FermToProp

Compact Exponential Cloverterm on GPU

Grid/algorithms/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
         });

Grid/algorithms/approx/Chebyshev.h

@@ -264,7 +264,7 @@ public:
       auto Tnp_v = Tnp->View();
       auto Tnm_v = Tnm->View();
       constexpr int Nsimd = vector_type::Nsimd();
-      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
+      accelerator_for(ss, in.Grid()->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
       });

Grid/communicator/Communicator_base.h

@@ -53,11 +53,10 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  int              _processor;       // linear processor rank
-  unsigned long    _ndimension;
-  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
   Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  int              _processor;       // linear processor rank
   Coordinate _processor_coor;  // linear processor coordinate
+  unsigned long    _ndimension;
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -98,9 +97,8 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
-  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                ProcessorCount(void)    ;
+  int                      ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -144,16 +142,16 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,int do_xmit,
+			       int xmit_to_rank,
 			       void *recv,
-			       int recv_from_rank,int do_recv,
+			       int recv_from_rank,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,int do_xmit,
+				    int xmit_to_rank,
 				    void *recv,
-				    int recv_from_rank,int do_recv,
+				    int recv_from_rank,
 				    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,_shm_processors);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,13 +124,12 @@ 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];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -155,7 +154,6 @@ 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
@@ -337,22 +335,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest, int dox,
+						     int dest,
 						     void *recv,
-						     int from, int dor,
+						     int from,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,int dox,
+							 int dest,
 							 void *recv,
-							 int from,int dor,
+							 int from,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -372,35 +370,32 @@ 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);
-      assert(ierr==0);
-      list.push_back(rrq);
-      off_node_bytes+=bytes;
-    }
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+    assert(ierr==0);
+    list.push_back(rrq);
+    off_node_bytes+=bytes;
   }
-  
-  if (dor) {
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
-      list.push_back(xrq);
-      off_node_bytes+=bytes;
-    } else {
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
-    }
-  }
-  
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-    this->StencilSendToRecvFromComplete(list,dir);
-    list.resize(0);
+
+  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+_processor*32;
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+    assert(ierr==0);
+    list.push_back(xrq);
+    off_node_bytes+=bytes;
+  } else {
+    // TODO : make a OMP loop on CPU, call threaded bcopy
+    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+    assert(shm!=NULL);
+    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
+    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
   }
 
+  //  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  //    this->StencilSendToRecvFromComplete(list,dir);
+  //  }
+
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)

Grid/communicator/Communicator_none.cc

@@ -45,14 +45,12 @@ 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);
@@ -113,18 +111,18 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,int dox,
+						     int xmit_to_rank,
 						     void *recv,
-						     int recv_from_rank,int dor,
+						     int recv_from_rank,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,int dox,
+							 int xmit_to_rank,
 							 void *recv,
-							 int recv_from_rank,int dor,
+							 int recv_from_rank,
 							 int bytes, int dir)
 {
   return 2.0*bytes;

Grid/communicator/SharedMemory.h

@@ -93,10 +93,9 @@ 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
-  // 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 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
   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,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   //////////////////////////////////////////////////////////////////////////////
   // 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,SHM);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
 }
 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,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -294,8 +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
   ////////////////////////////////////////////////////////////////
@@ -342,7 +341,7 @@ 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,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -354,8 +353,6 @@ 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

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

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_crc.h

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

Grid/lattice/Lattice_reduction.h

@@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
+#if defined(GRID_SYCL)
+#include <Grid/lattice/Lattice_reduction_sycl.h>
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -127,7 +130,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -136,7 +139,7 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -145,7 +148,7 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu_large(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -155,13 +158,13 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu(&arg_v[0],osites);
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  typename vobj::scalar_object ssum;
+  autoView( arg_v, arg, AcceleratorRead);
+  ssum= sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
   auto ssum= sum_cpu(&arg_v[0],osites);
 #endif  
   arg.Grid()->GlobalSum(ssum);
@@ -171,7 +174,7 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 template<class vobj>
 inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
   auto ssum= sum_gpu_large(&arg_v[0],osites);
@@ -235,11 +238,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-    
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
-
+    // This code could read coalesce
     // GPU - SIMT lane compliance...
     accelerator_for( ss, sites, 1,{
 	auto x_l = left_v[ss];

Grid/lattice/Lattice_reduction_sycl.h

@@ -0,0 +1,125 @@
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Possibly promote to double and sum
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_objectD sobjD;
+  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
+  sobj identity; zeroit(identity);
+  sobj ret ; 
+
+  Integer nsimd= vobj::Nsimd();
+  
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{osites},
+		      Reduction,
+		      [=] (cl::sycl::id<1> item, auto &sum) {
+      auto osite   = item[0];
+      sum +=Reduce(lat[osite]);
+     });
+   });
+  theGridAccelerator->wait();
+  ret = mysum[0];
+  free(mysum,*theGridAccelerator);
+  sobjD dret; convertType(dret,ret);
+  return dret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_tensor(lat,osites);
+}
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Return as same precision as input performing reduction in double precision though
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu(lat,osites);
+  return result;
+}
+
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu_large(lat,osites);
+  return result;
+}
+
+NAMESPACE_END(Grid);
+
+/*
+template<class Double> Double svm_reduce(Double *vec,uint64_t L)
+{
+  Double sumResult; zeroit(sumResult);
+  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
+  Double identity;  zeroit(identity);
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{L},
+		      Reduction,
+		      [=] (cl::sycl::id<1> index, auto &sum) {
+	 sum +=vec[index];
+     });
+   });
+  theGridAccelerator->wait();
+  Double ret = d_sum[0];
+  free(d_sum,*theGridAccelerator);
+  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
+  return ret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_type  scalar;
+
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobjD;
+
+  sobjD ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int nsimd = vobj::Nsimd();
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<scalar> buffer(osites*nsimd);
+  scalar *buf = &buffer[0];
+  vector *dat = (vector *)lat;
+
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,nsimd,{
+	int lane = acceleratorSIMTlane(nsimd);
+	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
+    });
+    //Precision change at this point is to late to gain precision
+    ret_p[w] = svm_reduce(buf,nsimd*osites);
+  }
+  return ret;
+}
+*/

Grid/log/Log.cc

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

Grid/log/Log.h

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

Grid/qcd/QCD.h

@@ -451,9 +451,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
+#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,AcceleratorWrite);
+  autoView(f_v,f,AcceleratorRead);
+  accelerator_for(idx,p_v.oSites(),1,{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -465,12 +476,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
+#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,AcceleratorWrite);
+  autoView(f_v,f,AcceleratorRead);
+  accelerator_for(idx,p_v.oSites(),1,{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -482,6 +504,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/ActionBase.h

@@ -40,29 +40,6 @@ class Action
 
 public:
   bool is_smeared = false;
-  RealD deriv_norm_sum;
-  RealD deriv_max_sum;
-  int   deriv_num;
-  RealD deriv_us;
-  RealD S_us;
-  RealD refresh_us;
-  void  reset_timer(void)        {
-    deriv_us = S_us = refresh_us = 0.0;
-    deriv_num=0;
-    deriv_norm_sum = deriv_max_sum=0.0;
-  }
-  void  deriv_log(RealD nrm, RealD max) { deriv_max_sum+=max; deriv_norm_sum+=nrm; deriv_num++;}
-  RealD deriv_max_average(void)         { return deriv_max_sum/deriv_num; };
-  RealD deriv_norm_average(void)        { return deriv_norm_sum/deriv_num; };
-  RealD deriv_timer(void)        { return deriv_us; };
-  RealD S_timer(void)            { return deriv_us; };
-  RealD refresh_timer(void)      { return deriv_us; };
-  void deriv_timer_start(void)   { deriv_us-=usecond(); }
-  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
-  void refresh_timer_start(void) { refresh_us-=usecond(); }
-  void refresh_timer_stop(void)  { refresh_us+=usecond(); }
-  void S_timer_start(void)       { S_us-=usecond(); }
-  void S_timer_stop(void)        { S_us+=usecond(); }
   // Heatbath?
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action

Grid/qcd/action/ActionCore.h

@@ -37,10 +37,6 @@ NAMESPACE_CHECK(ActionSet);
 #include <Grid/qcd/action/ActionParams.h>
 NAMESPACE_CHECK(ActionParams);
 
-#include <Grid/qcd/action/filters/MomentumFilter.h>
-#include <Grid/qcd/action/filters/DirichletFilter.h>
-#include <Grid/qcd/action/filters/DDHMCFilter.h>
-
 ////////////////////////////////////////////
 // Gauge Actions
 ////////////////////////////////////////////

Grid/qcd/action/ActionParams.h

@@ -37,32 +37,24 @@ NAMESPACE_BEGIN(Grid);
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
   Coordinate twists;
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  GparityWilsonImplParams() : twists(Nd, 0), dirichlet(Nd, 0) {};
+  GparityWilsonImplParams() : twists(Nd, 0) {};
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
-    dirichlet.resize(Nd,0);
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
-    dirichlet.resize(Nd,0);
   }
 };
 
 struct StaggeredImplParams {
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  StaggeredImplParams()
-  {
-    dirichlet.resize(Nd,0);
-  };
+  StaggeredImplParams()  {};
 };
   
   struct OneFlavourRationalParams : Serializable {

Grid/qcd/action/fermion/CloverHelpers.h

@@ -204,15 +204,18 @@ public:
   typedef WilsonCloverHelpers<Impl> Helpers;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
     Clover += diag_mass;
   }
 
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
-                          CloverTriangleField& Triangle,
-                          RealD csw_t, RealD diag_mass) {
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
 
-    // Do nothing
+    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
   }
 
   // TODO: implement Cmunu for better performances with compact layout, but don't do it
@@ -237,9 +240,17 @@ public:
   template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-    // do nothing!
-    // mass term is multiplied to exp(Clover) below
+  // Can this be avoided?
+  static void IdentityTimesC(const CloverField& in, RealD c) {
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
   }
 
   static int getNMAX(RealD prec, RealD R) {
@@ -254,175 +265,62 @@ public:
     return NMAX;
   }
 
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
-  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
 
-  	  typedef iMatrix<ComplexD,6> mat;
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
 
-  	  RealD qn[6];
-  	  RealD qnold[6];
-  	  RealD p[5];
-  	  RealD trA2, trA3, trA4;
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
 
-  	  mat A2, A3, A4, A5;
-  	  A2 = alpha * alpha * arg * arg;
-  	  A3 = alpha * arg * A2;
-  	  A4 = A2 * A2;
-  	  A5 = A2 * A3;
+    Clover *= (1.0/diag_mass);
 
-  	  trA2 = toReal( trace(A2) );
-  	  trA3 = toReal( trace(A3) );
-  	  trA4 = toReal( trace(A4));
-
-  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
-  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
-  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
-  	  p[3] = trA3 / 3.0;
-  	  p[4] = 0.5 * trA2;
-
-  	  qnold[0] = cN[Niter];
-  	  qnold[1] = 0.0;
-  	  qnold[2] = 0.0;
-  	  qnold[3] = 0.0;
-  	  qnold[4] = 0.0;
-  	  qnold[5] = 0.0;
-
-  	  for(int i = Niter-1; i >= 0; i--)
-  	  {
-  	   qn[0] = p[0] * qnold[5] + cN[i];
-  	   qn[1] = p[1] * qnold[5] + qnold[0];
-  	   qn[2] = p[2] * qnold[5] + qnold[1];
-  	   qn[3] = p[3] * qnold[5] + qnold[2];
-  	   qn[4] = p[4] * qnold[5] + qnold[3];
-  	   qn[5] = qnold[4];
-
-  	   qnold[0] = qn[0];
-  	   qnold[1] = qn[1];
-  	   qnold[2] = qn[2];
-  	   qnold[3] = qn[3];
-  	   qnold[4] = qn[4];
-  	   qnold[5] = qn[5];
-  	  }
-
-  	  mat unit(1.0);
-
-  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
-
-    }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Diagonal.Grid();
-    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
-
-    //
-    // Implementation completely in Daniel's layout
-    //
-
-    // Taylor expansion with Cayley-Hamilton recursion
-    // underlying Horner scheme as above
+    // Taylor expansion, slow but generic
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+    // qN = cN
+    // qn = cn + qn+1 X
     std::vector<RealD> cn(NMAX+1);
     cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++){
+    for (int i=1; i<=NMAX; i++)
       cn[i] = cn[i-1] / RealD(i);
-    }
 
-      // Taken over from Daniel's implementation
-      conformable(Diagonal, Triangle);
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
 
-      long lsites = grid->lSites();
-    {
-      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
-      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-      typedef iMatrix<ComplexD,6> mat;
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
 
-      autoView(diagonal_v,  Diagonal,  CpuRead);
-      autoView(triangle_v,  Triangle,  CpuRead);
-      autoView(diagonalExp_v, Diagonal, CpuWrite);
-      autoView(triangleExp_v, Triangle, CpuWrite);
+    Clover = ExpClover * diag_mass;
 
-      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
 
-    	  mat srcCloverOpUL(0.0); // upper left block
-    	  mat srcCloverOpLR(0.0); // lower right block
-    	  mat ExpCloverOp;
+    CloverInv = ExpClover * (1.0/diag_mass);
 
-        scalar_object_diagonal diagonal_tmp     = Zero();
-        scalar_object_diagonal diagonal_exp_tmp = Zero();
-        scalar_object_triangle triangle_tmp     = Zero();
-        scalar_object_triangle triangle_exp_tmp = Zero();
-
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-
-        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-        peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-        int block;
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-        		if (i == j){
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-        		}
-        		else{
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-        		}
-        	}
-        }
-        block = 1;
-        for(int i = 0; i < 6; i++){
-          	for(int j = 0; j < 6; j++){
-           		if (i == j){
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-           		}
-           		else{
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-           		}
-            }
-        }
-
-        // exp(Clover)
-
-        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-            	if (i == j){
-            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-            	}
-            	else if(i < j){
-            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-            	}
-           	}
-        }
-
-        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 1;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-              	if (i == j){
-              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-               	}
-               	else if(i < j){
-               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-               	}
-            }
-        }
-
-        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
-        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
-      });
-    }
-
-    Diagonal *= diag_mass;
-    Triangle *= diag_mass;
   }
 
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
+
+    if (fixedBoundaries)
+    {
+      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+    }
+    else
+    {
+      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
+    }
+  }
 
   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
     assert(0);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -225,7 +225,7 @@ public:
   RealD csw_t;
   RealD cF;
 
-  bool open_boundaries;
+  bool fixedBoundaries;
 
   CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
   CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;

Grid/qcd/action/fermion/FermionOperator.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -117,19 +117,19 @@ public:
     typedef decltype(coalescedRead(*in))    sobj;
     typedef decltype(coalescedRead(*out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
-    const vobj *vp1 = &in[m];
-    const vobj *vp = (lane&mask) ? vp1:vp0;
-    auto sa = coalescedRead(*vp,j0);
-    auto sb = coalescedRead(*vp,j1);
+    const vobj *vp0 = &in[k];  // out0[j] = merge low bit of type from in[k] and in[m] 
+    const vobj *vp1 = &in[m];  // out1[j] = merge hi  bit of type from in[k] and in[m]
+    const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
+    auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
+    auto sb = coalescedRead(*vp,j1); // lane to read for out 1
     hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);
-    projector::Proj(psb,sb,mu,dag);
+    projector::Proj(psa,sa,mu,dag);  // spin project the result0
+    projector::Proj(psb,sb,mu,dag);  // spin project the result1
     coalescedWrite(out0[j],psa);
     coalescedWrite(out1[j],psb);
 #else

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,10 +75,6 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
-  int Dirichlet;
-  Coordinate Block; 
-
-  /********** Deprecate timers **********/
   void Report(void);
   void ZeroCounters(void);
   double DhopCalls;
@@ -177,18 +173,7 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-
-  virtual void DirichletBlock(Coordinate & block)
-  {
-    assert(block.size()==Nd+1);
-    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
-      Dirichlet = 1;
-      Block = block;
-      Stencil.DirichletBlock(block); 
-      StencilEven.DirichletBlock(block); 
-      StencilOdd.DirichletBlock(block);
-    }
-  }
+    
   // Constructors
   /*
     WilsonFermion5D(int simd, 

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

@@ -48,7 +48,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , csw_r(_csw_r)
   , csw_t(_csw_t)
   , cF(_cF)
-  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
   , Diagonal(&Fgrid),        Triangle(&Fgrid)
   , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
   , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
@@ -67,7 +67,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries) {
+  if (fixedBoundaries) {
     this->BoundaryMaskEven.Checkerboard() = Even;
     this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
@@ -77,31 +77,31 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->open_boundaries) ApplyBoundaryMask(out);
+  if(this->fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
-  if(this->open_boundaries) {
+  if(this->fixedBoundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
@@ -112,7 +112,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in,
   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);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -121,19 +121,19 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& i
   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);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -147,7 +147,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -166,7 +166,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -186,7 +186,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!open_boundaries); // TODO check for changes required for open bc
+  assert(!fixedBoundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -305,6 +305,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   GridBase* grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
   CloverField TmpOriginal(grid);
+  CloverField TmpInverse(grid);
 
   // Compute the field strength terms mu>nu
   double t2 = usecond();
@@ -324,24 +325,27 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  // Handle mass term based on clover policy
-  CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
-  
-  // Convert the data layout of the clover term
+
+  // Instantiate the clover term
+  // - In case of the standard clover the mass term is added
+  // - In case of the exponential clover the clover term is exponentiated
   double t4 = usecond();
+  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
+
+  // Convert the data layout of the clover term
+  double t5 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Exponentiate the clover (nothing happens in case of the standard clover)
-  double t5 = usecond();
-  CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
-
-  // Possible modify the boundary values
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
   double t6 = usecond();
-  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
+  // Invert the Clover term
+  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
+  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
+  // TODO: For now this inversion is explictly done on the CPU
   double t7 = usecond();
-  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
 
   // Fill the remaining clover fields
   double t8 = usecond();
@@ -362,10 +366,10 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
   std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
   std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "convert =                    " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "exponentiation =             " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "boundaries =                 " << (t7 - t6) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "inversions =                 " << (t8 - t7) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
   std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
   std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
 }

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

@@ -60,8 +60,7 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   UmuOdd (_FourDimRedBlackGrid),
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid),
-  _tmp(&FiveDimRedBlackGrid),
-  Dirichlet(0)
+  _tmp(&FiveDimRedBlackGrid)
 {
   // some assertions
   assert(FiveDimGrid._ndimension==5);
@@ -219,14 +218,6 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
-  if ( Dirichlet ) {
-    std::cout << GridLogMessage << " Dirichlet BCs 5d " <<Block<<std::endl;
-    Coordinate GaugeBlock(Nd);
-    for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
-    std::cout << GridLogMessage << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
-    DirichletFilter<GaugeField> Filter(GaugeBlock);
-    Filter.applyFilter(HUmu);
-  }
   Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
   pickCheckerboard(Even,UmuEven,Umu);
   pickCheckerboard(Odd ,UmuOdd,Umu);

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

@@ -498,6 +498,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
+     acceleratorFenceComputeStream();
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
@@ -505,11 +506,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
+     acceleratorFenceComputeStream();
    }
    assert(0 && " Kernel optimisation case not covered ");
   }

Grid/qcd/action/filters/DDHMCFilter.h

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

Grid/qcd/action/filters/DirichletFilter.h

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

Grid/qcd/action/gauge/Photon.h

@@ -49,7 +49,7 @@ NAMESPACE_BEGIN(Grid);
     
     typedef Lattice<SiteLink>  LinkField;
     typedef Lattice<SiteField> Field;
-    typedef Field              ComplexField;
+    typedef LinkField          ComplexField;
   };
   
   typedef QedGImpl<vComplex> QedGImplR;

Grid/qcd/hmc/GenericHMCrunner.h

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

Grid/qcd/hmc/HMC.h

@@ -34,7 +34,6 @@ directory
 			    * @brief Classes for Hybrid Monte Carlo update
 			    *
 			    * @author Guido Cossu
-			    * @author Peter Boyle
 			    */
 			   //--------------------------------------------------------------------
 #pragma once
@@ -116,17 +115,22 @@ private:
 
     random(sRNG, rn_test);
 
-    std::cout << GridLogHMC << "--------------------------------------------------\n";
-    std::cout << GridLogHMC << "exp(-dH) = " << prob << "  Random = " << rn_test << "\n";
-    std::cout << GridLogHMC << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
+    std::cout << GridLogMessage
+              << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "exp(-dH) = " << prob
+              << "  Random = " << rn_test << "\n";
+    std::cout << GridLogMessage
+              << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
 
     if ((prob > 1.0) || (rn_test <= prob)) {  // accepted
-      std::cout << GridLogHMC << "Metropolis_test -- ACCEPTED\n";
-      std::cout << GridLogHMC << "--------------------------------------------------\n";
+      std::cout << GridLogMessage << "Metropolis_test -- ACCEPTED\n";
+      std::cout << GridLogMessage
+                << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
-      std::cout << GridLogHMC << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogHMC << "--------------------------------------------------\n";
+      std::cout << GridLogMessage << "Metropolis_test -- REJECTED\n";
+      std::cout << GridLogMessage
+                << "--------------------------------------------------\n";
       return false;
     }
   }
@@ -135,68 +139,19 @@ private:
   // Evolution
   /////////////////////////////////////////////////////////
   RealD evolve_hmc_step(Field &U) {
+    TheIntegrator.refresh(U, sRNG, pRNG);  // set U and initialize P and phi's
 
-    GridBase *Grid = U.Grid();
-
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // Mainly for DDHMC perform a random translation of U modulo volume
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Random shifting gauge field by [";
-    for(int d=0;d<Grid->Nd();d++) {
-
-      int L = Grid->GlobalDimensions()[d];
-
-      RealD rn_uniform;  random(sRNG, rn_uniform);
-
-      int shift = (int) (rn_uniform*L);
-
-      std::cout << shift;
-      if(d<Grid->Nd()-1) std::cout <<",";
-      else               std::cout <<"]\n";
-      
-      U = Cshift(U,d,shift);
-    }
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-
-    TheIntegrator.reset_timer();
-    
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // set U and initialize P and phi's
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Refresh momenta and pseudofermions";
-    TheIntegrator.refresh(U, sRNG, pRNG);  
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // initial state action
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Compute initial action";
-    RealD H0 = TheIntegrator.S(U);  
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    RealD H0 = TheIntegrator.S(U);  // initial state action
 
     std::streamsize current_precision = std::cout.precision();
     std::cout.precision(15);
-    std::cout << GridLogHMC << "Total H before trajectory = " << H0 << "\n";
+    std::cout << GridLogMessage << "Total H before trajectory = " << H0 << "\n";
     std::cout.precision(current_precision);
 
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << " Molecular Dynamics evolution ";
     TheIntegrator.integrate(U);
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // updated state action
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Compute final action";
-    RealD H1 = TheIntegrator.S(U);  
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    RealD H1 = TheIntegrator.S(U);  // updated state action
 
-
-    
     ///////////////////////////////////////////////////////////
     if(0){
       std::cout << "------------------------- Reversibility test" << std::endl;
@@ -208,16 +163,17 @@ private:
     }
     ///////////////////////////////////////////////////////////
 
+
     std::cout.precision(15);
-
-    std::cout << GridLogHMC << "--------------------------------------------------\n";
-    std::cout << GridLogHMC << "Total H after trajectory  = " << H1 << "  dH = " << H1 - H0 << "\n";
-    std::cout << GridLogHMC << "--------------------------------------------------\n";
-
+    std::cout << GridLogMessage << "Total H after trajectory  = " << H1
+	      << "  dH = " << H1 - H0 << "\n";
     std::cout.precision(current_precision);
     
     return (H1 - H0);
   }
+  
+
+  
 
 public:
   /////////////////////////////////////////
@@ -239,13 +195,10 @@ public:
 
     // Actual updates (evolve a copy Ucopy then copy back eventually)
     unsigned int FinalTrajectory = Params.Trajectories + Params.NoMetropolisUntil + Params.StartTrajectory;
-
     for (int traj = Params.StartTrajectory; traj < FinalTrajectory; ++traj) {
-
-      std::cout << GridLogHMC << "-- # Trajectory = " << traj << "\n";
-
+      std::cout << GridLogMessage << "-- # Trajectory = " << traj << "\n";
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
-      	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
+      	std::cout << GridLogMessage << "-- Thermalization" << std::endl;
       }
       
       double t0=usecond();
@@ -254,19 +207,20 @@ public:
       DeltaH = evolve_hmc_step(Ucopy);
       // Metropolis-Hastings test
       bool accept = true;
-      if (Params.MetropolisTest && traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
+      if (traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
         accept = metropolis_test(DeltaH);
       } else {
-      	std::cout << GridLogHMC << "Skipping Metropolis test" << std::endl;
+      	std::cout << GridLogMessage << "Skipping Metropolis test" << std::endl;
       }
 
       if (accept)
         Ucur = Ucopy; 
       
+     
+      
       double t1=usecond();
-      std::cout << GridLogHMC << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
+      std::cout << GridLogMessage << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
 
-      TheIntegrator.print_timer();
 
       for (int obs = 0; obs < Observables.size(); obs++) {
       	std::cout << GridLogDebug << "Observables # " << obs << std::endl;
@@ -274,7 +228,7 @@ public:
       	std::cout << GridLogDebug << "Observables pointer " << Observables[obs] << std::endl;
         Observables[obs]->TrajectoryComplete(traj + 1, Ucur, sRNG, pRNG);
       }
-      std::cout << GridLogHMC << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
+      std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
     }
   }
 

Grid/qcd/hmc/HMCResourceManager.h

@@ -72,8 +72,6 @@ class HMCResourceManager {
   typedef HMCModuleBase< BaseHmcCheckpointer<ImplementationPolicy> > CheckpointerBaseModule;
   typedef HMCModuleBase< HmcObservable<typename ImplementationPolicy::Field> > ObservableBaseModule;
   typedef ActionModuleBase< Action<typename ImplementationPolicy::Field>, GridModule > ActionBaseModule;
-  typedef typename ImplementationPolicy::Field MomentaField;
-  typedef typename ImplementationPolicy::Field Field;  
 
   // Named storage for grid pairs (std + red-black)
   std::unordered_map<std::string, GridModule> Grids;
@@ -82,9 +80,6 @@ class HMCResourceManager {
   // SmearingModule<ImplementationPolicy> Smearing;
   std::unique_ptr<CheckpointerBaseModule> CP;
 
-  // Momentum filter
-  std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> > Filter;
-  
   // A vector of HmcObservable modules
   std::vector<std::unique_ptr<ObservableBaseModule> > ObservablesList;
 
@@ -95,7 +90,6 @@ class HMCResourceManager {
 
   bool have_RNG;
   bool have_CheckPointer;
-  bool have_Filter;
 
   // NOTE: operator << is not overloaded for std::vector<string> 
   // so this function is necessary
@@ -107,7 +101,7 @@ class HMCResourceManager {
 
 
 public:
-  HMCResourceManager() : have_RNG(false), have_CheckPointer(false), have_Filter(false) {}
+  HMCResourceManager() : have_RNG(false), have_CheckPointer(false) {}
 
   template <class ReaderClass, class vector_type = vComplex >
   void initialize(ReaderClass &Read){
@@ -135,7 +129,6 @@ public:
     RNGModuleParameters RNGpar(Read);
     SetRNGSeeds(RNGpar);
 
-  
     // Observables
     auto &ObsFactory = HMC_ObservablesModuleFactory<observable_string, typename ImplementationPolicy::Field, ReaderClass>::getInstance(); 
     Read.push(observable_string);// here must check if existing...
@@ -215,16 +208,6 @@ public:
     AddGrid(s, Mod);
   }
 
-  void SetMomentumFilter( MomentumFilterBase<typename ImplementationPolicy::Field> * MomFilter) {
-    assert(have_Filter==false);
-    Filter = std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> >(MomFilter);
-    have_Filter = true;
-  }
-  MomentumFilterBase<typename ImplementationPolicy::Field> *GetMomentumFilter(void) {
-    if ( !have_Filter)
-      SetMomentumFilter(new MomentumFilterNone<typename ImplementationPolicy::Field>());
-    return Filter.get();
-  }
 
   GridCartesian* GetCartesian(std::string s = "") {
     if (s.empty()) s = Grids.begin()->first;
@@ -243,9 +226,6 @@ public:
   //////////////////////////////////////////////////////
   // Random number generators
   //////////////////////////////////////////////////////
-  
-  //Return true if the RNG objects have been instantiated
-  bool haveRNGs() const{ return have_RNG; }
 
   void AddRNGs(std::string s = "") {
     // Couple the RNGs to the GridModule tagged by s

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,6 +33,7 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
+#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -66,7 +67,6 @@ public:
 template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
-
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 
@@ -119,58 +119,36 @@ protected:
     }
   } update_P_hireps{};
 
- 
   void update_P(MomentaField& Mom, Field& U, int level, double ep) {
     // input U actually not used in the fundamental case
     // Fundamental updates, include smearing
 
     for (int a = 0; a < as[level].actions.size(); ++a) {
-
       double start_full = usecond();
       Field force(U.Grid());
       conformable(U.Grid(), Mom.Grid());
 
       Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
-      as[level].actions.at(a)->deriv_timer_start();
       as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
-      as[level].actions.at(a)->deriv_timer_stop();
 
       std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
-      auto name = as[level].actions.at(a)->action_name();
       if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
-
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
-
-      MomFilter->applyFilter(force);
-      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<< std::endl;
-      DumpSliceNorm("force ",force,Nd-1);
-      
-      Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
-      Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      Real force_max   = std::sqrt(maxLocalNorm2(force));
-      Real impulse_max = force_max * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      as[level].actions.at(a)->deriv_log(force_abs,force_max);
-      
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force average: " << force_abs <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force max    : " << force_max <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt average  : " << impulse_abs <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt max      : " << impulse_max <<" "<<name<<std::endl;
-
+      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
+      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << std::endl;
       Mom -= force * ep* HMC_MOMENTUM_DENOMINATOR;; 
       double end_full = usecond();
       double time_full  = (end_full - start_full) / 1e3;
       double time_force = (end_force - start_force) / 1e3;
       std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)"  << std::endl;
-
     }
 
     // Force from the other representations
     as[level].apply(update_P_hireps, Representations, Mom, U, ep);
 
+    MomFilter->applyFilter(Mom);
   }
 
   void update_U(Field& U, double ep) 
@@ -184,12 +162,8 @@ protected:
   
   void update_U(MomentaField& Mom, Field& U, double ep) 
   {
-    MomentaField MomFiltered(Mom.Grid());
-    MomFiltered = Mom;
-    MomFilter->applyFilter(MomFiltered);
-
     // exponential of Mom*U in the gauge fields case
-    FieldImplementation::update_field(MomFiltered, U, ep);
+    FieldImplementation::update_field(Mom, U, ep);
 
     // Update the smeared fields, can be implemented as observer
     Smearer.set_Field(U);
@@ -232,66 +206,6 @@ public:
   const MomentaField & getMomentum() const{ return P; }
   
 
-  void reset_timer(void)
-  {
-    for (int level = 0; level < as.size(); ++level) {
-      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
-        as[level].actions.at(actionID)->reset_timer();
-      }
-    }
-  }
-  void print_timer(void)
-  {
-    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::" << std::endl;
-    std::cout << GridLogMessage << " Refresh cumulative timings "<<std::endl;
-    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
-    for (int level = 0; level < as.size(); ++level) {
-      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
-	std::cout << GridLogMessage 
-		  << as[level].actions.at(actionID)->action_name()
-		  <<"["<<level<<"]["<< actionID<<"] "
-		  << as[level].actions.at(actionID)->refresh_us*1.0e-6<<" s"<< std::endl;
-      }
-    }
-    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
-    std::cout << GridLogMessage << " Action cumulative timings "<<std::endl;
-    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
-    for (int level = 0; level < as.size(); ++level) {
-      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
-	std::cout << GridLogMessage 
-		  << as[level].actions.at(actionID)->action_name()
-		  <<"["<<level<<"]["<< actionID<<"] "
-		  << as[level].actions.at(actionID)->S_us*1.0e-6<<" s"<< std::endl;
-      }
-    }
-    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
-    std::cout << GridLogMessage << " Force cumulative timings "<<std::endl;
-    std::cout << GridLogMessage << "------------------------- "<<std::endl;
-    for (int level = 0; level < as.size(); ++level) {
-      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
-	std::cout << GridLogMessage 
-		  << as[level].actions.at(actionID)->action_name()
-		  <<"["<<level<<"]["<< actionID<<"] "
-		  << as[level].actions.at(actionID)->deriv_us*1.0e-6<<" s"<< std::endl;
-      }
-    }
-    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
-    std::cout << GridLogMessage << " Force average size "<<std::endl;
-    std::cout << GridLogMessage << "------------------------- "<<std::endl;
-    for (int level = 0; level < as.size(); ++level) {
-      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
-	std::cout << GridLogMessage 
-		  << as[level].actions.at(actionID)->action_name()
-		  <<"["<<level<<"]["<< actionID<<"] : "
-		  <<" force max " << as[level].actions.at(actionID)->deriv_max_average()
-		  <<" norm "      << as[level].actions.at(actionID)->deriv_norm_average()
-		  <<" calls "     << as[level].actions.at(actionID)->deriv_num
-		  << std::endl;
-      }
-    }
-    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
-  }
-  
   void print_parameters()
   {
     std::cout << GridLogMessage << "[Integrator] Name : "<< integrator_name() << std::endl;
@@ -310,6 +224,7 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
+
   }
 
   void reverse_momenta()
@@ -352,19 +267,15 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
-	auto name = as[level].actions.at(actionID)->action_name();
-        std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
-
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
-	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
-	as[level].actions.at(actionID)->refresh_timer_stop();
       }
 
       // Refresh the higher representation actions
       as[level].apply(refresh_hireps, Representations, sRNG, pRNG);
     }
 
+    MomFilter->applyFilter(P);
   }
 
   // to be used by the actionlevel class to iterate
@@ -399,9 +310,7 @@ public:
         // based on the boolean is_smeared in actionID
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
-	        as[level].actions.at(actionID)->S_timer_start();
         Hterm = as[level].actions.at(actionID)->S(Us);
-   	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
       }

Grid/serialisation/JSON_IO.cc

@@ -26,7 +26,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef GRID_HIP
 
 NAMESPACE_BEGIN(Grid);
 
@@ -82,7 +82,7 @@ void JSONWriter::writeDefault(const std::string &s,	const std::string &x)
   if (s.size())
     ss_ << "\""<< s << "\" : \"" << os.str() << "\" ," ;
   else
-    ss_ << os.str() << " ," ;
+    ss_ << "\""<< os.str() << "\" ," ;
 }
 
 // Reader implementation ///////////////////////////////////////////////////////

Grid/serialisation/JSON_IO.h

@@ -54,7 +54,7 @@ namespace Grid
     void pop(void);
     template <typename U>
     void writeDefault(const std::string &s, const U &x);
-#ifdef __NVCC__
+#if defined(GRID_CUDA) || defined(GRID_HIP)
     void writeDefault(const std::string &s, const Grid::ComplexD &x) 
     { 
       std::complex<double> z(real(x),imag(x));
@@ -101,7 +101,7 @@ namespace Grid
     void readDefault(const std::string &s, std::vector<U> &output);
     template <typename U, typename P>
     void readDefault(const std::string &s, std::pair<U,P> &output);
-#ifdef __NVCC__
+#if defined(GRID_CUDA) || defined(GRID_HIP)
     void readDefault(const std::string &s, ComplexD &output)
     { 
       std::complex<double> z;

Grid/serialisation/Serialisation.h

@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include "BinaryIO.h"
 #include "TextIO.h"
 #include "XmlIO.h"
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef GRID_HIP
 #include "JSON_IO.h"
 #endif
 

Grid/stencil/SimpleCompressor.h

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

Grid/stencil/Stencil.h

@@ -80,11 +80,14 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
 ///////////////////////////////////////////////////////////////////
 template<class cobj,class vobj,class compressor>
 void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
-				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
+				 commVector<cobj *> pointers,
+				 int dimension,int plane,
+				 int cbmask,compressor &compress,int type) __attribute__((noinline));
 
 template<class cobj,class vobj,class compressor>
-void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
+				 const Lattice<vobj> &rhs,
+				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -92,14 +95,15 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const La
   int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
   auto rhs_v = rhs.View(AcceleratorRead);
+  auto rhs_p = &rhs_v[0];
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
   accelerator_forNB(j, num, vobj::Nsimd(), {
-      compress.CompressExchange(p0,p1, &rhs_v[0], j,
-			      so+tp[2*j  ].second,
-			      so+tp[2*j+1].second,
-			      type);
+      compress.CompressExchange(p0,p1, rhs_p, j,
+				so+tp[2*j  ].second,
+				so+tp[2*j+1].second,
+				type);
   });
   rhs_v.ViewClose();
 }
@@ -133,8 +137,6 @@ class CartesianStencilAccelerator {
   int           _osites;
   StencilVector _directions;
   StencilVector _distances;
-  StencilVector _comms_send;
-  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -228,14 +230,12 @@ public:
     void * recv_buf;
     Integer to_rank;
     Integer from_rank;
-    Integer do_send;
-    Integer do_recv;
     Integer bytes;
   };
   struct Merge {
     cobj * mpointer;
-    Vector<scalar_object *> rpointers;
-    Vector<cobj *> vpointers;
+    //    std::vector<scalar_object *> rpointers;
+    std::vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
   };
@@ -244,20 +244,7 @@ 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;
@@ -288,8 +275,7 @@ 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
   ///////////////////////////////////////////////////////////
@@ -302,6 +288,29 @@ 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
   ////////////////////////////////////////
@@ -328,12 +337,11 @@ 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
@@ -342,30 +350,67 @@ 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++){
-      _grid->StencilSendToRecvFromBegin(reqs[i],
-					Packets[i].send_buf,
-					Packets[i].to_rank,Packets[i].do_send,
-					Packets[i].recv_buf,
-					Packets[i].from_rank,Packets[i].do_recv,
-					Packets[i].bytes,i);
+      uint64_t bytes=_grid->StencilSendToRecvFromBegin(reqs[i],
+						     Packets[i].send_buf,
+						     Packets[i].to_rank,
+						     Packets[i].recv_buf,
+						     Packets[i].from_rank,
+						     Packets[i].bytes,i);
+      comms_bytes+=bytes;
+      shm_bytes  +=2*Packets[i].bytes-bytes;
     }
+    _grid->StencilBarrier();// Synch shared memory on a single nodes
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
@@ -373,34 +418,36 @@ public:
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
+    commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
-      /////////////////////////////////////////////////////////
-      // several way threaded on different communicators.
-      // Cannot combine with Dirichlet operators
-      // This scheme is needed on Intel Omnipath for best performance
-      // Deprecate once there are very few omnipath clusters
-      /////////////////////////////////////////////////////////
-      int nthreads = CartesianCommunicator::nCommThreads;
-      int old = GridThread::GetThreads();
-      GridThread::SetThreads(nthreads);
-      thread_for(i,Packets.size(),{
-	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-				       Packets[i].to_rank,Packets[i].do_send,
-				       Packets[i].recv_buf,
-				       Packets[i].from_rank,Packets[i].do_recv,
-				       Packets[i].bytes,i);
-      });
-      GridThread::SetThreads(old);
-    } else { 
-      /////////////////////////////////////////////////////////
-      // Concurrent and non-threaded asynch calls to MPI
-      /////////////////////////////////////////////////////////
+    if ( 0 ){
+      thread_region {
+	// must be called in parallel region
+	int mythread  = thread_num();
+	int maxthreads= thread_max();
+	int nthreads = CartesianCommunicator::nCommThreads;
+	assert(nthreads <= maxthreads);
+	if (nthreads == -1) nthreads = 1;
+	if (mythread < nthreads) {
+	  for (int i = mythread; i < Packets.size(); i += nthreads) {
+	    double start = usecond();
+	    uint64_t bytes= _grid->StencilSendToRecvFrom(Packets[i].send_buf,
+							 Packets[i].to_rank,
+							 Packets[i].recv_buf,
+							 Packets[i].from_rank,
+							 Packets[i].bytes,i);
+	    comm_bytes_thr[mythread] += bytes;
+	    shm_bytes_thr[mythread]  += Packets[i].bytes - bytes;
+	    comm_time_thr[mythread]  += usecond() - start;
+	  }
+	}
+      }
+    } else { // Concurrent and non-threaded asynch calls to MPI
       std::vector<std::vector<CommsRequest_t> > reqs;
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
@@ -442,23 +489,31 @@ public:
       sshift[1] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Odd);
       if ( sshift[0] == sshift[1] ) {
 	if (splice_dim) {
-	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx,point);
+	  splicetime-=usecond();
+	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
 	  is_same_node = is_same_node && tmp;
+	  splicetime+=usecond();
 	} else {
-	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx,point);
+	  nosplicetime-=usecond();
+	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
 	  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,point);
-	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx,point);
+	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx);
+	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx);
 	  is_same_node = is_same_node && tmp1 && tmp2;
+	  splicetime+=usecond();
 	} else {
-	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx,point);
-	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx,point);
+	  nosplicetime-=usecond();
+	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
+	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
 	  is_same_node = is_same_node && tmp1 && tmp2;
+	  nosplicetime+=usecond();
 	}
       }
     }
@@ -468,10 +523,13 @@ 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;
 
@@ -485,6 +543,7 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
+    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -497,70 +556,14 @@ public:
     Mergers.resize(0);
     MergersSHM.resize(0);
     Packets.resize(0);
-    CopyReceiveBuffers.resize(0);
-    CachedTransfers.resize(0);
+    calls++;
   }
-  void AddCopy(void *from,void * to, Integer bytes)
-  {
-    CopyReceiveBuffer obj;
-    obj.from_p = from;
-    obj.to_p = to;
-    obj.bytes= bytes;
-    CopyReceiveBuffers.push_back(obj);
-  }
-  void CommsCopy()
-  {
-    //    These are device resident MPI buffers.
-    for(int i=0;i<CopyReceiveBuffers.size();i++){
-      cobj *from=(cobj *)CopyReceiveBuffers[i].from_p;
-      cobj *to  =(cobj *)CopyReceiveBuffers[i].to_p;
-      Integer words = CopyReceiveBuffers[i].bytes/sizeof(cobj);
-
-      accelerator_forNB(j, words, cobj::Nsimd(), {
-	  coalescedWrite(to[j] ,coalescedRead(from [j]));
-      });
-    }
-  }
-  
-  Integer CheckForDuplicate(Integer direction, Integer OrthogPlane, Integer DestProc, void *recv_buf,Integer lane,Integer bytes,Integer cb)
-  {
-    CachedTransfer obj;
-    obj.direction   = direction;
-    obj.OrthogPlane = OrthogPlane;
-    obj.DestProc    = DestProc;
-    obj.recv_buf    = recv_buf;
-    obj.lane        = lane;
-    obj.bytes       = bytes;
-    obj.cb          = cb;
-
-    for(int i=0;i<CachedTransfers.size();i++){
-      if (   (CachedTransfers[i].direction  ==direction)
-	   &&(CachedTransfers[i].OrthogPlane==OrthogPlane)
-	   &&(CachedTransfers[i].DestProc   ==DestProc)
-	   &&(CachedTransfers[i].bytes      ==bytes)
-	   &&(CachedTransfers[i].lane       ==lane)
-	   &&(CachedTransfers[i].cb         ==cb)
-	     ){
-
-	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
-	return 1;
-      }
-    }
-
-    CachedTransfers.push_back(obj);
-    return 0;
-  }
-  void AddPacket(void *xmit,void * rcv,
-		 Integer to, Integer do_send,
-		 Integer from, Integer do_recv,
-		 Integer bytes){
+  void AddPacket(void *xmit,void * rcv, Integer to,Integer from,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);
   }
@@ -571,7 +574,7 @@ public:
     d.buffer_size = buffer_size;
     dv.push_back(d);
   }
-  void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
+  void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
     Merge m;
     m.type     = type;
     m.mpointer = merge_p;
@@ -580,17 +583,23 @@ 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();
+    accelerator_barrier();
     _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)
-  {
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
+
+    
+    mergetime-=usecond();
     for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
@@ -600,7 +609,9 @@ 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;
@@ -608,6 +619,7 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
+    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -644,54 +656,18 @@ public:
       }
     }
   }
-  /// Introduce a block structure and switch off comms on boundaries
-  void DirichletBlock(const Coordinate &dirichlet_block)
-  {
-    for(int ii=0;ii<this->_npoints;ii++){
-      int dimension    = this->_directions[ii];
-      int displacement = this->_distances[ii];
-      int gd = _grid->_gdimensions[dimension];
-      int fd = _grid->_fdimensions[dimension];
-      int pd = _grid->_processors [dimension];
-      int pc = _grid->_processor_coor[dimension];
-      int ld = fd/pd;
-      ///////////////////////////////////////////
-      // Figure out dirichlet send and receive
-      // on this leg of stencil.
-      ///////////////////////////////////////////
-      int comm_dim        = _grid->_processors[dimension] >1 ;
-      int block = dirichlet_block[dimension];
-      this->_comms_send[ii] = comm_dim;
-      this->_comms_recv[ii] = comm_dim;
-      if ( block && comm_dim ) {
-	assert(abs(displacement) < ld );
-      
-	if( displacement > 0 ) {
-	  // High side, low side
-	  // | <--B--->|
-	  // |    |    |
-	  //           noR
-	  // noS
-	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_recv[ii] = 0;
-	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_send[ii] = 0;
-	} else {
-	  // High side, low side
-	  // | <--B--->|
-	  // |    |    |
-	  //           noS
-	  // noR
-	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_send[ii] = 0;
-	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_recv[ii] = 0;
-	}
-      }
-    }
-  }
+
   CartesianStencil(GridBase *grid,
 		   int npoints,
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
 		   Parameters p)
+    : shm_bytes_thr(npoints),
+      comm_bytes_thr(npoints),
+      comm_enter_thr(npoints),
+      comm_leave_thr(npoints),
+      comm_time_thr(npoints)
   {
     face_table_computed=0;
     _grid    = grid;
@@ -705,12 +681,8 @@ 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);
 
-    if ( p.dirichlet.size() ) DirichletBlock(p.dirichlet); // comms send/recv set up
-
     _unified_buffer_size=0;
     surface_list.resize(0);
 
@@ -727,16 +699,15 @@ 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);
@@ -745,6 +716,7 @@ public:
       assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
 
       int sshift[2];
+
       //////////////////////////
       // Underlying approach. For each local site build
       // up a table containing the npoint "neighbours" and whether they
@@ -845,7 +817,6 @@ 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];
@@ -902,32 +873,25 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-
-      // Wrap locally dirichlet support case OR node local
-      if ( offnode==0 ) {
+      if (!offnode) {
 
 	int permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-	
+
       } else {
 
-	if ( comms_recv==0 ) {
-
-	  int permute_slice=1;
-	  CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-
-	} else { 
-
-	  ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
-
-	}
-
-      }
-      
-      if ( offnode ) {
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
+
+	//	int rank           = grid->_processor;
+	//	int recv_from_rank;
+	//	int xmit_to_rank;
+
+	int unified_buffer_offset = _unified_buffer_size;
 	_unified_buffer_size    += words;
+
+	ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
+
       }
     }
   }
@@ -1026,14 +990,11 @@ public:
   }
 
   template<class compressor>
-  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx, int point)
+  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
   {
     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());
 
@@ -1056,93 +1017,78 @@ public:
 
       int sx        = (x+sshift)%rd;
       int comm_proc = ((x+sshift)/rd)%pd;
-      
+
       if (comm_proc) {
-	
+
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
 
 	int bytes =  words * compress.CommDatumSize();
 
 	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
-	int comm_off = u_comm_offset;
+	if ( !face_table_computed ) {
+	  face_table.resize(face_idx+1);
+	  std::vector<std::pair<int,int> >  face_table_host ;
+	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
+	  face_table[face_idx].resize(face_table_host.size());
+	  acceleratorCopyToDevice(&face_table_host[0],
+				  &face_table[face_idx][0],
+				  face_table[face_idx].size()*sizeof(face_table_host[0]));
+	}
 
+	//      	int rank           = _grid->_processor;
 	int recv_from_rank;
 	int xmit_to_rank;
-	cobj *recv_buf;
-	cobj *send_buf;
 	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
 	assert (xmit_to_rank   != _grid->ThisRank());
 	assert (recv_from_rank != _grid->ThisRank());
 
-	if( comms_send ) {
-
-	  if ( !face_table_computed ) {
-	    face_table.resize(face_idx+1);
-	    std::vector<std::pair<int,int> >  face_table_host ;
-	    Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
-	    face_table[face_idx].resize(face_table_host.size());
-	    acceleratorCopyToDevice(&face_table_host[0],
-				    &face_table[face_idx][0],
-				    face_table[face_idx].size()*sizeof(face_table_host[0]));
-	  }
-
-
-	  if ( compress.DecompressionStep() ) {
-	    recv_buf=u_simd_recv_buf[0];
-	  } else {
-	    recv_buf=this->u_recv_buf_p;
-	  }
-
-	  send_buf = this->u_send_buf_p; // Gather locally, must send
-	
-	  ////////////////////////////////////////////////////////
-	  // Gather locally
-	  ////////////////////////////////////////////////////////
-	  assert(send_buf!=NULL);
-
-	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,comm_off,so);
+	cobj *recv_buf;
+	if ( compress.DecompressionStep() ) {
+	  recv_buf=u_simd_recv_buf[0];
+	} else {
+	  recv_buf=this->u_recv_buf_p;
 	}
 
-	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,bytes,cbmask);
-	if ( (!duplicate) ) { // Force comms for now
+	cobj *send_buf;
+	send_buf = this->u_send_buf_p; // Gather locally, must send
 
-	  ///////////////////////////////////////////////////////////
-	  // Build a list of things to do after we synchronise GPUs
-	  // Start comms now???
-	  ///////////////////////////////////////////////////////////
-	  AddPacket((void *)&send_buf[comm_off],
-		    (void *)&recv_buf[comm_off],
-		    xmit_to_rank, comms_send,
-		    recv_from_rank, comms_recv,
-		    bytes);
-	}
-	
-	if ( compress.DecompressionStep()  && comms_recv ) {
-	  AddDecompress(&this->u_recv_buf_p[comm_off],
-			&recv_buf[comm_off],
+	////////////////////////////////////////////////////////
+	// 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();
+
+	///////////////////////////////////////////////////////////
+	// Build a list of things to do after we synchronise GPUs
+	// Start comms now???
+	///////////////////////////////////////////////////////////
+	AddPacket((void *)&send_buf[u_comm_offset],
+		  (void *)&recv_buf[u_comm_offset],
+		  xmit_to_rank,
+		  recv_from_rank,
+		  bytes);
+
+	if ( compress.DecompressionStep() ) {
+	  AddDecompress(&this->u_recv_buf_p[u_comm_offset],
+			&recv_buf[u_comm_offset],
 			words,Decompressions);
 	}
-	
 	u_comm_offset+=words;
-	face_idx++;
-
       }
     }
     return 0;
   }
 
   template<class compressor>
-  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx,int point)
+  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
   {
     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];
@@ -1157,6 +1103,7 @@ public:
 
 
     int permute_type=_grid->PermuteType(dimension);
+    //    std::cout << "SimdNew permute type "<<permute_type<<std::endl;
 
     ///////////////////////////////////////////////
     // Simd direction uses an extract/merge pair
@@ -1173,8 +1120,8 @@ public:
     int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    Vector<cobj *> rpointers(maxl);
-    Vector<cobj *> spointers(maxl);
+    std::vector<cobj *> rpointers(maxl);
+    std::vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where
@@ -1190,9 +1137,8 @@ public:
 
       if ( any_offnode ) {
 
-	int comm_off = u_comm_offset;
 	for(int i=0;i<maxl;i++){
-	  spointers[i] = (cobj *) &u_simd_send_buf[i][comm_off];
+	  spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
 	}
 
 	int sx   = (x+sshift)%rd;
@@ -1201,17 +1147,18 @@ public:
 	  face_table.resize(face_idx+1);
 	  std::vector<std::pair<int,int> >  face_table_host ;
 				
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
+	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
 	  face_table[face_idx].resize(face_table_host.size());
 	  acceleratorCopyToDevice(&face_table_host[0],
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
 	}
+	gathermtime-=usecond();
 
-	if ( comms_send )
-	  Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
+	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
 
@@ -1228,8 +1175,8 @@ public:
 	  int nbr_plane = nbr_ic;
 	  assert (sx == nbr_ox);
 
-	  auto rp = &u_simd_recv_buf[i        ][comm_off];
-	  auto sp = &u_simd_send_buf[nbr_plane][comm_off];
+	  auto rp = &u_simd_recv_buf[i        ][u_comm_offset];
+	  auto sp = &u_simd_send_buf[nbr_plane][u_comm_offset];
 
 	  if(nbr_proc){
 
@@ -1240,13 +1187,8 @@ public:
 
 	    rpointers[i] = rp;
 
-	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,bytes,cbmask);
-	    if ( !duplicate  ) { 
-	      AddPacket((void *)sp,(void *)rp,
-			xmit_to_rank,comms_send,
-			recv_from_rank,comms_recv,
-			bytes);
-	    }
+	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
+
 
 	  } else {
 
@@ -1255,12 +1197,9 @@ public:
 	  }
 	}
 
-	if ( comms_recv ) {
-	  AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
-	}
+	AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
 
 	u_comm_offset     +=buffer_size;
-
       }
     }
     return 0;

Grid/tensors/Tensor_exp.h

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

Grid/threads/Accelerator.cc

@@ -195,12 +195,15 @@ void acceleratorInit(void)
 #ifdef GRID_SYCL
 
 cl::sycl::queue *theGridAccelerator;
+cl::sycl::queue *theCopyAccelerator;
 void acceleratorInit(void)
 {
   int nDevices = 1;
   cl::sycl::gpu_selector selector;
   cl::sycl::device selectedDevice { selector };
   theGridAccelerator = new sycl::queue (selectedDevice);
+  //  theCopyAccelerator = new sycl::queue (selectedDevice);
+  theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
 
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
   zeInit(0);

Grid/threads/Accelerator.h

@@ -247,7 +247,6 @@ inline int  acceleratorIsCommunicable(void *ptr)
 //////////////////////////////////////////////
 // SyCL acceleration
 //////////////////////////////////////////////
-
 #ifdef GRID_SYCL
 NAMESPACE_END(Grid);
 #include <CL/sycl.hpp>
@@ -262,6 +261,7 @@ NAMESPACE_END(Grid);
 NAMESPACE_BEGIN(Grid);
 
 extern cl::sycl::queue *theGridAccelerator;
+extern cl::sycl::queue *theCopyAccelerator;
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define GRID_SIMT
@@ -289,7 +289,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
       cgh.parallel_for(					\
       cl::sycl::nd_range<3>(global,local), \
       [=] (cl::sycl::nd_item<3> item) /*mutable*/     \
-      [[intel::reqd_sub_group_size(8)]]	      \
+      [[intel::reqd_sub_group_size(16)]]	      \
       {						      \
       auto iter1    = item.get_global_id(0);	      \
       auto iter2    = item.get_global_id(1);	      \
@@ -298,19 +298,19 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
      });	   			              \
     });
 
-#define accelerator_barrier(dummy) theGridAccelerator->wait();
+#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {
-  theGridAccelerator->memcpy(to,from,bytes);
-}
-inline void acceleratorCopySynchronise(void) {  theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
+
+inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {  theCopyAccelerator->memcpy(to,from,bytes);}
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
+
 inline int  acceleratorIsCommunicable(void *ptr)
 {
 #if 0
@@ -441,7 +441,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
-  hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);
+  hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToDevice,copyStream);
 }
 inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
 
@@ -461,8 +461,6 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
   accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, { __VA_ARGS__ } ); \
   accelerator_barrier(dummy);
 
-#define GRID_ACCELERATED
-
 #endif
 
 //////////////////////////////////////////////
@@ -513,7 +511,16 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
 inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 #endif
 
+//////////////////////////////////////////////
+// Fencing needed ONLY for SYCL
+//////////////////////////////////////////////
 
+#ifdef GRID_SYCL
+inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
+#else
+// Ordering within a stream guaranteed on Nvidia & AMD
+inline void acceleratorFenceComputeStream(void){ };
+#endif
 
 ///////////////////////////////////////////////////
 // Synchronise across local threads for divergence resynch

Grid/util/Sha.h

@@ -27,6 +27,7 @@
     /*  END LEGAL */
 extern "C" {
 #include <openssl/sha.h>
+#include <openssl/evp.h>
 }
 #ifdef USE_IPP
 #include "ipp.h"
@@ -70,10 +71,8 @@ public:
   static inline std::vector<unsigned char> sha256(const void *data,size_t bytes)
   {
     std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
-    SHA256_CTX sha256;
-    SHA256_Init  (&sha256);
-    SHA256_Update(&sha256, data,bytes);
-    SHA256_Final (&hash[0], &sha256);
+    auto digest = EVP_get_digestbyname("SHA256");
+    EVP_Digest(data, bytes, &hash[0], NULL, digest, NULL);
     return hash;
   }
   static inline std::vector<int> sha256_seeds(const std::string &s)

HMC/Mobius2p1f_DD_RHMC.cc

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

README.md

@@ -148,7 +148,7 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-mkl[=<path>]`: use Intel MKL for FFT (and LAPACK if enabled) routines. A UNIX prefix containing the library can be specified (optional).
 - `--enable-numa`: enable NUMA first touch optimisation
 - `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
-- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
+- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 64 bytes).
 - `--enable-comms=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.

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,1,
+					      xmit_to_rank,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,1,
+					      recv_from_rank,
 					      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,1,
+					      xmit_to_rank,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,1,
+					      recv_from_rank,
 					      bytes,mu+4);
 	  
 	  }
@@ -309,9 +309,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,1,
+					      xmit_to_rank,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,1,
+					      recv_from_rank,
 					      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,1,
+					      xmit_to_rank,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,1,
+					      recv_from_rank,
 					      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,1,
-					       (void *)&rbuf[dir][0], recv_from_rank,1, bytes,tid);
+	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
+					       (void *)&rbuf[dir][0], recv_from_rank, bytes,tid);
 
 	    thread_critical { dbytes+=tbytes; }
 	  }

benchmarks/Benchmark_dwf_fp32.cc

@@ -32,18 +32,18 @@
 using namespace std;
 using namespace Grid;
 
-////////////////////////
-/// Move to domains ////
-////////////////////////
-
-Gamma::Algebra Gmu [] = {
-			 Gamma::Algebra::GammaX,
-			 Gamma::Algebra::GammaY,
-			 Gamma::Algebra::GammaZ,
-			 Gamma::Algebra::GammaT
+template<class d>
+struct scal {
+  d internal;
 };
 
-void Benchmark(int Ls, Coordinate Dirichlet);
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
+  };
+
 
 int main (int argc, char ** argv)
 {
@@ -52,82 +52,24 @@ int main (int argc, char ** argv)
 
   int threads = GridThread::GetThreads();
 
+  Coordinate latt4 = GridDefaultLatt();
   int Ls=16;
-  for(int i=0;i<argc;i++) {
+  for(int i=0;i<argc;i++)
     if(std::string(argv[i]) == "-Ls"){
       std::stringstream ss(argv[i+1]); ss >> Ls;
     }
-  }
 
-  //////////////////
-  // With comms
-  //////////////////
-  Coordinate Dirichlet(Nd+1,0);
-
-  std::cout << "\n\n\n\n\n\n" <<std::endl;
-  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
-  std::cout << GridLogMessage<< " Testing with full communication " <<std::endl;
-  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
-  
-  Benchmark(Ls,Dirichlet);
-
-  //////////////////
-  // Domain decomposed
-  //////////////////
-  Coordinate latt4  = GridDefaultLatt();
-  Coordinate mpi    = GridDefaultMpi();
-  Coordinate CommDim(Nd);
-  Coordinate shm;
-  GlobalSharedMemory::GetShmDims(mpi,shm);
-
-
-  //////////////////////
-  // Node level
-  //////////////////////
-  std::cout << "\n\n\n\n\n\n" <<std::endl;
-  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
-  std::cout << GridLogMessage<< " Testing without internode communication " <<std::endl;
-  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
-
-  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
-  Dirichlet[0] = 0;
-  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
-  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
-  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
-  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
-
-  Benchmark(Ls,Dirichlet);
-
-  std::cout << "\n\n\n\n\n\n" <<std::endl;
-
-  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
-  std::cout << GridLogMessage<< " Testing without intranode communication " <<std::endl;
-  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
-
-  for(int d=0;d<Nd;d++) CommDim[d]= mpi[d]>1 ? 1 : 0;
-  Dirichlet[0] = 0;
-  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0];
-  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1];
-  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2];
-  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3];
-  
-  Benchmark(Ls,Dirichlet);
-
-  Grid_finalize();
-  exit(0);
-}
-void Benchmark(int Ls, Coordinate Dirichlet)
-{
-  Coordinate latt4 = GridDefaultLatt();
   GridLogLayout();
 
   long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
 
+
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
+  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
   GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
   GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
   GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
@@ -138,9 +80,9 @@ void Benchmark(int Ls, Coordinate Dirichlet)
 
   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
@@ -158,6 +100,7 @@ void Benchmark(int Ls, Coordinate Dirichlet)
   src = src*N2;
 #endif
 
+
   LatticeFermionF result(FGrid); result=Zero();
   LatticeFermionF    ref(FGrid);    ref=Zero();
   LatticeFermionF    tmp(FGrid);
@@ -165,31 +108,29 @@ void Benchmark(int Ls, Coordinate Dirichlet)
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeFieldF Umu(UGrid);
-  LatticeGaugeFieldF UmuCopy(UGrid);
   SU<Nc>::HotConfiguration(RNG4,Umu);
-  UmuCopy=Umu;
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+#if 0
+  Umu=1.0;
+  for(int mu=0;mu<Nd;mu++){
+    LatticeColourMatrixF ttmp(UGrid);
+    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
+    //    if (mu !=2 ) ttmp = 0;
+    //    ttmp = ttmp* pow(10.0,mu);
+    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
+  }
+  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
+#endif
 
-  ////////////////////////////////////
-  // Apply BCs
-  ////////////////////////////////////
-  Coordinate Block(4);
-  for(int d=0;d<4;d++)  Block[d]= Dirichlet[d+1];
-
-  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block5 " << Dirichlet << std::endl;
-  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block4 " << Block << std::endl;
-
-  DirichletFilter<LatticeGaugeFieldF> Filter(Block);
-  Filter.applyFilter(Umu);
-  
   ////////////////////////////////////
   // Naive wilson implementation
   ////////////////////////////////////
+  // replicate across fifth dimension
+  //  LatticeGaugeFieldF Umu5d(FGrid);
   std::vector<LatticeColourMatrixF> U(4,UGrid);
   for(int mu=0;mu<Nd;mu++){
     U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
   }
-
   std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
   if (1)
@@ -249,15 +190,12 @@ void Benchmark(int Ls, Coordinate Dirichlet)
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
-  DomainWallFermionF::ImplParams p;
-  p.dirichlet=Dirichlet;
-  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,p);
-  Dw.ImportGauge(Umu);
-  
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
   int ncall =300;
 
   if (1) {
     FGrid->Barrier();
+    Dw.ZeroCounters();
     Dw.Dhop(src,result,0);
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
@@ -282,20 +220,29 @@ void Benchmark(int Ls, Coordinate Dirichlet)
     double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
 
     std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
+    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
-    //    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
-    //    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
+    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
+    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
     err = ref-result;
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+    //exit(0);
 
     if(( norm2(err)>1.0e-4) ) {
+      /*
+      std::cout << "RESULT\n " << result<<std::endl;
+      std::cout << "REF   \n " << ref   <<std::endl;
+      std::cout << "ERR   \n " << err   <<std::endl;
+      */
       std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
       FGrid->Barrier();
       exit(-1);
     }
     assert (norm2(err)< 1.0e-4 );
+    Dw.Report();
   }
 
   if (1)
@@ -339,20 +286,21 @@ void Benchmark(int Ls, Coordinate Dirichlet)
     }
     ref = -0.5*ref;
   }
-
-  Dw.Dhop(src,result,DaggerYes);
-
-  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
+  //  dump=1;
+  Dw.Dhop(src,result,1);
   std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
-  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
-
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
   std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
   std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result;
   std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
-  assert((norm2(err)<1.0e-4));
-  
+  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;
+*/
+  }
   LatticeFermionF src_e (FrbGrid);
   LatticeFermionF src_o (FrbGrid);
   LatticeFermionF r_e   (FrbGrid);
@@ -382,6 +330,7 @@ void Benchmark(int Ls, Coordinate Dirichlet)
   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();
@@ -403,6 +352,7 @@ void Benchmark(int Ls, Coordinate Dirichlet)
     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);
@@ -417,7 +367,13 @@ void Benchmark(int Ls, Coordinate Dirichlet)
 
   err = r_eo-result;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  assert(norm2(err)<1.0e-4);
+  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;
+    */
+  }
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
@@ -426,4 +382,6 @@ void Benchmark(int Ls, Coordinate Dirichlet)
 
   assert(norm2(src_e)<1.0e-4);
   assert(norm2(src_o)<1.0e-4);
+  Grid_finalize();
+  exit(0);
 }

benchmarks/Benchmark_halo.cc

@@ -0,0 +1,131 @@
+ /*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid
+    Source file: ./benchmarks/Benchmark_dwf.cc
+    Copyright (C) 2015
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#ifdef GRID_CUDA
+#define CUDA_PROFILE
+#endif
+
+#ifdef CUDA_PROFILE
+#include <cuda_profiler_api.h>
+#endif
+
+using namespace std;
+using namespace Grid;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
+  };
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  Coordinate latt4= GridDefaultLatt();
+  Coordinate mpi  = GridDefaultMpi();
+  Coordinate simd = GridDefaultSimd(Nd,vComplexF::Nsimd());
+
+  GridLogLayout();
+
+  int Ls=16;
+  for(int i=0;i<argc;i++)
+    if(std::string(argv[i]) == "-Ls"){
+      std::stringstream ss(argv[i+1]); ss >> Ls;
+    }
+
+  
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,simd ,mpi);
+  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);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+  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);
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
+  LatticeGaugeFieldF Umu(UGrid);
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  RealD NP = UGrid->_Nprocessors;
+  RealD NN = UGrid->NodeCount();
+
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+
+  const int ncall = 500;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionF::HaloGatherOpt         "<<std::endl;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  {
+    typename DomainWallFermionF::Compressor compressor(0);
+    FGrid->Barrier();
+    Dw.Stencil.HaloExchangeOptGather(src,compressor);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      Dw.Stencil.HaloExchangeOptGather(src,compressor);
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+
+    double bytes=0.0;
+    if(mpi[0]) bytes+=latt4[1]*latt4[2]*latt4[3];
+    if(mpi[1]) bytes+=latt4[0]*latt4[2]*latt4[3];
+    if(mpi[2]) bytes+=latt4[0]*latt4[1]*latt4[3];
+    if(mpi[3]) bytes+=latt4[0]*latt4[1]*latt4[2];
+    bytes = bytes * Ls * 8.* (24.+12.)* 2.0;
+
+    std::cout<<GridLogMessage << "Gather us /call =   "<< (t1-t0)/ncall<<std::endl;
+    std::cout<<GridLogMessage << "Gather MBs /call =   "<< bytes*ncall/(t1-t0)<<std::endl;
+
+  }
+
+  Grid_finalize();
+  exit(0);
+}

examples/Example_Laplacian.cc

@@ -93,14 +93,14 @@ template<class Field> class FreeLaplacianStencil : public SparseMatrixBase<Field
 {
 public:
   typedef typename Field::vector_object siteObject;
-  typedef CartesianStencil<siteObject, siteObject, SimpleStencilParams> StencilImpl;
+  typedef CartesianStencil<siteObject, siteObject, int> StencilImpl;
 
   GridBase *grid;
   StencilImpl Stencil;
   SimpleCompressor<siteObject> Compressor;
   
   FreeLaplacianStencil(GridBase *_grid)
-    : Stencil    (_grid,6,Even,directions,displacements,SimpleStencilParams()), grid(_grid)
+    : Stencil    (_grid,6,Even,directions,displacements,0), grid(_grid)
   {  };
   
   virtual GridBase *Grid(void) { return grid; };
@@ -168,8 +168,7 @@ public:
   typedef iImplDoubledGaugeField<Simd> SiteDoubledGaugeField;
   typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
 
-  typedef CartesianStencil<siteObject, siteObject,SimpleStencilParams> StencilImpl;
-  SimpleStencilParams p;
+  typedef CartesianStencil<siteObject, siteObject, int> StencilImpl;
 
   GridBase *grid;
   StencilImpl Stencil;
@@ -178,7 +177,7 @@ public:
   CovariantLaplacianStencil(GaugeField &Umu)
     :
       grid(Umu.Grid()),
-      Stencil    (grid,6,Even,directions,displacements,p),
+      Stencil    (grid,6,Even,directions,displacements,0),
       Uds(grid)
   {
     for (int mu = 0; mu < Nd; mu++) {

systems/Crusher/dwf.slurm

@@ -3,28 +3,28 @@
 #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 8
-#SBATCH --exclusive
-#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
+#SBATCH -n 1
+#SBATCH --exclusive  
 
 DIR=.
 module list
-export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+#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=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 16 --grid 32.32.32.256 --mpi 1.1.1.8 --comms-overlap --shm 2048 --shm-mpi 0"
-echo $PARAMS
-srun --gpus-per-task 1 -n8  ./benchmarks/Benchmark_dwf_fp32 $PARAMS
-
+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/dwf8.slurm

@@ -6,43 +6,22 @@
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
-#SBATCH -N 8
-#SBATCH -n 64
+#SBATCH -N 1
+#SBATCH -n 8
 #SBATCH --exclusive
-#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
 
 DIR=.
 module list
-export MPICH_OFI_NIC_POLICY=GPU
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=NONE
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
-export MPICH_SMP_SINGLE_COPY_MODE=NONE
 export OMP_NUM_THREADS=1
 
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
+PARAMS=" --accelerator-threads 8 --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
 
-PARAMS=" --accelerator-threads 16 --grid 64.64.64.256 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
-echo $PARAMS
-#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.256.8node
-
-
-PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 1"
-echo $PARAMS
-srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node
-
-PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 0"
-echo $PARAMS
-#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node.shm0
-
-PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 1"
-echo $PARAMS
-#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node
-
-PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
-echo $PARAMS
-#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node_shm0
-
+srun --gpus-per-task 1 -n8 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
 
 

systems/Crusher/mpiwrapper.sh

@@ -1,11 +1,10 @@
 #!/bin/bash
 
 lrank=$SLURM_LOCALID
-lgpu=(0 1 2 3 7 6 5 4)
 
-export ROCR_VISIBLE_DEVICES=${lgpu[$lrank]}
+export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
 
-echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES "
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
 
 $*
 

systems/PVC/benchmarks/run-1tile.sh

@@ -0,0 +1,62 @@
+#!/bin/sh
+##SBATCH -p PVC-SPR-QZEH 
+##SBATCH -p PVC-ICX-QZNW
+#SBATCH -p QZ1J-ICX-PVC
+##SBATCH -p QZ1J-SPR-PVC-2C
+
+source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
+
+export NT=8
+
+export I_MPI_OFFLOAD=1
+export I_MPI_OFFLOAD_TOPOLIB=level_zero
+export I_MPI_OFFLOAD_DOMAIN_SIZE=-1
+
+# export IGC_EnableLSCFenceUGMBeforeEOT=0
+# export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file=False"
+export SYCL_DEVICE_FILTER=gpu,level_zero
+#export IGC_ShaderDumpEnable=1 
+#export IGC_DumpToCurrentDir=1
+export I_MPI_OFFLOAD_CELL=tile
+export EnableImplicitScaling=0
+export EnableWalkerPartition=0
+export ZE_AFFINITY_MASK=0.0
+mpiexec -launcher ssh -n 1 -host localhost  ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 --device-mem 32768
+
+export ZE_AFFINITY_MASK=0
+export I_MPI_OFFLOAD_CELL=device
+export EnableImplicitScaling=1
+export EnableWalkerPartition=1
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+#mpiexec -launcher ssh -n 2 -host localhost  vtune -collect gpu-hotspots -knob gpu-sampling-interval=1 -data-limit=0 -r ./vtune_run4 -- ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-overlap --shm-mpi 1
+
+#mpiexec  -launcher ssh -n 1 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-overlap --shm-mpi 1
+
+#mpiexec  -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1
+
+#mpiexec  -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-overlap --shm-mpi 1
+
+#mpiexec  -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0
+
+#mpirun -np 2 ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 16.32.32.64 --accelerator-threads $NT --comms-sequential --shm-mpi 0
+#mpirun -np 2 ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT --comms-sequential --shm-mpi 1
+

systems/PVC/benchmarks/run-2tile-mpi.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+##SBATCH -p PVC-SPR-QZEH 
+##SBATCH -p PVC-ICX-QZNW
+
+#SBATCH -p QZ1J-ICX-PVC
+
+source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
+
+export NT=16
+
+# export IGC_EnableLSCFenceUGMBeforeEOT=0
+# export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file=False"
+#export IGC_ShaderDumpEnable=1 
+#export IGC_DumpToCurrentDir=1
+export I_MPI_OFFLOAD=1
+export I_MPI_OFFLOAD_TOPOLIB=level_zero
+export I_MPI_OFFLOAD_DOMAIN_SIZE=-1
+export SYCL_DEVICE_FILTER=gpu,level_zero
+export I_MPI_OFFLOAD_CELL=tile
+export EnableImplicitScaling=0
+export EnableWalkerPartition=0
+export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
+
+for i in 0 
+do
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1  --device-mem 32768
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1  --device-mem 32768
+done
+#mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.1x2.log
+#mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.2x1.log
+
+

systems/PVC/benchmarks/wrap.sh

@@ -0,0 +1,14 @@
+#!/bin/sh
+
+export ZE_AFFINITY_MASK=0.$MPI_LOCALRANKID
+
+echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
+
+
+if [ $MPI_LOCALRANKID = "0" ] 
+then
+#  ~psteinbr/build_pti/ze_tracer -h $@
+  onetrace --chrome-device-timeline $@
+else
+  $@
+fi

systems/PVC/config-command

@@ -0,0 +1,16 @@
+INSTALL=/nfs/site/home/azusayax/install
+../../configure \
+	--enable-simd=GPU \
+	--enable-gen-simd-width=64 \
+	--enable-comms=mpi-auto \
+	--disable-accelerator-cshift \
+	--disable-gparity \
+	--disable-fermion-reps \
+	--enable-shm=nvlink \
+	--enable-accelerator=sycl \
+	--enable-unified=no \
+	MPICXX=mpicxx \
+	CXX=dpcpp \
+	LDFLAGS="-fsycl-device-code-split=per_kernel -fsycl-device-lib=all -lze_loader -L$INSTALL/lib" \
+	CXXFLAGS="-fsycl-unnamed-lambda -fsycl -no-fma -I$INSTALL/include -Wno-tautological-compare"
+

systems/PVC/setup.sh

@@ -0,0 +1,11 @@
+export https_proxy=http://proxy-chain.intel.com:911
+export LD_LIBRARY_PATH=/nfs/site/home/azusayax/install/lib:$LD_LIBRARY_PATH
+
+module load intel-release
+source /opt/intel/oneapi/PVC_setup.sh
+#source /opt/intel/oneapi/ATS_setup.sh
+module load intel/mpich/pvc45.3
+export PATH=~/ATS/pti-gpu/tools/onetrace/:$PATH
+
+#clsh embargo-ci-neo-022845
+#source /opt/intel/vtune_amplifier/amplxe-vars.sh

systems/Perlmutter/config-command

@@ -1,14 +1,9 @@
-DIR=`pwd`
-PREFIX=$DIR/../Prequisites/install/
 ../../configure \
     --enable-comms=mpi \
     --enable-simd=GPU \
     --enable-shm=nvlink \
     --enable-gen-simd-width=64 \
     --enable-accelerator=cuda \
-    --enable-setdevice \
-    --disable-accelerator-cshift \
-    --with-gmp=$PREFIX \
     --disable-fermion-reps \
     --disable-unified \
     --disable-gparity \

systems/Perlmutter/dwf4.slurm

@@ -1,27 +1,24 @@
 #!/bin/bash
-#SBATCH -A m3886_g
+#SBATCH -A mp13
 #SBATCH -C gpu
-#SBATCH -q debug
+#SBATCH -q regular
 #SBATCH -t 0:20:00
-#SBATCH -c 32
-#SBATCH -N 1
-#SBATCH -n 4
+#SBATCH -n 16
 #SBATCH --ntasks-per-node=4
-#SBATCH --gpus-per-task=1
+#SBATCH -c 32
 #SBATCH --exclusive
-#SBATCH --gpu-bind=none
+#SBATCH --gpus-per-task=1
+#SBATCH --gpu-bind=map_gpu:0,1,2,3
 
 export SLURM_CPU_BIND="cores"
-export MPICH_GPU_SUPPORT_ENABLED=1
 export MPICH_RDMA_ENABLED_CUDA=1
-export MPICH_GPU_IPC_ENABLED=1
-export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
-export MPICH_GPU_NO_ASYNC_MEMCPY=0
-#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_GPU_SUPPORT_ENABLED=1
+srun ./benchmarks/Benchmark_comms_host_device --mpi 2.2.2.2  --accelerator-threads 8 > comms.4node
 
-OPT="--comms-overlap --shm-mpi 1"
-VOL=64.64.32.32
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
-#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.1.1.4 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
-#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.8 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
+OPT="--comms-overlap --comms-concurrent --shm-mpi 0"
+srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 64.64.64.64 --accelerator-threads 8 --shm 2048 $OPT > dwf.64.64.64.64.4node.opt0
+srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 48.48.48.48 --accelerator-threads 8 --shm 2048 $OPT > dwf.48.48.48.48.4node.opt0
 
+OPT="--comms-overlap --comms-concurrent --shm-mpi 1"
+srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 64.64.64.64 --accelerator-threads 8 --shm 2048 $OPT > dwf.64.64.64.64.4node.opt1
+srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 48.48.48.48 --accelerator-threads 8 --shm 2048 $OPT > dwf.48.48.48.48.4node.opt1

systems/Perlmutter/sourceme.sh

@@ -1,4 +1,4 @@
 
 export CRAY_ACCEL_TARGET=nvidia80
 
-module load PrgEnv-gnu cpe-cuda cudatoolkit/11.4
+module load PrgEnv-gnu cpe-cuda cuda

systems/Spock/config-command

@@ -6,8 +6,6 @@
 --enable-simd=GPU \
 --disable-fermion-reps \
 --disable-gparity \
---with-gmp=$OLCF_GMP_ROOT \
---with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 CXX=hipcc MPICXX=mpicxx \
 CXXFLAGS="-fPIC -I/opt/rocm-4.3.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
 --prefix=/ccs/home/chulwoo/Grid \

systems/Spock/dwf8.slurm

@@ -1,7 +1,8 @@
 #!/bin/bash
 # Begin LSF Directives
 #SBATCH -A LGT104
-#SBATCH -t 3:00:00
+#SBATCH -t 01:00:00
+##SBATCH -U openmpThu
 #SBATCH -p ecp
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
@@ -13,12 +14,13 @@ DIR=.
 module list
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=CMA
-
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export OMP_NUM_THREADS=8
 
 AT=8
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads ${AT} --grid 16.16.16.48 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
-srun -N2 -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./HMC/Mobius2p1f_DD_RHMC $PARAMS
+PARAMS=" --accelerator-threads ${AT} --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+srun -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
 

systems/Spock/sourceme.sh

@@ -1,9 +1,5 @@
-module load emacs
 module load PrgEnv-gnu
-module load rocm/4.5.0
+module load rocm/4.3.0
 module load gmp
 module load cray-fftw
 module load craype-accel-amd-gfx908
-export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
-export MPICH_GPU_SUPPORT_ENABLED=1
-export LD_LIBRARY_PATH=/opt/cray/pe/gcc/mpfr/3.1.4/lib/:$LD_LIBRARY_PATH

systems/Tursa/dwf.4node.perf

@@ -1,25 +1,25 @@
-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
+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
 OPENMPI detected
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses
-AcceleratorCudaInit: a) IBM jsrun, or 
+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-setdevice=no 
+AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
+AcceleratorCudaInit: ================================================
 OPENMPI detected
 AcceleratorCudaInit[0]: ========================
 AcceleratorCudaInit[0]: Device Number    : 0
@@ -33,41 +33,11 @@ AcceleratorCudaInit[0]:   pciBusID: 3
 AcceleratorCudaInit[0]:   pciDeviceID: 0 
 AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses
-AcceleratorCudaInit: a) IBM jsrun, or 
+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-setdevice=no 
+AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
+AcceleratorCudaInit: ================================================
 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]: ========================
@@ -80,25 +50,43 @@ AcceleratorCudaInit[0]:   pciBusID: 3
 AcceleratorCudaInit[0]:   pciDeviceID: 0 
 AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses
-AcceleratorCudaInit: a) IBM jsrun, or 
+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-setdevice=no 
-local rank 1 device 0 bus id: 0000:44:00.0
+AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-local rank 0 device 0 bus id: 0000:03:00.0
+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: ================================================
+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 0x153960000000 for comms buffers 
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x7fcd80000000 for comms buffers 
 Setting up IPC
 
 __|__|__|__|__|__|__|__|__|__|__|__|__|__|__
@@ -128,7 +116,7 @@ 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=da06d15f73184ceb15d66d4e7e702b02fed7b940: (HEAD -> feature/dirichlet, develop) uncommited changes
+Current Grid git commit hash=9d2238148c56e3fbadfa95dcabf2b83d4bde14cd: (HEAD -> develop) uncommited changes
 
 Grid : Message : ================================================ 
 Grid : Message : MPI is initialised and logging filters activated 
@@ -136,102 +124,122 @@ 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 8 LARGE 2
+Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 32 LARGE 8
 Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
 Grid : Message : MemoryManager::Init() Using cudaMalloc
-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
+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

systems/Tursa/dwf4.slurm

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

tests/Test_simd.cc

@@ -793,6 +793,7 @@ int main (int argc, char ** argv)
     }
     std::cout <<" OK ! "<<std::endl;
 
+#ifdef USE_FP16
     // Double to Half
     std::cout << GridLogMessage<< "Double to half" ;
     precisionChange(&H[0],&D[0],Ndp);
@@ -822,6 +823,7 @@ int main (int argc, char ** argv)
       assert( tmp < 1.0e-3 );
     }
     std::cout <<" OK ! "<<std::endl;
+#endif
 
   }
   Grid_finalize();

tests/Test_stencil.cc

@@ -31,6 +31,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 using namespace std;
 using namespace Grid;
+ ;
 
 int main(int argc, char ** argv) {
   Grid_init(&argc, &argv);
@@ -79,8 +80,7 @@ int main(int argc, char ** argv) {
     Foo=lex;
   }
 
-  typedef CartesianStencil<vobj,vobj,SimpleStencilParams> Stencil;
-  SimpleStencilParams p;
+  typedef CartesianStencil<vobj,vobj,int> Stencil;
     for(int dir=0;dir<4;dir++){
       for(int disp=0;disp<Fine._fdimensions[dir];disp++){
 
@@ -90,7 +90,7 @@ int main(int argc, char ** argv) {
 	std::vector<int> directions(npoint,dir);
 	std::vector<int> displacements(npoint,disp);
 
-	Stencil myStencil(&Fine,npoint,0,directions,displacements,p);
+	Stencil myStencil(&Fine,npoint,0,directions,displacements,0);
 	Coordinate ocoor(4);
 	for(int o=0;o<Fine.oSites();o++){
 	  Fine.oCoorFromOindex(ocoor,o);
@@ -183,8 +183,8 @@ int main(int argc, char ** argv) {
 	std::vector<int> directions(npoint,dir);
 	std::vector<int> displacements(npoint,disp);
 
-	Stencil EStencil(&rbFine,npoint,Even,directions,displacements,p);
-	Stencil OStencil(&rbFine,npoint,Odd,directions,displacements,p);
+	Stencil EStencil(&rbFine,npoint,Even,directions,displacements,0);
+	Stencil OStencil(&rbFine,npoint,Odd,directions,displacements,0);
 
 	Coordinate ocoor(4);
 	for(int o=0;o<Fine.oSites();o++){