Bcopy on crusher compile

Feature/sumd npr

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/algorithms/CoarsenedMatrix.h

@@ -358,7 +358,7 @@ public:
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
     autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
     Vector<Aview> AcceleratorViewContainer;
@@ -380,7 +380,7 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int point=0;point<geom_v.npoint;point++){
+      for(int point=0;point<npoint;point++){
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);
 	  
@@ -424,7 +424,7 @@ public:
     autoView( in_v , in, AcceleratorRead);
     autoView( out_v , out, AcceleratorWrite);
     autoView( Stencil_v  , Stencil, AcceleratorRead);
-    auto& geom_v = geom;
+    int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
     Vector<Aview> AcceleratorViewContainer;
@@ -454,7 +454,7 @@ public:
       int ptype;
       StencilEntry *SE;
 
-      for(int p=0;p<geom_v.npoint;p++){
+      for(int p=0;p<npoint;p++){
         int point = points_p[p];
 
 	SE=Stencil_v.GetEntry(ptype,point,ss);

Grid/algorithms/LinearOperator.h

@@ -52,6 +52,7 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
+  virtual ~LinearOperatorBase(){};
 };
 
 
@@ -507,7 +508,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   virtual  void MpcDag   (const Field &in, Field &out){
     Mpc(in,out);
   }
-  virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+  virtual void MpcDagMpc(const Field &in, Field &out) {
     assert(0);// Never need with staggered
   }
 };
@@ -585,6 +586,7 @@ class HermOpOperatorFunction : public OperatorFunction<Field> {
 template<typename Field>
 class PlainHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   LinearOperatorBase<Field> &_Linop;
       
   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
@@ -598,6 +600,7 @@ public:
 template<typename Field>
 class FunctionHermOp : public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator(); 
   OperatorFunction<Field>   & _poly;
   LinearOperatorBase<Field> &_Linop;
       

Grid/algorithms/Preconditioner.h

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

Grid/algorithms/SparseMatrix.h

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

Grid/algorithms/iterative/BiCGSTABMixedPrec.h

@@ -37,6 +37,7 @@ template<class FieldD, class FieldF, typename std::enable_if< getPrecision<Field
 class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD> 
 {
   public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -36,6 +36,7 @@ NAMESPACE_BEGIN(Grid);
     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
   public:
+    using LinearFunction<FieldD>::operator();
     RealD   Tolerance;
     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;

Grid/algorithms/iterative/Deflation.h

@@ -33,16 +33,19 @@ namespace Grid {
 template<class Field>
 class ZeroGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
 };
 template<class Field>
 class DoNothingGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) {  };
 };
 template<class Field>
 class SourceGuesser: public LinearFunction<Field> {
 public:
+  using LinearFunction<Field>::operator();
   virtual void operator()(const Field &src, Field &guess) { guess = src; };
 };
 
@@ -57,6 +60,7 @@ private:
   const unsigned int       N;
 
 public:
+  using LinearFunction<Field>::operator();
 
   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
   : DeflatedGuesser(_evec, _eval, _evec.size())
@@ -87,6 +91,7 @@ private:
   const std::vector<RealD>       &eval_coarse;
 public:
   
+  using LinearFunction<FineField>::operator();
   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace,
 				const std::vector<CoarseField> &_evec_coarse,
 				const std::vector<RealD>       &_eval_coarse)

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -67,6 +67,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
@@ -97,6 +98,7 @@ public:
 template<class Fobj,class CComplex,int nbasis>
 class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
 public:
+  using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
   typedef iVector<CComplex,nbasis >           CoarseSiteVector;
   typedef Lattice<CoarseSiteVector>           CoarseField;
   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

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

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

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

Grid/allocator/MemoryManager.cc

@@ -159,7 +159,6 @@ void MemoryManager::Init(void)
 
   char * str;
   int Nc;
-  int NcS;
   
   str= getenv("GRID_ALLOC_NCACHE_LARGE");
   if ( str ) {

Grid/allocator/MemoryManager.h

@@ -170,6 +170,7 @@ private:
 
  public:
   static void Print(void);
+  static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);
   static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);

Grid/allocator/MemoryManagerCache.cc

@@ -474,6 +474,32 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
   }
 }
 
+void MemoryManager::PrintState(void* _CpuPtr)
+{
+  uint64_t CpuPtr = (uint64_t)_CpuPtr;
+
+  if ( EntryPresent(CpuPtr) ){
+    auto AccCacheIterator = EntryLookup(CpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+    if ( AccCache.state==EvictNext) str = std::string("EvictNext");
+
+    std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\t" << AccCache.cpuLock
+    << "\t" << AccCache.accLock
+    << "\t" << AccCache.LRU_valid<<std::endl;
+
+  } else {
+    std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl; 
+  }
+}
+
 NAMESPACE_END(Grid);
 
 #endif

Grid/allocator/MemoryManagerShared.cc

@@ -16,6 +16,10 @@ uint64_t  MemoryManager::DeviceToHostXfer;
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
+void  MemoryManager::PrintState(void* CpuPtr)
+{
+std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
+};
 void  MemoryManager::Print(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 

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;
@@ -97,8 +98,9 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                      ProcessorCount(void)    ;
+  int                ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -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();
@@ -370,26 +372,30 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+from*32;
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-    assert(ierr==0);
-    list.push_back(rrq);
-    off_node_bytes+=bytes;
+  if ( dox ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      list.push_back(rrq);
+      off_node_bytes+=bytes;
+    }
   }
   
-  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+_processor*32;
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-    assert(ierr==0);
-    list.push_back(xrq);
-    off_node_bytes+=bytes;
-  } else {
+  if (dor) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+_processor*32;
+      ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      list.push_back(xrq);
+      off_node_bytes+=bytes;
+    } else {
     // TODO : make a OMP loop on CPU, call threaded bcopy
-    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-    assert(shm!=NULL);
-    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
-    acceleratorCopySynchronise(); // MPI prob slower
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      assert(shm!=NULL);
+      //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+    }
   }
   
   if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
@@ -400,6 +406,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
+  //   std::cout << "Copy Synchronised\n"<<std::endl;
+  acceleratorCopySynchronise();
+
   int nreq=list.size();
 
   if (nreq==0) return;

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.h

@@ -46,3 +46,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_unary.h>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
+#include <Grid/lattice/Lattice_crc.h>

Grid/lattice/Lattice_base.h

@@ -88,6 +88,13 @@ public:
     LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
     accessor.ViewClose();
   }
+
+  // Helper function to print the state of this object in the AccCache
+  void PrintCacheState(void)
+  {
+    MemoryManager::PrintState(this->_odata);
+  }
+
   /////////////////////////////////////////////////////////////////////////////////
   // Return a view object that may be dereferenced in site loops.
   // The view is trivially copy constructible and may be copied to an accelerator device

Grid/lattice/Lattice_crc.h

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

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
 ////////////////////////////////////////////////////////////////////////////////////////////////////

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
@@ -37,14 +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 << "\tmaxThreadsPerBlock  = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  
+  */  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -52,10 +51,14 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
   
   // let the number of threads in a block be a multiple of 2, starting from warpSize
   threads = warpSize;
+  if ( threads*sizeofsobj > sharedMemPerBlock ) {
+    std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
+    return 0;
+  }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  
+  return 1;
 }
 
 template <class sobj, class Iterator>
@@ -195,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -204,7 +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);
@@ -215,6 +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
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -227,6 +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_transfer.h

@@ -85,6 +85,76 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
   });
 }
 
+template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
+{
+  half.Checkerboard() = cb;
+  autoView(half_v, half, AcceleratorWrite);
+  autoView(full_v, full, AcceleratorRead);
+  Coordinate rdim_full             = full.Grid()->_rdimensions;
+  Coordinate rdim_half             = half.Grid()->_rdimensions;
+  unsigned long ndim_half          = half.Grid()->_ndimension;
+  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
+  Coordinate ostride_half          = half.Grid()->_ostride;
+  accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{
+    
+    Coordinate coor;
+    int cbos;
+    int linear=0;
+
+    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
+    assert(coor.size()==ndim_half);
+
+    for(int d=0;d<ndim_half;d++){ 
+      if(checker_dim_mask_half[d]) linear += coor[d];
+    }
+    cbos = (linear&0x1);
+
+    if (cbos==cb) {
+      int ssh=0;
+      for(int d=0;d<ndim_half;d++) {
+        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
+        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
+      }
+      coalescedWrite(half_v[ssh],full_v(ss));
+    }
+  });
+}
+template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
+{
+  int cb = half.Checkerboard();
+  autoView(half_v , half, AcceleratorRead);
+  autoView(full_v , full, AcceleratorWrite);
+  Coordinate rdim_full             = full.Grid()->_rdimensions;
+  Coordinate rdim_half             = half.Grid()->_rdimensions;
+  unsigned long ndim_half          = half.Grid()->_ndimension;
+  Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
+  Coordinate ostride_half          = half.Grid()->_ostride;
+  accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{
+
+    Coordinate coor;
+    int cbos;
+    int linear=0;
+  
+    Lexicographic::CoorFromIndex(coor,ss,rdim_full);
+    assert(coor.size()==ndim_half);
+
+    for(int d=0;d<ndim_half;d++){ 
+      if(checker_dim_mask_half[d]) linear += coor[d];
+    }
+    cbos = (linear&0x1);
+
+    if (cbos==cb) {
+      int ssh=0;
+      for(int d=0;d<ndim_half;d++){
+        if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
+        else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
+      }
+      coalescedWrite(full_v[ss],half_v(ssh));
+    }
+
+  });
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Flexible Type Conversion for internal promotion to double as well as graceful
 // treatment of scalar-compatible types

Grid/log/Log.cc

@@ -69,6 +69,7 @@ GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(0);
@@ -79,6 +80,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogPerformance.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
+  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
     if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
@@ -87,7 +89,8 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -182,6 +182,7 @@ extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogHMC;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/IldgIO.h

@@ -576,6 +576,8 @@ class ScidacReader : public GridLimeReader {
     std::string rec_name(ILDG_BINARY_DATA);
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
       if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+  // in principle should do the line below, but that breaks backard compatibility with old data
+  // skipPastObjectRecord(std::string(GRID_FIELD_NORM));
 	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
 	return;
       }

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/fermion/CompactWilsonCloverFermion.h

@@ -0,0 +1,240 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h
+
+    Copyright (C) 2020 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Nils Meyer <nils.meyer@ur.de>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#pragma once
+
+#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+
+NAMESPACE_BEGIN(Grid);
+
+// see Grid/qcd/action/fermion/WilsonCloverFermion.h for description
+//
+// Modifications done here:
+//
+// Original: clover term = 12x12 matrix per site
+//
+// But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running)
+// Sufficient to store/transfer only the real parts of the diagonal and one triangular part
+// 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered
+//
+// Here: Above but diagonal as complex numbers, i.e., need to store/transfer
+// 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words
+//
+// Words per site and improvement compared to original (combined with the input and output spinors):
+//
+// - Original: 2*12 + 12*12 = 168 words -> 1.00 x less
+// - Minimal:  2*12 + 36    =  60 words -> 2.80 x less
+// - Here:     2*12 + 42    =  66 words -> 2.55 x less
+//
+// These improvements directly translate to wall-clock time
+//
+// Data layout:
+//
+// - diagonal and triangle part as separate lattice fields,
+//   this was faster than as 1 combined field on all tested machines
+// - diagonal: as expected
+// - triangle: store upper right triangle in row major order
+// - graphical:
+//        0  1  2  3  4
+//           5  6  7  8
+//              9 10 11 = upper right triangle indices
+//                12 13
+//                   14
+//     0
+//        1
+//           2
+//              3       = diagonal indices
+//                 4
+//                    5
+//     0
+//     1  5
+//     2  6  9          = lower left triangle indices
+//     3  7 10 12
+//     4  8 11 13 14
+//
+// Impact on total memory consumption:
+// - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site
+// - Here:     (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts =  24 complex words per site
+//           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
+//                                                                 =  84 complex words per site
+
+template<class Impl>
+class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
+                                   public WilsonCloverHelpers<Impl>,
+                                   public CompactWilsonCloverHelpers<Impl> {
+  /////////////////////////////////////////////
+  // Sizes
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_COMPACT_CLOVER_SIZES(Impl);
+
+  /////////////////////////////////////////////
+  // Type definitions
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  typedef WilsonFermion<Impl>              WilsonBase;
+  typedef WilsonCloverHelpers<Impl>        Helpers;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  /////////////////////////////////////////////
+  // Constructors
+  /////////////////////////////////////////////
+
+public:
+
+  CompactWilsonCloverFermion(GaugeField& _Umu,
+			    GridCartesian& Fgrid,
+			    GridRedBlackCartesian& Hgrid,
+			    const RealD _mass,
+			    const RealD _csw_r = 0.0,
+			    const RealD _csw_t = 0.0,
+			    const RealD _cF = 1.0,
+			    const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
+			    const ImplParams& impl_p = ImplParams());
+
+  /////////////////////////////////////////////
+  // Member functions (implementing interface)
+  /////////////////////////////////////////////
+
+public:
+
+  virtual void Instantiatable() {};
+  int          ConstEE()     override { return 0; };
+  int          isTrivialEE() override { return 0; };
+
+  void Dhop(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopOE(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopEO(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
+
+  void M(const FermionField& in, FermionField& out) override;
+
+  void Mdag(const FermionField& in, FermionField& out) override;
+
+  void Meooe(const FermionField& in, FermionField& out) override;
+
+  void MeooeDag(const FermionField& in, FermionField& out) override;
+
+  void Mooee(const FermionField& in, FermionField& out) override;
+
+  void MooeeDag(const FermionField& in, FermionField& out) override;
+
+  void MooeeInv(const FermionField& in, FermionField& out) override;
+
+  void MooeeInvDag(const FermionField& in, FermionField& out) override;
+
+  void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
+
+  void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
+
+  void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  /////////////////////////////////////////////
+  // Member functions (internals)
+  /////////////////////////////////////////////
+
+  void MooeeInternal(const FermionField&        in,
+                     FermionField&              out,
+                     const CloverDiagonalField& diagonal,
+                     const CloverTriangleField& triangle);
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+  void ImportGauge(const GaugeField& _Umu) override;
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+private:
+
+  template<class Field>
+  const MaskField* getCorrectMaskField(const Field &in) const {
+    if(in.Grid()->_isCheckerBoarded) {
+      if(in.Checkerboard() == Odd) {
+        return &this->BoundaryMaskOdd;
+      } else {
+        return &this->BoundaryMaskEven;
+      }
+    } else {
+      return &this->BoundaryMask;
+    }
+  }
+
+  template<class Field>
+  void ApplyBoundaryMask(Field& f) {
+    const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
+    assert(m != nullptr);
+    CompactHelpers::ApplyBoundaryMask(f, *m);
+  }
+
+  /////////////////////////////////////////////
+  // Member Data
+  /////////////////////////////////////////////
+
+public:
+
+  RealD csw_r;
+  RealD csw_t;
+  RealD cF;
+
+  bool open_boundaries;
+
+  CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
+  CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
+
+  CloverTriangleField Triangle,    TriangleEven,    TriangleOdd;
+  CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
+
+  FermionField Tmp;
+
+  MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -53,6 +53,7 @@ NAMESPACE_CHECK(Wilson);
 #include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(WilsonTM);
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
 NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
@@ -153,6 +154,23 @@ typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoInd
 typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
 typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
+// Compact Clover fermions
+typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
+typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
+
+typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
+typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
+
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
+
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
+
 // Domain Wall fermions
 typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;

Grid/qcd/action/fermion/FermionOperator.h

@@ -49,6 +49,8 @@ public:
 
   virtual FermionField &tmp(void) = 0;
 
+  virtual void DirichletBlock(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

@@ -4,10 +4,11 @@
 
     Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
 
-    Copyright (C) 2017
+    Copyright (C) 2017 - 2022
 
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: David Preti <>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -29,7 +30,8 @@
 
 #pragma once
 
-#include <Grid/Grid.h>
+#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -50,18 +52,15 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////
 
 template <class Impl>
-class WilsonCloverFermion : public WilsonFermion<Impl>
+class WilsonCloverFermion : public WilsonFermion<Impl>,
+                            public WilsonCloverHelpers<Impl>
 {
 public:
-  // Types definitions
   INHERIT_IMPL_TYPES(Impl);
-  template <typename vtype>
-  using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef iImplClover<Simd> SiteCloverType;
-  typedef Lattice<SiteCloverType> CloverFieldType;
+  INHERIT_CLOVER_TYPES(Impl);
 
-public:
-  typedef WilsonFermion<Impl> WilsonBase;
+  typedef WilsonFermion<Impl>       WilsonBase;
+  typedef WilsonCloverHelpers<Impl> Helpers;
 
   virtual int    ConstEE(void)     { return 0; };
   virtual void Instantiatable(void){};
@@ -72,42 +71,7 @@ public:
                       const RealD _csw_r = 0.0,
                       const RealD _csw_t = 0.0,
                       const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
-                      const ImplParams &impl_p = ImplParams()) : WilsonFermion<Impl>(_Umu,
-                                                                                     Fgrid,
-                                                                                     Hgrid,
-                                                                                     _mass, impl_p, clover_anisotropy),
-                                                                 CloverTerm(&Fgrid),
-                                                                 CloverTermInv(&Fgrid),
-                                                                 CloverTermEven(&Hgrid),
-                                                                 CloverTermOdd(&Hgrid),
-                                                                 CloverTermInvEven(&Hgrid),
-                                                                 CloverTermInvOdd(&Hgrid),
-                                                                 CloverTermDagEven(&Hgrid),
-                                                                 CloverTermDagOdd(&Hgrid),
-                                                                 CloverTermInvDagEven(&Hgrid),
-                                                                 CloverTermInvDagOdd(&Hgrid)
-  {
-    assert(Nd == 4); // require 4 dimensions
-
-    if (clover_anisotropy.isAnisotropic)
-    {
-      csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
-      diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
-    }
-    else
-    {
-      csw_r = _csw_r * 0.5;
-      diag_mass = 4.0 + _mass;
-    }
-    csw_t = _csw_t * 0.5;
-
-    if (csw_r == 0)
-      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
-    if (csw_t == 0)
-      std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
-
-    ImportGauge(_Umu);
-  }
+                      const ImplParams &impl_p = ImplParams());
 
   virtual void M(const FermionField &in, FermionField &out);
   virtual void Mdag(const FermionField &in, FermionField &out);
@@ -124,250 +88,21 @@ public:
   void ImportGauge(const GaugeField &_Umu);
 
   // Derivative parts unpreconditioned pseudofermions
-  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
-  {
-    conformable(X.Grid(), Y.Grid());
-    conformable(X.Grid(), force.Grid());
-    GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
-    GaugeField clover_force(force.Grid());
-    PropagatorField Lambda(force.Grid());
+  void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
 
-    // Guido: Here we are hitting some performance issues:
-    // need to extract the components of the DoubledGaugeField
-    // for each call
-    // Possible solution
-    // Create a vector object to store them? (cons: wasting space)
-    std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
-
-    Impl::extractLinkField(U, this->Umu);
-
-    force = Zero();
-    // Derivative of the Wilson hopping term
-    this->DhopDeriv(force, X, Y, dag);
-
-    ///////////////////////////////////////////////////////////
-    // Clover term derivative
-    ///////////////////////////////////////////////////////////
-    Impl::outerProductImpl(Lambda, X, Y);
-    //std::cout << "Lambda:" << Lambda << std::endl;
-
-    Gamma::Algebra sigma[] = {
-        Gamma::Algebra::SigmaXY,
-        Gamma::Algebra::SigmaXZ,
-        Gamma::Algebra::SigmaXT,
-        Gamma::Algebra::MinusSigmaXY,
-        Gamma::Algebra::SigmaYZ,
-        Gamma::Algebra::SigmaYT,
-        Gamma::Algebra::MinusSigmaXZ,
-        Gamma::Algebra::MinusSigmaYZ,
-        Gamma::Algebra::SigmaZT,
-        Gamma::Algebra::MinusSigmaXT,
-        Gamma::Algebra::MinusSigmaYT,
-        Gamma::Algebra::MinusSigmaZT};
-
-    /*
-      sigma_{\mu \nu}=
-      | 0         sigma[0]  sigma[1]  sigma[2] |
-      | sigma[3]    0       sigma[4]  sigma[5] |
-      | sigma[6]  sigma[7]     0      sigma[8] |
-      | sigma[9]  sigma[10] sigma[11]   0      |
-    */
-
-    int count = 0;
-    clover_force = Zero();
-    for (int mu = 0; mu < 4; mu++)
-    {
-      force_mu = Zero();
-      for (int nu = 0; nu < 4; nu++)
-      {
-        if (mu == nu)
-        continue;
-        
-        RealD factor;
-        if (nu == 4 || mu == 4)
-        {
-          factor = 2.0 * csw_t;
-        }
-        else
-        {
-          factor = 2.0 * csw_r;
-        }
-        PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
-        Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-        force_mu -= factor*Cmunu(U, lambda, mu, nu);                   // checked
-        count++;
-      }
-
-      pokeLorentz(clover_force, U[mu] * force_mu, mu);
-    }
-    //clover_force *= csw;
-    force += clover_force;
-  }
-
-  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
-  GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
-  {
-    conformable(lambda.Grid(), U[0].Grid());
-    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
-    // insertion in upper staple
-    // please check redundancy of shift operations
-
-    // C1+
-    tmp = lambda * U[nu];
-    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C2+
-    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
-
-    // C3+
-    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
-
-    // C4+
-    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
-
-    // insertion in lower staple
-    // C1-
-    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C2-
-    tmp = adj(lambda) * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
-
-    // C3-
-    tmp = lambda * U[nu];
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
-
-    // C4-
-    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
-
-    return out;
-  }
-
-protected:
+public:
   // here fixing the 4 dimensions, make it more general?
 
   RealD csw_r;                                               // Clover coefficient - spatial
   RealD csw_t;                                               // Clover coefficient - temporal
   RealD diag_mass;                                           // Mass term
-  CloverFieldType CloverTerm, CloverTermInv;                 // Clover term
-  CloverFieldType CloverTermEven, CloverTermOdd;             // Clover term EO
-  CloverFieldType CloverTermInvEven, CloverTermInvOdd;       // Clover term Inv EO
-  CloverFieldType CloverTermDagEven, CloverTermDagOdd;       // Clover term Dag EO
-  CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
-
- public:
-  // eventually these can be compressed into 6x6 blocks instead of the 12x12
-  // using the DeGrand-Rossi basis for the gamma matrices
-  CloverFieldType fillCloverYZ(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-    autoView(T_v,T,AcceleratorWrite);
-    autoView(F_v,F,AcceleratorRead);
-    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
-    {
-      T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
-      T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
-      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverXZ(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-    
-    autoView(T_v, T,AcceleratorWrite);
-    autoView(F_v, F,AcceleratorRead);
-    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
-    {
-      T_v[i]()(0, 1) = -F_v[i]()();
-      T_v[i]()(1, 0) = F_v[i]()();
-      T_v[i]()(2, 3) = -F_v[i]()();
-      T_v[i]()(3, 2) = F_v[i]()();
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverXY(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-
-    autoView(T_v,T,AcceleratorWrite);
-    autoView(F_v,F,AcceleratorRead);
-    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
-    {
-      T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
-      T_v[i]()(1, 1) = timesI(F_v[i]()());
-      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 3) = timesI(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverXT(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-
-    autoView( T_v , T, AcceleratorWrite);
-    autoView( F_v , F, AcceleratorRead);
-    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
-    {
-      T_v[i]()(0, 1) = timesI(F_v[i]()());
-      T_v[i]()(1, 0) = timesI(F_v[i]()());
-      T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverYT(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-    T = Zero();
-    
-    autoView( T_v ,T,AcceleratorWrite);
-    autoView( F_v ,F,AcceleratorRead);
-    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
-    {
-      T_v[i]()(0, 1) = -(F_v[i]()());
-      T_v[i]()(1, 0) = (F_v[i]()());
-      T_v[i]()(2, 3) = (F_v[i]()());
-      T_v[i]()(3, 2) = -(F_v[i]()());
-    });
-
-    return T;
-  }
-
-  CloverFieldType fillCloverZT(const GaugeLinkField &F)
-  {
-    CloverFieldType T(F.Grid());
-
-    T = Zero();
-
-    autoView( T_v , T,AcceleratorWrite);
-    autoView( F_v , F,AcceleratorRead);
-    accelerator_for(i, CloverTerm.Grid()->oSites(),1,
-    {
-      T_v[i]()(0, 0) = timesI(F_v[i]()());
-      T_v[i]()(1, 1) = timesMinusI(F_v[i]()());
-      T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
-      T_v[i]()(3, 3) = timesI(F_v[i]()());
-    });
-
-    return T;
-  }
+  CloverField CloverTerm, CloverTermInv;                     // Clover term
+  CloverField CloverTermEven, CloverTermOdd;                 // Clover term EO
+  CloverField CloverTermInvEven, CloverTermInvOdd;           // Clover term Inv EO
+  CloverField CloverTermDagEven, CloverTermDagOdd;           // Clover term Dag EO
+  CloverField CloverTermInvDagEven, CloverTermInvDagOdd;     // Clover term Inv Dag EO
 };
+
 NAMESPACE_END(Grid);
 
 

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -0,0 +1,761 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonCloverHelpers.h
+
+    Copyright (C) 2021 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#pragma once
+
+// Helper routines that implement common clover functionality
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Impl> class WilsonCloverHelpers {
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+
+  // Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
+  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
+  {
+    conformable(lambda.Grid(), U[0].Grid());
+    GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
+    // insertion in upper staple
+    // please check redundancy of shift operations
+
+    // C1+
+    tmp = lambda * U[nu];
+    out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C2+
+    tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
+
+    // C3+
+    tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
+
+    // C4+
+    out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
+
+    // insertion in lower staple
+    // C1-
+    out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C2-
+    tmp = adj(lambda) * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+
+    // C3-
+    tmp = lambda * U[nu];
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
+
+    // C4-
+    out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
+
+    return out;
+  }
+
+  static CloverField fillCloverYZ(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverXZ(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+    
+    autoView(T_v, T,AcceleratorWrite);
+    autoView(F_v, F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-F_v[i]()()));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead(F_v[i]()()));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead(-F_v[i]()()));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(F_v[i]()()));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverXY(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+
+    autoView(T_v,T,AcceleratorWrite);
+    autoView(F_v,F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverXT(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+
+    autoView( T_v , T, AcceleratorWrite);
+    autoView( F_v , F, AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverYT(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+    T = Zero();
+    
+    autoView( T_v ,T,AcceleratorWrite);
+    autoView( F_v ,F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 1), coalescedRead(-(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 0), coalescedRead((F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 3), coalescedRead((F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 2), coalescedRead(-(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  static CloverField fillCloverZT(const GaugeLinkField &F)
+  {
+    CloverField T(F.Grid());
+
+    T = Zero();
+
+    autoView( T_v , T,AcceleratorWrite);
+    autoView( F_v , F,AcceleratorRead);
+    accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
+    {
+      coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
+      coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
+    });
+
+    return T;
+  }
+
+  template<class _Spinor>
+  static accelerator_inline void multClover(_Spinor& phi, const SiteClover& C, const _Spinor& chi) {
+    auto CC = coalescedRead(C);
+    mult(&phi, &CC, &chi);
+  }
+
+  template<class _SpinorField>
+  inline void multCloverField(_SpinorField& out, const CloverField& C, const _SpinorField& phi) {
+    const int Nsimd = SiteSpinor::Nsimd();
+    autoView(out_v, out, AcceleratorWrite);
+    autoView(phi_v, phi, AcceleratorRead);
+    autoView(C_v,   C,   AcceleratorRead);
+    typedef decltype(coalescedRead(out_v[0])) calcSpinor;
+    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
+      calcSpinor tmp;
+      multClover(tmp,C_v[sss],phi_v(sss));
+      coalescedWrite(out_v[sss],tmp);
+    });
+  }
+};
+
+
+template<class Impl> class CompactWilsonCloverHelpers {
+public:
+
+  INHERIT_COMPACT_CLOVER_SIZES(Impl);
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  #if 0
+  static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
+    assert(i != j);
+    if(i < j) {
+      return triangle()(block)(triangle_index(i, j));
+    } else { // i > j
+      return conjugate(triangle()(block)(triangle_index(i, j)));
+    }
+  }
+  #else
+  template<typename vobj>
+  static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
+    assert(i != j);
+    if(i < j) {
+      return triangle()(block)(triangle_index(i, j));
+    } else { // i > j
+      return conjugate(triangle()(block)(triangle_index(i, j)));
+    }
+  }
+  #endif
+
+  static accelerator_inline int triangle_index(int i, int j) {
+    if(i == j)
+      return 0;
+    else if(i < j)
+      return Nred * (Nred - 1) / 2 - (Nred - i) * (Nred - i - 1) / 2 + j - i - 1;
+    else // i > j
+      return Nred * (Nred - 1) / 2 - (Nred - j) * (Nred - j - 1) / 2 + i - j - 1;
+  }
+
+  static void MooeeKernel_gpu(int                        Nsite,
+                              int                        Ls,
+                              const FermionField&        in,
+                              FermionField&              out,
+                              const CloverDiagonalField& diagonal,
+                              const CloverTriangleField& triangle) {
+    autoView(diagonal_v, diagonal, AcceleratorRead);
+    autoView(triangle_v, triangle, AcceleratorRead);
+    autoView(in_v,       in,       AcceleratorRead);
+    autoView(out_v,      out,      AcceleratorWrite);
+
+    typedef decltype(coalescedRead(out_v[0])) CalcSpinor;
+
+    const uint64_t NN = Nsite * Ls;
+
+    accelerator_for(ss, NN, Simd::Nsimd(), {
+      int sF = ss;
+      int sU = ss/Ls;
+      CalcSpinor res;
+      CalcSpinor in_t = in_v(sF);
+      auto diagonal_t = diagonal_v(sU);
+      auto triangle_t = triangle_v(sU);
+      for(int block=0; block<Nhs; block++) {
+        int s_start = block*Nhs;
+        for(int i=0; i<Nred; i++) {
+          int si = s_start + i/Nc, ci = i%Nc;
+          res()(si)(ci) = diagonal_t()(block)(i) * in_t()(si)(ci);
+          for(int j=0; j<Nred; j++) {
+            if (j == i) continue;
+            int sj = s_start + j/Nc, cj = j%Nc;
+            res()(si)(ci) = res()(si)(ci) + triangle_elem(triangle_t, block, i, j) * in_t()(sj)(cj);
+          };
+        };
+      };
+      coalescedWrite(out_v[sF], res);
+    });
+  }
+
+  static void MooeeKernel_cpu(int                        Nsite,
+                              int                        Ls,
+                              const FermionField&        in,
+                              FermionField&              out,
+                              const CloverDiagonalField& diagonal,
+                              const CloverTriangleField& triangle) {
+    autoView(diagonal_v, diagonal, CpuRead);
+    autoView(triangle_v, triangle, CpuRead);
+    autoView(in_v,       in,       CpuRead);
+    autoView(out_v,      out,      CpuWrite);
+
+    typedef SiteSpinor CalcSpinor;
+
+#if defined(A64FX) || defined(A64FXFIXEDSIZE)
+#define PREFETCH_CLOVER(BASE) {                                     \
+    uint64_t base;                                                  \
+    int pf_dist_L1 = 1;                                             \
+    int pf_dist_L2 = -5; /* -> penalty -> disable */                \
+                                                                    \
+    if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {           \
+      base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);               \
+      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL1STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL1STRM); \
+    }                                                               \
+                                                                    \
+    if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {           \
+      base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);               \
+      svprfd(svptrue_b64(), (int64_t*)(base +    0), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  256), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  512), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base +  768), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL2STRM); \
+      svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL2STRM); \
+    }                                                               \
+  }
+// TODO: Implement/generalize this for other architectures
+// I played around a bit on KNL (see below) but didn't bring anything
+// #elif defined(AVX512)
+// #define PREFETCH_CLOVER(BASE) {                              \
+//     uint64_t base;                                           \
+//     int pf_dist_L1 = 1;                                      \
+//     int pf_dist_L2 = +4;                                     \
+//                                                              \
+//     if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) {    \
+//       base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0);        \
+//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T0); \
+//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T0); \
+//     }                                                        \
+//                                                              \
+//     if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) {    \
+//       base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0);        \
+//       _mm_prefetch((const char*)(base +    0), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +   64), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  128), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  192), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  256), _MM_HINT_T1); \
+//       _mm_prefetch((const char*)(base +  320), _MM_HINT_T1); \
+//     }                                                        \
+//   }
+#else
+#define PREFETCH_CLOVER(BASE)
+#endif
+
+    const uint64_t NN = Nsite * Ls;
+
+    thread_for(ss, NN, {
+      int sF = ss;
+      int sU = ss/Ls;
+      CalcSpinor res;
+      CalcSpinor in_t = in_v[sF];
+      auto diag_t     = diagonal_v[sU]; // "diag" instead of "diagonal" here to make code below easier to read
+      auto triangle_t = triangle_v[sU];
+
+      // upper half
+      PREFETCH_CLOVER(0);
+
+      auto in_cc_0_0 = conjugate(in_t()(0)(0)); // Nils: reduces number
+      auto in_cc_0_1 = conjugate(in_t()(0)(1)); // of conjugates from
+      auto in_cc_0_2 = conjugate(in_t()(0)(2)); // 30 to 20
+      auto in_cc_1_0 = conjugate(in_t()(1)(0));
+      auto in_cc_1_1 = conjugate(in_t()(1)(1));
+
+      res()(0)(0) =               diag_t()(0)( 0) * in_t()(0)(0)
+                  +           triangle_t()(0)( 0) * in_t()(0)(1)
+                  +           triangle_t()(0)( 1) * in_t()(0)(2)
+                  +           triangle_t()(0)( 2) * in_t()(1)(0)
+                  +           triangle_t()(0)( 3) * in_t()(1)(1)
+                  +           triangle_t()(0)( 4) * in_t()(1)(2);
+
+      res()(0)(1) =           triangle_t()(0)( 0) * in_cc_0_0;
+      res()(0)(1) =               diag_t()(0)( 1) * in_t()(0)(1)
+                  +           triangle_t()(0)( 5) * in_t()(0)(2)
+                  +           triangle_t()(0)( 6) * in_t()(1)(0)
+                  +           triangle_t()(0)( 7) * in_t()(1)(1)
+                  +           triangle_t()(0)( 8) * in_t()(1)(2)
+                  + conjugate(       res()(0)( 1));
+
+      res()(0)(2) =           triangle_t()(0)( 1) * in_cc_0_0
+                  +           triangle_t()(0)( 5) * in_cc_0_1;
+      res()(0)(2) =               diag_t()(0)( 2) * in_t()(0)(2)
+                  +           triangle_t()(0)( 9) * in_t()(1)(0)
+                  +           triangle_t()(0)(10) * in_t()(1)(1)
+                  +           triangle_t()(0)(11) * in_t()(1)(2)
+                  + conjugate(       res()(0)( 2));
+
+      res()(1)(0) =           triangle_t()(0)( 2) * in_cc_0_0
+                  +           triangle_t()(0)( 6) * in_cc_0_1
+                  +           triangle_t()(0)( 9) * in_cc_0_2;
+      res()(1)(0) =               diag_t()(0)( 3) * in_t()(1)(0)
+                  +           triangle_t()(0)(12) * in_t()(1)(1)
+                  +           triangle_t()(0)(13) * in_t()(1)(2)
+                  + conjugate(       res()(1)( 0));
+
+      res()(1)(1) =           triangle_t()(0)( 3) * in_cc_0_0
+                  +           triangle_t()(0)( 7) * in_cc_0_1
+                  +           triangle_t()(0)(10) * in_cc_0_2
+                  +           triangle_t()(0)(12) * in_cc_1_0;
+      res()(1)(1) =               diag_t()(0)( 4) * in_t()(1)(1)
+                  +           triangle_t()(0)(14) * in_t()(1)(2)
+                  + conjugate(       res()(1)( 1));
+
+      res()(1)(2) =           triangle_t()(0)( 4) * in_cc_0_0
+                  +           triangle_t()(0)( 8) * in_cc_0_1
+                  +           triangle_t()(0)(11) * in_cc_0_2
+                  +           triangle_t()(0)(13) * in_cc_1_0
+                  +           triangle_t()(0)(14) * in_cc_1_1;
+      res()(1)(2) =               diag_t()(0)( 5) * in_t()(1)(2)
+                  + conjugate(       res()(1)( 2));
+
+      vstream(out_v[sF]()(0)(0), res()(0)(0));
+      vstream(out_v[sF]()(0)(1), res()(0)(1));
+      vstream(out_v[sF]()(0)(2), res()(0)(2));
+      vstream(out_v[sF]()(1)(0), res()(1)(0));
+      vstream(out_v[sF]()(1)(1), res()(1)(1));
+      vstream(out_v[sF]()(1)(2), res()(1)(2));
+
+      // lower half
+      PREFETCH_CLOVER(1);
+
+      auto in_cc_2_0 = conjugate(in_t()(2)(0));
+      auto in_cc_2_1 = conjugate(in_t()(2)(1));
+      auto in_cc_2_2 = conjugate(in_t()(2)(2));
+      auto in_cc_3_0 = conjugate(in_t()(3)(0));
+      auto in_cc_3_1 = conjugate(in_t()(3)(1));
+
+      res()(2)(0) =               diag_t()(1)( 0) * in_t()(2)(0)
+                  +           triangle_t()(1)( 0) * in_t()(2)(1)
+                  +           triangle_t()(1)( 1) * in_t()(2)(2)
+                  +           triangle_t()(1)( 2) * in_t()(3)(0)
+                  +           triangle_t()(1)( 3) * in_t()(3)(1)
+                  +           triangle_t()(1)( 4) * in_t()(3)(2);
+
+      res()(2)(1) =           triangle_t()(1)( 0) * in_cc_2_0;
+      res()(2)(1) =               diag_t()(1)( 1) * in_t()(2)(1)
+                  +           triangle_t()(1)( 5) * in_t()(2)(2)
+                  +           triangle_t()(1)( 6) * in_t()(3)(0)
+                  +           triangle_t()(1)( 7) * in_t()(3)(1)
+                  +           triangle_t()(1)( 8) * in_t()(3)(2)
+                  + conjugate(       res()(2)( 1));
+
+      res()(2)(2) =           triangle_t()(1)( 1) * in_cc_2_0
+                  +           triangle_t()(1)( 5) * in_cc_2_1;
+      res()(2)(2) =               diag_t()(1)( 2) * in_t()(2)(2)
+                  +           triangle_t()(1)( 9) * in_t()(3)(0)
+                  +           triangle_t()(1)(10) * in_t()(3)(1)
+                  +           triangle_t()(1)(11) * in_t()(3)(2)
+                  + conjugate(       res()(2)( 2));
+
+      res()(3)(0) =           triangle_t()(1)( 2) * in_cc_2_0
+                  +           triangle_t()(1)( 6) * in_cc_2_1
+                  +           triangle_t()(1)( 9) * in_cc_2_2;
+      res()(3)(0) =               diag_t()(1)( 3) * in_t()(3)(0)
+                  +           triangle_t()(1)(12) * in_t()(3)(1)
+                  +           triangle_t()(1)(13) * in_t()(3)(2)
+                  + conjugate(       res()(3)( 0));
+
+      res()(3)(1) =           triangle_t()(1)( 3) * in_cc_2_0
+                  +           triangle_t()(1)( 7) * in_cc_2_1
+                  +           triangle_t()(1)(10) * in_cc_2_2
+                  +           triangle_t()(1)(12) * in_cc_3_0;
+      res()(3)(1) =               diag_t()(1)( 4) * in_t()(3)(1)
+                  +           triangle_t()(1)(14) * in_t()(3)(2)
+                  + conjugate(       res()(3)( 1));
+
+      res()(3)(2) =           triangle_t()(1)( 4) * in_cc_2_0
+                  +           triangle_t()(1)( 8) * in_cc_2_1
+                  +           triangle_t()(1)(11) * in_cc_2_2
+                  +           triangle_t()(1)(13) * in_cc_3_0
+                  +           triangle_t()(1)(14) * in_cc_3_1;
+      res()(3)(2) =               diag_t()(1)( 5) * in_t()(3)(2)
+                  + conjugate(       res()(3)( 2));
+
+      vstream(out_v[sF]()(2)(0), res()(2)(0));
+      vstream(out_v[sF]()(2)(1), res()(2)(1));
+      vstream(out_v[sF]()(2)(2), res()(2)(2));
+      vstream(out_v[sF]()(3)(0), res()(3)(0));
+      vstream(out_v[sF]()(3)(1), res()(3)(1));
+      vstream(out_v[sF]()(3)(2), res()(3)(2));
+    });
+  }
+
+  static void MooeeKernel(int                        Nsite,
+                          int                        Ls,
+                          const FermionField&        in,
+                          FermionField&              out,
+                          const CloverDiagonalField& diagonal,
+                          const CloverTriangleField& triangle) {
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+    MooeeKernel_gpu(Nsite, Ls, in, out, diagonal, triangle);
+#else
+    MooeeKernel_cpu(Nsite, Ls, in, out, diagonal, triangle);
+#endif
+  }
+
+  static void Invert(const CloverDiagonalField& diagonal,
+                     const CloverTriangleField& triangle,
+                     CloverDiagonalField&       diagonalInv,
+                     CloverTriangleField&       triangleInv) {
+    conformable(diagonal, diagonalInv);
+    conformable(triangle, triangleInv);
+    conformable(diagonal, triangle);
+
+    diagonalInv.Checkerboard() = diagonal.Checkerboard();
+    triangleInv.Checkerboard() = triangle.Checkerboard();
+
+    GridBase* grid = diagonal.Grid();
+
+    long lsites = grid->lSites();
+
+    typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
+    typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
+
+    autoView(diagonal_v,  diagonal,  CpuRead);
+    autoView(triangle_v,  triangle,  CpuRead);
+    autoView(diagonalInv_v, diagonalInv, CpuWrite);
+    autoView(triangleInv_v, triangleInv, CpuWrite);
+
+    thread_for(site, lsites, { // NOTE: Not on GPU because of Eigen & (peek/poke)LocalSite
+      Eigen::MatrixXcd clover_inv_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
+      Eigen::MatrixXcd clover_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
+
+      scalar_object_diagonal diagonal_tmp     = Zero();
+      scalar_object_diagonal diagonal_inv_tmp = Zero();
+      scalar_object_triangle triangle_tmp     = Zero();
+      scalar_object_triangle triangle_inv_tmp = Zero();
+
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(site, lcoor);
+
+      peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
+      peekLocalSite(triangle_tmp, triangle_v, lcoor);
+
+      // TODO: can we save time here by inverting the two 6x6 hermitian matrices separately?
+      for (long s_row=0;s_row<Ns;s_row++) {
+        for (long s_col=0;s_col<Ns;s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for (long c_row=0;c_row<Nc;c_row++) {
+            for (long c_col=0;c_col<Nc;c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
+              else
+                clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(triangle_elem(triangle_tmp, block, i, j)));
+            }
+          }
+        }
+      }
+
+      clover_inv_eigen = clover_eigen.inverse();
+
+      for (long s_row=0;s_row<Ns;s_row++) {
+        for (long s_col=0;s_col<Ns;s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for (long c_row=0;c_row<Nc;c_row++) {
+            for (long c_col=0;c_col<Nc;c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                diagonal_inv_tmp()(block)(i) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
+              else if(i < j)
+                triangle_inv_tmp()(block)(triangle_index(i, j)) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
+              else
+                continue;
+            }
+          }
+        }
+      }
+
+      pokeLocalSite(diagonal_inv_tmp, diagonalInv_v, lcoor);
+      pokeLocalSite(triangle_inv_tmp, triangleInv_v, lcoor);
+    });
+  }
+
+  static void ConvertLayout(const CloverField&   full,
+                            CloverDiagonalField& diagonal,
+                            CloverTriangleField& triangle) {
+    conformable(full, diagonal);
+    conformable(full, triangle);
+
+    diagonal.Checkerboard() = full.Checkerboard();
+    triangle.Checkerboard() = full.Checkerboard();
+
+    autoView(full_v,     full,     AcceleratorRead);
+    autoView(diagonal_v, diagonal, AcceleratorWrite);
+    autoView(triangle_v, triangle, AcceleratorWrite);
+
+    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
+    accelerator_for(ss, full.Grid()->oSites(), 1, {
+      for(int s_row = 0; s_row < Ns; s_row++) {
+        for(int s_col = 0; s_col < Ns; s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for(int c_row = 0; c_row < Nc; c_row++) {
+            for(int c_col = 0; c_col < Nc; c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                diagonal_v[ss]()(block)(i) = full_v[ss]()(s_row, s_col)(c_row, c_col);
+              else if(i < j)
+                triangle_v[ss]()(block)(triangle_index(i, j)) = full_v[ss]()(s_row, s_col)(c_row, c_col);
+              else
+                continue;
+            }
+          }
+        }
+      }
+    });
+  }
+
+
+  static void ConvertLayout(const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverField&               full) {
+    conformable(full, diagonal);
+    conformable(full, triangle);
+
+    full.Checkerboard() = diagonal.Checkerboard();
+
+    full = Zero();
+
+    autoView(diagonal_v, diagonal, AcceleratorRead);
+    autoView(triangle_v, triangle, AcceleratorRead);
+    autoView(full_v,     full,     AcceleratorWrite);
+
+    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
+    accelerator_for(ss, full.Grid()->oSites(), 1, {
+      for(int s_row = 0; s_row < Ns; s_row++) {
+        for(int s_col = 0; s_col < Ns; s_col++) {
+          if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
+          int block       = s_row / Nhs;
+          int s_row_block = s_row % Nhs;
+          int s_col_block = s_col % Nhs;
+          for(int c_row = 0; c_row < Nc; c_row++) {
+            for(int c_col = 0; c_col < Nc; c_col++) {
+              int i = s_row_block * Nc + c_row;
+              int j = s_col_block * Nc + c_col;
+              if(i == j)
+                full_v[ss]()(s_row, s_col)(c_row, c_col) = diagonal_v[ss]()(block)(i);
+              else
+                full_v[ss]()(s_row, s_col)(c_row, c_col) = triangle_elem(triangle_v[ss], block, i, j);
+            }
+          }
+        }
+      }
+    });
+  }
+
+  static void ModifyBoundaries(CloverDiagonalField& diagonal, CloverTriangleField& triangle, RealD csw_t, RealD cF, RealD diag_mass) {
+    // Checks/grid
+    double t0 = usecond();
+    conformable(diagonal, triangle);
+    GridBase* grid = diagonal.Grid();
+
+    // Determine the boundary coordinates/sites
+    double t1 = usecond();
+    int t_dir = Nd - 1;
+    Lattice<iScalar<vInteger>> t_coor(grid);
+    LatticeCoordinate(t_coor, t_dir);
+    int T = grid->GlobalDimensions()[t_dir];
+
+    // Set off-diagonal parts at boundary to zero -- OK
+    double t2 = usecond();
+    CloverTriangleField zeroTriangle(grid);
+    zeroTriangle.Checkerboard() = triangle.Checkerboard();
+    zeroTriangle = Zero();
+    triangle = where(t_coor == 0,   zeroTriangle, triangle);
+    triangle = where(t_coor == T-1, zeroTriangle, triangle);
+
+    // Set diagonal to unity (scaled correctly) -- OK
+    double t3 = usecond();
+    CloverDiagonalField tmp(grid);
+    tmp.Checkerboard() = diagonal.Checkerboard();
+    tmp                = -1.0 * csw_t + diag_mass;
+    diagonal           = where(t_coor == 0,   tmp, diagonal);
+    diagonal           = where(t_coor == T-1, tmp, diagonal);
+
+    // Correct values next to boundary
+    double t4 = usecond();
+    if(cF != 1.0) {
+      tmp = cF - 1.0;
+      tmp += diagonal;
+      diagonal = where(t_coor == 1,   tmp, diagonal);
+      diagonal = where(t_coor == T-2, tmp, diagonal);
+    }
+
+    // Report timings
+    double t5 = usecond();
+#if 0
+    std::cout << GridLogMessage << "CompactWilsonCloverHelpers::ModifyBoundaries timings:"
+              << " checks = "          << (t1 - t0) / 1e6
+              << ", coordinate = "     << (t2 - t1) / 1e6
+              << ", off-diag zero = "  << (t3 - t2) / 1e6
+              << ", diagonal unity = " << (t4 - t3) / 1e6
+              << ", near-boundary = "  << (t5 - t4) / 1e6
+              << ", total = "          << (t5 - t0) / 1e6
+              << std::endl;
+#endif
+  }
+
+  template<class Field, class Mask>
+  static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
+    conformable(f, m);
+    auto grid  = f.Grid();
+    const uint32_t Nsite = grid->oSites();
+    const uint32_t Nsimd = grid->Nsimd();
+    autoView(f_v, f, AcceleratorWrite);
+    autoView(m_v, m, AcceleratorRead);
+    // NOTE: this function cannot be 'private' since nvcc forbids this for kernels
+    accelerator_for(ss, Nsite, Nsimd, {
+      coalescedWrite(f_v[ss], m_v(ss) * f_v(ss));
+    });
+  }
+
+  template<class MaskField>
+  static void SetupMasks(MaskField& full, MaskField& even, MaskField& odd) {
+    assert(even.Grid()->_isCheckerBoarded && even.Checkerboard() == Even);
+    assert(odd.Grid()->_isCheckerBoarded  && odd.Checkerboard()  == Odd);
+    assert(!full.Grid()->_isCheckerBoarded);
+
+    GridBase* grid = full.Grid();
+    int t_dir = Nd-1;
+    Lattice<iScalar<vInteger>> t_coor(grid);
+    LatticeCoordinate(t_coor, t_dir);
+    int T = grid->GlobalDimensions()[t_dir];
+
+    MaskField zeroMask(grid); zeroMask = Zero();
+    full = 1.0;
+    full = where(t_coor == 0,   zeroMask, full);
+    full = where(t_coor == T-1, zeroMask, full);
+
+    pickCheckerboard(Even, even, full);
+    pickCheckerboard(Odd,  odd,  full);
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -0,0 +1,92 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/WilsonCloverTypes.h
+
+    Copyright (C) 2021 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<class Impl>
+class WilsonCloverTypes {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
+
+  typedef iImplClover<Simd> SiteClover;
+
+  typedef Lattice<SiteClover> CloverField;
+};
+
+template<class Impl>
+class CompactWilsonCloverTypes {
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+  static_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
+  static constexpr int Ntriangle = (Nred - 1) * Nc; // 15
+
+  template<typename vtype> using iImplCloverDiagonal = iScalar<iVector<iVector<vtype, Ndiagonal>, Nblock>>;
+  template<typename vtype> using iImplCloverTriangle = iScalar<iVector<iVector<vtype, Ntriangle>, Nblock>>;
+
+  typedef iImplCloverDiagonal<Simd> SiteCloverDiagonal;
+  typedef iImplCloverTriangle<Simd> SiteCloverTriangle;
+  typedef iSinglet<Simd>            SiteMask;
+
+  typedef Lattice<SiteCloverDiagonal> CloverDiagonalField;
+  typedef Lattice<SiteCloverTriangle> CloverTriangleField;
+  typedef Lattice<SiteMask>           MaskField;
+};
+
+#define INHERIT_CLOVER_TYPES(Impl)                                 \
+  typedef typename WilsonCloverTypes<Impl>::SiteClover SiteClover; \
+  typedef typename WilsonCloverTypes<Impl>::CloverField CloverField;
+
+#define INHERIT_COMPACT_CLOVER_TYPES(Impl) \
+  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverDiagonal  SiteCloverDiagonal; \
+  typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverTriangle  SiteCloverTriangle; \
+  typedef typename CompactWilsonCloverTypes<Impl>::SiteMask            SiteMask; \
+  typedef typename CompactWilsonCloverTypes<Impl>::CloverDiagonalField CloverDiagonalField; \
+  typedef typename CompactWilsonCloverTypes<Impl>::CloverTriangleField CloverTriangleField; \
+  typedef typename CompactWilsonCloverTypes<Impl>::MaskField           MaskField; \
+  /* ugly duplication but needed inside functionality classes */ \
+  template<typename vtype> using iImplCloverDiagonal = \
+    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ndiagonal>, CompactWilsonCloverTypes<Impl>::Nblock>>; \
+  template<typename vtype> using iImplCloverTriangle = \
+    iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ntriangle>, CompactWilsonCloverTypes<Impl>::Nblock>>;
+
+#define INHERIT_COMPACT_CLOVER_SIZES(Impl)                                    \
+  static constexpr int Nred      = CompactWilsonCloverTypes<Impl>::Nred;      \
+  static constexpr int Nblock    = CompactWilsonCloverTypes<Impl>::Nblock;    \
+  static constexpr int Ndiagonal = CompactWilsonCloverTypes<Impl>::Ndiagonal; \
+  static constexpr int Ntriangle = CompactWilsonCloverTypes<Impl>::Ntriangle;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/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 +178,17 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
 
+  virtual void DirichletBlock(Coordinate & block)
+  {
+    assert(block.size()==Nd+1);
+    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
+      Dirichlet = 1;
+      Block = block;
+      Stencil.DirichletBlock(block); 
+      StencilEven.DirichletBlock(block); 
+      StencilOdd.DirichletBlock(block);
+    }
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

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

@@ -828,6 +828,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
 #if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
+  unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
   // GENERAL CAYLEY CASE
   ////////////////////////////////////////////////
@@ -880,7 +881,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 
   std::vector<RealD> G_s(Ls,1.0);
-  RealD sign = 1; // sign flip for vector/tadpole
+  RealD sign = 1.0; // sign flip for vector/tadpole
   if ( curr_type == Current::Axial ) {
     for(int s=0;s<Ls/2;s++){
       G_s[s] = -1.0;
@@ -890,7 +891,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
     auto b=this->_b;
     auto c=this->_c;
     if ( b == 1 && c == 0 ) {
-      sign = -1;    
+      sign = -1.0;    
     }
     else {
       std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
@@ -934,7 +935,13 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
     tmp    = Cshift(tmp,mu,-1);
     Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
     tmp = -G_s[s]*( Utmp + gmu*Utmp );
-    tmp    = where((lcoor>=tmin+tshift),tmp,zz); // Mask the time 
+    // 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);
+    }
     L_Q   += where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
 
     InsertSlice(L_Q, q_out, s , 0);

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

@@ -0,0 +1,363 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermionImplementation.h
+
+    Copyright (C) 2017 - 2022
+
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Grid.h>
+#include <Grid/qcd/spin/Dirac.h>
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
+
+NAMESPACE_BEGIN(Grid);
+template<class Impl>
+CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
+                                                             GridCartesian& Fgrid,
+                                                             GridRedBlackCartesian& Hgrid,
+                                                             const RealD _mass,
+                                                             const RealD _csw_r,
+                                                             const RealD _csw_t,
+                                                             const RealD _cF,
+                                                             const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                                             const ImplParams& impl_p)
+  : WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
+  , csw_r(_csw_r)
+  , csw_t(_csw_t)
+  , cF(_cF)
+  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , Diagonal(&Fgrid),        Triangle(&Fgrid)
+  , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
+  , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
+  , DiagonalInv(&Fgrid),     TriangleInv(&Fgrid)
+  , DiagonalInvEven(&Hgrid), TriangleInvEven(&Hgrid)
+  , DiagonalInvOdd(&Hgrid),  TriangleInvOdd(&Hgrid)
+  , Tmp(&Fgrid)
+  , BoundaryMask(&Fgrid)
+  , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
+{
+  csw_r *= 0.5;
+  csw_t *= 0.5;
+  if (clover_anisotropy.isAnisotropic)
+    csw_r /= clover_anisotropy.xi_0;
+
+  ImportGauge(_Umu);
+  if (open_boundaries)
+    CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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) {
+  WilsonBase::DhopOE(in, out, dag);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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) {
+  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) {
+  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) {
+  out.Checkerboard() = in.Checkerboard();
+  WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
+  Mooee(in, Tmp);
+  axpy(out, 1.0, out, Tmp);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
+  out.Checkerboard() = in.Checkerboard();
+  WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
+  MooeeDag(in, Tmp);
+  axpy(out, 1.0, out, Tmp);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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) {
+  WilsonBase::MeooeDag(in, out);
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
+  if(in.Grid()->_isCheckerBoarded) {
+    if(in.Checkerboard() == Odd) {
+      MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
+    } else {
+      MooeeInternal(in, out, DiagonalEven, TriangleEven);
+    }
+  } else {
+    MooeeInternal(in, out, Diagonal, Triangle);
+  }
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
+  Mooee(in, out); // blocks are hermitian
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
+  if(in.Grid()->_isCheckerBoarded) {
+    if(in.Checkerboard() == Odd) {
+      MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
+    } else {
+      MooeeInternal(in, out, DiagonalInvEven, TriangleInvEven);
+    }
+  } else {
+    MooeeInternal(in, out, DiagonalInv, TriangleInv);
+  }
+  if(open_boundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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) {
+  DhopDir(in, out, dir, disp);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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) {
+  assert(!open_boundaries); // TODO check for changes required for open bc
+
+  // NOTE: code copied from original clover term
+  conformable(X.Grid(), Y.Grid());
+  conformable(X.Grid(), force.Grid());
+  GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
+  GaugeField clover_force(force.Grid());
+  PropagatorField Lambda(force.Grid());
+
+  // Guido: Here we are hitting some performance issues:
+  // need to extract the components of the DoubledGaugeField
+  // for each call
+  // Possible solution
+  // Create a vector object to store them? (cons: wasting space)
+  std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
+
+  Impl::extractLinkField(U, this->Umu);
+
+  force = Zero();
+  // Derivative of the Wilson hopping term
+  this->DhopDeriv(force, X, Y, dag);
+
+  ///////////////////////////////////////////////////////////
+  // Clover term derivative
+  ///////////////////////////////////////////////////////////
+  Impl::outerProductImpl(Lambda, X, Y);
+  //std::cout << "Lambda:" << Lambda << std::endl;
+
+  Gamma::Algebra sigma[] = {
+      Gamma::Algebra::SigmaXY,
+      Gamma::Algebra::SigmaXZ,
+      Gamma::Algebra::SigmaXT,
+      Gamma::Algebra::MinusSigmaXY,
+      Gamma::Algebra::SigmaYZ,
+      Gamma::Algebra::SigmaYT,
+      Gamma::Algebra::MinusSigmaXZ,
+      Gamma::Algebra::MinusSigmaYZ,
+      Gamma::Algebra::SigmaZT,
+      Gamma::Algebra::MinusSigmaXT,
+      Gamma::Algebra::MinusSigmaYT,
+      Gamma::Algebra::MinusSigmaZT};
+
+  /*
+    sigma_{\mu \nu}=
+    | 0         sigma[0]  sigma[1]  sigma[2] |
+    | sigma[3]    0       sigma[4]  sigma[5] |
+    | sigma[6]  sigma[7]     0      sigma[8] |
+    | sigma[9]  sigma[10] sigma[11]   0      |
+  */
+
+  int count = 0;
+  clover_force = Zero();
+  for (int mu = 0; mu < 4; mu++)
+  {
+    force_mu = Zero();
+    for (int nu = 0; nu < 4; nu++)
+    {
+      if (mu == nu)
+        continue;
+
+      RealD factor;
+      if (nu == 4 || mu == 4)
+      {
+        factor = 2.0 * csw_t;
+      }
+      else
+      {
+        factor = 2.0 * csw_r;
+      }
+      PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
+      Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);   // checked
+      count++;
+    }
+
+    pokeLorentz(clover_force, U[mu] * force_mu, mu);
+  }
+  //clover_force *= csw;
+  force += clover_force;
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::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) {
+  assert(0);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&        in,
+                    FermionField&              out,
+                    const CloverDiagonalField& diagonal,
+                    const CloverTriangleField& triangle) {
+  assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
+  out.Checkerboard() = in.Checkerboard();
+  conformable(in, out);
+  conformable(in, diagonal);
+  conformable(in, triangle);
+
+  CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
+}
+
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
+  // NOTE: parts copied from original implementation
+
+  // Import gauge into base class
+  double t0 = usecond();
+  WilsonBase::ImportGauge(_Umu); // NOTE: called here and in wilson constructor -> performed twice, but can't avoid that
+
+  // Initialize temporary variables
+  double t1 = usecond();
+  conformable(_Umu.Grid(), this->GaugeGrid());
+  GridBase* grid = _Umu.Grid();
+  typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
+  CloverField TmpOriginal(grid);
+
+  // Compute the field strength terms mu>nu
+  double t2 = usecond();
+  WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
+  WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
+  WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
+
+  // Compute the Clover Operator acting on Colour and Spin
+  // multiply here by the clover coefficients for the anisotropy
+  double t3 = usecond();
+  TmpOriginal  = Helpers::fillCloverYZ(Bx) * csw_r;
+  TmpOriginal += Helpers::fillCloverXZ(By) * csw_r;
+  TmpOriginal += Helpers::fillCloverXY(Bz) * csw_r;
+  TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
+  TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
+  TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
+  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
+  double t5 = 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();
+  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+
+  // Fill the remaining clover fields
+  double t7 = usecond();
+  pickCheckerboard(Even, DiagonalEven,    Diagonal);
+  pickCheckerboard(Even, TriangleEven,    Triangle);
+  pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
+  pickCheckerboard(Odd,  TriangleOdd,     Triangle);
+  pickCheckerboard(Even, DiagonalInvEven, DiagonalInv);
+  pickCheckerboard(Even, TriangleInvEven, TriangleInv);
+  pickCheckerboard(Odd,  DiagonalInvOdd,  DiagonalInv);
+  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
+}
+
+NAMESPACE_END(Grid);

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

@@ -2,12 +2,13 @@
 
     Grid physics library, www.github.com/paboyle/Grid
 
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
+    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
 
-    Copyright (C) 2017
+    Copyright (C) 2017 - 2022
 
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -33,6 +34,45 @@
 
 NAMESPACE_BEGIN(Grid);
 
+template<class Impl>
+WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField&                         _Umu,
+                                               GridCartesian&                      Fgrid,
+                                               GridRedBlackCartesian&              Hgrid,
+                                               const RealD                         _mass,
+                                               const RealD                         _csw_r,
+                                               const RealD                         _csw_t,
+                                               const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                               const ImplParams&                   impl_p)
+  : WilsonFermion<Impl>(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
+  , CloverTerm(&Fgrid)
+  , CloverTermInv(&Fgrid)
+  , CloverTermEven(&Hgrid)
+  , CloverTermOdd(&Hgrid)
+  , CloverTermInvEven(&Hgrid)
+  , CloverTermInvOdd(&Hgrid)
+  , CloverTermDagEven(&Hgrid)
+  , CloverTermDagOdd(&Hgrid)
+  , CloverTermInvDagEven(&Hgrid)
+  , CloverTermInvDagOdd(&Hgrid) {
+  assert(Nd == 4); // require 4 dimensions
+
+  if(clover_anisotropy.isAnisotropic) {
+    csw_r     = _csw_r * 0.5 / clover_anisotropy.xi_0;
+    diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
+  } else {
+    csw_r     = _csw_r * 0.5;
+    diag_mass = 4.0 + _mass;
+  }
+  csw_t = _csw_t * 0.5;
+
+  if(csw_r == 0)
+    std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
+  if(csw_t == 0)
+    std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
+
+  ImportGauge(_Umu);
+}
+
 // *NOT* EO
 template <class Impl>
 void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
@@ -67,10 +107,13 @@ void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 template <class Impl>
 void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
+  double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
+  double t1 = usecond();
   GridBase *grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
 
+  double t2 = usecond();
   // Compute the field strength terms mu>nu
   WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
   WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
@@ -79,19 +122,22 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
   WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
 
+  double t3 = usecond();
   // Compute the Clover Operator acting on Colour and Spin
   // multiply here by the clover coefficients for the anisotropy
-  CloverTerm  = fillCloverYZ(Bx) * csw_r;
-  CloverTerm += fillCloverXZ(By) * csw_r;
-  CloverTerm += fillCloverXY(Bz) * csw_r;
-  CloverTerm += fillCloverXT(Ex) * csw_t;
-  CloverTerm += fillCloverYT(Ey) * csw_t;
-  CloverTerm += fillCloverZT(Ez) * csw_t;
+  CloverTerm  = Helpers::fillCloverYZ(Bx) * csw_r;
+  CloverTerm += Helpers::fillCloverXZ(By) * csw_r;
+  CloverTerm += Helpers::fillCloverXY(Bz) * csw_r;
+  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;
 
+  double t5 = usecond();
   {
     autoView(CTv,CloverTerm,CpuRead);
     autoView(CTIv,CloverTermInv,CpuWrite);
@@ -100,7 +146,7 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
       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 SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
+      typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
       peekLocalSite(Qx, CTv, lcoor);
       //if (csw!=0){
       for (int j = 0; j < Ns; j++)
@@ -125,6 +171,7 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
     });
   }
 
+  double t6 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@@ -137,6 +184,20 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 
   pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
   pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
+  double t7 = 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
 }
 
 template <class Impl>
@@ -167,7 +228,7 @@ template <class Impl>
 void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
 {
   out.Checkerboard() = in.Checkerboard();
-  CloverFieldType *Clover;
+  CloverField *Clover;
   assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
 
   if (dag)
@@ -182,12 +243,12 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
       {
         Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
       }
-      out = *Clover * in;
+      Helpers::multCloverField(out, *Clover, in);
     }
     else
     {
       Clover = (inv) ? &CloverTermInv : &CloverTerm;
-      out = adj(*Clover) * in;
+      Helpers::multCloverField(out, *Clover, in); // don't bother with adj, hermitian anyway
     }
   }
   else
@@ -205,18 +266,98 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
         //  std::cout << "Calling clover term Even" << std::endl;
         Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
       }
-      out = *Clover * in;
+      Helpers::multCloverField(out, *Clover, in);
       //  std::cout << GridLogMessage << "*Clover.Checkerboard() "  << (*Clover).Checkerboard() << std::endl;
     }
     else
     {
       Clover = (inv) ? &CloverTermInv : &CloverTerm;
-      out = *Clover * in;
+      Helpers::multCloverField(out, *Clover, in);
     }
   }
-
 } // MooeeInternal
 
+// Derivative parts unpreconditioned pseudofermions
+template <class Impl>
+void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
+{
+  conformable(X.Grid(), Y.Grid());
+  conformable(X.Grid(), force.Grid());
+  GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
+  GaugeField clover_force(force.Grid());
+  PropagatorField Lambda(force.Grid());
+
+  // Guido: Here we are hitting some performance issues:
+  // need to extract the components of the DoubledGaugeField
+  // for each call
+  // Possible solution
+  // Create a vector object to store them? (cons: wasting space)
+  std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
+
+  Impl::extractLinkField(U, this->Umu);
+
+  force = Zero();
+  // Derivative of the Wilson hopping term
+  this->DhopDeriv(force, X, Y, dag);
+
+  ///////////////////////////////////////////////////////////
+  // Clover term derivative
+  ///////////////////////////////////////////////////////////
+  Impl::outerProductImpl(Lambda, X, Y);
+  //std::cout << "Lambda:" << Lambda << std::endl;
+
+  Gamma::Algebra sigma[] = {
+      Gamma::Algebra::SigmaXY,
+      Gamma::Algebra::SigmaXZ,
+      Gamma::Algebra::SigmaXT,
+      Gamma::Algebra::MinusSigmaXY,
+      Gamma::Algebra::SigmaYZ,
+      Gamma::Algebra::SigmaYT,
+      Gamma::Algebra::MinusSigmaXZ,
+      Gamma::Algebra::MinusSigmaYZ,
+      Gamma::Algebra::SigmaZT,
+      Gamma::Algebra::MinusSigmaXT,
+      Gamma::Algebra::MinusSigmaYT,
+      Gamma::Algebra::MinusSigmaZT};
+
+  /*
+    sigma_{\mu \nu}=
+    | 0         sigma[0]  sigma[1]  sigma[2] |
+    | sigma[3]    0       sigma[4]  sigma[5] |
+    | sigma[6]  sigma[7]     0      sigma[8] |
+    | sigma[9]  sigma[10] sigma[11]   0      |
+  */
+
+  int count = 0;
+  clover_force = Zero();
+  for (int mu = 0; mu < 4; mu++)
+  {
+    force_mu = Zero();
+    for (int nu = 0; nu < 4; nu++)
+    {
+      if (mu == nu)
+      continue;
+
+      RealD factor;
+      if (nu == 4 || mu == 4)
+      {
+        factor = 2.0 * csw_t;
+      }
+      else
+      {
+        factor = 2.0 * csw_r;
+      }
+      PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
+      Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);                   // checked
+      count++;
+    }
+
+    pokeLorentz(clover_force, U[mu] * force_mu, mu);
+  }
+  //clover_force *= csw;
+  force += clover_force;
+}
 
 // Derivative parts
 template <class Impl>

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);
@@ -218,6 +219,14 @@ 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/WilsonKernelsHandImplementation.h

@@ -77,23 +77,23 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define REGISTER
 
 #ifdef GRID_SIMT
-#define LOAD_CHIMU(ptype)		\
+#define LOAD_CHIMU(Ptype)		\
   {const SiteSpinor & ref (in[offset]);	\
-    Chimu_00=coalescedReadPermute<ptype>(ref()(0)(0),perm,lane);	\
-    Chimu_01=coalescedReadPermute<ptype>(ref()(0)(1),perm,lane);		\
-    Chimu_02=coalescedReadPermute<ptype>(ref()(0)(2),perm,lane);		\
-    Chimu_10=coalescedReadPermute<ptype>(ref()(1)(0),perm,lane);		\
-    Chimu_11=coalescedReadPermute<ptype>(ref()(1)(1),perm,lane);		\
-    Chimu_12=coalescedReadPermute<ptype>(ref()(1)(2),perm,lane);		\
-    Chimu_20=coalescedReadPermute<ptype>(ref()(2)(0),perm,lane);		\
-    Chimu_21=coalescedReadPermute<ptype>(ref()(2)(1),perm,lane);		\
-    Chimu_22=coalescedReadPermute<ptype>(ref()(2)(2),perm,lane);		\
-    Chimu_30=coalescedReadPermute<ptype>(ref()(3)(0),perm,lane);		\
-    Chimu_31=coalescedReadPermute<ptype>(ref()(3)(1),perm,lane);		\
-    Chimu_32=coalescedReadPermute<ptype>(ref()(3)(2),perm,lane);	}
+    Chimu_00=coalescedReadPermute<Ptype>(ref()(0)(0),perm,lane);	\
+    Chimu_01=coalescedReadPermute<Ptype>(ref()(0)(1),perm,lane);		\
+    Chimu_02=coalescedReadPermute<Ptype>(ref()(0)(2),perm,lane);		\
+    Chimu_10=coalescedReadPermute<Ptype>(ref()(1)(0),perm,lane);		\
+    Chimu_11=coalescedReadPermute<Ptype>(ref()(1)(1),perm,lane);		\
+    Chimu_12=coalescedReadPermute<Ptype>(ref()(1)(2),perm,lane);		\
+    Chimu_20=coalescedReadPermute<Ptype>(ref()(2)(0),perm,lane);		\
+    Chimu_21=coalescedReadPermute<Ptype>(ref()(2)(1),perm,lane);		\
+    Chimu_22=coalescedReadPermute<Ptype>(ref()(2)(2),perm,lane);		\
+    Chimu_30=coalescedReadPermute<Ptype>(ref()(3)(0),perm,lane);		\
+    Chimu_31=coalescedReadPermute<Ptype>(ref()(3)(1),perm,lane);		\
+    Chimu_32=coalescedReadPermute<Ptype>(ref()(3)(2),perm,lane);	}
 #define PERMUTE_DIR(dir) ;
 #else
-#define LOAD_CHIMU(ptype)		\
+#define LOAD_CHIMU(Ptype)		\
   {const SiteSpinor & ref (in[offset]);	\
     Chimu_00=ref()(0)(0);\
     Chimu_01=ref()(0)(1);\
@@ -109,12 +109,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     Chimu_32=ref()(3)(2);}
 
 #define PERMUTE_DIR(dir)			\
-  permute##dir(Chi_00,Chi_00);	\
-      permute##dir(Chi_01,Chi_01);\
-      permute##dir(Chi_02,Chi_02);\
-      permute##dir(Chi_10,Chi_10);	\
-      permute##dir(Chi_11,Chi_11);\
-      permute##dir(Chi_12,Chi_12);
+  permute##dir(Chi_00,Chi_00);			\
+  permute##dir(Chi_01,Chi_01);			\
+  permute##dir(Chi_02,Chi_02);			\
+  permute##dir(Chi_10,Chi_10);			\
+  permute##dir(Chi_11,Chi_11);			\
+  permute##dir(Chi_12,Chi_12);
 
 #endif
 
@@ -371,88 +371,91 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   result_32-= UChi_12;
 
 #define HAND_STENCIL_LEGB(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
-  if ( local ) {				\
-    LOAD_CHIMU(PERM);				\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else {					\
-    LOAD_CHI;					\
-  }						\
-  acceleratorSynchronise();			\
-  MULT_2SPIN(DIR);				\
-  RECON;					
+  {int ptype;					\
+   SE=st.GetEntry(ptype,DIR,ss);		\
+   auto offset = SE->_offset;			\
+   auto local  = SE->_is_local;			\
+   auto perm   = SE->_permute;			\
+   if ( local ) {				\
+     LOAD_CHIMU(PERM);				\
+     PROJ;					\
+     if ( perm) {				\
+       PERMUTE_DIR(PERM);			\
+     }						\
+   } else {					\
+     LOAD_CHI;					\
+   }						\
+   acceleratorSynchronise();			\
+   MULT_2SPIN(DIR);				\
+   RECON;					}
 
-#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
-  SE=&st_p[DIR+8*ss];				\
-  ptype=st_perm[DIR];				\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
-  if ( local ) {				\
-    LOAD_CHIMU(PERM);				\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else {					\
-    LOAD_CHI;					\
-  }						\
-  acceleratorSynchronise();			\
-  MULT_2SPIN(DIR);				\
-  RECON;					
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)		\
+  { SE=&st_p[DIR+8*ss];						\
+  auto ptype=st_perm[DIR];					\
+  auto offset = SE->_offset;					\
+  auto local  = SE->_is_local;					\
+  auto perm   = SE->_permute;					\
+  if ( local ) {						\
+    LOAD_CHIMU(PERM);						\
+    PROJ;							\
+    if ( perm) {						\
+      PERMUTE_DIR(PERM);					\
+    }								\
+  } else {							\
+    LOAD_CHI;							\
+  }								\
+  acceleratorSynchronise();					\
+  MULT_2SPIN(DIR);						\
+  RECON;					}
 
 #define HAND_STENCIL_LEGA(PROJ,PERM,DIR,RECON)				\
-  SE=&st_p[DIR+8*ss];							\
-  ptype=st_perm[DIR];							\
- /*SE=st.GetEntry(ptype,DIR,ss);*/					\
-  offset = SE->_offset;				\
-  perm   = SE->_permute;			\
-  LOAD_CHIMU(PERM);				\
-  PROJ;						\
-  MULT_2SPIN(DIR);				\
-  RECON;					
+  { SE=&st_p[DIR+8*ss];							\
+    auto ptype=st_perm[DIR];						\
+    /*SE=st.GetEntry(ptype,DIR,ss);*/					\
+    auto offset = SE->_offset;						\
+    auto perm   = SE->_permute;						\
+    LOAD_CHIMU(PERM);							\
+    PROJ;								\
+    MULT_2SPIN(DIR);							\
+    RECON;					}
 
 #define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
-  if ( local ) {				\
-    LOAD_CHIMU(PERM);				\
-    PROJ;					\
-    if ( perm) {				\
-      PERMUTE_DIR(PERM);			\
-    }						\
-  } else if ( st.same_node[DIR] ) {		\
-    LOAD_CHI;					\
-  }						\
-  acceleratorSynchronise();			\
-  if (local || st.same_node[DIR] ) {		\
-    MULT_2SPIN(DIR);				\
-    RECON;					\
-  }						\
-  acceleratorSynchronise();			
+  { int ptype;						\
+  SE=st.GetEntry(ptype,DIR,ss);				\
+  auto offset = SE->_offset;					\
+  auto local  = SE->_is_local;					\
+  auto perm   = SE->_permute;					\
+  if ( local ) {						\
+    LOAD_CHIMU(PERM);						\
+    PROJ;							\
+    if ( perm) {						\
+      PERMUTE_DIR(PERM);					\
+    }								\
+  } else if ( st.same_node[DIR] ) {				\
+    LOAD_CHI;							\
+  }								\
+  acceleratorSynchronise();					\
+  if (local || st.same_node[DIR] ) {				\
+    MULT_2SPIN(DIR);						\
+    RECON;							\
+  }								\
+  acceleratorSynchronise();			}
 
 #define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON)	\
-  SE=st.GetEntry(ptype,DIR,ss);			\
-  offset = SE->_offset;				\
-  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {	\
-    LOAD_CHI;					\
-    MULT_2SPIN(DIR);				\
-    RECON;					\
-    nmu++;					\
-  }						\
-  acceleratorSynchronise();			
+  { int ptype;						\
+  SE=st.GetEntry(ptype,DIR,ss);				\
+  auto offset = SE->_offset;				\
+  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {		\
+    LOAD_CHI;						\
+    MULT_2SPIN(DIR);					\
+    RECON;						\
+    nmu++;						\
+  }							\
+  acceleratorSynchronise();			}
 
-#define HAND_RESULT(ss)				\
-  {						\
-    SiteSpinor & ref (out[ss]);			\
+#define HAND_RESULT(ss)					\
+  {							\
+    SiteSpinor & ref (out[ss]);				\
     coalescedWrite(ref()(0)(0),result_00,lane);		\
     coalescedWrite(ref()(0)(1),result_01,lane);		\
     coalescedWrite(ref()(0)(2),result_02,lane);		\
@@ -563,7 +566,6 @@ WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p,
 
   HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
   HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
   HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
@@ -593,9 +595,7 @@ WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,Site
 
   HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
-
   HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
   HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
   HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
@@ -623,8 +623,6 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView
   HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
-  int offset,local,perm, ptype;
-  
   HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
   HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
   HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
@@ -640,8 +638,8 @@ template<class Impl>  accelerator_inline void
 WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -652,7 +650,6 @@ WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,Si
 
   HAND_DECLARATIONS(Simt);
 
-  int offset,local,perm, ptype;
   StencilEntry *SE;
   ZERO_RESULT;
   HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
@@ -670,8 +667,8 @@ template<class Impl> accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
   typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
@@ -682,7 +679,6 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldVi
   HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
-  int offset,local,perm, ptype;
   ZERO_RESULT;
   HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
   HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
@@ -699,8 +695,8 @@ template<class Impl>  accelerator_inline void
 WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor  *buf,
 					  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
 // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
@@ -711,7 +707,7 @@ WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,Si
 
   HAND_DECLARATIONS(Simt);
 
-  int offset, ptype;
+  //  int offset, ptype;
   StencilEntry *SE;
   int nmu=0;
   ZERO_RESULT;
@@ -730,8 +726,8 @@ template<class Impl>  accelerator_inline
 void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
 						  int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
 {
-  auto st_p = st._entries_p;						
-  auto st_perm = st._permute_type;					
+  //  auto st_p = st._entries_p;						
+  //  auto st_perm = st._permute_type;					
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
   typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
@@ -742,7 +738,7 @@ void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldVi
   HAND_DECLARATIONS(Simt);
 
   StencilEntry *SE;
-  int offset, ptype;
+  //  int offset, ptype;
   int nmu=0;
   ZERO_RESULT;
   HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);

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

@@ -0,0 +1,41 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/ qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master
+
+    Copyright (C) 2017 - 2022
+
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Grid.h>
+#include <Grid/qcd/spin/Dirac.h>
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
+#include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class CompactWilsonCloverFermion<IMPLEMENTATION>; 
+
+NAMESPACE_END(Grid);

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

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

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

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

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

@@ -40,7 +40,7 @@ EOF
 
 done
 
-CC_LIST="WilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
+CC_LIST="WilsonCloverFermionInstantiation CompactWilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
 
 for impl in $WILSON_IMPL_LIST
 do

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 << " 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/gauge/GaugeImplTypes.h

@@ -78,6 +78,8 @@ public:
   typedef Lattice<SiteLink>    LinkField; 
   typedef Lattice<SiteField>   Field;
 
+  typedef SU<Nrepresentation> Group;
+
   // Guido: we can probably separate the types from the HMC functions
   // this will create 2 kind of implementations
   // probably confusing the users
@@ -118,7 +120,7 @@ public:
     LinkField Pmu(P.Grid());
     Pmu = Zero();
     for (int mu = 0; mu < Nd; mu++) {
-      SU<Nrepresentation>::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
+      Group::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
       RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR) ;
       Pmu = Pmu*scale;
       PokeIndex<LorentzIndex>(P, Pmu, mu);
@@ -159,15 +161,15 @@ public:
   }
 
   static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
-    SU<Nc>::HotConfiguration(pRNG, U);
+    Group::HotConfiguration(pRNG, U);
   }
 
   static inline void TepidConfiguration(GridParallelRNG &pRNG, Field &U) {
-    SU<Nc>::TepidConfiguration(pRNG, U);
+    Group::TepidConfiguration(pRNG, U);
   }
 
   static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
-    SU<Nc>::ColdConfiguration(pRNG, U);
+    Group::ColdConfiguration(pRNG, U);
   }
 };
 

Grid/qcd/hmc/GenericHMCrunner.h

@@ -129,18 +129,10 @@ public:
     Runner(S);
   }
 
-  //////////////////////////////////////////////////////////////////
-
-private:
-  template <class SmearingPolicy>
-  void Runner(SmearingPolicy &Smearing) {
-    auto UGrid = Resources.GetCartesian();
-    Resources.AddRNGs();
-    Field U(UGrid);
-
-    // Can move this outside?
-    typedef IntegratorType<SmearingPolicy> TheIntegrator;
-    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
+  //Use the checkpointer to initialize the RNGs and the gauge field, writing the resulting gauge field into U.
+  //This is called automatically by Run but may be useful elsewhere, e.g. for integrator tuning experiments
+  void initializeGaugeFieldAndRNGs(Field &U){
+    if(!Resources.haveRNGs()) Resources.AddRNGs();
 
     if (Parameters.StartingType == "HotStart") {
       // Hot start
@@ -167,6 +159,25 @@ private:
 	<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart]\n";
       exit(1);
     }
+  }
+
+
+
+  //////////////////////////////////////////////////////////////////
+
+private:
+  template <class SmearingPolicy>
+  void Runner(SmearingPolicy &Smearing) {
+    auto UGrid = Resources.GetCartesian();
+    Field U(UGrid);
+
+    initializeGaugeFieldAndRNGs(U);
+
+    typedef IntegratorType<SmearingPolicy> TheIntegrator;
+    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
+
+    // Sets the momentum filter
+    MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
 
     Smearing.set_Field(U);
 

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 << GridLogMessage
-              << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "exp(-dH) = " << prob
-              << "  Random = " << rn_test << "\n";
-    std::cout << GridLogMessage
-              << "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 << GridLogMessage << "Metropolis_test -- ACCEPTED\n";
-      std::cout << GridLogMessage
-                << "--------------------------------------------------\n";
+      std::cout << GridLogHMC << "Metropolis_test -- ACCEPTED\n";
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
-      std::cout << GridLogMessage << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogMessage
-                << "--------------------------------------------------\n";
+      std::cout << GridLogHMC << "Metropolis_test -- REJECTED\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 << GridLogMessage << "Total H before trajectory = " << H0 << "\n";
+    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 << GridLogMessage << "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
@@ -195,10 +239,13 @@ 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 << GridLogMessage << "-- # Trajectory = " << traj << "\n";
+
+      std::cout << GridLogHMC << "-- # Trajectory = " << traj << "\n";
+
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
-      	std::cout << GridLogMessage << "-- Thermalization" << std::endl;
+      	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
       }
       
       double t0=usecond();
@@ -207,20 +254,19 @@ public:
       DeltaH = evolve_hmc_step(Ucopy);
       // Metropolis-Hastings test
       bool accept = true;
-      if (traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
+      if (Params.MetropolisTest && traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
         accept = metropolis_test(DeltaH);
       } else {
-      	std::cout << GridLogMessage << "Skipping Metropolis test" << std::endl;
+      	std::cout << GridLogHMC << "Skipping Metropolis test" << std::endl;
       }
 
       if (accept)
         Ucur = Ucopy; 
       
-     
-      
       double t1=usecond();
-      std::cout << GridLogMessage << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
+      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;
@@ -228,7 +274,7 @@ public:
       	std::cout << GridLogDebug << "Observables pointer " << Observables[obs] << std::endl;
         Observables[obs]->TrajectoryComplete(traj + 1, Ucur, sRNG, pRNG);
       }
-      std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
+      std::cout << GridLogHMC << ":::::::::::::::::::::::::::::::::::::::::::" << 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;
@@ -227,6 +244,9 @@ public:
   // Random number generators
   //////////////////////////////////////////////////////
   
+  //Return true if the RNG objects have been instantiated
+  bool haveRNGs() const{ return have_RNG; }
+
   void AddRNGs(std::string s = "") {
     // Couple the RNGs to the GridModule tagged by s
     // the default is the first grid registered

Grid/qcd/hmc/UsingHMC.md

@@ -1,61 +1,63 @@
-Using HMC in Grid version 0.5.1
+# Using HMC in Grid
 
-These are the instructions to use the Generalised HMC on Grid version 0.5.1.
-Disclaimer: GRID is still under active development so any information here can be changed in future releases.
+These are the instructions to use the Generalised HMC on Grid as of commit `749b802`.
+Disclaimer: Grid is still under active development so any information here can be changed in future releases.
 
 
-Command line options
-===================
-(relevant file GenericHMCrunner.h)
+## Command line options
+
+(relevant file `GenericHMCrunner.h`)
 The initial configuration can be changed at the command line using 
---StartType <your choice>
-valid choices, one among these
-HotStart, ColdStart, TepidStart, CheckpointStart
-default: HotStart
+`--StartingType STARTING_TYPE`, where `STARTING_TYPE` is one of
+`HotStart`, `ColdStart`, `TepidStart`, and `CheckpointStart`.
+Default: `--StartingType HotStart`
 
-example
-./My_hmc_exec  --StartType HotStart
+Example:
+```
+./My_hmc_exec  --StartingType HotStart
+```
 
-The CheckpointStart option uses the prefix for the configurations and rng seed files defined in your executable and the initial configuration is specified by
---StartTrajectory <integer>
-default: 0
+The `CheckpointStart` option uses the prefix for the configurations and rng seed files defined in your executable and the initial configuration is specified by
+`--StartingTrajectory STARTING_TRAJECTORY`, where `STARTING_TRAJECTORY` is an integer.
+Default: `--StartingTrajectory 0`
 
 The number of trajectories for a specific run are specified at command line by
---Trajectories <integer>
-default: 1
+`--Trajectories TRAJECTORIES`, where `TRAJECTORIES` is an integer.
+Default: `--Trajectories 1`
 
 The number of thermalization steps (i.e. steps when the Metropolis acceptance check is turned off) is specified by
---Thermalizations <integer>
-default: 10
-
+`--Thermalizations THERMALIZATIONS`, where `THERMALIZATIONS` is an integer.
+Default: `--Thermalizations 10`
 
 Any other parameter is defined in the source for the executable.
 
-HMC controls
-===========
+## HMC controls
 
 The lines 
 
+```
   std::vector<int> SerSeed({1, 2, 3, 4, 5});
   std::vector<int> ParSeed({6, 7, 8, 9, 10});
+```
 
 define the seeds for the serial and the parallel RNG.
 
 The line 
 
+```
   TheHMC.MDparameters.set(20, 1.0);// MDsteps, traj length
+```
 
 declares the number of molecular dynamics steps and the total trajectory length.
 
 
-Actions
-======
+## Actions
 
-Action names are defined in the file
-lib/qcd/Actions.h
+Action names are defined in the directory `Grid/qcd/action`.
 
-Gauge actions list:
+Gauge actions list (from `Grid/qcd/action/gauge/Gauge.h`):
 
+```
 WilsonGaugeActionR;
 WilsonGaugeActionF;
 WilsonGaugeActionD;
@@ -68,8 +70,9 @@ IwasakiGaugeActionD;
 SymanzikGaugeActionR;
 SymanzikGaugeActionF;
 SymanzikGaugeActionD;
+```
 
-
+```
 ConjugateWilsonGaugeActionR;
 ConjugateWilsonGaugeActionF;
 ConjugateWilsonGaugeActionD;
@@ -82,26 +85,23 @@ ConjugateIwasakiGaugeActionD;
 ConjugateSymanzikGaugeActionR;
 ConjugateSymanzikGaugeActionF;
 ConjugateSymanzikGaugeActionD;
+```
 
+Each of these action accepts one single parameter at creation time (beta).
+Example for creating a Symanzik action with beta=4.0
 
+```
+  SymanzikGaugeActionR(4.0)
+```
+
+Scalar actions list (from `Grid/qcd/action/scalar/Scalar.h`):
+
+```
 ScalarActionR;
 ScalarActionF;
 ScalarActionD;
+```
 
-
-each of these action accept one single parameter at creation time (beta).
-Example for creating a Symanzik action with beta=4.0
-
-	SymanzikGaugeActionR(4.0)
-
-The suffixes R,F,D in the action names refer to the Real
-(the precision is defined at compile time by the --enable-precision flag in the configure),
-Float and Double, that force the precision of the action to be 32, 64 bit respectively.
-
-
-
-
-
-
-
-
+The suffixes `R`, `F`, `D` in the action names refer to the `Real`
+(the precision is defined at compile time by the `--enable-precision` flag in the configure),
+`Float` and `Double`, that force the precision of the action to be 32, 64 bit respectively.

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;
 
@@ -119,36 +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();
-      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
-      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << std::endl;
+
+      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 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;
+
       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) 
@@ -162,8 +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);
@@ -206,6 +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;
@@ -224,7 +310,6 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
-
   }
 
   void reverse_momenta()
@@ -267,15 +352,19 @@ 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
+	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
@@ -310,7 +399,9 @@ 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/stencil/Stencil.h

@@ -131,8 +131,11 @@ class CartesianStencilAccelerator {
   int           _checkerboard;
   int           _npoints; // Move to template param?
   int           _osites;
+  int           _dirichlet;
   StencilVector _directions;
   StencilVector _distances;
+  StencilVector _comms_send;
+  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -226,6 +229,8 @@ public:
     void * recv_buf;
     Integer to_rank;
     Integer from_rank;
+    Integer do_send;
+    Integer do_recv;
     Integer bytes;
   };
   struct Merge {
@@ -240,7 +245,20 @@ 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 +289,8 @@ 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 +303,6 @@ 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
   ////////////////////////////////////////
@@ -322,8 +318,8 @@ public:
     int simd_layout     = _grid->_simd_layout[dimension];
     int comm_dim        = _grid->_processors[dimension] >1 ;
 
-    int recv_from_rank;
-    int xmit_to_rank;
+    //    int recv_from_rank;
+    //    int xmit_to_rank;
 
     if ( ! comm_dim ) return 1;
     if ( displacement == 0 ) return 1;
@@ -333,11 +329,12 @@ 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 +343,29 @@ 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],
-						     Packets[i].send_buf,
-						     Packets[i].to_rank,
-						     Packets[i].recv_buf,
-						     Packets[i].from_rank,
-						     Packets[i].bytes,i);
-      comms_bytes+=bytes;
-      shm_bytes  +=2*Packets[i].bytes-bytes;
+      _grid->StencilSendToRecvFromBegin(reqs[i],
+					Packets[i].send_buf,
+					Packets[i].to_rank,Packets[i].do_send,
+					Packets[i].recv_buf,
+					Packets[i].from_rank,Packets[i].do_recv,
+					Packets[i].bytes,i);
     }
   }
 
@@ -413,7 +374,6 @@ 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 +381,27 @@ 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();
-	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,
-							 Packets[i].recv_buf,
-							 Packets[i].from_rank,
-							 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
+      /////////////////////////////////////////////////////////
+      // 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;
+      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].do_recv,
+				       Packets[i].bytes,i);
+      });
+      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 +443,23 @@ 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 +469,10 @@ 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 +486,6 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
-    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -551,14 +498,72 @@ 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)
+  {
+    //    std::cout << "Adding CopyReceiveBuffer "<<std::hex<<from<<" "<<to<<std::dec<<" "<<bytes<<std::endl;
+    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);
+      //    std::cout << "CopyReceiveBuffer "<<std::hex<<from<<" "<<to<<std::dec<<" "<<words*sizeof(cobj)<<std::endl;
+      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)
+	     ){
+	//	std::cout << "Found duplicate plane dir "<<direction<<" plane "<< OrthogPlane<< " simd "<<lane << " relproc "<<DestProc<< " bytes "<<bytes <<std::endl;
+	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
+	return 1;
+      }
+    }
+
+    //    std::cout << "No duplicate plane dir "<<direction<<" plane "<< OrthogPlane<< " simd "<<lane << " relproc "<<DestProc<<"  bytes "<<bytes<<std::endl;
+    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 +583,17 @@ 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 +603,7 @@ 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 +611,6 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
-    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -650,19 +647,58 @@ public:
       }
     }
   }
+  /// Introduce a block structure and switch off comms on boundaries
+  void DirichletBlock(const Coordinate &dirichlet_block)
+  {
+    this->_dirichlet = 1;
+    for(int ii=0;ii<this->_npoints;ii++){
+      int dimension    = this->_directions[ii];
+      int displacement = this->_distances[ii];
+      int shift = displacement;
+      int gd = _grid->_gdimensions[dimension];
+      int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      int ld = gd/pd;
+      int pc = _grid->_processor_coor[dimension];
+      ///////////////////////////////////////////
+      // 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 ) {
+	assert(abs(displacement) < ld );
       
+	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)
   {
+    this->_dirichlet = 0;
     face_table_computed=0;
     _grid    = grid;
     this->parameters=p;
@@ -675,6 +711,8 @@ 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);
 
     _unified_buffer_size=0;
@@ -693,24 +731,27 @@ 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);
       int rotate_dim      = _grid->_simd_layout[dimension]>2;
 
+      this->_comms_send[ii] = comm_dim;
+      this->_comms_recv[ii] = comm_dim;
+
       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 +852,7 @@ 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];
@@ -867,7 +909,9 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-      if (!offnode) {
+
+      // Wrap locally dirichlet support case OR node local
+      if ( (offnode==0) || (comms_recv==0)  ) {
 
 	int permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
@@ -984,11 +1028,14 @@ 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());
 
@@ -1014,6 +1061,8 @@ public:
       
       if (comm_proc) {
 
+	
+	
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
 
@@ -1045,44 +1094,53 @@ public:
 	  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();
+	if ( comms_send ) 
+	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so);
+	face_idx++;
 
-	///////////////////////////////////////////////////////////
-	// 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,
-		  bytes);
+	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[u_comm_offset],0,bytes,cbmask);
+	if ( (!duplicate) ) { // Force comms for now
 
-	if ( compress.DecompressionStep() ) {
+	  ///////////////////////////////////////////////////////////
+	  // 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, 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],
 			words,Decompressions);
 	}
 	u_comm_offset+=words;
-      }
+	}
     }
     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];
@@ -1147,12 +1205,11 @@ public:
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
 	}
-	gathermtime-=usecond();
 
+	//	if ( comms_send )
 	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
 
@@ -1181,8 +1238,13 @@ 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 {
 

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/tensors/Tensor_traits.h

@@ -47,20 +47,20 @@ NAMESPACE_BEGIN(Grid);
   class TypePair {
   public:
     T _internal[2];
-    TypePair<T>& operator=(const Grid::Zero& o) {
+    accelerator TypePair<T>& operator=(const Grid::Zero& o) {
       _internal[0] = Zero();
       _internal[1] = Zero();
       return *this;
     }
 
-    TypePair<T> operator+(const TypePair<T>& o) const {
+    accelerator TypePair<T> operator+(const TypePair<T>& o) const {
       TypePair<T> r;
       r._internal[0] = _internal[0] + o._internal[0];
       r._internal[1] = _internal[1] + o._internal[1];
       return r;
     }
 
-    TypePair<T>& operator+=(const TypePair<T>& o) {
+    accelerator TypePair<T>& operator+=(const TypePair<T>& o) {
       _internal[0] += o._internal[0];
       _internal[1] += o._internal[1];
       return *this;

Grid/threads/Accelerator.cc

@@ -84,7 +84,8 @@ void acceleratorInit(void)
   // IBM Jsrun makes cuda Device numbering screwy and not match rank
   if ( world_rank == 0 ) {
     printf("AcceleratorCudaInit: using default device \n");
-    printf("AcceleratorCudaInit: assume user either uses a) IBM jsrun, or \n");
+    printf("AcceleratorCudaInit: assume user either uses\n");
+    printf("AcceleratorCudaInit: a) IBM jsrun, or \n");
     printf("AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
     printf("AcceleratorCudaInit: Configure options --enable-setdevice=no \n");
   }
@@ -95,7 +96,7 @@ void acceleratorInit(void)
 #endif
 
   cudaSetDevice(device);
-
+  cudaStreamCreate(&copyStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -109,6 +110,7 @@ void acceleratorInit(void)
 
 #ifdef GRID_HIP
 hipDeviceProp_t *gpu_props;
+hipStream_t copyStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -166,16 +168,25 @@ void acceleratorInit(void)
 #ifdef GRID_DEFAULT_GPU
   if ( world_rank == 0 ) {
     printf("AcceleratorHipInit: using default device \n");
-    printf("AcceleratorHipInit: assume user either uses a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
-    printf("AcceleratorHipInit: Configure options --enable-summit, --enable-select-gpu=no \n");
+    printf("AcceleratorHipInit: assume user or srun sets ROCR_VISIBLE_DEVICES and numa binding \n");
+    printf("AcceleratorHipInit: Configure options --enable-setdevice=no \n");
   }
+  int device = 0;
 #else
   if ( world_rank == 0 ) {
     printf("AcceleratorHipInit: rank %d setting device to node rank %d\n",world_rank,rank);
-    printf("AcceleratorHipInit: Configure options --enable-select-gpu=yes \n");
+    printf("AcceleratorHipInit: Configure options --enable-setdevice=yes \n");
   }
-  hipSetDevice(rank);
+  int device = rank;
 #endif
+  hipSetDevice(device);
+  hipStreamCreate(&copyStream);
+  const int len=64;
+  char busid[len];
+  if( rank == world_rank ) { 
+    hipDeviceGetPCIBusId(busid, len, device);
+    printf("local rank %d device %d bus id: %s\n", rank, device, busid);
+  }
   if ( world_rank == 0 )  printf("AcceleratorHipInit: ================================================\n");
 }
 #endif

Grid/threads/Accelerator.h

@@ -95,6 +95,7 @@ void     acceleratorInit(void);
 //////////////////////////////////////////////
 
 #ifdef GRID_CUDA
+
 #include <cuda.h>
 
 #ifdef __CUDA_ARCH__
@@ -133,11 +134,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_threads(nsimd,acceleratorThreads(),1);			\
     dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
-    std::cout << "========================== CUDA KERNEL CALL\n";	\
-    cuda_mem();								\
     LambdaApply<<<cu_blocks,cu_threads>>>(num1,num2,nsimd,lambda);	\
-    cuda_mem();								\
-    std::cout << "========================== CUDA KERNEL DONE\n";	\
   }
 
 #define accelerator_for6dNB(iter1, num1,				\
@@ -233,6 +230,7 @@ inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes
   cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
 }
 inline void acceleratorCopySynchronise(void) { cudaStreamSynchronize(copyStream); };
+
 inline int  acceleratorIsCommunicable(void *ptr)
 {
   //  int uvm=0;
@@ -309,7 +307,7 @@ inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {
   theGridAccelerator->memcpy(to,from,bytes);
 }
-inline void acceleratorCopySynchronise(void) {  theGridAccelerator->wait(); }
+inline void acceleratorCopySynchronise(void) {  theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
@@ -340,10 +338,11 @@ NAMESPACE_BEGIN(Grid);
 #define accelerator        __host__ __device__
 #define accelerator_inline __host__ __device__ inline
 
+extern hipStream_t copyStream;
 /*These routines define mapping from thread grid to loop & vector lane indexing */
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
-  return hipThreadIdx_z; 
+  return hipThreadIdx_x; 
 #else
   return 0;
 #endif
@@ -357,19 +356,41 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
       { __VA_ARGS__;}							\
     };									\
     int nt=acceleratorThreads();					\
-    dim3 hip_threads(nt,1,nsimd);					\
-    dim3 hip_blocks ((num1+nt-1)/nt,num2,1);				\
-    hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
-		       0,0,						\
-		       num1,num2,nsimd,lambda);				\
+    dim3 hip_threads(nsimd, nt, 1);					 \
+    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,						\
+            num1,num2,nsimd, lambda);				\
+    } else { \
+      hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
+            0,0,						\
+            num1,num2,nsimd, lambda);				\
+    } \
   }
 
+
 template<typename lambda>  __global__
+__launch_bounds__(64,1)
+void LambdaApply64(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
+{
+  // Following the same scheme as CUDA for now
+  uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
+  uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
+  uint64_t z = threadIdx.x;
+  if ( (x < numx) && (y<numy) && (z<numz) ) {
+    Lambda(x,y,z);
+  }
+}
+
+template<typename lambda>  __global__
+__launch_bounds__(1024,1)
 void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 {
-  uint64_t x = hipThreadIdx_x + hipBlockDim_x*hipBlockIdx_x;
-  uint64_t y = hipThreadIdx_y + hipBlockDim_y*hipBlockIdx_y;
-  uint64_t z = hipThreadIdx_z ;//+ hipBlockDim_z*hipBlockIdx_z;
+  // Following the same scheme as CUDA for now
+  uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
+  uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
+  uint64_t z = threadIdx.x;
   if ( (x < numx) && (y<numy) && (z<numz) ) {
     Lambda(x,y,z);
   }
@@ -414,10 +435,16 @@ inline void acceleratorFreeShared(void *ptr){ hipFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ hipFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
-inline void acceleratorCopySynchronise(void) {  }
+//inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
+//inline void acceleratorCopySynchronise(void) {  }
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(base,value,bytes);}
 
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
+{
+  hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);
+}
+inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
+
 #endif
 
 //////////////////////////////////////////////
@@ -434,6 +461,8 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
   accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, { __VA_ARGS__ } ); \
   accelerator_barrier(dummy);
 
+#define GRID_ACCELERATED
+
 #endif
 
 //////////////////////////////////////////////
@@ -454,9 +483,10 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 #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; }
@@ -488,18 +518,12 @@ inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 ///////////////////////////////////////////////////
 // Synchronise across local threads for divergence resynch
 ///////////////////////////////////////////////////
-accelerator_inline void acceleratorSynchronise(void) 
+accelerator_inline void acceleratorSynchronise(void)  // Only Nvidia needs 
 {
 #ifdef GRID_SIMT
 #ifdef GRID_CUDA
   __syncwarp();
 #endif
-#ifdef GRID_SYCL
-  //cl::sycl::detail::workGroupBarrier();
-#endif
-#ifdef GRID_HIP
-  __syncthreads();
-#endif
 #endif
   return;
 }

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

Grid/util/Coordinate.h

@@ -88,7 +88,7 @@ public:
 // Coordinate class, maxdims = 8 for now.
 ////////////////////////////////////////////////////////////////
 #define GRID_MAX_LATTICE_DIMENSION (8)
-#define GRID_MAX_SIMD              (16)
+#define GRID_MAX_SIMD              (32)
 
 static constexpr int MaxDims = GRID_MAX_LATTICE_DIMENSION;
 

Grid/util/Init.cc

@@ -167,6 +167,13 @@ void GridCmdOptionInt(std::string &str,int & val)
   return;
 }
 
+void GridCmdOptionFloat(std::string &str,float & val)
+{
+  std::stringstream ss(str);
+  ss>>val;
+  return;
+}
+
 
 void GridParseLayout(char **argv,int argc,
 		     Coordinate &latt_c,
@@ -527,6 +534,7 @@ void Grid_init(int *argc,char ***argv)
 void Grid_finalize(void)
 {
 #if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
+  MPI_Barrier(MPI_COMM_WORLD);
   MPI_Finalize();
   Grid_unquiesce_nodes();
 #endif

Grid/util/Init.h

@@ -57,6 +57,7 @@ void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec);
 template<class VectorInt>
 void GridCmdOptionIntVector(const std::string &str,VectorInt & vec);
 void GridCmdOptionInt(std::string &str,int & val);
+void GridCmdOptionFloat(std::string &str,float & val);
 
 
 void GridParseLayout(char **argv,int argc,

HMC/Mobius2p1f_DD_RHMC.cc

@@ -0,0 +1,232 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_hmc_EODWFRatio.cc
+
+Copyright (C) 2015-2016
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Guido Cossu <guido.cossu@ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/Grid.h>
+
+int main(int argc, char **argv) {
+  using namespace Grid;
+
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+
+   // Typedefs to simplify notation
+  typedef WilsonImplR FermionImplPolicy;
+  typedef MobiusFermionR FermionAction;
+  typedef typename FermionAction::FermionField FermionField;
+
+  typedef Grid::XmlReader       Serialiser;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  //  typedef GenericHMCRunner<LeapFrog> HMCWrapper;
+  //  MD.name    = std::string("Leap Frog");
+  //  typedef GenericHMCRunner<ForceGradient> HMCWrapper;
+  //  MD.name    = std::string("Force Gradient");
+  typedef GenericHMCRunner<MinimumNorm2> HMCWrapper;
+  MD.name    = std::string("MinimumNorm2");
+  MD.MDsteps =  4;
+  MD.trajL   = 1.0;
+
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = 8;
+  HMCparams.Trajectories     = 200;
+  HMCparams.NoMetropolisUntil=  0;
+  // "[HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+  //  HMCparams.StartingType     =std::string("ColdStart");
+  HMCparams.StartingType     =std::string("CheckpointStart");
+  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_EODWF_lat";
+  CPparams.rng_prefix    = "ckpoint_EODWF_rng";
+  CPparams.saveInterval  = 1;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = "1 2 3 4 5";
+  RNGpar.parallel_seeds = "6 7 8 9 10";
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  // Construct observables
+  // here there is too much indirection
+  typedef PlaquetteMod<HMCWrapper::ImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+
+  const int Ls      = 16;
+  Real beta         = 2.13;
+  Real light_mass   = 0.01;
+  Real strange_mass = 0.04;
+  Real pv_mass      = 1.0;
+  RealD M5  = 1.8;
+  RealD b   = 1.0;
+  RealD c   = 0.0;
+
+  // FIXME:
+  // Same in MC and MD
+  // Need to mix precision too
+  OneFlavourRationalParams OFRp;
+  OFRp.lo       = 4.0e-3;
+  OFRp.hi       = 30.0;
+  OFRp.MaxIter  = 10000;
+  OFRp.tolerance= 1.0e-10;
+  OFRp.degree   = 16;
+  OFRp.precision= 50;
+
+  std::vector<Real> hasenbusch({ 0.01, 0.04, 0.2 , pv_mass });
+  std::vector<bool> dirichlet ({ true, true, true });
+
+  auto GridPtr   = TheHMC.Resources.GetCartesian();
+  auto GridRBPtr = TheHMC.Resources.GetRBCartesian();
+
+  ////////////////////////////////////////////////////////////////
+  // Domain decomposed
+  ////////////////////////////////////////////////////////////////
+  Coordinate latt4  = GridPtr->GlobalDimensions();
+  Coordinate mpi    = GridPtr->ProcessorGrid();
+  Coordinate shm;
+
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  
+  Coordinate CommDim(Nd);
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+
+  Coordinate Dirichlet(Nd+1,0);
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Coordinate Block4(Nd);
+  Block4[0] = Dirichlet[1];
+  Block4[1] = Dirichlet[2];
+  Block4[2] = Dirichlet[3];
+  Block4[3] = Dirichlet[4];
+  TheHMC.Resources.SetMomentumFilter(new DDHMCFilter<WilsonImplR::Field>(Block4));
+
+  //////////////////////////
+  // Fermion Grid
+  //////////////////////////
+  auto FGrid     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr);
+  auto FrbGrid   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr);
+
+  IwasakiGaugeActionR GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeField U(GridPtr);
+
+  // These lines are unecessary if BC are all periodic
+  std::vector<Complex> boundary = {1,1,1,-1};
+  FermionAction::ImplParams Params(boundary);
+
+  double StoppingCondition = 1e-10;
+  double MaxCGIterations = 30000;
+  ConjugateGradient<FermionField>  CG(StoppingCondition,MaxCGIterations);
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1);
+  ActionLevel<HMCWrapper::Field> Level2(2);
+  ActionLevel<HMCWrapper::Field> Level3(8);
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionAction StrangeOp (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params);
+  FermionAction StrangePauliVillarsOp(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, Params);
+
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermion(StrangePauliVillarsOp,StrangeOp,OFRp);
+  //  Level1.push_back(&StrangePseudoFermion);
+
+  ////////////////////////////////////
+  // up down action
+  ////////////////////////////////////
+  std::vector<Real> light_den;
+  std::vector<Real> light_num;
+  std::vector<int> dirichlet_den;
+  std::vector<int> dirichlet_num;
+
+  int n_hasenbusch = hasenbusch.size();
+  light_den.push_back(light_mass);
+  dirichlet_den.push_back(0);
+  for(int h=0;h<n_hasenbusch;h++){
+    light_den.push_back(hasenbusch[h]);
+    light_num.push_back(hasenbusch[h]);
+    dirichlet_num.push_back(1);
+    dirichlet_den.push_back(1);
+  }
+  light_num.push_back(pv_mass);
+  dirichlet_num.push_back(0);
+
+  std::vector<FermionAction *> Numerators;
+  std::vector<FermionAction *> Denominators;
+  std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
+  
+  for(int h=0;h<n_hasenbusch+1;h++){
+    std::cout << GridLogMessage << " 2f quotient Action  "<< light_num[h]<< " (" << dirichlet_num[h]
+	      <<") / " << light_den[h]<< " (" << dirichlet_den[h]<<")"<< std::endl;
+    Numerators.push_back  (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, Params));
+    Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, Params));
+    Quotients.push_back   (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],CG,CG));
+    if ( dirichlet_den[h]==1) Denominators[h]->DirichletBlock(Dirichlet);
+    if ( dirichlet_num[h]==1) Numerators[h]->DirichletBlock(Dirichlet);
+  }
+
+  int nquo=Quotients.size();
+  Level1.push_back(Quotients[0]);
+  Level1.push_back(Quotients[nquo-1]);
+  for(int h=1;h<nquo-1;h++){
+    Level2.push_back(Quotients[h]);
+  }
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+
+  /////////////////////////////////////////////////////////////
+
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.Run();  // no smearing
+
+  Grid_finalize();
+} // main
+
+
+

benchmarks/Benchmark_IO.cc

@@ -137,7 +137,7 @@ int main (int argc, char ** argv)
 
   Eigen::MatrixXd mean(nVol, 4), stdDev(nVol, 4), rob(nVol, 4);
   Eigen::VectorXd avMean(4), avStdDev(4), avRob(4);
-  double          n = BENCH_IO_NPASS;
+  //  double          n = BENCH_IO_NPASS;
 
   stats(mean, stdDev, perf);
   stats(avMean, avStdDev, avPerf);
@@ -164,7 +164,7 @@ int main (int argc, char ** argv)
                 mean(volInd(l), gWrite), stdDev(volInd(l), gWrite));
   }
   MSG << std::endl;
-  MSG << "Robustness of individual results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
+  MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)" << std::endl;
   MSG << std::endl;
   grid_printf("%4s %12s %12s %12s %12s\n",
               "L", "std read", "std write", "Grid read", "Grid write");
@@ -185,7 +185,7 @@ int main (int argc, char ** argv)
               avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite),
               avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
   MSG << std::endl;
-  MSG << "Robustness of volume-averaged results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
+  MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)" << std::endl;
   MSG << std::endl;
   grid_printf("%12s %12s %12s %12s\n",
               "std read", "std write", "Grid read", "Grid write");

benchmarks/Benchmark_ITT.cc

@@ -142,7 +142,7 @@ public:
 	  //	  bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
 	}
 
-	int ncomm;
+	//	int ncomm;
 	double dbytes;
 
         for(int dir=0;dir<8;dir++) {
@@ -290,7 +290,7 @@ public:
       LatticeSU4 z(&Grid); z=Zero();
       LatticeSU4 x(&Grid); x=Zero();
       LatticeSU4 y(&Grid); y=Zero();
-      double a=2.0;
+      //      double a=2.0;
 
       uint64_t Nloop=NLOOP;
 

benchmarks/Benchmark_comms.cc

@@ -217,9 +217,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu);
 	
 	    comm_proc = mpi_layout[mu]-1;
@@ -228,9 +228,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu+4][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu+4);
 	  
 	  }
@@ -309,9 +309,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu);
 	    Grid.StencilSendToRecvFromComplete(requests,mu);
 	    requests.resize(0);
@@ -322,9 +322,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu+4][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu+4);
 	    Grid.StencilSendToRecvFromComplete(requests,mu+4);
 	    requests.resize(0);
@@ -411,8 +411,8 @@ int main (int argc, char ** argv)
 	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	    }
             int tid = omp_get_thread_num();
-	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
-					       (void *)&rbuf[dir][0], recv_from_rank, bytes,tid);
+	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,1,
+					       (void *)&rbuf[dir][0], recv_from_rank,1, bytes,tid);
 
 	    thread_critical { dbytes+=tbytes; }
 	  }

benchmarks/Benchmark_comms_host_device.cc

@@ -72,7 +72,7 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
   std::vector<double> t_time(Nloop);
-  time_statistics timestat;
+  //  time_statistics timestat;
 
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from host memory "<<std::endl;

benchmarks/Benchmark_dwf_fp32.cc

@@ -32,18 +32,18 @@
 using namespace std;
 using namespace Grid;
 
-template<class d>
-struct scal {
-  d internal;
+////////////////////////
+/// Move to domains ////
+////////////////////////
+
+Gamma::Algebra Gmu [] = {
+			 Gamma::Algebra::GammaX,
+			 Gamma::Algebra::GammaY,
+			 Gamma::Algebra::GammaZ,
+			 Gamma::Algebra::GammaT
 };
 
-  Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT
-  };
-
+void Benchmark(int Ls, Coordinate Dirichlet);
 
 int main (int argc, char ** argv)
 {
@@ -52,24 +52,82 @@ int main (int argc, char ** argv)
 
   int threads = GridThread::GetThreads();
 
-  Coordinate latt4 = GridDefaultLatt();
   int Ls=16;
-  for(int i=0;i<argc;i++)
+  for(int i=0;i<argc;i++) {
     if(std::string(argv[i]) == "-Ls"){
       std::stringstream ss(argv[i+1]); ss >> Ls;
     }
+  }
 
+  //////////////////
+  // With comms
+  //////////////////
+  Coordinate Dirichlet(Nd+1,0);
+
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing with full communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  
+  Benchmark(Ls,Dirichlet);
+
+  //////////////////
+  // Domain decomposed
+  //////////////////
+  Coordinate latt4  = GridDefaultLatt();
+  Coordinate mpi    = GridDefaultMpi();
+  Coordinate CommDim(Nd);
+  Coordinate shm;
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+
+
+  //////////////////////
+  // Node level
+  //////////////////////
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing without internode communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+  Dirichlet[0] = 0;
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Benchmark(Ls,Dirichlet);
+
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing without intranode communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+
+  for(int d=0;d<Nd;d++) CommDim[d]= mpi[d]>1 ? 1 : 0;
+  Dirichlet[0] = 0;
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3];
+  
+  Benchmark(Ls,Dirichlet);
+
+  Grid_finalize();
+  exit(0);
+}
+void Benchmark(int Ls, Coordinate Dirichlet)
+{
+  Coordinate latt4 = GridDefaultLatt();
   GridLogLayout();
 
   long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
 
-
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
   GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
   GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
   GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
@@ -80,9 +138,9 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
   GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
+
   std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
   GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
-  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
 
   LatticeFermionF src   (FGrid); random(RNG5,src);
 #if 0
@@ -100,7 +158,6 @@ int main (int argc, char ** argv)
   src = src*N2;
 #endif
 
-
   LatticeFermionF result(FGrid); result=Zero();
   LatticeFermionF    ref(FGrid);    ref=Zero();
   LatticeFermionF    tmp(FGrid);
@@ -108,38 +165,31 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeFieldF Umu(UGrid);
+  LatticeGaugeFieldF UmuCopy(UGrid);
   SU<Nc>::HotConfiguration(RNG4,Umu);
+  UmuCopy=Umu;
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
-#if 0
-  Umu=1.0;
-  for(int mu=0;mu<Nd;mu++){
-    LatticeColourMatrixF ttmp(UGrid);
-    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
-    //    if (mu !=2 ) ttmp = 0;
-    //    ttmp = ttmp* pow(10.0,mu);
-    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
-  }
-  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
-#endif
+
+  ////////////////////////////////////
+  // Apply BCs
+  ////////////////////////////////////
+  Coordinate Block(4);
+  for(int d=0;d<4;d++)  Block[d]= Dirichlet[d+1];
+
+  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block5 " << Dirichlet << std::endl;
+  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block4 " << Block << std::endl;
+
+  DirichletFilter<LatticeGaugeFieldF> Filter(Block);
+  Filter.applyFilter(Umu);
   
   ////////////////////////////////////
   // Naive wilson implementation
   ////////////////////////////////////
-  // replicate across fifth dimension
-  LatticeGaugeFieldF Umu5d(FGrid);
-  std::vector<LatticeColourMatrixF> U(4,FGrid);
-  {
-    autoView( Umu5d_v, Umu5d, CpuWrite);
-    autoView( Umu_v  , Umu  , CpuRead);
-    for(int ss=0;ss<Umu.Grid()->oSites();ss++){
-      for(int s=0;s<Ls;s++){
-	Umu5d_v[Ls*ss+s] = Umu_v[ss];
-      }
-    }
-  }
+  std::vector<LatticeColourMatrixF> U(4,UGrid);
   for(int mu=0;mu<Nd;mu++){
-    U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
   }
+
   std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
   if (1)
@@ -147,10 +197,28 @@ int main (int argc, char ** argv)
     ref = Zero();
     for(int mu=0;mu<Nd;mu++){
 
-      tmp = U[mu]*Cshift(src,mu+1,1);
+      tmp = Cshift(src,mu+1,1);
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = U_v[ss]*tmp_v[Ls*ss+s];
+	  }
+	}
+      }
       ref=ref + tmp - Gamma(Gmu[mu])*tmp;
 
-      tmp =adj(U[mu])*src;
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	autoView( src_v, src    , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
+	  }
+	}
+      }
       tmp =Cshift(tmp,mu+1,-1);
       ref=ref + tmp + Gamma(Gmu[mu])*tmp;
     }
@@ -182,11 +250,13 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
   DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  int ncall =3000;
+  Dw.DirichletBlock(Dirichlet);
+  Dw.ImportGauge(Umu);
+  
+  int ncall =300;
 
   if (1) {
     FGrid->Barrier();
-    Dw.ZeroCounters();
     Dw.Dhop(src,result,0);
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
@@ -211,29 +281,20 @@ int main (int argc, char ** argv)
     double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
 
     std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
-    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
-    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
+    //    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
+    //    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
     err = ref-result;
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-    //exit(0);
 
     if(( norm2(err)>1.0e-4) ) {
-      /*
-      std::cout << "RESULT\n " << result<<std::endl;
-      std::cout << "REF   \n " << ref   <<std::endl;
-      std::cout << "ERR   \n " << err   <<std::endl;
-      */
       std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
       FGrid->Barrier();
       exit(-1);
     }
     assert (norm2(err)< 1.0e-4 );
-    Dw.Report();
   }
 
   if (1)
@@ -242,16 +303,30 @@ int main (int argc, char ** argv)
     for(int mu=0;mu<Nd;mu++){
 
       //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
-      tmp = U[mu]*Cshift(src,mu+1,1);
+      tmp = Cshift(src,mu+1,1);
       {
 	autoView( ref_v, ref, CpuWrite);
 	autoView( tmp_v, tmp, CpuRead);
-	for(int i=0;i<ref_v.size();i++){
-	  ref_v[i]+= tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]; ;
+	autoView( U_v  , U[mu]  , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    int i=s+Ls*ss;
+	    ref_v[i]+= U_v[ss]*(tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]); ;
+	  }
 	}
       }
       
-      tmp =adj(U[mu])*src;
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	autoView( src_v, src    , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
+	  }
+	}
+      }
+      //      tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu+1,-1);
       {
 	autoView( ref_v, ref, CpuWrite);
@@ -263,21 +338,20 @@ int main (int argc, char ** argv)
     }
     ref = -0.5*ref;
   }
-  //  dump=1;
-  Dw.Dhop(src,result,1);
+
+  Dw.Dhop(src,result,DaggerYes);
+
+  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
   std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
+  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
+
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
   std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
   std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result;
   std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-/*
-	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
-	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
-	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
-*/
-  }
+  assert((norm2(err)<1.0e-4));
+  
   LatticeFermionF src_e (FrbGrid);
   LatticeFermionF src_o (FrbGrid);
   LatticeFermionF r_e   (FrbGrid);
@@ -307,7 +381,6 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
   {
-    Dw.ZeroCounters();
     FGrid->Barrier();
     Dw.DhopEO(src_o,r_e,DaggerNo);
     double t0=usecond();
@@ -329,7 +402,6 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
-    Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
   Dw.DhopOE(src_e,r_o,DaggerNo);
@@ -344,13 +416,7 @@ int main (int argc, char ** argv)
 
   err = r_eo-result;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-    /*
-	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
-	std::cout<< "Deo REF\n " <<result  << std::endl;
-	std::cout<< "Deo ERR   \n " << err <<std::endl;
-    */
-  }
+  assert(norm2(err)<1.0e-4);
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
@@ -359,6 +425,4 @@ int main (int argc, char ** argv)
 
   assert(norm2(src_e)<1.0e-4);
   assert(norm2(src_o)<1.0e-4);
-  Grid_finalize();
-  exit(0);
 }

benchmarks/Benchmark_memory_bandwidth.cc

@@ -184,8 +184,10 @@ int main (int argc, char ** argv)
       
       double bytes=1.0*vol*Nvec*sizeof(Real);
       double flops=vol*Nvec*2;// mul,add
-      std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"  \t\t"<<bytes/time<<"\t\t"<<flops/time<< "\t\t"<<(stop-start)/1000./1000.<< "\t\t " <<std::endl;
-
+      std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"
+	       <<bytes<<"  \t\t"<<bytes/time<<"\t\t"<<flops/time<< "\t\t"
+	       <<(stop-start)/1000./1000.<< "\t\t " <<std::endl;
+      assert(nn==nn);
   }    
 
   Grid_finalize();

examples/Example_Laplacian_solver.cc

@@ -4,7 +4,7 @@ using namespace Grid;
 template<class Field>
 void SimpleConjugateGradient(LinearOperatorBase<Field> &HPDop,const Field &b, Field &x)
 {
-    RealD cp, c, alpha, d, beta, ssq, qq;
+    RealD cp, c, alpha, d, beta, ssq;
     RealD Tolerance=1.0e-10;
     int MaxIterations=10000;
     

examples/Example_wall_wall_3pt.cc

@@ -0,0 +1,539 @@
+/*
+ * Warning: This code illustrative only: not well tested, and not meant for production use
+ * without regression / tests being applied
+ */
+
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+typedef SpinColourMatrix Propagator;
+typedef SpinColourVector Fermion;
+typedef PeriodicGimplR   GimplR;
+
+template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
+{
+public:
+  INHERIT_GIMPL_TYPES(Gimpl);
+
+  GridBase *grid;
+  GaugeField U;
+  
+  CovariantLaplacianCshift(GaugeField &_U)    :
+    grid(_U.Grid()),
+    U(_U) {  };
+
+  virtual GridBase *Grid(void) { return grid; };
+
+  virtual void  M    (const Field &in, Field &out)
+  {
+    out=Zero();
+    for(int mu=0;mu<Nd-1;mu++) {
+      GaugeLinkField Umu = PeekIndex<LorentzIndex>(U, mu); // NB: Inefficent
+      out = out - Gimpl::CovShiftForward(Umu,mu,in);    
+      out = out - Gimpl::CovShiftBackward(Umu,mu,in);    
+      out = out + 2.0*in;
+    }
+  };
+  virtual void  Mdag (const Field &in, Field &out) { M(in,out);}; // Laplacian is hermitian
+  virtual  void Mdiag    (const Field &in, Field &out)                  {assert(0);}; // Unimplemented need only for multigrid
+  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);}; // Unimplemented need only for multigrid
+  virtual  void MdirAll  (const Field &in, std::vector<Field> &out)     {assert(0);}; // Unimplemented need only for multigrid
+};
+
+void MakePhase(Coordinate mom,LatticeComplex &phase)
+{
+  GridBase *grid = phase.Grid();
+  auto latt_size = grid->GlobalDimensions();
+  ComplexD ci(0.0,1.0);
+  phase=Zero();
+
+  LatticeComplex coor(phase.Grid());
+  for(int mu=0;mu<Nd;mu++){
+    RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
+    LatticeCoordinate(coor,mu);
+    phase = phase + (TwoPiL * mom[mu]) * coor;
+  }
+  phase = exp(phase*ci);
+}
+void LinkSmear(int nstep, RealD rho,LatticeGaugeField &Uin,LatticeGaugeField &Usmr)
+{
+  Smear_Stout<GimplR> Stout(rho);
+  LatticeGaugeField Utmp(Uin.Grid());
+  Utmp = Uin;
+  for(int i=0;i<nstep;i++){
+    Stout.smear(Usmr,Utmp);
+    Utmp = Usmr;
+  }
+}
+void PointSource(Coordinate &coor,LatticePropagator &source)
+{
+  //  Coordinate coor({0,0,0,0});
+  source=Zero();
+  SpinColourMatrix kronecker; kronecker=1.0;
+  pokeSite(kronecker,source,coor);
+}
+void GFWallSource(int tslice,LatticePropagator &source)
+{
+  GridBase *grid = source.Grid();
+  LatticeComplex one(grid); one = ComplexD(1.0,0.0);
+  LatticeComplex zz(grid); zz=Zero();
+  LatticeInteger t(grid);
+  LatticeCoordinate(t,Tdir);
+  one = where(t==Integer(tslice), one, zz);
+  source = 1.0;
+  source = source * one;
+}
+
+void Z2WallSource(GridParallelRNG &RNG,int tslice,LatticePropagator &source)
+{
+  GridBase *grid = source.Grid();
+  LatticeComplex noise(grid);
+  LatticeComplex zz(grid); zz=Zero();
+  LatticeInteger t(grid);
+
+  RealD nrm=1.0/sqrt(2);
+  bernoulli(RNG, noise); // 0,1 50:50
+
+  noise = (2.*noise - Complex(1,1))*nrm;
+
+  LatticeCoordinate(t,Tdir);
+  noise = where(t==Integer(tslice), noise, zz);
+
+  source = 1.0;
+  source = source*noise;
+  std::cout << " Z2 wall " << norm2(source) << std::endl;
+}
+void GaugeFix(LatticeGaugeField &U,LatticeGaugeField &Ufix)
+{
+  Real alpha=0.05;
+
+  Real plaq=WilsonLoops<GimplR>::avgPlaquette(U);
+
+  std::cout << " Initial plaquette "<<plaq << std::endl;
+
+  LatticeColourMatrix   xform(U.Grid()); 
+  Ufix = U;
+  int orthog=Nd-1;
+  FourierAcceleratedGaugeFixer<GimplR>::SteepestDescentGaugeFix(Ufix,xform,alpha,100000,1.0e-14, 1.0e-14,true,orthog);
+  
+  plaq=WilsonLoops<GimplR>::avgPlaquette(Ufix);
+
+  std::cout << " Final plaquette "<<plaq << std::endl;
+}
+template<class Field>
+void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
+{
+  typedef CovariantLaplacianCshift <GimplR,Field> Laplacian_t;
+  Laplacian_t Laplacian(U);
+
+  Integer Iterations = 40;
+  Real width = 2.0;
+  Real coeff = (width*width) / Real(4*Iterations);
+
+  Field tmp(U.Grid());
+  smeared=unsmeared;
+  //  chi = (1-p^2/2N)^N kronecker
+  for(int n = 0; n < Iterations; ++n) {
+    Laplacian.M(smeared,tmp);
+    smeared = smeared - coeff*tmp;
+    std::cout << " smear iter " << n<<" " <<norm2(smeared)<<std::endl;
+  }
+}
+void GaussianSource(Coordinate &site,LatticeGaugeField &U,LatticePropagator &source)
+{
+  LatticePropagator tmp(source.Grid());
+  PointSource(site,source);
+  std::cout << " GaussianSource Kronecker "<< norm2(source)<<std::endl;
+  tmp = source;
+  GaussianSmear(U,tmp,source);
+  std::cout << " GaussianSource Smeared "<< norm2(source)<<std::endl;
+}
+void GaussianWallSource(GridParallelRNG &RNG,int tslice,LatticeGaugeField &U,LatticePropagator &source)
+{
+  Z2WallSource(RNG,tslice,source);
+  auto tmp = source;
+  GaussianSmear(U,tmp,source);
+}
+void SequentialSource(int tslice,Coordinate &mom,LatticePropagator &spectator,LatticePropagator &source)
+{
+  assert(mom.size()==Nd);
+  assert(mom[Tdir] == 0);
+
+  GridBase * grid = spectator.Grid();
+
+  LatticeInteger ts(grid);
+  LatticeCoordinate(ts,Tdir);
+  source = Zero();
+  source = where(ts==Integer(tslice),spectator,source); // Stick in a slice of the spectator, zero everywhere else
+
+  LatticeComplex phase(grid);
+  MakePhase(mom,phase);
+
+  source = source *phase;
+}
+template<class Action>
+void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
+{
+  GridBase *UGrid = D.GaugeGrid();
+  GridBase *FGrid = D.FermionGrid();
+
+  LatticeFermion src4  (UGrid); 
+  LatticeFermion src5  (FGrid); 
+  LatticeFermion result5(FGrid);
+  LatticeFermion result4(UGrid);
+  
+  ConjugateGradient<LatticeFermion> CG(1.0e-12,100000);
+  SchurRedBlackDiagTwoSolve<LatticeFermion> schur(CG);
+  ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
+  for(int s=0;s<Nd;s++){
+    for(int c=0;c<Nc;c++){
+      PropToFerm<Action>(src4,source,s,c);
+
+      D.ImportPhysicalFermionSource(src4,src5);
+
+      result5=Zero();
+      schur(D,src5,result5,ZG);
+      std::cout<<GridLogMessage
+	       <<"spin "<<s<<" color "<<c
+	       <<" norm2(src5d) "   <<norm2(src5)
+               <<" norm2(result5d) "<<norm2(result5)<<std::endl;
+
+      D.ExportPhysicalFermionSolution(result5,result4);
+
+      FermToProp<Action>(propagator,result4,s,c);
+    }
+  }
+}
+
+class MesonFile: Serializable {
+public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(MesonFile, std::vector<std::vector<Complex> >, data);
+};
+
+void MesonTrace(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
+{
+  const int nchannel=4;
+  Gamma::Algebra Gammas[nchannel][2] = {
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::Gamma5},
+    {Gamma::Algebra::GammaTGamma5,Gamma::Algebra::GammaTGamma5},
+    {Gamma::Algebra::GammaTGamma5,Gamma::Algebra::Gamma5},
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::GammaTGamma5}
+  };
+
+  Gamma G5(Gamma::Algebra::Gamma5);
+
+  LatticeComplex meson_CF(q1.Grid());
+  MesonFile MF;
+
+  for(int ch=0;ch<nchannel;ch++){
+
+    Gamma Gsrc(Gammas[ch][0]);
+    Gamma Gsnk(Gammas[ch][1]);
+
+    meson_CF = trace(G5*adj(q1)*G5*Gsnk*q2*adj(Gsrc));
+
+    std::vector<TComplex> meson_T;
+    sliceSum(meson_CF,meson_T, Tdir);
+
+    int nt=meson_T.size();
+
+    std::vector<Complex> corr(nt);
+    for(int t=0;t<nt;t++){
+      corr[t] = TensorRemove(meson_T[t]); // Yes this is ugly, not figured a work around
+      std::cout << " channel "<<ch<<" t "<<t<<" " <<corr[t]<<std::endl;
+    }
+    MF.data.push_back(corr);
+  }
+
+  {
+    XmlWriter WR(file);
+    write(WR,"MesonFile",MF);
+  }
+}
+
+void Meson3pt(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
+{
+  const int nchannel=4;
+  Gamma::Algebra Gammas[nchannel][2] = {
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::GammaX},
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::GammaY},
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::GammaZ},
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::GammaT}
+  };
+
+  Gamma G5(Gamma::Algebra::Gamma5);
+
+  LatticeComplex meson_CF(q1.Grid());
+  MesonFile MF;
+
+  for(int ch=0;ch<nchannel;ch++){
+
+    Gamma Gsrc(Gammas[ch][0]);
+    Gamma Gsnk(Gammas[ch][1]);
+
+    meson_CF = trace(G5*adj(q1)*G5*Gsnk*q2*adj(Gsrc));
+
+    std::vector<TComplex> meson_T;
+    sliceSum(meson_CF,meson_T, Tdir);
+
+    int nt=meson_T.size();
+
+    std::vector<Complex> corr(nt);
+    for(int t=0;t<nt;t++){
+      corr[t] = TensorRemove(meson_T[t]); // Yes this is ugly, not figured a work around
+      std::cout << " channel "<<ch<<" t "<<t<<" " <<corr[t]<<std::endl;
+    }
+    MF.data.push_back(corr);
+  }
+
+  {
+    XmlWriter WR(file);
+    write(WR,"MesonFile",MF);
+  }
+}
+
+
+void WallSinkMesonTrace(std::string file,std::vector<Propagator> &q1,std::vector<Propagator> &q2)
+{
+  const int nchannel=4;
+  Gamma::Algebra Gammas[nchannel][2] = {
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::Gamma5},
+    {Gamma::Algebra::GammaTGamma5,Gamma::Algebra::GammaTGamma5},
+    {Gamma::Algebra::GammaTGamma5,Gamma::Algebra::Gamma5},
+    {Gamma::Algebra::Gamma5      ,Gamma::Algebra::GammaTGamma5}
+  };
+
+  Gamma G5(Gamma::Algebra::Gamma5);
+  int nt=q1.size();
+  std::vector<Complex> meson_CF(nt);
+  MesonFile MF;
+
+  for(int ch=0;ch<nchannel;ch++){
+
+    Gamma Gsrc(Gammas[ch][0]);
+    Gamma Gsnk(Gammas[ch][1]);
+
+    std::vector<Complex> corr(nt);
+    for(int t=0;t<nt;t++){
+      meson_CF[t] = trace(G5*adj(q1[t])*G5*Gsnk*q2[t]*adj(Gsrc));
+      corr[t] = TensorRemove(meson_CF[t]); // Yes this is ugly, not figured a work around
+      std::cout << " channel "<<ch<<" t "<<t<<" " <<corr[t]<<std::endl;
+    }
+    MF.data.push_back(corr);
+  }
+
+  {
+    XmlWriter WR(file);
+    write(WR,"MesonFile",MF);
+  }
+}
+int make_idx(int p, int m,int nmom)
+{
+  if (m==0) return p;
+  assert(p==0);
+  return nmom + m - 1;
+}
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  // Double precision grids
+  auto latt = GridDefaultLatt();
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), 
+								   GridDefaultSimd(Nd,vComplex::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+
+  LatticeGaugeField Umu(UGrid);
+  LatticeGaugeField Utmp(UGrid);
+  LatticeGaugeField Usmr(UGrid);
+  std::string config;
+  if( argc > 1 && argv[1][0] != '-' )
+  {
+    std::cout<<GridLogMessage <<"Loading configuration from "<<argv[1]<<std::endl;
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu, header, argv[1]);
+    config=argv[1];
+  }
+  else
+  {
+    std::cout<<GridLogMessage <<"Using hot configuration"<<std::endl;
+    SU<Nc>::ColdConfiguration(Umu);
+    config="ColdConfig";
+  }
+  //  GaugeFix(Umu,Utmp);
+  //  Umu=Utmp;
+
+  int nsmr=3;
+  RealD rho=0.1;
+  LinkSmear(nsmr,rho,Umu,Usmr);
+
+
+  std::vector<int>   smeared_link({ 0,0,1} ); 
+  std::vector<RealD> masses({ 0.004,0.02477,0.447} ); // u/d, s, c ??
+  std::vector<RealD> M5s   ({ 1.8,1.8,1.0} ); 
+  std::vector<RealD> bs   ({ 1.0,1.0,1.5} );  // DDM
+  std::vector<RealD> cs   ({ 0.0,0.0,0.5} );  // DDM
+  std::vector<int>   Ls_s ({ 16,16,12} );
+  std::vector<GridCartesian *> FGrids;
+  std::vector<GridRedBlackCartesian *> FrbGrids;
+
+  std::vector<Coordinate> momenta;
+  momenta.push_back(Coordinate({0,0,0,0}));
+  momenta.push_back(Coordinate({1,0,0,0}));
+  momenta.push_back(Coordinate({2,0,0,0}));
+
+  int nmass = masses.size();
+  int nmom  = momenta.size();
+
+  std::vector<MobiusFermionR *> FermActs;
+  
+  std::cout<<GridLogMessage <<"======================"<<std::endl;
+  std::cout<<GridLogMessage <<"MobiusFermion action as Scaled Shamir kernel"<<std::endl;
+  std::cout<<GridLogMessage <<"======================"<<std::endl;
+
+  std::vector<Complex> boundary = {1,1,1,-1};
+  typedef MobiusFermionR FermionAction;
+  FermionAction::ImplParams Params(boundary);
+
+  for(int m=0;m<masses.size();m++) {
+
+    RealD mass = masses[m];
+    RealD M5   = M5s[m];
+    RealD b    = bs[m];
+    RealD c    = cs[m];
+    int   Ls   = Ls_s[m];
+
+    if ( smeared_link[m] ) Utmp = Usmr;
+    else                   Utmp = Umu;
+    
+    FGrids.push_back(SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid));
+    FrbGrids.push_back(SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid));
+
+    FermActs.push_back(new MobiusFermionR(Utmp,*FGrids[m],*FrbGrids[m],*UGrid,*UrbGrid,mass,M5,b,c,Params));
+  }
+
+  LatticePropagator z2wall_source(UGrid);
+  LatticePropagator gfwall_source(UGrid);
+  LatticePropagator phased_prop(UGrid);
+
+  int tslice = 0;
+  int tseq=(tslice+16)%latt[Nd-1];
+  //////////////////////////////////////////////////////////////////////
+  // RNG seeded for Z2 wall
+  //////////////////////////////////////////////////////////////////////
+  // You can manage seeds however you like.
+  // Recommend SeedUniqueString.
+  //////////////////////////////////////////////////////////////////////
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString("Study2-Source_Z2_p_0_0_0_t_0-880");
+  Z2WallSource  (RNG4,tslice,z2wall_source);
+  GFWallSource  (tslice,gfwall_source);
+
+  std::vector<LatticeComplex> phase(nmom,UGrid);
+  for(int m=0;m<nmom;m++){
+    MakePhase(momenta[m],phase[m]);
+  }
+
+  std::vector<LatticePropagator> Z2Props   (nmom+nmass-1,UGrid);
+  std::vector<LatticePropagator> GFProps   (nmom+nmass-1,UGrid);
+  for(int p=0;p<nmom;p++) {
+    int m=0;
+    int idx = make_idx(p,m,nmom);
+    phased_prop = z2wall_source * phase[p];
+    Solve(*FermActs[m],phased_prop  ,Z2Props[idx]);
+
+    phased_prop = gfwall_source * phase[p];
+    Solve(*FermActs[m],phased_prop  ,GFProps[idx]);
+  }
+  for(int m=1;m<nmass;m++) {
+    int p=0;
+    int idx = make_idx(p,m,nmom);
+    phased_prop = z2wall_source;
+    Solve(*FermActs[m],phased_prop  ,Z2Props[idx]);
+
+    phased_prop = gfwall_source;
+    Solve(*FermActs[m],phased_prop  ,GFProps[idx]);
+  }
+
+  std::vector<std::vector<Propagator> > wsnk_z2Props(nmom+nmass-1);
+  std::vector<std::vector<Propagator> > wsnk_gfProps(nmom+nmass-1);
+
+  // Non-zero kaon and point and D two point
+  // WW stick momentum on m1 (lighter)
+  //     zero momentum on m2
+  for(int m1=0;m1<nmass;m1++) {
+  for(int m2=m1;m2<nmass;m2++) {
+    int pmax = (m1==0)? nmom:1;
+    for(int p=0;p<pmax;p++){
+
+      std::stringstream ssg,ssz;
+      std::stringstream wssg,wssz;
+
+      int idx1 = make_idx(p,m1,nmom);
+      int idx2 = make_idx(0,m2,nmom);
+
+      /// Point sinks
+      ssg<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_p_gf_meson.xml";
+      ssz<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_p_z2_meson.xml";
+      MesonTrace(ssz.str(),Z2Props[idx1],Z2Props[idx2],phase[p]); // Q1 is conjugated
+      MesonTrace(ssg.str(),GFProps[idx1],GFProps[idx2],phase[p]); 
+      
+      /// Wall sinks
+      wssg<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_w_gf_meson.xml";
+      wssz<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_w_z2_meson.xml";
+      
+      phased_prop = GFProps[m2] * phase[p];
+      sliceSum(phased_prop,wsnk_gfProps[m1],Tdir);
+      sliceSum(GFProps[m1],wsnk_gfProps[m2],Tdir);
+      WallSinkMesonTrace(wssg.str(),wsnk_gfProps[m1],wsnk_gfProps[m2]);
+
+      phased_prop = Z2Props[m2] * phase[p];
+      sliceSum(phased_prop,wsnk_gfProps[m1],Tdir);
+      sliceSum(Z2Props[m1],wsnk_gfProps[m2],Tdir);
+      WallSinkMesonTrace(wssz.str(),wsnk_z2Props[m1],wsnk_z2Props[m2]);
+    }
+  }}
+
+
+  /////////////////////////////////////
+  // Sequential solves
+  /////////////////////////////////////
+  LatticePropagator  seq_wsnk_z2src(UGrid);
+  LatticePropagator  seq_wsnk_gfsrc(UGrid);
+  LatticePropagator  seq_psnk_z2src(UGrid);
+  LatticePropagator  seq_psnk_gfsrc(UGrid);
+  LatticePropagator source(UGrid);
+  for(int m=0;m<nmass-1;m++){
+    int spect_idx = make_idx(0,m,nmom);
+    int charm=nmass-1;
+
+    SequentialSource(tseq,momenta[0],GFProps[spect_idx],source);
+    Solve(*FermActs[charm],source,seq_psnk_gfsrc);
+    
+    SequentialSource(tseq,momenta[0],Z2Props[spect_idx],source);
+    Solve(*FermActs[charm],source,seq_psnk_z2src);
+
+    // Todo need wall sequential solve
+    for(int p=0;p<nmom;p++){
+      int active_idx = make_idx(p,0,nmom);
+      std::stringstream seq_3pt_p_z2;
+      std::stringstream seq_3pt_p_gf;
+      std::stringstream seq_3pt_w_z2;
+      std::stringstream seq_3pt_w_gf;
+      seq_3pt_p_z2  <<config<<"_3pt_p"<<p<< "_m" << m << "_p_z2_meson.xml";
+      seq_3pt_p_gf  <<config<<"_3pt_p"<<p<< "_m" << m << "_p_gf_meson.xml";
+      seq_3pt_w_z2  <<config<<"_3pt_p"<<p<< "_m" << m << "_w_z2_meson.xml";
+      seq_3pt_w_gf  <<config<<"_3pt_p"<<p<< "_m" << m << "_w_gf_meson.xml";
+      Meson3pt(seq_3pt_p_gf.str(),GFProps[active_idx],seq_psnk_gfsrc,phase[p]);
+      Meson3pt(seq_3pt_p_z2.str(),Z2Props[active_idx],seq_psnk_z2src,phase[p]);
+    }    
+  }
+  
+  Grid_finalize();
+}
+
+
+

examples/Example_wall_wall_spectrum.cc

@@ -9,6 +9,7 @@ using namespace std;
 using namespace Grid;
 typedef SpinColourMatrix Propagator;
 typedef SpinColourVector Fermion;
+typedef PeriodicGimplR   GimplR;
 
 template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
 {
@@ -55,6 +56,16 @@ void MakePhase(Coordinate mom,LatticeComplex &phase)
   }
   phase = exp(phase*ci);
 }
+void LinkSmear(int nstep, RealD rho,LatticeGaugeField &Uin,LatticeGaugeField &Usmr)
+{
+  Smear_Stout<GimplR> Stout(rho);
+  LatticeGaugeField Utmp(Uin.Grid());
+  Utmp = Uin;
+  for(int i=0;i<nstep;i++){
+    Stout.smear(Usmr,Utmp);
+    Utmp = Usmr;
+  }
+}
 void PointSource(Coordinate &coor,LatticePropagator &source)
 {
   //  Coordinate coor({0,0,0,0});
@@ -97,23 +108,23 @@ void GaugeFix(LatticeGaugeField &U,LatticeGaugeField &Ufix)
 {
   Real alpha=0.05;
 
-  Real plaq=WilsonLoops<PeriodicGimplR>::avgPlaquette(U);
+  Real plaq=WilsonLoops<GimplR>::avgPlaquette(U);
 
   std::cout << " Initial plaquette "<<plaq << std::endl;
 
   LatticeColourMatrix   xform(U.Grid()); 
   Ufix = U;
   int orthog=Nd-1;
-  FourierAcceleratedGaugeFixer<PeriodicGimplR>::SteepestDescentGaugeFix(Ufix,xform,alpha,10000,1.0e-12, 1.0e-12,true,orthog);
+  FourierAcceleratedGaugeFixer<GimplR>::SteepestDescentGaugeFix(Ufix,xform,alpha,100000,1.0e-14, 1.0e-14,true,orthog);
   
-  plaq=WilsonLoops<PeriodicGimplR>::avgPlaquette(Ufix);
+  plaq=WilsonLoops<GimplR>::avgPlaquette(Ufix);
 
   std::cout << " Final plaquette "<<plaq << std::endl;
 }
 template<class Field>
 void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
 {
-  typedef CovariantLaplacianCshift <PeriodicGimplR,Field> Laplacian_t;
+  typedef CovariantLaplacianCshift <GimplR,Field> Laplacian_t;
   Laplacian_t Laplacian(U);
 
   Integer Iterations = 40;
@@ -167,19 +178,21 @@ void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
   GridBase *UGrid = D.GaugeGrid();
   GridBase *FGrid = D.FermionGrid();
 
-  LatticeFermion src4  (UGrid); 
+  LatticeFermion src4  (UGrid); src4 = Zero();
   LatticeFermion src5  (FGrid); 
   LatticeFermion result5(FGrid);
   LatticeFermion result4(UGrid);
   
-  ConjugateGradient<LatticeFermion> CG(1.0e-8,100000);
-  SchurRedBlackDiagMooeeSolve<LatticeFermion> schur(CG);
+  ConjugateGradient<LatticeFermion> CG(1.0e-12,100000);
+  SchurRedBlackDiagTwoSolve<LatticeFermion> schur(CG);
   ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
+  std::cout<<GridLogMessage<< " source4 "<<norm2(source)<<std::endl;
   for(int s=0;s<Nd;s++){
     for(int c=0;c<Nc;c++){
       PropToFerm<Action>(src4,source,s,c);
-
+      std::cout<<GridLogMessage<< s<<c<<" src4 "<<norm2(src4)<<std::endl;
       D.ImportPhysicalFermionSource(src4,src5);
+      std::cout<<GridLogMessage<< s<<c<<" src5 "<<norm2(src5)<<std::endl;
 
       result5=Zero();
       schur(D,src5,result5,ZG);
@@ -287,15 +300,10 @@ int main (int argc, char ** argv)
 								   GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
 
-  //////////////////////////////////////////////////////////////////////
-  // You can manage seeds however you like.
-  // Recommend SeedUniqueString.
-  //////////////////////////////////////////////////////////////////////
-  std::vector<int> seeds4({1,2,3,4}); 
-  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
 
   LatticeGaugeField Umu(UGrid);
-  LatticeGaugeField Ufixed(UGrid);
+  LatticeGaugeField Utmp(UGrid);
+  LatticeGaugeField Usmr(UGrid);
   std::string config;
   if( argc > 1 && argv[1][0] != '-' )
   {
@@ -308,13 +316,20 @@ int main (int argc, char ** argv)
   {
     std::cout<<GridLogMessage <<"Using hot configuration"<<std::endl;
     SU<Nc>::ColdConfiguration(Umu);
-    //    SU<Nc>::HotConfiguration(RNG4,Umu);
-    config="HotConfig";
+    config="ColdConfig";
   }
-  GaugeFix(Umu,Ufixed);
-  Umu=Ufixed;
+  //  GaugeFix(Umu,Utmp);
+  //  Umu=Utmp;
 
+  int nsmr=3;
+  RealD rho=0.1;
+  RealD plaq_gf =WilsonLoops<GimplR>::avgPlaquette(Umu);
+  LinkSmear(nsmr,rho,Umu,Usmr);
+  RealD plaq_smr=WilsonLoops<GimplR>::avgPlaquette(Usmr);
+  std::cout << GridLogMessage << " GF Plaquette " <<plaq_gf<<std::endl;
+  std::cout << GridLogMessage << " SM Plaquette " <<plaq_smr<<std::endl;
 
+  std::vector<int>   smeared_link({ 0,0,1} ); 
   std::vector<RealD> masses({ 0.004,0.02477,0.447} ); // u/d, s, c ??
   std::vector<RealD> M5s   ({ 1.8,1.8,1.0} ); 
   std::vector<RealD> bs   ({ 1.0,1.0,1.5} );  // DDM
@@ -330,6 +345,9 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage <<"======================"<<std::endl;
   std::cout<<GridLogMessage <<"MobiusFermion action as Scaled Shamir kernel"<<std::endl;
   std::cout<<GridLogMessage <<"======================"<<std::endl;
+  std::vector<Complex> boundary = {1,1,1,-1};
+  typedef MobiusFermionR FermionAction;
+  FermionAction::ImplParams Params(boundary);
 
   for(int m=0;m<masses.size();m++) {
 
@@ -339,31 +357,41 @@ int main (int argc, char ** argv)
     RealD c    = cs[m];
     int   Ls   = Ls_s[m];
 
+    if ( smeared_link[m] ) Utmp = Usmr;
+    else                   Utmp = Umu;
+    
     FGrids.push_back(SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid));
     FrbGrids.push_back(SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid));
 
-    FermActs.push_back(new MobiusFermionR(Umu,*FGrids[m],*FrbGrids[m],*UGrid,*UrbGrid,mass,M5,b,c));
+    FermActs.push_back(new MobiusFermionR(Utmp,*FGrids[m],*FrbGrids[m],*UGrid,*UrbGrid,mass,M5,b,c,Params));
   }
 
-  LatticePropagator point_source(UGrid);
   LatticePropagator z2wall_source(UGrid);
   LatticePropagator gfwall_source(UGrid);
 
-  Coordinate Origin({0,0,0,0});
-  PointSource   (Origin,point_source);
-  Z2WallSource  (RNG4,0,z2wall_source);
-  GFWallSource  (0,gfwall_source);
+  int tslice = 0;
+  //////////////////////////////////////////////////////////////////////
+  // RNG seeded for Z2 wall
+  //////////////////////////////////////////////////////////////////////
+  // You can manage seeds however you like.
+  // Recommend SeedUniqueString.
+  //////////////////////////////////////////////////////////////////////
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString("Study2-Source_Z2_p_0_0_0_t_0-880");
+  Z2WallSource  (RNG4,tslice,z2wall_source);
+  GFWallSource  (tslice,gfwall_source);
 
-  std::vector<LatticePropagator> PointProps(nmass,UGrid);
-  std::vector<LatticePropagator> GaussProps(nmass,UGrid);
   std::vector<LatticePropagator> Z2Props   (nmass,UGrid);
   std::vector<LatticePropagator> GFProps   (nmass,UGrid);
 
   for(int m=0;m<nmass;m++) {
 
+    std::cout << GridLogMessage << " Mass " <<m << " z2wall source "<<norm2(z2wall_source)<<std::endl;
     Solve(*FermActs[m],z2wall_source    ,Z2Props[m]);
+    std::cout << GridLogMessage << " Mass " <<m << " gfwall source "<<norm2(gfwall_source)<<std::endl;
     Solve(*FermActs[m],gfwall_source    ,GFProps[m]);
 
+    std::cout << GridLogMessage << " Mass " <<m << " z2wall source "<<norm2(z2wall_source)<< " " << norm2(gfwall_source)<<std::endl;
+  
   }
 
   LatticeComplex phase(UGrid);
@@ -383,14 +411,15 @@ int main (int argc, char ** argv)
     std::stringstream wssg,wssz;
 
     /// Point sinks
-    ssg<<config<< "_m" << m1 << "_m"<< m2 << "p_gf_meson.xml";
-    ssz<<config<< "_m" << m1 << "_m"<< m2 << "p_z2_meson.xml";
+    ssg<<config<< "_m" << m1 << "_m"<< m2 << "_p_gf_meson.xml";
+    ssz<<config<< "_m" << m1 << "_m"<< m2 << "_p_z2_meson.xml";
 
     MesonTrace(ssz.str(),Z2Props[m1],Z2Props[m2],phase);
+    MesonTrace(ssg.str(),GFProps[m1],GFProps[m2],phase);
 
     /// Wall sinks
-    wssg<<config<< "_m" << m1 << "_m"<< m2 << "w_gf_meson.xml";
-    wssz<<config<< "_m" << m1 << "_m"<< m2 << "w_z2_meson.xml";
+    wssg<<config<< "_m" << m1 << "_m"<< m2 << "_w_gf_meson.xml";
+    wssz<<config<< "_m" << m1 << "_m"<< m2 << "_w_z2_meson.xml";
 
     WallSinkMesonTrace(wssg.str(),wsnk_gfProps[m1],wsnk_gfProps[m2]);
     WallSinkMesonTrace(wssz.str(),wsnk_z2Props[m1],wsnk_z2Props[m2]);

systems/Crusher/config-command

@@ -0,0 +1,12 @@
+../../configure --enable-comms=mpi-auto \
+--enable-unified=no \
+--enable-shm=nvlink \
+--enable-accelerator=hip \
+--enable-gen-simd-width=64 \
+--enable-simd=GPU \
+--disable-fermion-reps \
+--disable-gparity \
+CXX=hipcc MPICXX=mpicxx \
+CXXFLAGS="-fPIC -I/opt/rocm-4.5.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
+ LDFLAGS=" -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa "
+HIPFLAGS = --amdgpu-target=gfx90a

systems/Crusher/dwf.slurm

@@ -0,0 +1,30 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 1
+#SBATCH -n 8
+#SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export OMP_NUM_THREADS=1
+
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+
+PARAMS=" --accelerator-threads 16 --grid 32.32.32.256 --mpi 1.1.1.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+srun --gpus-per-task 1 -n8  ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+
+
+

systems/Crusher/dwf4.slurm

@@ -0,0 +1,27 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 1
+#SBATCH -n 4
+#SBATCH --exclusive
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export OMP_NUM_THREADS=4
+
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads 8 --grid 32.32.64.64 --mpi 1.1.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+
+srun --gpus-per-task 1 -n4 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+
+

systems/Crusher/dwf8.slurm

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 8
+#SBATCH -n 64
+#SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
+
+DIR=.
+module list
+export MPICH_OFI_NIC_POLICY=GPU
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+export OMP_NUM_THREADS=1
+
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.256 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.256.8node
+
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 1"
+echo $PARAMS
+srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node.shm0
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 1"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node_shm0
+
+
+

systems/Crusher/mpiwrapper.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+lrank=$SLURM_LOCALID
+lgpu=(0 1 2 3 7 6 5 4)
+
+export ROCR_VISIBLE_DEVICES=${lgpu[$lrank]}
+
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES "
+
+$*
+
+
+

systems/Crusher/sourceme.sh

@@ -0,0 +1,5 @@
+module load PrgEnv-gnu
+module load rocm/4.5.0
+module load gmp
+module load cray-fftw
+module load craype-accel-amd-gfx90a

systems/Spock/comms.slurm

@@ -0,0 +1,26 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -p ecp
+#SBATCH -J comms
+#SBATCH -o comms.%J
+#SBATCH -e comms.%J
+#SBATCH -N 1
+#SBATCH -n 2
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+export OMP_NUM_THREADS=8
+
+AT=8
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads ${AT} --grid 64.64.32.32 --mpi 2.1.1.1 "
+srun -n2 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_comms_host_device $PARAMS
+

systems/Spock/config-command

@@ -0,0 +1,14 @@
+../../configure --enable-comms=mpi-auto \
+--enable-unified=no \
+--enable-shm=nvlink \
+--enable-accelerator=hip \
+--enable-gen-simd-width=64 \
+--enable-simd=GPU \
+--disable-fermion-reps \
+--disable-gparity \
+--with-gmp=$OLCF_GMP_ROOT \
+--with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
+CXX=hipcc MPICXX=mpicxx \
+CXXFLAGS="-fPIC -I/opt/rocm-4.3.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
+--prefix=/ccs/home/chulwoo/Grid \
+ LDFLAGS=" -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa "

systems/Spock/dwf.slurm

@@ -0,0 +1,26 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -p ecp
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 1
+#SBATCH -n 1
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=NONE
+export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export OMP_NUM_THREADS=8
+
+AT=8
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads ${AT} --grid 32.32.32.32 --mpi 1.1.1.1 --comms-overlap"
+srun -n1 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+

systems/Spock/dwf4.slurm

@@ -0,0 +1,26 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
+#SBATCH -p ecp
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 1
+#SBATCH -n 4
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export OMP_NUM_THREADS=8
+
+AT=8
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads ${AT} --grid 32.32.64.64 --mpi 1.1.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+srun -n4 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+

systems/Spock/dwf8.slurm

@@ -0,0 +1,24 @@
+#!/bin/bash
+# Begin LSF Directives
+#SBATCH -A LGT104
+#SBATCH -t 3:00:00
+#SBATCH -p ecp
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 2
+#SBATCH -n 8
+
+DIR=.
+module list
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_SMP_SINGLE_COPY_MODE=CMA
+
+export OMP_NUM_THREADS=8
+
+AT=8
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads ${AT} --grid 16.16.16.48 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+srun -N2 -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./HMC/Mobius2p1f_DD_RHMC $PARAMS
+

systems/Spock/mpiwrapper.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+lrank=$SLURM_LOCALID
+
+export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
+
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
+
+$*
+
+
+

systems/Spock/sourceme.sh

@@ -0,0 +1,9 @@
+module load emacs
+module load PrgEnv-gnu
+module load rocm/4.5.0
+module load gmp
+module load cray-fftw
+module load craype-accel-amd-gfx908
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+export LD_LIBRARY_PATH=/opt/cray/pe/gcc/mpfr/3.1.4/lib/:$LD_LIBRARY_PATH

systems/Summit/comms.4node

@@ -0,0 +1,179 @@
+OPENMPI detected
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device Number    : 0
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device identifier: Tesla V100-SXM2-16GB
+AcceleratorCudaInit[0]:   totalGlobalMem: 16911433728 
+AcceleratorCudaInit[0]:   managedMemory: 1 
+AcceleratorCudaInit[0]:   isMultiGpuBoard: 0 
+AcceleratorCudaInit[0]:   warpSize: 32 
+AcceleratorCudaInit[0]:   pciBusID: 4 
+AcceleratorCudaInit[0]:   pciDeviceID: 0 
+AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
+AcceleratorCudaInit: rank 0 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 0 device 0 bus id: 0004:04:00.0
+AcceleratorCudaInit: ================================================
+SharedMemoryMpi:  World communicator of size 24
+SharedMemoryMpi:  Node  communicator of size 6
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 1073741824bytes at 0x200060000000 for comms buffers 
+Setting up IPC
+
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|_ |  |  |  |  |  |  |  |  |  |  |  | _|__
+__|_                                    _|__
+__|_   GGGG    RRRR    III    DDDD      _|__
+__|_  G        R   R    I     D   D     _|__
+__|_  G        R   R    I     D    D    _|__
+__|_  G  GG    RRRR     I     D    D    _|__
+__|_  G   G    R  R     I     D   D     _|__
+__|_   GGGG    R   R   III    DDDD      _|__
+__|_                                    _|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+  |  |  |  |  |  |  |  |  |  |  |  |  |  |  
+
+
+Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
+
+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.
+Current Grid git commit hash=7cb1ff7395a5833ded6526c43891bd07a0436290: (HEAD -> develop, origin/develop, origin/HEAD) clean
+
+Grid : Message : ================================================ 
+Grid : Message : MPI is initialised and logging filters activated 
+Grid : Message : ================================================ 
+Grid : Message : Requested 1073741824 byte stencil comms buffers 
+AcceleratorCudaInit: rank 1 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 1 device 1 bus id: 0004:05:00.0
+AcceleratorCudaInit: rank 2 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 2 device 2 bus id: 0004:06:00.0
+AcceleratorCudaInit: rank 5 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 5 device 5 bus id: 0035:05:00.0
+AcceleratorCudaInit: rank 4 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 4 device 4 bus id: 0035:04:00.0
+AcceleratorCudaInit: rank 3 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 3 device 3 bus id: 0035:03:00.0
+Grid : Message : MemoryManager Cache 13529146982 bytes 
+Grid : Message : MemoryManager::Init() setting up
+Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
+Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
+Grid : Message : MemoryManager::Init() Using cudaMalloc
+Grid : Message : 2.137929 s : Grid is setup to use 6 threads
+Grid : Message : 2.137941 s : Number of iterations to average: 250
+Grid : Message : 2.137950 s : ====================================================================================================
+Grid : Message : 2.137958 s : = Benchmarking sequential halo exchange from host memory 
+Grid : Message : 2.137966 s : ====================================================================================================
+Grid : Message : 2.137974 s :  L  	 Ls  	    bytes		MB/s uni	MB/s bidi
+AcceleratorCudaInit: rank 22 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 10 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 15 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 21 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 20 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 7 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 9 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 11 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 8 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 6 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 19 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 23 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 18 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 12 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 16 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 13 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 14 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 17 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+Grid : Message : 2.604949 s :    8	8	     393216       89973.9  		179947.8
+Grid : Message : 2.668249 s :    8	8	     393216       18650.3  		37300.5
+Grid : Message : 2.732288 s :    8	8	     393216       18428.5  		36857.1
+Grid : Message : 2.753565 s :    8	8	     393216       55497.2  		110994.4
+Grid : Message : 2.808960 s :   12	8	    1327104       100181.5  		200363.0
+Grid : Message : 3.226900 s :   12	8	    1327104       20600.5  		41201.0
+Grid : Message : 3.167459 s :   12	8	    1327104       24104.6  		48209.2
+Grid : Message : 3.227660 s :   12	8	    1327104       66156.7  		132313.5
+Grid : Message : 3.413570 s :   16	8	    3145728       56174.4  		112348.8
+Grid : Message : 3.802697 s :   16	8	    3145728       24255.9  		48511.7
+Grid : Message : 4.190498 s :   16	8	    3145728       24336.7  		48673.4
+Grid : Message : 4.385171 s :   16	8	    3145728       48484.1  		96968.2
+Grid : Message : 4.805284 s :   20	8	    6144000       46380.5  		92761.1
+Grid : Message : 5.562975 s :   20	8	    6144000       24328.5  		48656.9
+Grid : Message : 6.322562 s :   20	8	    6144000       24266.7  		48533.4
+Grid : Message : 6.773598 s :   20	8	    6144000       40868.5  		81736.9
+Grid : Message : 7.600999 s :   24	8	   10616832       40198.3  		80396.6
+Grid : Message : 8.912917 s :   24	8	   10616832       24279.5  		48559.1
+Grid : Message : 10.220961 s :   24	8	   10616832       24350.2  		48700.4
+Grid : Message : 11.728250 s :   24	8	   10616832       37390.9  		74781.8
+Grid : Message : 12.497258 s :   28	8	   16859136       36792.2  		73584.5
+Grid : Message : 14.585387 s :   28	8	   16859136       24222.2  		48444.3
+Grid : Message : 16.664783 s :   28	8	   16859136       24323.4  		48646.8
+Grid : Message : 17.955238 s :   28	8	   16859136       39194.7  		78389.4
+Grid : Message : 20.136479 s :   32	8	   25165824       35718.3  		71436.5
+Grid : Message : 23.241958 s :   32	8	   25165824       24311.4  		48622.9
+Grid : Message : 26.344810 s :   32	8	   25165824       24331.9  		48663.7
+Grid : Message : 28.384420 s :   32	8	   25165824       37016.3  		74032.7
+Grid : Message : 28.388879 s : ====================================================================================================
+Grid : Message : 28.388894 s : = Benchmarking sequential halo exchange from GPU memory 
+Grid : Message : 28.388909 s : ====================================================================================================
+Grid : Message : 28.388924 s :  L  	 Ls  	    bytes		MB/s uni	MB/s bidi
+Grid : Message : 28.553993 s :    8	8	     393216       8272.4  		16544.7
+Grid : Message : 28.679592 s :    8	8	     393216       9395.4  		18790.8
+Grid : Message : 28.811112 s :    8	8	     393216       8971.0  		17942.0
+Grid : Message : 28.843770 s :    8	8	     393216       36145.6  		72291.2
+Grid : Message : 28.981754 s :   12	8	    1327104       49591.6  		99183.2
+Grid : Message : 29.299764 s :   12	8	    1327104       12520.8  		25041.7
+Grid : Message : 29.620288 s :   12	8	    1327104       12422.2  		24844.4
+Grid : Message : 29.657645 s :   12	8	    1327104       106637.5  		213275.1
+Grid : Message : 29.952933 s :   16	8	    3145728       43939.2  		87878.5
+Grid : Message : 30.585411 s :   16	8	    3145728       14922.1  		29844.2
+Grid : Message : 31.219781 s :   16	8	    3145728       14877.2  		29754.4
+Grid : Message : 31.285017 s :   16	8	    3145728       144724.3  		289448.7
+Grid : Message : 31.706443 s :   20	8	    6144000       54676.2  		109352.4
+Grid : Message : 32.739205 s :   20	8	    6144000       17848.0  		35696.1
+Grid : Message : 33.771852 s :   20	8	    6144000       17849.9  		35699.7
+Grid : Message : 33.871981 s :   20	8	    6144000       184141.4  		368282.8
+Grid : Message : 34.536808 s :   24	8	   10616832       55784.3  		111568.6
+Grid : Message : 36.275648 s :   24	8	   10616832       18317.6  		36635.3
+Grid : Message : 37.997181 s :   24	8	   10616832       18501.7  		37003.4
+Grid : Message : 38.140442 s :   24	8	   10616832       222383.9  		444767.9
+Grid : Message : 39.177222 s :   28	8	   16859136       56609.7  		113219.4
+Grid : Message : 41.874755 s :   28	8	   16859136       18749.9  		37499.8
+Grid : Message : 44.529381 s :   28	8	   16859136       19052.9  		38105.8
+Grid : Message : 44.742192 s :   28	8	   16859136       237717.1  		475434.2
+Grid : Message : 46.184000 s :   32	8	   25165824       57091.2  		114182.4
+Grid : Message : 50.734740 s :   32	8	   25165824       19411.0  		38821.9
+Grid : Message : 53.931228 s :   32	8	   25165824       19570.6  		39141.2
+Grid : Message : 54.238467 s :   32	8	   25165824       245765.6  		491531.2
+Grid : Message : 54.268664 s : ====================================================================================================
+Grid : Message : 54.268680 s : = All done; Bye Bye
+Grid : Message : 54.268691 s : ====================================================================================================

systems/Summit/dwf.24.4node

@@ -0,0 +1,206 @@
+OPENMPI detected
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device Number    : 0
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device identifier: Tesla V100-SXM2-16GB
+AcceleratorCudaInit[0]:   totalGlobalMem: 16911433728 
+AcceleratorCudaInit[0]:   managedMemory: 1 
+AcceleratorCudaInit[0]:   isMultiGpuBoard: 0 
+AcceleratorCudaInit[0]:   warpSize: 32 
+AcceleratorCudaInit[0]:   pciBusID: 4 
+AcceleratorCudaInit[0]:   pciDeviceID: 0 
+AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
+AcceleratorCudaInit: rank 0 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 0 device 0 bus id: 0004:04:00.0
+AcceleratorCudaInit: ================================================
+SharedMemoryMpi:  World communicator of size 24
+SharedMemoryMpi:  Node  communicator of size 6
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x200080000000 for comms buffers 
+AcceleratorCudaInit: rank 3 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 3 device 3 bus id: 0035:03:00.0
+AcceleratorCudaInit: rank 5 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 5 device 5 bus id: 0035:05:00.0
+Setting up IPC
+
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|_ |  |  |  |  |  |  |  |  |  |  |  | _|__
+__|_                                    _|__
+__|_   GGGG    RRRR    III    DDDD      _|__
+__|_  G        R   R    I     D   D     _|__
+__|_  G        R   R    I     D    D    _|__
+__|_  G  GG    RRRR     I     D    D    _|__
+__|_  G   G    R  R     I     D   D     _|__
+__|_   GGGG    R   R   III    DDDD      _|__
+__|_                                    _|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+  |  |  |  |  |  |  |  |  |  |  |  |  |  |  
+
+
+Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
+
+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
+AcceleratorCudaInit: rank 4 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 4 device 4 bus id: 0035:04:00.0
+AcceleratorCudaInit: rank 1 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 1 device 1 bus id: 0004:05:00.0
+AcceleratorCudaInit: rank 2 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 2 device 2 bus id: 0004:06:00.0
+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.
+Current Grid git commit hash=7cb1ff7395a5833ded6526c43891bd07a0436290: (HEAD -> develop, origin/develop, origin/HEAD) clean
+
+Grid : Message : ================================================ 
+Grid : Message : MPI is initialised and logging filters activated 
+Grid : Message : ================================================ 
+Grid : Message : Requested 2147483648 byte stencil comms buffers 
+Grid : Message : MemoryManager Cache 8388608000 bytes 
+Grid : Message : MemoryManager::Init() setting up
+Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
+Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
+Grid : Message : MemoryManager::Init() Using cudaMalloc
+Grid : Message : 1.731905 s : Grid Layout
+Grid : Message : 1.731915 s : 	Global lattice size  : 48 48 48 72 
+Grid : Message : 1.731928 s : 	OpenMP threads       : 6
+Grid : Message : 1.731938 s : 	MPI tasks            : 2 2 2 3 
+AcceleratorCudaInit: rank 9 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 23 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 22 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 21 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 18 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 6 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 7 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 10 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 8 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 11 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 20 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 19 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 13 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 12 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 14 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 16 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 15 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 17 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+Grid : Message : 2.683494 s : Making s innermost grids
+Grid : Message : 2.780034 s : Initialising 4d RNG
+Grid : Message : 2.833099 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 2.833121 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 2.916841 s : Initialising 5d RNG
+Grid : Message : 3.762880 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 3.762902 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 5.264345 s : Initialised RNGs
+Grid : Message : 6.489904 s : Drawing gauge field
+Grid : Message : 6.729262 s : Random gauge initialised 
+Grid : Message : 7.781273 s : Setting up Cshift based reference 
+Grid : Message : 8.725313 s : *****************************************************************
+Grid : Message : 8.725332 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 8.725342 s : *****************************************************************
+Grid : Message : 8.725352 s : *****************************************************************
+Grid : Message : 8.725362 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 8.725372 s : * Vectorising space-time by 4
+Grid : Message : 8.725383 s : * VComplexF size is 32 B
+Grid : Message : 8.725395 s : * SINGLE precision 
+Grid : Message : 8.725405 s : * Using Overlapped Comms/Compute
+Grid : Message : 8.725415 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 8.725425 s : *****************************************************************
+Grid : Message : 9.465229 s : Called warmup
+Grid : Message : 58.646066 s : Called Dw 3000 times in 4.91764e+07 us
+Grid : Message : 58.646121 s : mflop/s =   1.02592e+07
+Grid : Message : 58.646134 s : mflop/s per rank =  427468
+Grid : Message : 58.646145 s : mflop/s per node =  2.56481e+06
+Grid : Message : 58.646156 s : RF  GiB/s (base 2) =   20846.5
+Grid : Message : 58.646166 s : mem GiB/s (base 2) =   13029.1
+Grid : Message : 58.648008 s : norm diff   1.04778e-13
+Grid : Message : 58.734885 s : #### Dhop calls report 
+Grid : Message : 58.734897 s : WilsonFermion5D Number of DhopEO Calls   : 6002
+Grid : Message : 58.734909 s : WilsonFermion5D TotalTime   /Calls        : 8217.71 us
+Grid : Message : 58.734922 s : WilsonFermion5D CommTime    /Calls        : 7109.5 us
+Grid : Message : 58.734933 s : WilsonFermion5D FaceTime    /Calls        : 446.623 us
+Grid : Message : 58.734943 s : WilsonFermion5D ComputeTime1/Calls        : 18.0558 us
+Grid : Message : 58.734953 s : WilsonFermion5D ComputeTime2/Calls        : 731.097 us
+Grid : Message : 58.734979 s : Average mflops/s per call                : 4.8157e+09
+Grid : Message : 58.734989 s : Average mflops/s per call per rank       : 2.00654e+08
+Grid : Message : 58.734999 s : Average mflops/s per call per node       : 1.20393e+09
+Grid : Message : 58.735008 s : Average mflops/s per call (full)         : 1.04183e+07
+Grid : Message : 58.735017 s : Average mflops/s per call per rank (full): 434094
+Grid : Message : 58.735026 s : Average mflops/s per call per node (full): 2.60456e+06
+Grid : Message : 58.735035 s : WilsonFermion5D Stencil
+Grid : Message : 58.735043 s : WilsonFermion5D StencilEven
+Grid : Message : 58.735051 s : WilsonFermion5D StencilOdd
+Grid : Message : 58.735059 s : WilsonFermion5D Stencil     Reporti()
+Grid : Message : 58.735067 s : WilsonFermion5D StencilEven Reporti()
+Grid : Message : 58.735075 s : WilsonFermion5D StencilOdd  Reporti()
+Grid : Message : 64.934380 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 64.934740 s : Called DwDag
+Grid : Message : 64.934870 s : norm dag result 12.0422
+Grid : Message : 64.120756 s : norm dag ref    12.0422
+Grid : Message : 64.149389 s : norm dag diff   7.6644e-14
+Grid : Message : 64.317786 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 64.465331 s : src_e0.499995
+Grid : Message : 64.524653 s : src_o0.500005
+Grid : Message : 64.558706 s : *********************************************************
+Grid : Message : 64.558717 s : * Benchmarking DomainWallFermionF::DhopEO                
+Grid : Message : 64.558727 s : * Vectorising space-time by 4
+Grid : Message : 64.558737 s : * SINGLE precision 
+Grid : Message : 64.558745 s : * Using Overlapped Comms/Compute
+Grid : Message : 64.558753 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 64.558761 s : *********************************************************
+Grid : Message : 92.702145 s : Deo mflop/s =   8.97692e+06
+Grid : Message : 92.702185 s : Deo mflop/s per rank   374038
+Grid : Message : 92.702198 s : Deo mflop/s per node   2.24423e+06
+Grid : Message : 92.702209 s : #### Dhop calls report 
+Grid : Message : 92.702223 s : WilsonFermion5D Number of DhopEO Calls   : 3001
+Grid : Message : 92.702240 s : WilsonFermion5D TotalTime   /Calls        : 9377.88 us
+Grid : Message : 92.702257 s : WilsonFermion5D CommTime    /Calls        : 8221.84 us
+Grid : Message : 92.702277 s : WilsonFermion5D FaceTime    /Calls        : 543.548 us
+Grid : Message : 92.702301 s : WilsonFermion5D ComputeTime1/Calls        : 20.936 us
+Grid : Message : 92.702322 s : WilsonFermion5D ComputeTime2/Calls        : 732.33 us
+Grid : Message : 92.702376 s : Average mflops/s per call                : 4.13001e+09
+Grid : Message : 92.702387 s : Average mflops/s per call per rank       : 1.72084e+08
+Grid : Message : 92.702397 s : Average mflops/s per call per node       : 1.0325e+09
+Grid : Message : 92.702407 s : Average mflops/s per call (full)         : 9.12937e+06
+Grid : Message : 92.702416 s : Average mflops/s per call per rank (full): 380391
+Grid : Message : 92.702426 s : Average mflops/s per call per node (full): 2.28234e+06
+Grid : Message : 92.702435 s : WilsonFermion5D Stencil
+Grid : Message : 92.702443 s : WilsonFermion5D StencilEven
+Grid : Message : 92.702451 s : WilsonFermion5D StencilOdd
+Grid : Message : 92.702459 s : WilsonFermion5D Stencil     Reporti()
+Grid : Message : 92.702467 s : WilsonFermion5D StencilEven Reporti()
+Grid : Message : 92.702475 s : WilsonFermion5D StencilOdd  Reporti()
+Grid : Message : 92.772983 s : r_e6.02121
+Grid : Message : 92.786384 s : r_o6.02102
+Grid : Message : 92.799622 s : res12.0422
+Grid : Message : 93.860500 s : norm diff   0
+Grid : Message : 93.162026 s : norm diff even  0
+Grid : Message : 93.197529 s : norm diff odd   0

systems/Summit/dwf.32.4node

@@ -0,0 +1,206 @@
+OPENMPI detected
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device Number    : 0
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device identifier: Tesla V100-SXM2-16GB
+AcceleratorCudaInit[0]:   totalGlobalMem: 16911433728 
+AcceleratorCudaInit[0]:   managedMemory: 1 
+AcceleratorCudaInit[0]:   isMultiGpuBoard: 0 
+AcceleratorCudaInit[0]:   warpSize: 32 
+AcceleratorCudaInit[0]:   pciBusID: 4 
+AcceleratorCudaInit[0]:   pciDeviceID: 0 
+AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
+AcceleratorCudaInit: rank 0 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 0 device 0 bus id: 0004:04:00.0
+AcceleratorCudaInit: ================================================
+SharedMemoryMpi:  World communicator of size 24
+SharedMemoryMpi:  Node  communicator of size 6
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x200080000000 for comms buffers 
+Setting up IPC
+
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|_ |  |  |  |  |  |  |  |  |  |  |  | _|__
+__|_                                    _|__
+__|_   GGGG    RRRR    III    DDDD      _|__
+__|_  G        R   R    I     D   D     _|__
+__|_  G        R   R    I     D    D    _|__
+__|_  G  GG    RRRR     I     D    D    _|__
+__|_  G   G    R  R     I     D   D     _|__
+__|_   GGGG    R   R   III    DDDD      _|__
+__|_                                    _|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+  |  |  |  |  |  |  |  |  |  |  |  |  |  |  
+
+
+Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
+
+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
+AcceleratorCudaInit: rank 2 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 2 device 2 bus id: 0004:06:00.0
+AcceleratorCudaInit: rank 1 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 1 device 1 bus id: 0004:05:00.0
+AcceleratorCudaInit: rank 4 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 4 device 4 bus id: 0035:04:00.0
+AcceleratorCudaInit: rank 3 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 3 device 3 bus id: 0035:03:00.0
+AcceleratorCudaInit: rank 5 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+local rank 5 device 5 bus id: 0035:05:00.0
+GNU General Public License for more details.
+Current Grid git commit hash=7cb1ff7395a5833ded6526c43891bd07a0436290: (HEAD -> develop, origin/develop, origin/HEAD) clean
+
+Grid : Message : ================================================ 
+Grid : Message : MPI is initialised and logging filters activated 
+Grid : Message : ================================================ 
+Grid : Message : Requested 2147483648 byte stencil comms buffers 
+Grid : Message : MemoryManager Cache 8388608000 bytes 
+Grid : Message : MemoryManager::Init() setting up
+Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
+Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
+Grid : Message : MemoryManager::Init() Using cudaMalloc
+Grid : Message : 1.544984 s : Grid Layout
+Grid : Message : 1.544992 s : 	Global lattice size  : 64 64 64 96 
+Grid : Message : 1.545003 s : 	OpenMP threads       : 6
+Grid : Message : 1.545011 s : 	MPI tasks            : 2 2 2 3 
+AcceleratorCudaInit: rank 8 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 6 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 11 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 16 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 17 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 13 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 12 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 21 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 23 setting device to node rank 5
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 22 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 19 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 18 setting device to node rank 0
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 7 setting device to node rank 1
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 10 setting device to node rank 4
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 9 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 14 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 15 setting device to node rank 3
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+AcceleratorCudaInit: rank 20 setting device to node rank 2
+AcceleratorCudaInit: Configure options --enable-setdevice=yes 
+Grid : Message : 2.994920 s : Making s innermost grids
+Grid : Message : 2.232502 s : Initialising 4d RNG
+Grid : Message : 2.397047 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 2.397069 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 2.653140 s : Initialising 5d RNG
+Grid : Message : 5.285347 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 5.285369 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 9.994738 s : Initialised RNGs
+Grid : Message : 13.153426 s : Drawing gauge field
+Grid : Message : 13.825697 s : Random gauge initialised 
+Grid : Message : 18.537657 s : Setting up Cshift based reference 
+Grid : Message : 22.296755 s : *****************************************************************
+Grid : Message : 22.296781 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 22.296791 s : *****************************************************************
+Grid : Message : 22.296800 s : *****************************************************************
+Grid : Message : 22.296809 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 22.296818 s : * Vectorising space-time by 4
+Grid : Message : 22.296828 s : * VComplexF size is 32 B
+Grid : Message : 22.296838 s : * SINGLE precision 
+Grid : Message : 22.296847 s : * Using Overlapped Comms/Compute
+Grid : Message : 22.296855 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 22.296863 s : *****************************************************************
+Grid : Message : 24.746452 s : Called warmup
+Grid : Message : 137.525756 s : Called Dw 3000 times in 1.12779e+08 us
+Grid : Message : 137.525818 s : mflop/s =   1.41383e+07
+Grid : Message : 137.525831 s : mflop/s per rank =  589097
+Grid : Message : 137.525843 s : mflop/s per node =  3.53458e+06
+Grid : Message : 137.525854 s : RF  GiB/s (base 2) =   28728.7
+Grid : Message : 137.525864 s : mem GiB/s (base 2) =   17955.5
+Grid : Message : 137.693645 s : norm diff   1.04885e-13
+Grid : Message : 137.965585 s : #### Dhop calls report 
+Grid : Message : 137.965598 s : WilsonFermion5D Number of DhopEO Calls   : 6002
+Grid : Message : 137.965612 s : WilsonFermion5D TotalTime   /Calls        : 18899.7 us
+Grid : Message : 137.965624 s : WilsonFermion5D CommTime    /Calls        : 16041.4 us
+Grid : Message : 137.965634 s : WilsonFermion5D FaceTime    /Calls        : 859.705 us
+Grid : Message : 137.965644 s : WilsonFermion5D ComputeTime1/Calls        : 70.5881 us
+Grid : Message : 137.965654 s : WilsonFermion5D ComputeTime2/Calls        : 2094.8 us
+Grid : Message : 137.965682 s : Average mflops/s per call                : 3.87638e+09
+Grid : Message : 137.965692 s : Average mflops/s per call per rank       : 1.61516e+08
+Grid : Message : 137.965702 s : Average mflops/s per call per node       : 9.69095e+08
+Grid : Message : 137.965712 s : Average mflops/s per call (full)         : 1.43168e+07
+Grid : Message : 137.965721 s : Average mflops/s per call per rank (full): 596533
+Grid : Message : 137.965730 s : Average mflops/s per call per node (full): 3.5792e+06
+Grid : Message : 137.965740 s : WilsonFermion5D Stencil
+Grid : Message : 137.965748 s : WilsonFermion5D StencilEven
+Grid : Message : 137.965756 s : WilsonFermion5D StencilOdd
+Grid : Message : 137.965764 s : WilsonFermion5D Stencil     Reporti()
+Grid : Message : 137.965772 s : WilsonFermion5D StencilEven Reporti()
+Grid : Message : 137.965780 s : WilsonFermion5D StencilOdd  Reporti()
+Grid : Message : 156.554605 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 156.554632 s : Called DwDag
+Grid : Message : 156.554642 s : norm dag result 12.0421
+Grid : Message : 156.639265 s : norm dag ref    12.0421
+Grid : Message : 156.888281 s : norm dag diff   7.62057e-14
+Grid : Message : 157.609797 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 158.208630 s : src_e0.499996
+Grid : Message : 158.162447 s : src_o0.500004
+Grid : Message : 158.267780 s : *********************************************************
+Grid : Message : 158.267791 s : * Benchmarking DomainWallFermionF::DhopEO                
+Grid : Message : 158.267801 s : * Vectorising space-time by 4
+Grid : Message : 158.267811 s : * SINGLE precision 
+Grid : Message : 158.267820 s : * Using Overlapped Comms/Compute
+Grid : Message : 158.267828 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 158.267836 s : *********************************************************
+Grid : Message : 216.487829 s : Deo mflop/s =   1.37283e+07
+Grid : Message : 216.487869 s : Deo mflop/s per rank   572011
+Grid : Message : 216.487881 s : Deo mflop/s per node   3.43206e+06
+Grid : Message : 216.487893 s : #### Dhop calls report 
+Grid : Message : 216.487903 s : WilsonFermion5D Number of DhopEO Calls   : 3001
+Grid : Message : 216.487913 s : WilsonFermion5D TotalTime   /Calls        : 19399.6 us
+Grid : Message : 216.487923 s : WilsonFermion5D CommTime    /Calls        : 16475.4 us
+Grid : Message : 216.487933 s : WilsonFermion5D FaceTime    /Calls        : 972.393 us
+Grid : Message : 216.487943 s : WilsonFermion5D ComputeTime1/Calls        : 49.8474 us
+Grid : Message : 216.487953 s : WilsonFermion5D ComputeTime2/Calls        : 2089.93 us
+Grid : Message : 216.488001 s : Average mflops/s per call                : 5.39682e+09
+Grid : Message : 216.488011 s : Average mflops/s per call per rank       : 2.24867e+08
+Grid : Message : 216.488020 s : Average mflops/s per call per node       : 1.3492e+09
+Grid : Message : 216.488030 s : Average mflops/s per call (full)         : 1.39479e+07
+Grid : Message : 216.488039 s : Average mflops/s per call per rank (full): 581162
+Grid : Message : 216.488048 s : Average mflops/s per call per node (full): 3.48697e+06
+Grid : Message : 216.488057 s : WilsonFermion5D Stencil
+Grid : Message : 216.488065 s : WilsonFermion5D StencilEven
+Grid : Message : 216.488073 s : WilsonFermion5D StencilOdd
+Grid : Message : 216.488081 s : WilsonFermion5D Stencil     Reporti()
+Grid : Message : 216.488089 s : WilsonFermion5D StencilEven Reporti()
+Grid : Message : 216.488097 s : WilsonFermion5D StencilOdd  Reporti()
+Grid : Message : 217.384495 s : r_e6.02113
+Grid : Message : 217.426121 s : r_o6.02096
+Grid : Message : 217.472636 s : res12.0421
+Grid : Message : 218.200068 s : norm diff   0
+Grid : Message : 218.645673 s : norm diff even  0
+Grid : Message : 218.816561 s : norm diff odd   0

systems/Summit/dwf16.lsf

@@ -1,7 +1,7 @@
 #!/bin/bash
 #BSUB -P LGT104
 #BSUB -W 2:00
-#BSUB -nnodes 4
+#BSUB -nnodes 16
 #BSUB -J DWF
 
 export OMP_NUM_THREADS=6
@@ -9,14 +9,14 @@ export PAMI_IBV_ADAPTER_AFFINITY=1
 export PAMI_ENABLE_STRIPING=1
 export OPT="--comms-concurrent --comms-overlap "
 
-APP="./benchmarks/Benchmark_comms_host_device  --mpi 2.2.2.3 "
-jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP
+APP="./benchmarks/Benchmark_comms_host_device  --mpi 4.4.4.3 "
+jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > comms.16node.log
 
-APP="./benchmarks/Benchmark_dwf_fp32 --grid 48.48.48.72 --mpi 2.2.2.3 --shm 1024 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
-jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP
+APP="./benchmarks/Benchmark_dwf_fp32 --grid 96.96.96.72 --mpi 4.4.4.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
+jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.16node.24.log
 
-APP="./benchmarks/Benchmark_dwf_fp32 --grid 64.64.64.96 --mpi 2.2.2.3 --shm 1024 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
-jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP
+APP="./benchmarks/Benchmark_dwf_fp32 --grid 128.128.128.96 --mpi 4.4.4.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
+jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.16node.32.log
 
 
 

systems/Summit/dwf4.lsf

@@ -10,13 +10,13 @@ export PAMI_ENABLE_STRIPING=1
 export OPT="--comms-concurrent --comms-overlap "
 #export GRID_ALLOC_NCACHE_LARGE=1
 export APP="./benchmarks/Benchmark_comms_host_device  --mpi 2.2.2.3 "
-jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP
+jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > comms.4node
 
-APP="./benchmarks/Benchmark_dwf_fp32 --grid 48.48.48.72 --mpi 2.2.2.3 --shm 1024 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
-jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP
+APP="./benchmarks/Benchmark_dwf_fp32 --grid 48.48.48.72 --mpi 2.2.2.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
+jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.24.4node
 
-APP="./benchmarks/Benchmark_dwf_fp32 --grid 64.64.64.96 --mpi 2.2.2.3 --shm 1024 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
-jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP
+APP="./benchmarks/Benchmark_dwf_fp32 --grid 64.64.64.96 --mpi 2.2.2.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
+jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.32.4node
 
 
 

systems/Tursa/config-command

@@ -5,7 +5,7 @@
     --enable-gen-simd-width=64 \
     --enable-accelerator=cuda \
     --with-lime=/mnt/lustre/tursafs1/home/tc002/tc002/dc-boyl1/spack/spack/opt/spack/linux-rhel8-zen/gcc-8.4.1/c-lime-2-3-9-e6wxqrid6rqmd45z7n32dxkvkykpvyez \
-    --disable-accelerator-cshift \
+    --enable-accelerator-cshift \
     --disable-unified \
     CXX=nvcc \
     LDFLAGS="-cudart shared " \

systems/Tursa/dwf.4node.perf

@@ -1,25 +1,25 @@
-tu-c0r0n00 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n00 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n09 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n00 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n06 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n06 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n09 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n09 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n03 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n06 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n09 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n00 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n03 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n03 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n06 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n03 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n00 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n00 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n00 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n00 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n06 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n06 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n06 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n06 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n03 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n03 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n03 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n03 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n09 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n09 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n09 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n09 - 3 device=3 binding=--interleave=6,7
 OPENMPI detected
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
 AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
-AcceleratorCudaInit: ================================================
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
 OPENMPI detected
 AcceleratorCudaInit[0]: ========================
 AcceleratorCudaInit[0]: Device Number    : 0
@@ -33,11 +33,41 @@ AcceleratorCudaInit[0]:   pciBusID: 3
 AcceleratorCudaInit[0]:   pciDeviceID: 0 
 AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
 AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
-AcceleratorCudaInit: ================================================
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
 OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
 AcceleratorCudaInit[0]: ========================
 AcceleratorCudaInit[0]: Device Number    : 0
 AcceleratorCudaInit[0]: ========================
@@ -50,43 +80,25 @@ AcceleratorCudaInit[0]:   pciBusID: 3
 AcceleratorCudaInit[0]:   pciDeviceID: 0 
 AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
 AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+local rank 1 device 0 bus id: 0000:44:00.0
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
+local rank 0 device 0 bus id: 0000:03:00.0
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
+local rank 0 device 0 bus id: 0000:03:00.0
+AcceleratorCudaInit: ================================================
+AcceleratorCudaInit: ================================================
+local rank 2 device 0 bus id: 0000:84:00.0
 SharedMemoryMpi:  World communicator of size 16
 SharedMemoryMpi:  Node  communicator of size 4
-0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x7fcd80000000 for comms buffers 
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x153960000000 for comms buffers 
 Setting up IPC
 
 __|__|__|__|__|__|__|__|__|__|__|__|__|__|__
@@ -116,7 +128,7 @@ 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.
-Current Grid git commit hash=9d2238148c56e3fbadfa95dcabf2b83d4bde14cd: (HEAD -> develop) uncommited changes
+Current Grid git commit hash=da06d15f73184ceb15d66d4e7e702b02fed7b940: (HEAD -> feature/dirichlet, develop) uncommited changes
 
 Grid : Message : ================================================ 
 Grid : Message : MPI is initialised and logging filters activated 
@@ -124,122 +136,102 @@ Grid : Message : ================================================
 Grid : Message : Requested 2147483648 byte stencil comms buffers 
 Grid : Message : MemoryManager Cache 34004218675 bytes 
 Grid : Message : MemoryManager::Init() setting up
-Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 32 LARGE 8
+Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
 Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
 Grid : Message : MemoryManager::Init() Using cudaMalloc
-Grid : Message : 1.198523 s : Grid Layout
-Grid : Message : 1.198530 s : 	Global lattice size  : 64 64 64 64 
-Grid : Message : 1.198534 s : 	OpenMP threads       : 4
-Grid : Message : 1.198535 s : 	MPI tasks            : 2 2 2 2 
-Grid : Message : 1.397615 s : Making s innermost grids
-Grid : Message : 1.441828 s : Initialising 4d RNG
-Grid : Message : 1.547973 s : Intialising parallel RNG with unique string 'The 4D RNG'
-Grid : Message : 1.547998 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
-Grid : Message : 1.954777 s : Initialising 5d RNG
-Grid : Message : 3.633825 s : Intialising parallel RNG with unique string 'The 5D RNG'
-Grid : Message : 3.633869 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
-Grid : Message : 12.162710 s : Initialised RNGs
-Grid : Message : 15.882520 s : Drawing gauge field
-Grid : Message : 15.816362 s : Random gauge initialised 
-Grid : Message : 17.279671 s : Setting up Cshift based reference 
-Grid : Message : 26.331426 s : *****************************************************************
-Grid : Message : 26.331452 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
-Grid : Message : 26.331454 s : *****************************************************************
-Grid : Message : 26.331456 s : *****************************************************************
-Grid : Message : 26.331458 s : * Benchmarking DomainWallFermionR::Dhop                  
-Grid : Message : 26.331459 s : * Vectorising space-time by 8
-Grid : Message : 26.331463 s : * VComplexF size is 64 B
-Grid : Message : 26.331465 s : * SINGLE precision 
-Grid : Message : 26.331467 s : * Using Overlapped Comms/Compute
-Grid : Message : 26.331468 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 26.331469 s : *****************************************************************
-Grid : Message : 28.413717 s : Called warmup
-Grid : Message : 56.418423 s : Called Dw 3000 times in 2.80047e+07 us
-Grid : Message : 56.418476 s : mflop/s =   3.79581e+07
-Grid : Message : 56.418479 s : mflop/s per rank =  2.37238e+06
-Grid : Message : 56.418481 s : mflop/s per node =  9.48953e+06
-Grid : Message : 56.418483 s : RF  GiB/s (base 2) =   77130
-Grid : Message : 56.418485 s : mem GiB/s (base 2) =   48206.3
-Grid : Message : 56.422076 s : norm diff   1.03481e-13
-Grid : Message : 56.456894 s : #### Dhop calls report 
-Grid : Message : 56.456899 s : WilsonFermion5D Number of DhopEO Calls   : 6002
-Grid : Message : 56.456903 s : WilsonFermion5D TotalTime   /Calls        : 4710.93 us
-Grid : Message : 56.456905 s : WilsonFermion5D CommTime    /Calls        : 3196.15 us
-Grid : Message : 56.456908 s : WilsonFermion5D FaceTime    /Calls        : 494.392 us
-Grid : Message : 56.456910 s : WilsonFermion5D ComputeTime1/Calls        : 44.4107 us
-Grid : Message : 56.456912 s : WilsonFermion5D ComputeTime2/Calls        : 1037.75 us
-Grid : Message : 56.456921 s : Average mflops/s per call                : 3.55691e+09
-Grid : Message : 56.456925 s : Average mflops/s per call per rank       : 2.22307e+08
-Grid : Message : 56.456928 s : Average mflops/s per call per node       : 8.89228e+08
-Grid : Message : 56.456930 s : Average mflops/s per call (full)         : 3.82915e+07
-Grid : Message : 56.456933 s : Average mflops/s per call per rank (full): 2.39322e+06
-Grid : Message : 56.456952 s : Average mflops/s per call per node (full): 9.57287e+06
-Grid : Message : 56.456954 s : WilsonFermion5D Stencil
-Grid : Message : 56.457016 s :  Stencil calls 3001
-Grid : Message : 56.457022 s :  Stencil halogtime 0
-Grid : Message : 56.457024 s :  Stencil gathertime 55.9154
-Grid : Message : 56.457026 s :  Stencil gathermtime 20.1073
-Grid : Message : 56.457028 s :  Stencil mergetime 18.5585
-Grid : Message : 56.457030 s :  Stencil decompresstime 0.0639787
-Grid : Message : 56.457032 s :  Stencil comms_bytes 4.02653e+08
-Grid : Message : 56.457034 s :  Stencil commtime 6379.93
-Grid : Message : 56.457036 s :  Stencil 63.1124 GB/s per rank
-Grid : Message : 56.457038 s :  Stencil 252.45 GB/s per node
-Grid : Message : 56.457040 s : WilsonFermion5D StencilEven
-Grid : Message : 56.457048 s : WilsonFermion5D StencilOdd
-Grid : Message : 56.457062 s : WilsonFermion5D Stencil     Reporti()
-Grid : Message : 56.457065 s : WilsonFermion5D StencilEven Reporti()
-Grid : Message : 56.457066 s : WilsonFermion5D StencilOdd  Reporti()
-Grid : Message : 79.259261 s : Compare to naive wilson implementation Dag to verify correctness
-Grid : Message : 79.259287 s : Called DwDag
-Grid : Message : 79.259288 s : norm dag result 12.0421
-Grid : Message : 79.271740 s : norm dag ref    12.0421
-Grid : Message : 79.287759 s : norm dag diff   7.63236e-14
-Grid : Message : 79.328100 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
-Grid : Message : 79.955951 s : src_e0.499997
-Grid : Message : 80.633620 s : src_o0.500003
-Grid : Message : 80.164163 s : *********************************************************
-Grid : Message : 80.164168 s : * Benchmarking DomainWallFermionF::DhopEO                
-Grid : Message : 80.164170 s : * Vectorising space-time by 8
-Grid : Message : 80.164172 s : * SINGLE precision 
-Grid : Message : 80.164174 s : * Using Overlapped Comms/Compute
-Grid : Message : 80.164177 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 80.164178 s : *********************************************************
-Grid : Message : 93.797635 s : Deo mflop/s =   3.93231e+07
-Grid : Message : 93.797670 s : Deo mflop/s per rank   2.45769e+06
-Grid : Message : 93.797672 s : Deo mflop/s per node   9.83077e+06
-Grid : Message : 93.797674 s : #### Dhop calls report 
-Grid : Message : 93.797675 s : WilsonFermion5D Number of DhopEO Calls   : 3001
-Grid : Message : 93.797677 s : WilsonFermion5D TotalTime   /Calls        : 4542.83 us
-Grid : Message : 93.797679 s : WilsonFermion5D CommTime    /Calls        : 2978.97 us
-Grid : Message : 93.797681 s : WilsonFermion5D FaceTime    /Calls        : 602.287 us
-Grid : Message : 93.797683 s : WilsonFermion5D ComputeTime1/Calls        : 67.1416 us
-Grid : Message : 93.797685 s : WilsonFermion5D ComputeTime2/Calls        : 1004.07 us
-Grid : Message : 93.797713 s : Average mflops/s per call                : 3.30731e+09
-Grid : Message : 93.797717 s : Average mflops/s per call per rank       : 2.06707e+08
-Grid : Message : 93.797719 s : Average mflops/s per call per node       : 8.26827e+08
-Grid : Message : 93.797721 s : Average mflops/s per call (full)         : 3.97084e+07
-Grid : Message : 93.797727 s : Average mflops/s per call per rank (full): 2.48178e+06
-Grid : Message : 93.797732 s : Average mflops/s per call per node (full): 9.92711e+06
-Grid : Message : 93.797735 s : WilsonFermion5D Stencil
-Grid : Message : 93.797746 s : WilsonFermion5D StencilEven
-Grid : Message : 93.797758 s : WilsonFermion5D StencilOdd
-Grid : Message : 93.797769 s :  Stencil calls 3001
-Grid : Message : 93.797773 s :  Stencil halogtime 0
-Grid : Message : 93.797776 s :  Stencil gathertime 56.7458
-Grid : Message : 93.797780 s :  Stencil gathermtime 22.6504
-Grid : Message : 93.797782 s :  Stencil mergetime 21.1913
-Grid : Message : 93.797786 s :  Stencil decompresstime 0.0556481
-Grid : Message : 93.797788 s :  Stencil comms_bytes 2.01327e+08
-Grid : Message : 93.797791 s :  Stencil commtime 2989.33
-Grid : Message : 93.797795 s :  Stencil 67.3484 GB/s per rank
-Grid : Message : 93.797798 s :  Stencil 269.394 GB/s per node
-Grid : Message : 93.797801 s : WilsonFermion5D Stencil     Reporti()
-Grid : Message : 93.797803 s : WilsonFermion5D StencilEven Reporti()
-Grid : Message : 93.797805 s : WilsonFermion5D StencilOdd  Reporti()
-Grid : Message : 93.873429 s : r_e6.02111
-Grid : Message : 93.879931 s : r_o6.02102
-Grid : Message : 93.885912 s : res12.0421
-Grid : Message : 94.876555 s : norm diff   0
-Grid : Message : 95.485643 s : norm diff even  0
-Grid : Message : 95.581236 s : norm diff odd   0
+Grid : Message : 1.875883 s : Grid Layout
+Grid : Message : 1.875893 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 1.875897 s : 	OpenMP threads       : 4
+Grid : Message : 1.875898 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 1.993571 s : Initialising 4d RNG
+Grid : Message : 2.881990 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 2.882370 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 2.495044 s : Initialising 5d RNG
+Grid : Message : 4.120900 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 4.121350 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 15.268010 s : Drawing gauge field
+Grid : Message : 16.234025 s : Random gauge initialised 
+Grid : Message : 16.234057 s : Applying BCs 
+Grid : Message : 16.365565 s : Setting up Cshift based reference 
+Grid : Message : 44.512418 s : *****************************************************************
+Grid : Message : 44.512448 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 44.512450 s : *****************************************************************
+Grid : Message : 44.512451 s : *****************************************************************
+Grid : Message : 44.512452 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 44.512453 s : * Vectorising space-time by 8
+Grid : Message : 44.512454 s : * VComplexF size is 64 B
+Grid : Message : 44.512456 s : * SINGLE precision 
+Grid : Message : 44.512459 s : * Using Overlapped Comms/Compute
+Grid : Message : 44.512460 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 44.512461 s : *****************************************************************
+Grid : Message : 46.389070 s : Called warmup
+Grid : Message : 49.211265 s : Called Dw 300 times in 2.82203e+06 us
+Grid : Message : 49.211295 s : mflop/s =   3.76681e+07
+Grid : Message : 49.211297 s : mflop/s per rank =  2.35425e+06
+Grid : Message : 49.211299 s : mflop/s per node =  9.41702e+06
+Grid : Message : 49.211301 s : RF  GiB/s (base 2) =   76540.6
+Grid : Message : 49.211308 s : mem GiB/s (base 2) =   47837.9
+Grid : Message : 49.214868 s : norm diff   1.06409e-13
+Grid : Message : 92.647781 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 92.647816 s : Called DwDag
+Grid : Message : 92.647817 s : norm dag result 12.0421
+Grid : Message : 92.801806 s : norm dag ref    12.0421
+Grid : Message : 92.817724 s : norm dag diff   7.21921e-14
+Grid : Message : 92.858973 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 93.210378 s : src_e0.499997
+Grid : Message : 93.583286 s : src_o0.500003
+Grid : Message : 93.682468 s : *********************************************************
+Grid : Message : 93.682471 s : * Benchmarking DomainWallFermionF::DhopEO                
+Grid : Message : 93.682472 s : * Vectorising space-time by 8
+Grid : Message : 93.682473 s : * SINGLE precision 
+Grid : Message : 93.682475 s : * Using Overlapped Comms/Compute
+Grid : Message : 93.682476 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 93.682477 s : *********************************************************
+Grid : Message : 95.162342 s : Deo mflop/s =   3.92487e+07
+Grid : Message : 95.162387 s : Deo mflop/s per rank   2.45305e+06
+Grid : Message : 95.162389 s : Deo mflop/s per node   9.81219e+06
+Grid : Message : 95.232801 s : r_e6.02111
+Grid : Message : 95.240061 s : r_o6.02102
+Grid : Message : 95.245975 s : res12.0421
+Grid : Message : 95.833402 s : norm diff   0
+Grid : Message : 96.573829 s : norm diff even  0
+Grid : Message : 96.868272 s : norm diff odd   0
+ Dirichlet block [0 64 64 32 32]
+Grid : Message : 97.756909 s : Grid Layout
+Grid : Message : 97.756911 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 97.756921 s : 	OpenMP threads       : 4
+Grid : Message : 97.756922 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 97.897085 s : Initialising 4d RNG
+Grid : Message : 97.965061 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 97.965097 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 98.367431 s : Initialising 5d RNG
+Grid : Message : 99.752745 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 99.752790 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 111.290148 s : Drawing gauge field
+Grid : Message : 112.349289 s : Random gauge initialised 
+Grid : Message : 112.349320 s : Applying BCs 
+Grid : Message : 113.948740 s : Setting up Cshift based reference 
+Grid : Message : 140.320415 s : *****************************************************************
+Grid : Message : 140.320443 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 140.320444 s : *****************************************************************
+Grid : Message : 140.320445 s : *****************************************************************
+Grid : Message : 140.320446 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 140.320447 s : * Vectorising space-time by 8
+Grid : Message : 140.320448 s : * VComplexF size is 64 B
+Grid : Message : 140.320450 s : * SINGLE precision 
+Grid : Message : 140.320451 s : * Using Overlapped Comms/Compute
+Grid : Message : 140.320452 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 140.320453 s : *****************************************************************
+Grid : Message : 142.296150 s : Called warmup
+Grid : Message : 144.397678 s : Called Dw 300 times in 2.36719e+06 us
+Grid : Message : 144.397700 s : mflop/s =   4.49058e+07
+Grid : Message : 144.397702 s : mflop/s per rank =  2.80661e+06
+Grid : Message : 144.397704 s : mflop/s per node =  1.12265e+07
+Grid : Message : 144.397706 s : RF  GiB/s (base 2) =   91247.6
+Grid : Message : 144.397708 s : mem GiB/s (base 2) =   57029.7
+Grid : Message : 144.401269 s : norm diff   9.78944e-14
+Grid : Message : 186.885460 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 186.885492 s : Called DwDag
+Grid : Message : 186.885493 s : norm dag result 10.4157
+Grid : Message : 186.897154 s : norm dag ref    11.2266
+Grid : Message : 186.912538 s : norm dag diff   0.484633

systems/Tursa/dwf4.slurm

@@ -1,14 +1,13 @@
 #!/bin/bash
 #SBATCH -J dslash
-#SBATCH -A tc002
-#SBATCH -t 2:20:00
-#SBATCH --nodelist=tu-c0r0n[00,03,06,09]
+#SBATCH -A dp207
 #SBATCH --exclusive
 #SBATCH --nodes=4
 #SBATCH --ntasks=16
+#SBATCH --qos=standard
 #SBATCH --ntasks-per-node=4
 #SBATCH --cpus-per-task=8
-#SBATCH --time=12:00:00
+#SBATCH --time=0:05:00
 #SBATCH --partition=gpu
 #SBATCH --gres=gpu:4
 #SBATCH --output=%x.%j.out

systems/Tursa/sourceme.sh

@@ -1,2 +1,6 @@
-spack load c-lime
-module load cuda/11.4.1 openmpi/4.1.1 ucx/1.10.1
+module load cuda/11.4.1  openmpi/4.1.1-cuda11.4.1  ucx/1.12.0-cuda11.4.1  
+#module load cuda/11.4.1 openmpi/4.1.1 ucx/1.10.1
+export PREFIX=/home/tc002/tc002/shared/env/prefix/
+export LD_LIBRARY_PATH=$PREFIX/lib/:$LD_LIBRARY_PATH
+unset SBATCH_EXPORT
+

systems/mac-arm/config-command-mpi

@@ -0,0 +1 @@
+CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi

tests/core/Test_compact_wilson_clover_speedup.cc

@@ -0,0 +1,226 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./tests/core/Test_compact_wilson_clover_speedup.cc
+
+    Copyright (C) 2020 - 2022
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Nils Meyer       <nils.meyer@ur.de>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Grid.h>
+
+using namespace Grid;
+
+NAMESPACE_BEGIN(CommandlineHelpers);
+
+static bool checkPresent(int* argc, char*** argv, const std::string& option) {
+  return GridCmdOptionExists(*argv, *argv + *argc, option);
+}
+
+static std::string getContent(int* argc, char*** argv, const std::string& option) {
+  return GridCmdOptionPayload(*argv, *argv + *argc, option);
+}
+
+static int readInt(int* argc, char*** argv, std::string&& option, int defaultValue) {
+  std::string arg;
+  int         ret = defaultValue;
+  if(checkPresent(argc, argv, option)) {
+    arg = getContent(argc, argv, option);
+    GridCmdOptionInt(arg, ret);
+  }
+  return ret;
+}
+
+static float readFloat(int* argc, char*** argv, std::string&& option, float defaultValue) {
+  std::string arg;
+  float       ret = defaultValue;
+  if(checkPresent(argc, argv, option)) {
+    arg = getContent(argc, argv, option);
+    GridCmdOptionFloat(arg, ret);
+  }
+  return ret;
+}
+
+NAMESPACE_END(CommandlineHelpers);
+
+
+#define _grid_printf(LOGGER, ...) \
+  { \
+    if((LOGGER).isActive()) { /* this makes it safe to put, e.g., norm2 in the calling code w.r.t. performance */ \
+      char _printf_buf[1024]; \
+      std::sprintf(_printf_buf, __VA_ARGS__); \
+      std::cout << (LOGGER) << _printf_buf; \
+      fflush(stdout); \
+    } \
+  }
+#define grid_printf_msg(...) _grid_printf(GridLogMessage, __VA_ARGS__)
+
+
+template<typename Field>
+bool resultsAgree(const Field& ref, const Field& res, const std::string& name) {
+  RealD checkTolerance = (getPrecision<Field>::value == 2) ? 1e-15 : 1e-7;
+  Field diff(ref.Grid());
+  diff = ref - res;
+  auto absDev = norm2(diff);
+  auto relDev = absDev / norm2(ref);
+  std::cout << GridLogMessage
+            << "norm2(reference), norm2(" << name << "), abs. deviation, rel. deviation: " << norm2(ref) << " "
+            << norm2(res) << " " << absDev << " " << relDev << " -> check "
+            << ((relDev < checkTolerance) ? "passed" : "failed") << std::endl;
+
+  return relDev <= checkTolerance;
+}
+
+
+template<typename vCoeff_t>
+void runBenchmark(int* argc, char*** argv) {
+  // read from command line
+  const int   nIter        = CommandlineHelpers::readInt(     argc, argv, "--niter", 1000);
+  const RealD mass         = CommandlineHelpers::readFloat(   argc, argv, "--mass",  0.5);
+  const RealD csw          = CommandlineHelpers::readFloat(   argc, argv, "--csw",   1.0);
+  const RealD cF           = CommandlineHelpers::readFloat(   argc, argv, "--cF",    1.0);
+  const bool  antiPeriodic = CommandlineHelpers::checkPresent(argc, argv, "--antiperiodic");
+
+  // precision
+  static_assert(getPrecision<vCoeff_t>::value == 2 || getPrecision<vCoeff_t>::value == 1, "Incorrect precision"); // double or single
+  std::string precision = (getPrecision<vCoeff_t>::value == 2 ? "double" : "single");
+
+  // setup grids
+  GridCartesian*         UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vCoeff_t::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  // clang-format on
+
+  // setup rng
+  std::vector<int> seeds({1, 2, 3, 4});
+  GridParallelRNG  pRNG(UGrid);
+  pRNG.SeedFixedIntegers(seeds);
+
+  // type definitions
+  typedef WilsonImpl<vCoeff_t, FundamentalRepresentation, CoeffReal> WImpl;
+  typedef WilsonCloverFermion<WImpl> WilsonCloverOperator;
+  typedef CompactWilsonCloverFermion<WImpl> CompactWilsonCloverOperator;
+  typedef typename WilsonCloverOperator::FermionField Fermion;
+  typedef typename WilsonCloverOperator::GaugeField Gauge;
+
+  // setup fields
+  Fermion src(UGrid); random(pRNG, src);
+  Fermion ref(UGrid); ref = Zero();
+  Fermion res(UGrid); res = Zero();
+  Fermion hop(UGrid); hop = Zero();
+  Fermion diff(UGrid); diff = Zero();
+  Gauge   Umu(UGrid); SU3::HotConfiguration(pRNG, Umu);
+
+  // setup boundary phases
+  typename WilsonCloverOperator::ImplParams implParams;
+  std::vector<Complex> boundary_phases(Nd, 1.);
+  if(antiPeriodic) boundary_phases[Nd-1] = -1.;
+  implParams.boundary_phases = boundary_phases;
+  WilsonAnisotropyCoefficients anisParams;
+
+  // misc stuff needed for benchmarks
+  double volume=1.0; for(int mu=0; mu<Nd; mu++) volume*=UGrid->_fdimensions[mu];
+
+  // setup fermion operators
+  WilsonCloverOperator        Dwc(        Umu, *UGrid, *UrbGrid, mass, csw, csw,     anisParams, implParams);
+  CompactWilsonCloverOperator Dwc_compact(Umu, *UGrid, *UrbGrid, mass, csw, csw, cF, anisParams, implParams);
+
+  // now test the conversions
+  typename CompactWilsonCloverOperator::CloverField         tmp_ref(UGrid);  tmp_ref  = Dwc.CloverTerm;
+  typename CompactWilsonCloverOperator::CloverField         tmp_res(UGrid);  tmp_res  = Zero();
+  typename CompactWilsonCloverOperator::CloverField         tmp_diff(UGrid); tmp_diff = Zero();
+  typename CompactWilsonCloverOperator::CloverDiagonalField diagonal(UGrid); diagonal = Zero();
+  typename CompactWilsonCloverOperator::CloverTriangleField triangle(UGrid); diagonal = Zero();
+  CompactWilsonCloverOperator::CompactHelpers::ConvertLayout(tmp_ref, diagonal, triangle);
+  CompactWilsonCloverOperator::CompactHelpers::ConvertLayout(diagonal, triangle, tmp_res);
+  tmp_diff = tmp_ref - tmp_res;
+  std::cout << GridLogMessage << "conversion: ref, res, diff, eps"
+            << " " << norm2(tmp_ref)
+            << " " << norm2(tmp_res)
+            << " " << norm2(tmp_diff)
+            << " " << norm2(tmp_diff) / norm2(tmp_ref)
+            << std::endl;
+
+  // performance per site (use minimal values necessary)
+  double hop_flop_per_site            = 1320; // Rich's Talk + what Peter uses
+  double hop_byte_per_site            = (8 * 9 + 9 * 12) * 2 * getPrecision<vCoeff_t>::value * 4;
+  double clov_flop_per_site           = 504; // Rich's Talk and 1412.2629
+  double clov_byte_per_site           = (2 * 18 + 12 + 12) * 2 * getPrecision<vCoeff_t>::value * 4;
+  double clov_flop_per_site_performed = 1128;
+  double clov_byte_per_site_performed = (12 * 12 + 12 + 12) * 2 * getPrecision<vCoeff_t>::value * 4;
+
+  // total performance numbers
+  double hop_gflop_total            = volume * nIter * hop_flop_per_site / 1e9;
+  double hop_gbyte_total            = volume * nIter * hop_byte_per_site / 1e9;
+  double clov_gflop_total           = volume * nIter * clov_flop_per_site / 1e9;
+  double clov_gbyte_total           = volume * nIter * clov_byte_per_site / 1e9;
+  double clov_gflop_performed_total = volume * nIter * clov_flop_per_site_performed / 1e9;
+  double clov_gbyte_performed_total = volume * nIter * clov_byte_per_site_performed / 1e9;
+
+  // warmup + measure dhop
+  for(auto n : {1, 2, 3, 4, 5}) Dwc.Dhop(src, hop, 0);
+  double t0 = usecond();
+  for(int n = 0; n < nIter; n++) Dwc.Dhop(src, hop, 0);
+  double t1 = usecond();
+  double secs_hop = (t1-t0)/1e6;
+  grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n",
+              "hop", precision.c_str(), secs_hop, hop_gflop_total/secs_hop, hop_gbyte_total/secs_hop, 0.0, secs_hop/secs_hop);
+
+#define BENCH_CLOVER_KERNEL(KERNEL) { \
+  /* warmup + measure reference clover */ \
+  for(auto n : {1, 2, 3, 4, 5}) Dwc.KERNEL(src, ref); \
+  double t2 = usecond(); \
+  for(int n = 0; n < nIter; n++) Dwc.KERNEL(src, ref); \
+  double t3 = usecond(); \
+  double secs_ref = (t3-t2)/1e6; \
+  grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n", \
+                  "reference_"#KERNEL, precision.c_str(), secs_ref, clov_gflop_total/secs_ref, clov_gbyte_total/secs_ref, secs_ref/secs_ref, secs_ref/secs_hop); \
+  grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n", /* to see how well the ET performs */  \
+                  "reference_"#KERNEL"_performed", precision.c_str(), secs_ref, clov_gflop_performed_total/secs_ref, clov_gbyte_performed_total/secs_ref, secs_ref/secs_ref, secs_ref/secs_hop); \
+\
+  /* warmup + measure compact clover */ \
+  for(auto n : {1, 2, 3, 4, 5}) Dwc_compact.KERNEL(src, res); \
+  double t4 = usecond(); \
+  for(int n = 0; n < nIter; n++) Dwc_compact.KERNEL(src, res); \
+  double t5 = usecond(); \
+  double secs_res = (t5-t4)/1e6; \
+  grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n", \
+                  "compact_"#KERNEL, precision.c_str(), secs_res, clov_gflop_total/secs_res, clov_gbyte_total/secs_res, secs_ref/secs_res, secs_res/secs_hop); \
+  assert(resultsAgree(ref, res, #KERNEL)); \
+}
+
+  BENCH_CLOVER_KERNEL(Mooee);
+  BENCH_CLOVER_KERNEL(MooeeDag);
+  BENCH_CLOVER_KERNEL(MooeeInv);
+  BENCH_CLOVER_KERNEL(MooeeInvDag);
+
+  grid_printf_msg("finalize %s\n", precision.c_str());
+}
+
+int main(int argc, char** argv) {
+  Grid_init(&argc, &argv);
+
+  runBenchmark<vComplexD>(&argc, &argv);
+  runBenchmark<vComplexF>(&argc, &argv);
+
+  Grid_finalize();
+}