mean, __stddev));
-        __x._V_hot_ = _V_hot;
-        __x._V_ = _Vp;
-    }
-    __is.flags(__save_flags);
-    return __is;
-}
-}

Grid/stencil/SimpleCompressor.h

@@ -52,6 +52,11 @@ public:
     return arg;
   }
 };
+class SimpleStencilParams{
+public:
+  Coordinate dirichlet;
+  SimpleStencilParams() {};
+};
 
 NAMESPACE_END(Grid);
 

Grid/stencil/Stencil.h

@@ -133,6 +133,8 @@ class CartesianStencilAccelerator {
   int           _osites;
   StencilVector _directions;
   StencilVector _distances;
+  StencilVector _comms_send;
+  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -226,6 +228,8 @@ public:
     void * recv_buf;
     Integer to_rank;
     Integer from_rank;
+    Integer do_send;
+    Integer do_recv;
     Integer bytes;
   };
   struct Merge {
@@ -240,7 +244,20 @@ public:
     cobj * mpi_p;
     Integer buffer_size;
   };
-
+  struct CopyReceiveBuffer {
+    void * from_p;
+    void * to_p;
+    Integer bytes;
+  };
+  struct CachedTransfer {
+    Integer direction;
+    Integer OrthogPlane;
+    Integer DestProc;
+    Integer bytes;
+    Integer lane;
+    Integer cb;
+    void *recv_buf;
+  };
 
 protected:
   GridBase *                        _grid;
@@ -271,7 +288,8 @@ public:
   std::vector<Merge> MergersSHM;
   std::vector<Decompress> Decompressions;
   std::vector<Decompress> DecompressionsSHM;
-
+  std::vector<CopyReceiveBuffer> CopyReceiveBuffers ;
+  std::vector<CachedTransfer> CachedTransfers;
   ///////////////////////////////////////////////////////////
   // Unified Comms buffers for all directions
   ///////////////////////////////////////////////////////////
@@ -284,29 +302,6 @@ public:
   int u_comm_offset;
   int _unified_buffer_size;
 
-  /////////////////////////////////////////
-  // Timing info; ugly; possibly temporary
-  /////////////////////////////////////////
-  double commtime;
-  double mpi3synctime;
-  double mpi3synctime_g;
-  double shmmergetime;
-  double gathertime;
-  double gathermtime;
-  double halogtime;
-  double mergetime;
-  double decompresstime;
-  double comms_bytes;
-  double shm_bytes;
-  double splicetime;
-  double nosplicetime;
-  double calls;
-  std::vector<double> comm_bytes_thr;
-  std::vector<double> shm_bytes_thr;
-  std::vector<double> comm_time_thr;
-  std::vector<double> comm_enter_thr;
-  std::vector<double> comm_leave_thr;
-
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
@@ -333,11 +328,12 @@ public:
   //////////////////////////////////////////
   // Comms packet queue for asynch thread
   // Use OpenMP Tasks for cleaner ???
+  // must be called *inside* parallel region
   //////////////////////////////////////////
+  /*
   void CommunicateThreaded()
   {
 #ifdef GRID_OMP
-    // must be called in parallel region
     int mythread = omp_get_thread_num();
     int nthreads = CartesianCommunicator::nCommThreads;
 #else
@@ -346,65 +342,29 @@ public:
 #endif
     if (nthreads == -1) nthreads = 1;
     if (mythread < nthreads) {
-      comm_enter_thr[mythread] = usecond();
       for (int i = mythread; i < Packets.size(); i += nthreads) {
 	uint64_t bytes = _grid->StencilSendToRecvFrom(Packets[i].send_buf,
 						      Packets[i].to_rank,
 						      Packets[i].recv_buf,
 						      Packets[i].from_rank,
 						      Packets[i].bytes,i);
-	comm_bytes_thr[mythread] += bytes;
-	shm_bytes_thr[mythread] += 2*Packets[i].bytes-bytes; // Send + Recv.
-
-      }
-      comm_leave_thr[mythread]= usecond();
-      comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
       }
     }
-
-  void CollateThreads(void)
-  {
-    int nthreads = CartesianCommunicator::nCommThreads;
-    double first=0.0;
-    double last =0.0;
-
-    for(int t=0;t<nthreads;t++) {
-
-      double t0 = comm_enter_thr[t];
-      double t1 = comm_leave_thr[t];
-      comms_bytes+=comm_bytes_thr[t];
-      shm_bytes  +=shm_bytes_thr[t];
-
-      comm_enter_thr[t] = 0.0;
-      comm_leave_thr[t] = 0.0;
-      comm_time_thr[t]   = 0.0;
-      comm_bytes_thr[t]=0;
-      shm_bytes_thr[t]=0;
-
-      if ( first == 0.0 ) first = t0;                   // first is t0
-      if ( (t0 > 0.0) && ( t0 < first ) ) first = t0;   // min time seen
-
-      if ( t1 > last ) last = t1;                       // max time seen
-
-    }
-    commtime+= last-first;
   }
+  */
   ////////////////////////////////////////////////////////////////////////
   // Non blocking send and receive. Necessarily parallel.
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
     reqs.resize(Packets.size());
-    commtime-=usecond();
     for(int i=0;i<Packets.size();i++){
-      uint64_t bytes=_grid->StencilSendToRecvFromBegin(reqs[i],
+      _grid->StencilSendToRecvFromBegin(reqs[i],
 					Packets[i].send_buf,
-						     Packets[i].to_rank,
+					Packets[i].to_rank,Packets[i].do_send,
 					Packets[i].recv_buf,
-						     Packets[i].from_rank,
+					Packets[i].from_rank,Packets[i].do_recv,
 					Packets[i].bytes,i);
-      comms_bytes+=bytes;
-      shm_bytes  +=2*Packets[i].bytes-bytes;
     }
   }
 
@@ -413,7 +373,6 @@ public:
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
-    commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
@@ -421,28 +380,27 @@ public:
   void Communicate(void)
   {
     if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
-      thread_region {
-	// must be called in parallel region
-	int mythread  = thread_num();
-	int maxthreads= thread_max();
+      /////////////////////////////////////////////////////////
+      // several way threaded on different communicators.
+      // Cannot combine with Dirichlet operators
+      // This scheme is needed on Intel Omnipath for best performance
+      // Deprecate once there are very few omnipath clusters
+      /////////////////////////////////////////////////////////
       int nthreads = CartesianCommunicator::nCommThreads;
-	assert(nthreads <= maxthreads);
-	if (nthreads == -1) nthreads = 1;
-	if (mythread < nthreads) {
-	  for (int i = mythread; i < Packets.size(); i += nthreads) {
-	    double start = usecond();
-	    uint64_t bytes= _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-							 Packets[i].to_rank,
+      int old = GridThread::GetThreads();
+      GridThread::SetThreads(nthreads);
+      thread_for(i,Packets.size(),{
+	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
+				       Packets[i].to_rank,Packets[i].do_send,
 				       Packets[i].recv_buf,
-							 Packets[i].from_rank,
+				       Packets[i].from_rank,Packets[i].do_recv,
 				       Packets[i].bytes,i);
-	    comm_bytes_thr[mythread] += bytes;
-	    shm_bytes_thr[mythread]  += Packets[i].bytes - bytes;
-	    comm_time_thr[mythread]  += usecond() - start;
-	  }
-	}
-      }
-    } else { // Concurrent and non-threaded asynch calls to MPI
+      });
+      GridThread::SetThreads(old);
+    } else { 
+      /////////////////////////////////////////////////////////
+      // Concurrent and non-threaded asynch calls to MPI
+      /////////////////////////////////////////////////////////
       std::vector<std::vector<CommsRequest_t> > reqs;
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
@@ -484,31 +442,23 @@ public:
       sshift[1] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Odd);
       if ( sshift[0] == sshift[1] ) {
 	if (splice_dim) {
-	  splicetime-=usecond();
-	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
+	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
-	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
+	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp;
-	  nosplicetime+=usecond();
 	}
       } else {
 	if(splice_dim){
-	  splicetime-=usecond();
 	  // if checkerboard is unfavourable take two passes
 	  // both with block stride loop iteration
-	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx);
-	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx);
+	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx,point);
+	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
-	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
-	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
+	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx,point);
+	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  nosplicetime+=usecond();
 	}
       }
     }
@@ -518,13 +468,10 @@ public:
   template<class compressor>
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
-    mpi3synctime_g-=usecond();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime_g+=usecond();
 
     // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
-    halogtime-=usecond();
 
     u_comm_offset=0;
 
@@ -538,7 +485,6 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
-    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -551,14 +497,70 @@ public:
     Mergers.resize(0);
     MergersSHM.resize(0);
     Packets.resize(0);
-    calls++;
+    CopyReceiveBuffers.resize(0);
+    CachedTransfers.resize(0);
   }
-  void AddPacket(void *xmit,void * rcv, Integer to,Integer from,Integer bytes){
+  void AddCopy(void *from,void * to, Integer bytes)
+  {
+    CopyReceiveBuffer obj;
+    obj.from_p = from;
+    obj.to_p = to;
+    obj.bytes= bytes;
+    CopyReceiveBuffers.push_back(obj);
+  }
+  void CommsCopy()
+  {
+    //    These are device resident MPI buffers.
+    for(int i=0;i<CopyReceiveBuffers.size();i++){
+      cobj *from=(cobj *)CopyReceiveBuffers[i].from_p;
+      cobj *to  =(cobj *)CopyReceiveBuffers[i].to_p;
+      Integer words = CopyReceiveBuffers[i].bytes/sizeof(cobj);
+
+      accelerator_forNB(j, words, cobj::Nsimd(), {
+	  coalescedWrite(to[j] ,coalescedRead(from [j]));
+      });
+    }
+  }
+  
+  Integer CheckForDuplicate(Integer direction, Integer OrthogPlane, Integer DestProc, void *recv_buf,Integer lane,Integer bytes,Integer cb)
+  {
+    CachedTransfer obj;
+    obj.direction   = direction;
+    obj.OrthogPlane = OrthogPlane;
+    obj.DestProc    = DestProc;
+    obj.recv_buf    = recv_buf;
+    obj.lane        = lane;
+    obj.bytes       = bytes;
+    obj.cb          = cb;
+
+    for(int i=0;i<CachedTransfers.size();i++){
+      if (   (CachedTransfers[i].direction  ==direction)
+	   &&(CachedTransfers[i].OrthogPlane==OrthogPlane)
+	   &&(CachedTransfers[i].DestProc   ==DestProc)
+	   &&(CachedTransfers[i].bytes      ==bytes)
+	   &&(CachedTransfers[i].lane       ==lane)
+	   &&(CachedTransfers[i].cb         ==cb)
+	     ){
+
+	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
+	return 1;
+      }
+    }
+
+    CachedTransfers.push_back(obj);
+    return 0;
+  }
+  void AddPacket(void *xmit,void * rcv,
+		 Integer to, Integer do_send,
+		 Integer from, Integer do_recv,
+		 Integer bytes){
     Packet p;
     p.send_buf = xmit;
     p.recv_buf = rcv;
     p.to_rank  = to;
     p.from_rank= from;
+    p.do_send  = do_send;
+    p.do_recv  = do_recv;
     p.bytes    = bytes;
     Packets.push_back(p);
   }
@@ -578,22 +580,17 @@ public:
     mv.push_back(m);
   }
   template<class decompressor>  void CommsMerge(decompressor decompress)    {
+    CommsCopy();
     CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    mpi3synctime-=usecond();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime+=usecond();
-    shmmergetime-=usecond();
     CommsMerge(decompress,MergersSHM,DecompressionsSHM);
-    shmmergetime+=usecond();
   }
 
   template<class decompressor>
-  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
-
-    
-    mergetime-=usecond();
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd)
+  {
     for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
@@ -603,9 +600,7 @@ public:
 	  decompress.Exchange(mp,vp0,vp1,type,o);
       });
     }
-    mergetime+=usecond();
 
-    decompresstime-=usecond();
     for(int i=0;i<dd.size();i++){
       auto kp = dd[i].kernel_p;
       auto mp = dd[i].mpi_p;
@@ -613,7 +608,6 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
-    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -646,22 +640,60 @@ public:
 	}
       }
       if(local == 0) {
-	surface_list.push_back(site);
+	for(int s=0;s<Ls;s++){
+	  surface_list.push_back(site*Ls+s);
+	}
+      }
+    }
+  }
+  /// Introduce a block structure and switch off comms on boundaries
+  void DirichletBlock(const Coordinate &dirichlet_block)
+  {
+    for(int ii=0;ii<this->_npoints;ii++){
+      int dimension    = this->_directions[ii];
+      int displacement = this->_distances[ii];
+      int gd = _grid->_gdimensions[dimension];
+      int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      int pc = _grid->_processor_coor[dimension];
+      int ld = fd/pd;
+      ///////////////////////////////////////////
+      // Figure out dirichlet send and receive
+      // on this leg of stencil.
+      ///////////////////////////////////////////
+      int comm_dim        = _grid->_processors[dimension] >1 ;
+      int block = dirichlet_block[dimension];
+      this->_comms_send[ii] = comm_dim;
+      this->_comms_recv[ii] = comm_dim;
+      if ( block && comm_dim ) {
+	assert(abs(displacement) < ld );
+	// Quiesce communication across block boundaries
+	if( displacement > 0 ) {
+	  // High side, low side
+	  // | <--B--->|
+	  // |    |    |
+	  //           noR
+	  // noS
+	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_recv[ii] = 0;
+	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_send[ii] = 0;
+	} else {
+	  // High side, low side
+	  // | <--B--->|
+	  // |    |    |
+	  //           noS
+	  // noR
+	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_send[ii] = 0;
+	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_recv[ii] = 0;
+	}
       }
     }
   }
-
   CartesianStencil(GridBase *grid,
 		   int npoints,
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
 		   Parameters p)
-    : shm_bytes_thr(npoints),
-      comm_bytes_thr(npoints),
-      comm_enter_thr(npoints),
-      comm_leave_thr(npoints),
-      comm_time_thr(npoints)
   {
     face_table_computed=0;
     _grid    = grid;
@@ -675,8 +707,12 @@ public:
     this->_simd_layout = _grid->_simd_layout; // copy simd_layout to give access to Accelerator Kernels
     this->_directions = StencilVector(directions);
     this->_distances  = StencilVector(distances);
+    this->_comms_send.resize(npoints); 
+    this->_comms_recv.resize(npoints); 
     this->same_node.resize(npoints);
 
+    if ( p.dirichlet.size() ) DirichletBlock(p.dirichlet); // comms send/recv set up
+
     _unified_buffer_size=0;
     surface_list.resize(0);
 
@@ -693,15 +729,16 @@ public:
       int displacement = distances[i];
       int shift = displacement;
 
+      int gd = _grid->_gdimensions[dimension];
       int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      //      int ld = gd/pd;
       int rd = _grid->_rdimensions[dimension];
+      int pc = _grid->_processor_coor[dimension];
       this->_permute_type[point]=_grid->PermuteType(dimension);
 
       this->_checkerboard = checkerboard;
 
-      //////////////////////////
-      // the permute type
-      //////////////////////////
       int simd_layout     = _grid->_simd_layout[dimension];
       int comm_dim        = _grid->_processors[dimension] >1 ;
       int splice_dim      = _grid->_simd_layout[dimension]>1 && (comm_dim);
@@ -710,7 +747,6 @@ public:
       assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
 
       int sshift[2];
-
       //////////////////////////
       // Underlying approach. For each local site build
       // up a table containing the npoint "neighbours" and whether they
@@ -811,6 +847,7 @@ public:
     GridBase *grid=_grid;
     const int Nsimd = grid->Nsimd();
 
+    int comms_recv      = this->_comms_recv[point];
     int fd              = _grid->_fdimensions[dimension];
     int ld              = _grid->_ldimensions[dimension];
     int rd              = _grid->_rdimensions[dimension];
@@ -836,12 +873,14 @@ public:
     for(int x=0;x<rd;x++){
 
       int permute_type=grid->PermuteType(dimension);
+      int permute_slice;
 
       int sx        =  (x+sshift)%rd;
 
       int offnode = 0;
       if ( simd_layout > 1 ) {
 
+	permute_slice=1;
 	for(int i=0;i<Nsimd;i++){
 
 	  int inner_bit = (Nsimd>>(permute_type+1));
@@ -858,6 +897,7 @@ public:
       } else {
 	int comm_proc = ((x+sshift)/rd)%pd;
 	offnode = (comm_proc!= 0);
+	permute_slice=0;
       }
 
       int wraparound=0;
@@ -867,25 +907,31 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-      if (!offnode) {
 
-	int permute_slice=0;
+      // Wrap locally dirichlet support case OR node local
+      if ( offnode==0 ) {
+
+	permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
 	
       } else {
 
+	if ( comms_recv ) {
+
+	  ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
+
+	} else { 
+
+	  CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
+
+	}
+
+      }
+      
+      if ( offnode ) {
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
-
-	//	int rank           = grid->_processor;
-	//	int recv_from_rank;
-	//	int xmit_to_rank;
-
-	int unified_buffer_offset = _unified_buffer_size;
 	_unified_buffer_size    += words;
-
-	ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
-
       }
     }
   }
@@ -984,11 +1030,14 @@ public:
   }
 
   template<class compressor>
-  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
+  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx, int point)
   {
     typedef typename cobj::vector_type vector_type;
     typedef typename cobj::scalar_type scalar_type;
 
+    int comms_send   = this->_comms_send[point] ;
+    int comms_recv   = this->_comms_recv[point] ;
+
     assert(rhs.Grid()==_grid);
     //	  conformable(_grid,rhs.Grid());
 
@@ -1020,69 +1069,84 @@ public:
 	int bytes =  words * compress.CommDatumSize();
 
 	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
+	int comm_off = u_comm_offset;
+
+	int recv_from_rank;
+	int xmit_to_rank;
+	cobj *recv_buf;
+	cobj *send_buf;
+	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+	assert (xmit_to_rank   != _grid->ThisRank());
+	assert (recv_from_rank != _grid->ThisRank());
+
+	if( comms_send ) {
+
 	  if ( !face_table_computed ) {
 	    face_table.resize(face_idx+1);
 	    std::vector<std::pair<int,int> >  face_table_host ;
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
+	    Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
 	    face_table[face_idx].resize(face_table_host.size());
 	    acceleratorCopyToDevice(&face_table_host[0],
 				    &face_table[face_idx][0],
 				    face_table[face_idx].size()*sizeof(face_table_host[0]));
 	  }
 
-	//      	int rank           = _grid->_processor;
-	int recv_from_rank;
-	int xmit_to_rank;
-	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
-	assert (xmit_to_rank   != _grid->ThisRank());
-	assert (recv_from_rank != _grid->ThisRank());
-
-	cobj *recv_buf;
 	  if ( compress.DecompressionStep() ) {
 	    recv_buf=u_simd_recv_buf[0];
 	  } else {
 	    recv_buf=this->u_recv_buf_p;
 	  }
 
-	cobj *send_buf;
 	  send_buf = this->u_send_buf_p; // Gather locally, must send
 	
 	  ////////////////////////////////////////////////////////
 	  // Gather locally
 	  ////////////////////////////////////////////////////////
-	gathertime-=usecond();
 	  assert(send_buf!=NULL);
-	Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so); face_idx++;
-	gathertime+=usecond();
+
+	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,comm_off,so);
+	}
+
+	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,bytes,cbmask);
+	if ( (!duplicate) ) { // Force comms for now
 
 	  ///////////////////////////////////////////////////////////
 	  // Build a list of things to do after we synchronise GPUs
 	  // Start comms now???
 	  ///////////////////////////////////////////////////////////
-	AddPacket((void *)&send_buf[u_comm_offset],
-		  (void *)&recv_buf[u_comm_offset],
-		  xmit_to_rank,
-		  recv_from_rank,
+	  AddPacket((void *)&send_buf[comm_off],
+		    (void *)&recv_buf[comm_off],
+		    xmit_to_rank, comms_send,
+		    recv_from_rank, comms_recv,
 		    bytes);
+	}
 	
-	if ( compress.DecompressionStep() ) {
-	  AddDecompress(&this->u_recv_buf_p[u_comm_offset],
-			&recv_buf[u_comm_offset],
+	if ( compress.DecompressionStep()  && comms_recv ) {
+	  AddDecompress(&this->u_recv_buf_p[comm_off],
+			&recv_buf[comm_off],
 			words,Decompressions);
 	}
+	
 	u_comm_offset+=words;
+	face_idx++;
+
       }
     }
     return 0;
   }
 
   template<class compressor>
-  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
+  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx,int point)
   {
     const int Nsimd = _grid->Nsimd();
 
     const int maxl =2;// max layout in a direction
+
+    int comms_send   = this->_comms_send[point] ;
+    int comms_recv   = this->_comms_recv[point] ;
+
     int fd = _grid->_fdimensions[dimension];
     int rd = _grid->_rdimensions[dimension];
     int ld = _grid->_ldimensions[dimension];
@@ -1097,7 +1161,6 @@ public:
 
 
     int permute_type=_grid->PermuteType(dimension);
-    //    std::cout << "SimdNew permute type "<<permute_type<<std::endl;
 
     ///////////////////////////////////////////////
     // Simd direction uses an extract/merge pair
@@ -1131,8 +1194,9 @@ public:
 
       if ( any_offnode ) {
 
+	int comm_off = u_comm_offset;
 	for(int i=0;i<maxl;i++){
-	  spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
+	  spointers[i] = (cobj *) &u_simd_send_buf[i][comm_off];
 	}
 
 	int sx   = (x+sshift)%rd;
@@ -1141,18 +1205,17 @@ public:
 	  face_table.resize(face_idx+1);
 	  std::vector<std::pair<int,int> >  face_table_host ;
 				
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
+	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
 	  face_table[face_idx].resize(face_table_host.size());
 	  acceleratorCopyToDevice(&face_table_host[0],
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
 	}
-	gathermtime-=usecond();
 
+	if ( comms_send || comms_recv )
 	  Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
 	face_idx++;
 
-	gathermtime+=usecond();
 	//spointers[0] -- low
 	//spointers[1] -- high
 
@@ -1169,8 +1232,8 @@ public:
 	  int nbr_plane = nbr_ic;
 	  assert (sx == nbr_ox);
 
-	  auto rp = &u_simd_recv_buf[i        ][u_comm_offset];
-	  auto sp = &u_simd_send_buf[nbr_plane][u_comm_offset];
+	  auto rp = &u_simd_recv_buf[i        ][comm_off];
+	  auto sp = &u_simd_send_buf[nbr_plane][comm_off];
 
 	  if(nbr_proc){
 
@@ -1181,8 +1244,13 @@ public:
 
 	    rpointers[i] = rp;
 
-	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
-
+	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,bytes,cbmask);
+	    if ( !duplicate  ) { 
+	      AddPacket((void *)sp,(void *)rp,
+			xmit_to_rank,comms_send,
+			recv_from_rank,comms_recv,
+			bytes);
+	    }
 
 	  } else {
 
@@ -1191,9 +1259,12 @@ public:
 	  }
 	}
 
-	AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
+	if ( comms_recv ) {
+	  AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
+	}
 
 	u_comm_offset     +=buffer_size;
+
       }
     }
     return 0;

Grid/tensors/Tensor_exp.h

@@ -55,7 +55,7 @@ template<class vtype, int N> accelerator_inline iVector<vtype, N> Exponentiate(c
 
 
 // Specialisation: Cayley-Hamilton exponential for SU(3)
-#ifndef GRID_CUDA
+#ifndef GRID_ACCELERATED
 template<class vtype, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0>::type * =nullptr> 
 accelerator_inline iMatrix<vtype,3> Exponentiate(const iMatrix<vtype,3> &arg, RealD alpha  , Integer Nexp = DEFAULT_MAT_EXP )
 {

Grid/threads/Accelerator.cc

@@ -6,9 +6,17 @@ uint32_t accelerator_threads=2;
 uint32_t acceleratorThreads(void)       {return accelerator_threads;};
 void     acceleratorThreads(uint32_t t) {accelerator_threads = t;};
 
+#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
+#define ENV_LOCAL_RANK_SLURM   "SLURM_LOCALID"
+#define ENV_RANK_SLURM         "SLURM_PROCID"
+#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
+#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
+
 #ifdef GRID_CUDA
 cudaDeviceProp *gpu_props;
 cudaStream_t copyStream;
+cudaStream_t cpuStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -17,12 +25,6 @@ void acceleratorInit(void)
 
   char * localRankStr = NULL;
   int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
-#define ENV_LOCAL_RANK_SLURM   "SLURM_LOCALID"
-#define ENV_RANK_SLURM         "SLURM_PROCID"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_SLURM  )) != NULL) { world_rank = atoi(localRankStr);}
@@ -97,6 +99,7 @@ void acceleratorInit(void)
 
   cudaSetDevice(device);
   cudaStreamCreate(&copyStream);
+  cudaStreamCreate(&cpuStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -111,6 +114,7 @@ void acceleratorInit(void)
 #ifdef GRID_HIP
 hipDeviceProp_t *gpu_props;
 hipStream_t copyStream;
+hipStream_t cpuStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -119,10 +123,6 @@ void acceleratorInit(void)
 
   char * localRankStr = NULL;
   int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
-#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
   // We extract the local rank initialization using an environment variable
   if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL)
   {
@@ -134,7 +134,9 @@ void acceleratorInit(void)
   }
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
+  if ((localRankStr = getenv(ENV_RANK_SLURM  )) != NULL) { world_rank = atoi(localRankStr);}
 
+  if ( world_rank == 0 ) 
     printf("world_rank %d has %d devices\n",world_rank,nDevices);
   size_t totalDeviceMem=0;
   for (int i = 0; i < nDevices; i++) {
@@ -181,6 +183,7 @@ void acceleratorInit(void)
 #endif
   hipSetDevice(device);
   hipStreamCreate(&copyStream);
+  hipStreamCreate(&cpuStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -208,10 +211,7 @@ void acceleratorInit(void)
   
   char * localRankStr = NULL;
   int rank = 0, world_rank=0; 
-#define ENV_LOCAL_RANK_OMPI    "OMPI_COMM_WORLD_LOCAL_RANK"
-#define ENV_LOCAL_RANK_MVAPICH "MV2_COMM_WORLD_LOCAL_RANK"
-#define ENV_RANK_OMPI          "OMPI_COMM_WORLD_RANK"
-#define ENV_RANK_MVAPICH       "MV2_COMM_WORLD_RANK"
+
   // We extract the local rank initialization using an environment variable
   if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL)
   {

Grid/threads/Accelerator.h

@@ -107,6 +107,7 @@ void     acceleratorInit(void);
 
 extern int acceleratorAbortOnGpuError;
 extern cudaStream_t copyStream;
+extern cudaStream_t cpuStream;
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
@@ -134,7 +135,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_threads(nsimd,acceleratorThreads(),1);			\
     dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
-    LambdaApply<<<cu_blocks,cu_threads>>>(num1,num2,nsimd,lambda);	\
+    LambdaApply<<<cu_blocks,cu_threads,0,cpuStream>>>(num1,num2,nsimd,lambda);	\
   }
 
 #define accelerator_for6dNB(iter1, num1,				\
@@ -153,7 +154,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_blocks (num1,num2,num3);					\
     dim3 cu_threads(num4,num5,num6);					\
-    Lambda6Apply<<<cu_blocks,cu_threads>>>(num1,num2,num3,num4,num5,num6,lambda); \
+    Lambda6Apply<<<cu_blocks,cu_threads,0,cpuStream>>>(num1,num2,num3,num4,num5,num6,lambda); \
   }
 
 template<typename lambda>  __global__
@@ -189,7 +190,7 @@ void Lambda6Apply(uint64_t num1, uint64_t num2, uint64_t num3,
 
 #define accelerator_barrier(dummy)					\
   {									\
-    cudaDeviceSynchronize();						\
+    cudaStreamSynchronize(cpuStream);					\
     cudaError err = cudaGetLastError();					\
     if ( cudaSuccess != err ) {						\
       printf("accelerator_barrier(): Cuda error %s \n",			\
@@ -206,8 +207,7 @@ inline void *acceleratorAllocShared(size_t bytes)
   auto err = cudaMallocManaged((void **)&ptr,bytes);
   if( err != cudaSuccess ) {
     ptr = (void *) NULL;
-    printf(" cudaMallocManaged failed for %lu %s \n",bytes,cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
+    printf(" cudaMallocManaged failed for %d %s \n",bytes,cudaGetErrorString(err));
   }
   return ptr;
 };
@@ -217,47 +217,15 @@ inline void *acceleratorAllocDevice(size_t bytes)
   auto err = cudaMalloc((void **)&ptr,bytes);
   if( err != cudaSuccess ) {
     ptr = (void *) NULL;
-    printf(" cudaMalloc failed for %lu %s \n",bytes,cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
+    printf(" cudaMalloc failed for %d %s \n",bytes,cudaGetErrorString(err));
   }
   return ptr;
 };
-inline void acceleratorFreeShared(void *ptr){
-  auto err = cudaFree(ptr);
-  if( err != cudaSuccess ) {
-    printf(" cudaFree(Shared) failed %s \n",cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
-  }
-};
-inline void acceleratorFreeDevice(void *ptr){
-  auto err = cudaFree(ptr);
-  if( err != cudaSuccess ) {
-    printf(" cudaFree(Device) failed %s \n",cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
-  }
-};
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  {
-  auto err = cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);
-  if( err != cudaSuccess ) {
-    printf(" cudaMemcpy(host->device) failed for %lu %s \n",bytes,cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
-  }
-}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){
-  auto err = cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);
-  if( err != cudaSuccess ) {
-    printf(" cudaMemcpy(device->host) failed for %lu %s \n",bytes,cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
-  }
-}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) {
-  auto err = cudaMemset(base,value,bytes);
-  if( err != cudaSuccess ) {
-    printf(" cudaMemSet failed for %lu %s \n",bytes,cudaGetErrorString(err)); fflush(stdout);
-    if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);
-  }
-}
-
+inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
+inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
+inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
   cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
@@ -372,6 +340,7 @@ NAMESPACE_BEGIN(Grid);
 #define accelerator_inline __host__ __device__ inline
 
 extern hipStream_t copyStream;
+extern hipStream_t cpuStream;
 /*These routines define mapping from thread grid to loop & vector lane indexing */
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
@@ -393,11 +362,11 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
     dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
     if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
       hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads,		\
-            0,0,						\
+			 0,cpuStream,					\
 			 num1,num2,nsimd, lambda);			\
     } else { \
       hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
-            0,0,						\
+			 0,cpuStream,					\
 			 num1,num2,nsimd, lambda);			\
     } \
   }
@@ -431,7 +400,7 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 
 #define accelerator_barrier(dummy)				\
   {								\
-    hipDeviceSynchronize();					\
+    hipStreamSynchronize(cpuStream);					\
     auto err = hipGetLastError();				\
     if ( err != hipSuccess ) {					\
       printf("After hipDeviceSynchronize() : HIP error %s \n", hipGetErrorString( err )); \
@@ -474,7 +443,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
-  hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToDevice,copyStream);
+  hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);
 }
 inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
 
@@ -494,6 +463,8 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
   accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, { __VA_ARGS__ } ); \
   accelerator_barrier(dummy);
 
+#define GRID_ACCELERATED
+
 #endif
 
 //////////////////////////////////////////////
@@ -514,9 +485,10 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
 #define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) thread_for2d(iter1,num1,iter2,num2,{ __VA_ARGS__ });
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
+
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes);}
 inline void acceleratorCopySynchronise(void) {};
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }

Grid/threads/Threads.h

@@ -72,3 +72,20 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define thread_region                                       DO_PRAGMA(omp parallel)
 #define thread_critical                                     DO_PRAGMA(omp critical)
 
+#ifdef GRID_OMP
+inline void thread_bcopy(void *from, void *to,size_t bytes)
+{
+  uint64_t *ufrom = (uint64_t *)from;
+  uint64_t *uto   = (uint64_t *)to;
+  assert(bytes%8==0);
+  uint64_t words=bytes/8;
+  thread_for(w,words,{
+      uto[w] = ufrom[w];
+  });
+}
+#else
+inline void thread_bcopy(void *from, void *to,size_t bytes)
+{
+  bcopy(from,to,bytes);
+}
+#endif

HMC/DWF2p1fIwasakiGparity.cc

@@ -1,473 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./HMC/DWF2p1fIwasakiGparity.cc
-
-Copyright (C) 2015-2016
-
-Author: Christopher Kelly <ckelly@bnl.gov>
-Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
-
-
-This program is free software; you can redistribute it and/or modify
-it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2 of the License, or
-(at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License along
-with this program; if not, write to the Free Software Foundation, Inc.,
-51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-See the full license in the file "LICENSE" in the top level distribution
-directory
-*************************************************************************************/
-/*  END LEGAL */
-#include <Grid/Grid.h>
-
-using namespace Grid;
-
-//2+1f DWF+I ensemble with G-parity BCs
-//designed to reproduce ensembles in https://arxiv.org/pdf/1908.08640.pdf
-struct RatQuoParameters: Serializable {
-  GRID_SERIALIZABLE_CLASS_MEMBERS(RatQuoParameters,
-				  double, bnd_lo,
-				  double, bnd_hi,
-				  Integer, action_degree,
-				  double, action_tolerance,
-				  Integer, md_degree,
-				  double, md_tolerance,
-				  Integer, reliable_update_freq,
-				  Integer, bnd_check_freq);
-  RatQuoParameters() { 
-    bnd_lo = 1e-2;
-    bnd_hi = 30;
-    action_degree = 10;
-    action_tolerance = 1e-10;
-    md_degree = 10;
-    md_tolerance = 1e-8;
-    bnd_check_freq = 20;
-    reliable_update_freq = 50;
-  }
-
-  void Export(RationalActionParams &into) const{
-    into.lo = bnd_lo;
-    into.hi = bnd_hi;
-    into.action_degree = action_degree;
-    into.action_tolerance = action_tolerance;
-    into.md_degree = md_degree;
-    into.md_tolerance = md_tolerance;
-    into.BoundsCheckFreq = bnd_check_freq;
-  }
-};
-
-
-struct EvolParameters: Serializable {
-  GRID_SERIALIZABLE_CLASS_MEMBERS(EvolParameters,
-                                  Integer, StartTrajectory,
-                                  Integer, Trajectories,
-				  Integer, SaveInterval,
-				  Integer, Steps,
-                                  bool, MetropolisTest,
-				  std::string, StartingType,
-				  std::vector<Integer>, GparityDirs,
-				  RatQuoParameters, rat_quo_l,
-				  RatQuoParameters, rat_quo_s);
-
-  EvolParameters() {
-    //For initial thermalization; afterwards user should switch Metropolis on and use StartingType=CheckpointStart
-    MetropolisTest    = false;
-    StartTrajectory   = 0;
-    Trajectories      = 50;
-    SaveInterval = 5;
-    StartingType      = "ColdStart";
-    GparityDirs.resize(3, 1); //1 for G-parity, 0 for periodic
-    Steps = 5;
-  }
-};
-
-bool fileExists(const std::string &fn){
-  std::ifstream f(fn);
-  return f.good();
-}
-
-
-
-
-struct LanczosParameters: Serializable {
-  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
-				  double, alpha,
-				  double, beta,
-				  double, mu,
-				  int, ord,
-				  int, n_stop,
-				  int, n_want,
-				  int, n_use,
-				  double, tolerance);
-
-  LanczosParameters() {
-    alpha = 35;
-    beta = 5;
-    mu = 0;
-    ord = 100;
-    n_stop = 10;
-    n_want = 10;
-    n_use = 15;
-    tolerance = 1e-6;
-  }
-};
-
-
-
-template<typename FermionActionD, typename FermionFieldD>
-void computeEigenvalues(std::string param_file,
-			GridCartesian* Grid, GridRedBlackCartesian* rbGrid, const LatticeGaugeFieldD &latt,  //expect lattice to have been initialized to something
-			FermionActionD &action, GridParallelRNG &rng){
-  
-  LanczosParameters params;
-  if(fileExists(param_file)){
-    std::cout << GridLogMessage << " Reading " << param_file << std::endl;
-    Grid::XmlReader rd(param_file);
-    read(rd, "LanczosParameters", params);
-  }else if(!GlobalSharedMemory::WorldRank){
-    std::cout << GridLogMessage << " File " << param_file << " does not exist" << std::endl;
-    std::cout << GridLogMessage << " Writing xml template to " << param_file << ".templ" << std::endl;
-    Grid::XmlWriter wr(param_file + ".templ");
-    write(wr, "LanczosParameters", params);
-  }
-
-  FermionFieldD gauss_o(rbGrid);
-  FermionFieldD gauss(Grid);
-  gaussian(rng, gauss);
-  pickCheckerboard(Odd, gauss_o, gauss);
-
-  action.ImportGauge(latt);
-
-  SchurDiagMooeeOperator<FermionActionD, FermionFieldD> hermop(action);
-  PlainHermOp<FermionFieldD> hermop_wrap(hermop);
-  //ChebyshevLanczos<FermionFieldD> Cheb(params.alpha, params.beta, params.mu, params.ord);
-  assert(params.mu == 0.0);
-
-  Chebyshev<FermionFieldD> Cheb(params.beta*params.beta, params.alpha*params.alpha, params.ord+1);
-  FunctionHermOp<FermionFieldD> Cheb_wrap(Cheb, hermop);
-
-  std::cout << "IRL: alpha=" << params.alpha << " beta=" << params.beta << " mu=" << params.mu << " ord=" << params.ord << std::endl;
-  ImplicitlyRestartedLanczos<FermionFieldD> IRL(Cheb_wrap, hermop_wrap, params.n_stop, params.n_want, params.n_use, params.tolerance, 10000);
-
-  std::vector<RealD> eval(params.n_use);
-  std::vector<FermionFieldD> evec(params.n_use, rbGrid);
-  int Nconv;
-  IRL.calc(eval, evec, gauss_o, Nconv);
-
-  std::cout << "Eigenvalues:" << std::endl;
-  for(int i=0;i<params.n_want;i++){
-    std::cout << i << " " << eval[i] << std::endl;
-  }
-}
-
-
-//Check the quality of the RHMC approx
-template<typename FermionActionD, typename FermionFieldD, typename RHMCtype>
-void checkRHMC(GridCartesian* Grid, GridRedBlackCartesian* rbGrid, const LatticeGaugeFieldD &latt,  //expect lattice to have been initialized to something
-	       FermionActionD &numOp, FermionActionD &denOp, RHMCtype &rhmc, GridParallelRNG &rng,
-	       int inv_pow, const std::string &quark_descr){
-
-  FermionFieldD gauss_o(rbGrid);
-  FermionFieldD gauss(Grid);
-  gaussian(rng, gauss);
-  pickCheckerboard(Odd, gauss_o, gauss);
-
-  numOp.ImportGauge(latt);
-  denOp.ImportGauge(latt);
-
-  typedef typename FermionActionD::Impl_t FermionImplPolicyD;
-  SchurDifferentiableOperator<FermionImplPolicyD> MdagM(numOp);
-  SchurDifferentiableOperator<FermionImplPolicyD> VdagV(denOp);
-      
-  std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
-  InversePowerBoundsCheck(inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegPowerAction); //use large tolerance to prevent exit on fail; we are trying to tune here!
-  std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
-  InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegHalfPowerAction);
-  std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
-  InversePowerBoundsCheck(inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegPowerAction);
-  std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
-  InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegHalfPowerAction);
-  std::cout << "Finished: Checking quality of RHMC action approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
-
-  std::cout << "-------------------------------------------------------------------------------" << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
-  InversePowerBoundsCheck(inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegPowerMD); 
-  std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << inv_pow << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
-  InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, MdagM,gauss_o, rhmc.ApproxNegHalfPowerMD);
-  std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark numerator and power -1/" << 2*inv_pow << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
-  InversePowerBoundsCheck(inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegPowerMD);
-  std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << inv_pow << std::endl;
-
-  std::cout << "Starting: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
-  InversePowerBoundsCheck(2*inv_pow, 10000, 1e16, VdagV,gauss_o, rhmc.ApproxNegHalfPowerMD);
-  std::cout << "Finished: Checking quality of RHMC MD approx for " << quark_descr << " quark denominator and power -1/" << 2*inv_pow << std::endl;
-}
-
-
-
-
-
-
-
-
-
-int main(int argc, char **argv) {
-  Grid_init(&argc, &argv);
-  int threads = GridThread::GetThreads();
-  // here make a routine to print all the relevant information on the run
-  std::cout << GridLogMessage << "Grid is setup to use " << threads << " threads" << std::endl;
-
-  std::string param_file = "params.xml";
-  bool file_load_check = false;
-  for(int i=1;i<argc;i++){
-    std::string sarg(argv[i]);
-    if(sarg == "--param_file"){
-      assert(i!=argc-1);
-      param_file = argv[i+1];
-    }else if(sarg == "--read_check"){ //check the fields load correctly and pass checksum/plaquette repro
-      file_load_check = true;
-    }
-  }
-
-  //Read the user parameters
-  EvolParameters user_params;
-  
-  if(fileExists(param_file)){
-    std::cout << GridLogMessage << " Reading " << param_file << std::endl;
-    Grid::XmlReader rd(param_file);
-    read(rd, "Params", user_params);
-  }else if(!GlobalSharedMemory::WorldRank){
-    std::cout << GridLogMessage << " File " << param_file << " does not exist" << std::endl;
-    std::cout << GridLogMessage << " Writing xml template to " << param_file << ".templ" << std::endl;
-    Grid::XmlWriter wr(param_file + ".templ");
-    write(wr, "Params", user_params);
-
-    std::cout << GridLogMessage << " Done" << std::endl;
-    Grid_finalize();
-    return 0;
-  }
-
-  //Check the parameters
-  if(user_params.GparityDirs.size() != Nd-1){
-    std::cerr << "Error in input parameters: expect GparityDirs to have size = " << Nd-1 << std::endl;
-    exit(1);
-  }
-  for(int i=0;i<Nd-1;i++)
-    if(user_params.GparityDirs[i] != 0 && user_params.GparityDirs[i] != 1){
-      std::cerr << "Error in input parameters: expect GparityDirs values to be 0 (periodic) or 1 (G-parity)" << std::endl;
-      exit(1);
-    }
-
-   // Typedefs to simplify notation
-  typedef GparityDomainWallFermionD FermionActionD;
-  typedef typename FermionActionD::Impl_t FermionImplPolicyD;
-  typedef typename FermionActionD::FermionField FermionFieldD;
-
-  typedef GparityDomainWallFermionF FermionActionF;
-  typedef typename FermionActionF::Impl_t FermionImplPolicyF;
-  typedef typename FermionActionF::FermionField FermionFieldF;
-
-  typedef GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicyD,FermionImplPolicyF> MixedPrecRHMC;
-  typedef GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicyD> DoublePrecRHMC;
-
-  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
-  IntegratorParameters MD;
-  typedef ConjugateHMCRunnerD<MinimumNorm2> HMCWrapper; //NB: This is the "Omelyan integrator"
-  typedef HMCWrapper::ImplPolicy GaugeImplPolicy;
-  MD.name    = std::string("MinimumNorm2");
-  MD.MDsteps = user_params.Steps;
-  MD.trajL   = 1.0;
-
-  HMCparameters HMCparams;
-  HMCparams.StartTrajectory  = user_params.StartTrajectory;
-  HMCparams.Trajectories     = user_params.Trajectories;
-  HMCparams.NoMetropolisUntil= 0;
-  HMCparams.StartingType     = user_params.StartingType;
-  HMCparams.MetropolisTest = user_params.MetropolisTest;
-  HMCparams.MD = MD;
-  HMCWrapper TheHMC(HMCparams);
-
-  // Grid from the command line arguments --grid and --mpi
-  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
-
-  CheckpointerParameters CPparams;
-  CPparams.config_prefix = "ckpoint_lat";
-  CPparams.rng_prefix    = "ckpoint_rng";
-  CPparams.saveInterval  = user_params.SaveInterval;
-  CPparams.format        = "IEEE64BIG";
-  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
-
-  //Note that checkpointing saves the RNG state so that this initialization is required only for the very first configuration
-  RNGModuleParameters RNGpar;
-  RNGpar.serial_seeds = "1 2 3 4 5";
-  RNGpar.parallel_seeds = "6 7 8 9 10";
-  TheHMC.Resources.SetRNGSeeds(RNGpar);
-
-  typedef PlaquetteMod<GaugeImplPolicy> PlaqObs;
-  TheHMC.Resources.AddObservable<PlaqObs>();
-  //////////////////////////////////////////////
-
-  const int Ls      = 16;
-  Real beta         = 2.13;
-  Real light_mass   = 0.01;
-  Real strange_mass = 0.032;
-  Real pv_mass      = 1.0;
-  RealD M5  = 1.8;
-
-  //Setup the Grids
-  auto GridPtrD   = TheHMC.Resources.GetCartesian();
-  auto GridRBPtrD = TheHMC.Resources.GetRBCartesian();
-  auto FGridD     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtrD);
-  auto FrbGridD   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtrD);
-
-  GridCartesian* GridPtrF = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
-  GridRedBlackCartesian* GridRBPtrF = SpaceTimeGrid::makeFourDimRedBlackGrid(GridPtrF);
-  auto FGridF     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtrF);
-  auto FrbGridF   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtrF);
-
-  ConjugateIwasakiGaugeActionD GaugeAction(beta);
-
-  // temporarily need a gauge field
-  LatticeGaugeFieldD Ud(GridPtrD);
-  LatticeGaugeFieldF Uf(GridPtrF);
- 
-  //Setup the BCs
-  FermionActionD::ImplParams Params;
-  for(int i=0;i<Nd-1;i++) Params.twists[i] = user_params.GparityDirs[i]; //G-parity directions
-  Params.twists[Nd-1] = 1; //APBC in time direction
-
-  std::vector<int> dirs4(Nd);
-  for(int i=0;i<Nd-1;i++) dirs4[i] = user_params.GparityDirs[i];
-  dirs4[Nd-1] = 0; //periodic gauge BC in time
-
-  GaugeImplPolicy::setDirections(dirs4); //gauge BC
-
-  //Run optional gauge field checksum checker and exit
-  if(file_load_check){
-    TheHMC.initializeGaugeFieldAndRNGs(Ud);
-    std::cout << GridLogMessage << " Done" << std::endl;
-    Grid_finalize();
-    return 0;
-  }
-
-
-  ////////////////////////////////////
-  // Collect actions
-  ////////////////////////////////////
-  ActionLevel<HMCWrapper::Field> Level1(1); //light quark + strange quark
-  ActionLevel<HMCWrapper::Field> Level2(8); //gauge (8 increments per step)
-
-
-  /////////////////////////////////////////////////////////////
-  // Light action
-  /////////////////////////////////////////////////////////////
-
-  FermionActionD Numerator_lD(Ud,*FGridD,*FrbGridD,*GridPtrD,*GridRBPtrD, light_mass,M5,Params);
-  FermionActionD Denominator_lD(Ud,*FGridD,*FrbGridD,*GridPtrD,*GridRBPtrD, pv_mass,M5,Params);
-
-  FermionActionF Numerator_lF(Uf,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF, light_mass,M5,Params);
-  FermionActionF Denominator_lF(Uf,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF, pv_mass,M5,Params);
-
-  RationalActionParams rat_act_params_l;
-  rat_act_params_l.inv_pow  = 2; // (M^dag M)^{1/2}
-  rat_act_params_l.precision= 60;
-  rat_act_params_l.MaxIter  = 10000;
-  user_params.rat_quo_l.Export(rat_act_params_l);
-  std::cout << GridLogMessage << " Light quark bounds check every " << rat_act_params_l.BoundsCheckFreq << " trajectories (avg)" << std::endl;
- 
-  MixedPrecRHMC Quotient_l(Denominator_lD, Numerator_lD, Denominator_lF, Numerator_lF, rat_act_params_l, user_params.rat_quo_l.reliable_update_freq);
-  //DoublePrecRHMC Quotient_l(Denominator_lD, Numerator_lD, rat_act_params_l);
-  Level1.push_back(&Quotient_l);
-
-
-  ////////////////////////////////////
-  // Strange action
-  ////////////////////////////////////
-  FermionActionD Numerator_sD(Ud,*FGridD,*FrbGridD,*GridPtrD,*GridRBPtrD,strange_mass,M5,Params);
-  FermionActionD Denominator_sD(Ud,*FGridD,*FrbGridD,*GridPtrD,*GridRBPtrD, pv_mass,M5,Params);
-
-  FermionActionF Numerator_sF(Uf,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,strange_mass,M5,Params);
-  FermionActionF Denominator_sF(Uf,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF, pv_mass,M5,Params);
-
-  RationalActionParams rat_act_params_s;
-  rat_act_params_s.inv_pow  = 4; // (M^dag M)^{1/4}
-  rat_act_params_s.precision= 60;
-  rat_act_params_s.MaxIter  = 10000;
-  user_params.rat_quo_s.Export(rat_act_params_s);
-  std::cout << GridLogMessage << " Heavy quark bounds check every " << rat_act_params_l.BoundsCheckFreq << " trajectories (avg)" << std::endl;
-
-  MixedPrecRHMC Quotient_s(Denominator_sD, Numerator_sD, Denominator_sF, Numerator_sF, rat_act_params_s, user_params.rat_quo_s.reliable_update_freq); 
-  //DoublePrecRHMC Quotient_s(Denominator_sD, Numerator_sD, rat_act_params_s); 
-  Level1.push_back(&Quotient_s);  
-
-
-  /////////////////////////////////////////////////////////////
-  // Gauge action
-  /////////////////////////////////////////////////////////////
-  Level2.push_back(&GaugeAction);
-  TheHMC.TheAction.push_back(Level1);
-  TheHMC.TheAction.push_back(Level2);
-  std::cout << GridLogMessage << " Action complete "<< std::endl;
-
-
-  //Action tuning
-  bool tune_rhmc_l=false, tune_rhmc_s=false, eigenrange_l=false, eigenrange_s=false; 
-  std::string lanc_params_l, lanc_params_s;
-  for(int i=1;i<argc;i++){
-    std::string sarg(argv[i]);
-    if(sarg == "--tune_rhmc_l") tune_rhmc_l=true;
-    else if(sarg == "--tune_rhmc_s") tune_rhmc_s=true;
-    else if(sarg == "--eigenrange_l"){
-      assert(i < argc-1);
-      eigenrange_l=true;
-      lanc_params_l = argv[i+1];
-    }
-    else if(sarg == "--eigenrange_s"){
-      assert(i < argc-1);
-      eigenrange_s=true;
-      lanc_params_s = argv[i+1];
-    }
-  }
-  if(tune_rhmc_l || tune_rhmc_s || eigenrange_l || eigenrange_s){
-    TheHMC.initializeGaugeFieldAndRNGs(Ud);
-    if(eigenrange_l) computeEigenvalues<FermionActionD, FermionFieldD>(lanc_params_l, FGridD, FrbGridD, Ud, Numerator_lD, TheHMC.Resources.GetParallelRNG());
-    if(eigenrange_s) computeEigenvalues<FermionActionD, FermionFieldD>(lanc_params_s, FGridD, FrbGridD, Ud, Numerator_sD, TheHMC.Resources.GetParallelRNG());
-    if(tune_rhmc_l) checkRHMC<FermionActionD, FermionFieldD, decltype(Quotient_l)>(FGridD, FrbGridD, Ud, Numerator_lD, Denominator_lD, Quotient_l, TheHMC.Resources.GetParallelRNG(), 2, "light");
-    if(tune_rhmc_s) checkRHMC<FermionActionD, FermionFieldD, decltype(Quotient_s)>(FGridD, FrbGridD, Ud, Numerator_sD, Denominator_sD, Quotient_s, TheHMC.Resources.GetParallelRNG(), 4, "strange");
-
-    std::cout << GridLogMessage << " Done" << std::endl;
-    Grid_finalize();
-    return 0;
-  }
-
-
-  //Run the HMC
-  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
-  TheHMC.Run();
-
-  std::cout << GridLogMessage << " Done" << std::endl;
-  Grid_finalize();
-  return 0;
-} // main
-