Dirichlet first cut - wrong answers on dagger multiply.

Struggling to get a compute node so changing systems

Grid/allocator/MemoryManager.h

@@ -114,11 +114,6 @@ private:
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
  
-  static uint64_t     DeviceAccesses;
-  static uint64_t     HostAccesses;
-  static uint64_t     DeviceAccessBytes;
-  static uint64_t     HostAccessBytes;
- 
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////
@@ -157,7 +152,6 @@ private:
 
   //  static void  LRUupdate(AcceleratorViewEntry &AccCache);
   static void  LRUinsert(AcceleratorViewEntry &AccCache);
-  static void  LRUinsertback(AcceleratorViewEntry &AccCache);
   static void  LRUremove(AcceleratorViewEntry &AccCache);
   
   // manage entries in the table

Grid/allocator/MemoryManagerCache.cc

@@ -23,11 +23,6 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
-uint64_t  MemoryManager::DeviceAccesses;
-uint64_t  MemoryManager::HostAccesses;
-uint64_t  MemoryManager::DeviceAccessBytes;
-uint64_t  MemoryManager::HostAccessBytes;
- 
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -91,14 +86,6 @@ void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
   AccCache.LRU_valid = 1;
   DeviceLRUBytes+=AccCache.bytes;
 }
-void  MemoryManager::LRUinsertback(AcceleratorViewEntry &AccCache)
-{
-  assert(AccCache.LRU_valid==0);
-  LRU.push_back(AccCache.CpuPtr);
-  AccCache.LRU_entry = --LRU.end();
-  AccCache.LRU_valid = 1;
-  DeviceLRUBytes+=AccCache.bytes;
-}
 void  MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
 {
   assert(AccCache.LRU_valid==1);
@@ -142,7 +129,6 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
-  
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
@@ -245,9 +231,6 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
     EntryCreate(CpuPtr,bytes,mode,hint);
   }
 
-  DeviceAccesses++;
-  DeviceAccessBytes+=bytes;
-
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
   if (!AccCache.AccPtr) {
@@ -366,10 +349,6 @@ void MemoryManager::CpuViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock==0);
 
   AccCache.cpuLock--;
-
-  if(AccCache.cpuLock==0) {
-    LRUinsertback(AccCache);
-  }
 }
 /*
  *  Action  State   StateNext         Flush    Clone
@@ -392,9 +371,6 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
     EntryCreate(CpuPtr,bytes,mode,transient);
   }
 
-  HostAccesses++;
-  HostAccessBytes+=bytes;
-  
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
@@ -440,12 +416,6 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
 
   AccCache.transient= transient? EvictNext : 0;
 
-  // If view is opened on host remove from LRU
-  // Host close says evict next from device
-  if(AccCache.LRU_valid==1){
-    LRUremove(AccCache);
-  }
-  
   return AccCache.CpuPtr;
 }
 void  MemoryManager::NotifyDeletion(void *_ptr)

Grid/allocator/MemoryManagerShared.cc

@@ -12,10 +12,6 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
-uint64_t  MemoryManager::DeviceAccesses;
-uint64_t  MemoryManager::HostAccesses;
-uint64_t  MemoryManager::DeviceAccessBytes;
-uint64_t  MemoryManager::HostAccessBytes;
 
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };

Grid/communicator/Communicator_base.h

@@ -53,10 +53,11 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
+  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processor_coor;  // linear processor coordinate
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -97,6 +98,7 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
   int                ProcessorCount(void)    ;
 
@@ -142,16 +144,16 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,
+			       int xmit_to_rank,int do_xmit,
 			       void *recv,
-			       int recv_from_rank,
+			       int recv_from_rank,int do_recv,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,
+				    int xmit_to_rank,int do_xmit,
 				    void *recv,
-				    int recv_from_rank,
+				    int recv_from_rank,int do_recv,
 				    int bytes,int dir);
   
   

Grid/communicator/Communicator_mpi3.cc

@@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
   // Remap using the shared memory optimising routine
   // The remap creates a comm which must be freed
   ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,12 +124,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
-
+  Coordinate shm_processors       (_ndimension,1);
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
   for(int d=0;d<parent_ndimension;d++){
     parent_processor_coor[pad+d]=parent._processor_coor[d];
     parent_processors    [pad+d]=parent._processors[d];
+    shm_processors       [pad+d]=parent._shm_processors[d];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -154,6 +155,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     ccoor[d] = parent_processor_coor[d] % processors[d];
     scoor[d] = parent_processor_coor[d] / processors[d];
     ssize[d] = parent_processors[d]     / processors[d];
+    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@@ -335,22 +337,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest,
+						     int dest, int dox,
 						     void *recv,
-						     int from,
+						     int from, int dor,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,
+							 int dest,int dox,
 							 void *recv,
-							 int from,
+							 int from,int dor,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -370,6 +372,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
+  if ( 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);
@@ -377,7 +380,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(rrq);
       off_node_bytes+=bytes;
     }
+  }
   
+  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);
@@ -391,6 +396,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
     }
+  }
   
   if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
     this->StencilSendToRecvFromComplete(list,dir);

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/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/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 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 int Nsite = grid->oSites();
+    const int 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

@@ -174,6 +174,11 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
 
+  void DirichletBlock(std::vector<int> & block){
+    Stencil.DirichletBlock(block); 
+    StencilEven.DirichletBlock(block); 
+    StencilOdd.DirichletBlock(block); 
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

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/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/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
   ////////////////////////////////////////
@@ -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],
+      _grid->StencilSendToRecvFromBegin(reqs[i],
 					Packets[i].send_buf,
-						     Packets[i].to_rank,
+					Packets[i].to_rank,Packets[i].do_send,
 					Packets[i].recv_buf,
-						     Packets[i].from_rank,
+					Packets[i].from_rank,Packets[i].do_recv,
 					Packets[i].bytes,i);
-      comms_bytes+=bytes;
-      shm_bytes  +=2*Packets[i].bytes-bytes;
     }
   }
 
@@ -413,7 +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();
+      /////////////////////////////////////////////////////////
+      // several way threaded on different communicators.
+      // Cannot combine with Dirichlet operators
+      // This scheme is needed on Intel Omnipath for best performance
+      // Deprecate once there are very few omnipath clusters
+      /////////////////////////////////////////////////////////
       int nthreads = CartesianCommunicator::nCommThreads;
-	assert(nthreads <= maxthreads);
-	if (nthreads == -1) nthreads = 1;
-	if (mythread < nthreads) {
-	  for (int i = mythread; i < Packets.size(); i += nthreads) {
-	    double start = usecond();
-	    uint64_t bytes= _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-							 Packets[i].to_rank,
+      int old = GridThread::GetThreads();
+      GridThread::SetThreads(nthreads);
+      thread_for(i,Packets.size(),{
+	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
+				       Packets[i].to_rank,Packets[i].do_send,
 				       Packets[i].recv_buf,
-							 Packets[i].from_rank,
+				       Packets[i].from_rank,Packets[i].do_recv,
 				       Packets[i].bytes,i);
-	    comm_bytes_thr[mythread] += bytes;
-	    shm_bytes_thr[mythread]  += Packets[i].bytes - bytes;
-	    comm_time_thr[mythread]  += usecond() - start;
-	  }
-	}
-      }
-    } else { // Concurrent and non-threaded asynch calls to MPI
+      });
+      GridThread::SetThreads(old);
+    } else { 
+      /////////////////////////////////////////////////////////
+      // Concurrent and non-threaded asynch calls to MPI
+      /////////////////////////////////////////////////////////
       std::vector<std::vector<CommsRequest_t> > reqs;
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
@@ -484,31 +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 std::vector<int> &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;
 
@@ -1045,16 +1094,20 @@ 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++;
+
+	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[u_comm_offset],0,bytes,cbmask);
+	if ( (!duplicate) ) { // Force comms for now
 
 	  ///////////////////////////////////////////////////////////
 	  // Build a list of things to do after we synchronise GPUs
@@ -1062,11 +1115,12 @@ public:
 	  ///////////////////////////////////////////////////////////
 	  AddPacket((void *)&send_buf[u_comm_offset],
 		    (void *)&recv_buf[u_comm_offset],
-		  xmit_to_rank,
-		  recv_from_rank,
+		    xmit_to_rank, comms_send,
+		    recv_from_rank, comms_recv,
 		    bytes);
+	}
 	
-	if ( compress.DecompressionStep() ) {
+	if ( compress.DecompressionStep() && comms_recv ) {
 	  AddDecompress(&this->u_recv_buf_p[u_comm_offset],
 			&recv_buf[u_comm_offset],
 			words,Decompressions);
@@ -1078,11 +1132,15 @@ public:
   }
 
   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) ) { // Force comms for now
+	      AddPacket((void *)sp,(void *)rp,
+			xmit_to_rank,comms_send,
+			recv_from_rank,comms_recv,
+			bytes);
+	    }
 
 	  } else {
 
@@ -1191,7 +1253,9 @@ public:
 	  }
 	}
 
+	if ( comms_recv ) {
 	  AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
+	}
 
 	u_comm_offset     +=buffer_size;
       }

Grid/threads/Accelerator.h

@@ -342,7 +342,7 @@ 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
@@ -356,19 +356,41 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
       { __VA_ARGS__;}							\
     };									\
     int nt=acceleratorThreads();					\
-    dim3 hip_threads(nt,1,nsimd);					\
+    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);
   }

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,

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,

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_dwf_fp32.cc

@@ -32,10 +32,103 @@
 using namespace std;
 using namespace Grid;
 
-template<class d>
-struct scal {
-  d internal;
+////////////////////////
+/// Move to domains ////
+////////////////////////
+
+struct DomainDecomposition
+{
+  Coordinate Block;
+
+  DomainDecomposition(const Coordinate &_Block): Block(_Block){ assert(Block.size()==Nd);};
+  
+  template<class Field>
+  void ProjectDomain(Field &f,Integer domain)
+  {
+    GridBase *grid = f.Grid();
+    int dims = grid->Nd();
+    int isDWF= (dims==Nd+1);
+    assert((dims==Nd)||(dims==Nd+1));
+
+    Field   zz(grid);  zz = Zero();
+    LatticeInteger coor(grid);
+    LatticeInteger domaincoor(grid);
+    LatticeInteger mask(grid); mask = Integer(1);
+    LatticeInteger zi(grid);     zi = Integer(0);
+    for(int d=0;d<Nd;d++){
+      Integer B= Block[d];
+      if ( B ) {
+	LatticeCoordinate(coor,d+isDWF);
+	domaincoor = mod(coor,B);
+	mask = where(domaincoor==Integer(0),zi,mask);
+	mask = where(domaincoor==Integer(B-1),zi,mask);
+      }
+    }
+    if ( !domain )
+      f = where(mask==Integer(1),f,zz);
+    else 
+      f = where(mask==Integer(0),f,zz);
   };
+};
+
+template<typename MomentaField>
+struct DirichletFilter: public MomentumFilterBase<MomentaField>
+{
+  Coordinate Block;
+  
+  DirichletFilter(const Coordinate &_Block): Block(_Block) {}
+
+  // Edge detect using domain projectors
+  void applyFilter (MomentaField &U) const override
+  {
+    DomainDecomposition Domains(Block);
+    GridBase *grid = U.Grid();
+    LatticeInteger  coor(grid);
+    LatticeInteger  face(grid);
+    LatticeInteger  one(grid);   one = 1;
+    LatticeInteger  zero(grid); zero = 0;
+    LatticeInteger  omega(grid);
+    LatticeInteger  omegabar(grid);
+    LatticeInteger  tmp(grid);
+
+    omega=one;    Domains.ProjectDomain(omega,0);
+    omegabar=one; Domains.ProjectDomain(omegabar,1);
+    
+    LatticeInteger nface(grid); nface=Zero();
+    
+    MomentaField projected(grid); projected=Zero();
+    typedef decltype(PeekIndex<LorentzIndex>(U,0)) MomentaLinkField;
+    MomentaLinkField  Umu(grid);
+    MomentaLinkField   zz(grid); zz=Zero();
+
+    int dims = grid->Nd();
+    Coordinate Global=grid->GlobalDimensions();
+    assert(dims==Nd);
+
+    for(int mu=0;mu<Nd;mu++){
+
+      if ( Block[mu]!=0 ) {
+
+	Umu = PeekIndex<LorentzIndex>(U,mu);
+
+	// Upper face 
+ 	tmp = Cshift(omegabar,mu,1);
+	tmp = tmp + omega;
+	face = where(tmp == Integer(2),one,zero );
+
+ 	tmp = Cshift(omega,mu,1);
+	tmp = tmp + omegabar;
+	face = where(tmp == Integer(2),one,face );
+
+	Umu = where(face,zz,Umu);
+
+	PokeIndex<LorentzIndex>(U, Umu, mu);
+      }
+    }
+  }
+};
+
+
 
 Gamma::Algebra Gmu [] = {
 			 Gamma::Algebra::GammaX,
@@ -44,6 +137,7 @@ struct scal {
 			 Gamma::Algebra::GammaT
 };
 
+void Benchmark(int Ls, std::vector<int> Dirichlet);
 
 int main (int argc, char ** argv)
 {
@@ -52,24 +146,48 @@ 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;
     }
+  }
+  std::vector<int> Dirichlet(5,0);
+  Benchmark(Ls,Dirichlet);
+  Coordinate latt4  = GridDefaultLatt();
+  Coordinate mpi    = GridDefaultMpi();
+  Coordinate shm;
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  /*
+  Dirichlet = std::vector<int>({0,
+				latt4[0]/mpi[0] * shm[0],
+				latt4[1]/mpi[1] * shm[1],
+				latt4[2]/mpi[2] * shm[2],
+				latt4[3]/mpi[3] * shm[3]});
+  */
+  Dirichlet = std::vector<int>({0,
+				latt4[0]/mpi[0] ,
+				latt4[1]/mpi[1] ,
+				latt4[2]/mpi[2] ,
+				latt4[3]/mpi[3] });
   
+  std::cout << " Dirichlet block "<< Dirichlet<< std::endl;
+  Benchmark(Ls,Dirichlet);
+  Grid_finalize();
+  exit(0);
+}
+void Benchmark(int Ls, std::vector<int> 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);
@@ -80,26 +198,13 @@ 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
-  src = Zero();
-  {
-    Coordinate origin({0,0,0,latt4[2]-1,0});
-    SpinColourVectorF tmp;
-    tmp=Zero();
-    tmp()(0)(0)=Complex(-2.0,0.0);
-    std::cout << " source site 0 " << tmp<<std::endl;
-    pokeSite(tmp,src,origin);
-  }
-#else
   RealD N2 = 1.0/::sqrt(norm2(src));
   src = src*N2;
-#endif
-
 
   LatticeFermionF result(FGrid); result=Zero();
   LatticeFermionF    ref(FGrid);    ref=Zero();
@@ -110,17 +215,17 @@ int main (int argc, char ** argv)
   LatticeGaugeFieldF Umu(UGrid);
   SU<Nc>::HotConfiguration(RNG4,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
+  ////////////////////////////////////
+  std::cout << GridLogMessage << "Applying BCs " << std::endl;
+  Coordinate Block(4);
+  for(int d=0;d<4;d++)  Block[d]= Dirichlet[d+1];
+
+  std::cout << GridLogMessage << "Dirichlet Block " << Block<< std::endl;
+  DirichletFilter<LatticeGaugeFieldF> Filter(Block);
+  Filter.applyFilter(Umu);
   
   ////////////////////////////////////
   // Naive wilson implementation
@@ -191,11 +296,11 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
   DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  Dw.DirichletBlock(Dirichlet);
   int ncall =300;
 
   if (1) {
     FGrid->Barrier();
-    Dw.ZeroCounters();
     Dw.Dhop(src,result,0);
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
@@ -220,8 +325,6 @@ 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;
@@ -229,20 +332,13 @@ int main (int argc, char ** argv)
     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)
@@ -294,13 +390,14 @@ int main (int argc, char ** argv)
   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;
-*/
+
+  if (  norm2(err) > 1.0e-4 ) {
+    std::cout << "Error vector is\n" <<err << std::endl;
+    std::cout << "Ref   vector is\n" <<ref << std::endl;
+    std::cout << "Result  vector is\n" <<result << std::endl;
   }
+  assert((norm2(err)<1.0e-4));
+  
   LatticeFermionF src_e (FrbGrid);
   LatticeFermionF src_o (FrbGrid);
   LatticeFermionF r_e   (FrbGrid);
@@ -330,7 +427,6 @@ 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();
@@ -352,7 +448,6 @@ 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);
@@ -367,13 +462,7 @@ 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);
@@ -382,6 +471,4 @@ int main (int argc, char ** argv)
 
   assert(norm2(src_e)<1.0e-4);
   assert(norm2(src_o)<1.0e-4);
-  Grid_finalize();
-  exit(0);
 }

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 -p ecp
+#SBATCH -J DWF
+#SBATCH -o DWF.%J
+#SBATCH -e DWF.%J
+#SBATCH -N 1
+#SBATCH -n 1
+#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=1
+
+AT=8
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+
+PARAMS=" --accelerator-threads ${AT} --grid 24.24.24.24 --shm-mpi 0 --mpi 1.1.1.1"
+
+srun --gpus-per-task 1 -n1 ./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 +1,27 @@
+#!/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 8
+#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=1
+
+echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads 8 --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+
+srun --gpus-per-task 1 -n8 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+
+

systems/Crusher/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/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/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

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();
+}