Merge branch 'release/ISC-freeze-2'

Conflicts:
	lib/qcd/action/fermion/ImprovedStaggeredFermion.cc

.travis.yml

@@ -19,6 +19,8 @@ before_install:
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc; fi
     
 install:
+    - export CWD=`pwd`
+    - echo $CWD
     - export CC=$CC$VERSION
     - export CXX=$CXX$VERSION
     - echo $PATH
@@ -36,11 +38,22 @@ script:
     - ./bootstrap.sh
     - mkdir build
     - cd build
-    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none
+    - mkdir lime
+    - cd lime
+    - mkdir build
+    - cd build
+    - wget http://usqcd-software.github.io/downloads/c-lime/lime-1.3.2.tar.gz
+    - tar xf lime-1.3.2.tar.gz
+    - cd lime-1.3.2
+    - ./configure --prefix=$CWD/build/lime/install
+    - make -j4
+    - make install
+    - cd $CWD/build
+    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install
     - make -j4 
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - echo make clean
-    - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
+    - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install
     - make -j4
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - make check

benchmarks/Benchmark_ITT.cc

@@ -158,8 +158,10 @@ public:
 
 	  dbytes=0;
 	  ncomm=0;
-
-	  parallel_for(int dir=0;dir<8;dir++){
+#ifdef GRID_OMP
+#pragma omp parallel for num_threads(Grid::CartesianCommunicator::nCommThreads)
+#endif
+	  for(int dir=0;dir<8;dir++){
 
 	    double tbytes;
 	    int mu =dir % 4;
@@ -175,9 +177,14 @@ public:
 		int comm_proc = mpi_layout[mu]-1;
 		Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	      }
+#ifdef GRID_OMP
+	int tid = omp_get_thread_num(); 
+#else 
+        int tid = dir;
+#endif
 	      tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
 						 (void *)&rbuf[dir][0], recv_from_rank,
-						 bytes,dir);
+						 bytes,tid);
 	  
 #ifdef GRID_OMP
 #pragma omp atomic

benchmarks/Benchmark_comms.cc

@@ -169,7 +169,11 @@ int main (int argc, char ** argv)
   for(int lat=4;lat<=maxlat;lat+=4){
     for(int Ls=8;Ls<=8;Ls*=2){
 
-      std::vector<int> latt_size  ({lat,lat,lat,lat});
+      std::vector<int> latt_size  ({lat*mpi_layout[0],
+                                    lat*mpi_layout[1],
+                                    lat*mpi_layout[2],
+                                    lat*mpi_layout[3]});
+
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
       RealD Nrank = Grid._Nprocessors;
@@ -446,7 +450,7 @@ int main (int argc, char ** argv)
   }    
 
 
-
+#ifdef GRID_OMP
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking threaded STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
@@ -485,7 +489,8 @@ int main (int argc, char ** argv)
 	dbytes=0;
 	ncomm=0;
 
-	parallel_for(int dir=0;dir<8;dir++){
+#pragma omp parallel for num_threads(Grid::CartesianCommunicator::nCommThreads)
+	for(int dir=0;dir<8;dir++){
 
 	  double tbytes;
 	  int mu =dir % 4;
@@ -502,9 +507,9 @@ int main (int argc, char ** argv)
 	      int comm_proc = mpi_layout[mu]-1;
 	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	    }
-
+            int tid = omp_get_thread_num();
 	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
-					       (void *)&rbuf[dir][0], recv_from_rank, bytes,dir);
+					       (void *)&rbuf[dir][0], recv_from_rank, bytes,tid);
 
 #pragma omp atomic
 	    dbytes+=tbytes;
@@ -532,7 +537,7 @@ int main (int argc, char ** argv)
  
     }
   }    
-
+#endif
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= All done; Bye Bye"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;

configure.ac

@@ -340,7 +340,7 @@ case ${ac_PRECISION} in
 esac
 
 ######################  Shared memory allocation technique under MPI3
-AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmopen|hugetlbfs|shmnone],
+AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmopen|shmget|hugetlbfs|shmnone],
               [Select SHM allocation technique])],[ac_SHM=${enable_shm}],[ac_SHM=shmopen])
 
 case ${ac_SHM} in
@@ -349,6 +349,10 @@ case ${ac_SHM} in
      AC_DEFINE([GRID_MPI3_SHMOPEN],[1],[GRID_MPI3_SHMOPEN] )
      ;;
 
+     shmget)
+     AC_DEFINE([GRID_MPI3_SHMGET],[1],[GRID_MPI3_SHMGET] )
+     ;;
+
      shmnone)
      AC_DEFINE([GRID_MPI3_SHM_NONE],[1],[GRID_MPI3_SHM_NONE] )
      ;;
@@ -366,7 +370,7 @@ esac
 AC_ARG_ENABLE([shmpath],[AC_HELP_STRING([--enable-shmpath=path],
               [Select SHM mmap base path for hugetlbfs])],
 	      [ac_SHMPATH=${enable_shmpath}],
-	      [ac_SHMPATH=/var/lib/hugetlbfs/pagesize-2MB/])
+	      [ac_SHMPATH=/var/lib/hugetlbfs/global/pagesize-2MB/])
 AC_DEFINE_UNQUOTED([GRID_SHM_PATH],["$ac_SHMPATH"],[Path to a hugetlbfs filesystem for MMAPing])
 
 ############### communication type selection

lib/algorithms/LinearOperator.h

@@ -309,36 +309,59 @@ namespace Grid {
       class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
     protected:
       Matrix &_Mat;
+      Field tmp;
+      RealD mass;
+      double tMpc;
+      double tIP;
+      double tMeo;
+      double taxpby_norm;
+      uint64_t ncall;
     public:
-      SchurStaggeredOperator (Matrix &Mat): _Mat(Mat){};
+      void Report(void)
+      {
+	std::cout << GridLogMessage << " HermOpAndNorm.Mpc "<< tMpc/ncall<<" usec "<<std::endl;
+	std::cout << GridLogMessage << " HermOpAndNorm.IP  "<< tIP /ncall<<" usec "<<std::endl;
+	std::cout << GridLogMessage << " Mpc.MeoMoe        "<< tMeo/ncall<<" usec "<<std::endl;
+	std::cout << GridLogMessage << " Mpc.axpby_norm    "<< taxpby_norm/ncall<<" usec "<<std::endl;
+      }
+      SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
+      { 
+	assert( _Mat.isTrivialEE() );
+	mass = _Mat.Mass();
+	tMpc=0;
+	tIP =0;
+        tMeo=0;
+        taxpby_norm=0;
+	ncall=0;
+      }
       virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	GridLogIterative.TimingMode(1);
-	std::cout << GridLogIterative << " HermOpAndNorm "<<std::endl;
+	ncall++;
+	tMpc-=usecond();
 	n2 = Mpc(in,out);
-	std::cout << GridLogIterative << " HermOpAndNorm.Mpc "<<std::endl;
+	tMpc+=usecond();
+	tIP-=usecond();
 	ComplexD dot= innerProduct(in,out);
-	std::cout << GridLogIterative << " HermOpAndNorm.innerProduct "<<std::endl;
+	tIP+=usecond();
 	n1 = real(dot);
       }
       virtual void HermOp(const Field &in, Field &out){
-	std::cout << GridLogIterative << " HermOp "<<std::endl;
-	Mpc(in,out);
+	ncall++;
+	tMpc-=usecond();
+	_Mat.Meooe(in,out);
+	_Mat.Meooe(out,tmp);
+	tMpc+=usecond();
+	taxpby_norm-=usecond();
+	axpby(out,-1.0,mass*mass,tmp,in);
+	taxpby_norm+=usecond();
       }
       virtual  RealD Mpc      (const Field &in, Field &out) {
-	Field tmp(in._grid);
-	Field tmp2(in._grid);
-
-	std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
-	_Mat.Mooee(in,out);
-	_Mat.Mooee(out,tmp);
-	std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
-
+	tMeo-=usecond();
 	_Mat.Meooe(in,out);
-	_Mat.Meooe(out,tmp2);
-	std::cout << GridLogIterative << " HermOp.MeooeMeooe "<<std::endl;
-
-	RealD nn=axpy_norm(out,-1.0,tmp2,tmp);
-	std::cout << GridLogIterative << " HermOp.axpy_norm "<<std::endl;
+	_Mat.Meooe(out,tmp);
+	tMeo+=usecond();
+	taxpby_norm-=usecond();
+	RealD nn=axpby_norm(out,-1.0,mass*mass,tmp,in);
+	taxpby_norm+=usecond();
 	return nn;
       }
       virtual  RealD MpcDag   (const Field &in, Field &out){

lib/algorithms/iterative/ConjugateGradient.h

@@ -54,6 +54,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
+
     psi.checkerboard = src.checkerboard;
     conformable(psi, src);
 
@@ -70,7 +71,6 @@ class ConjugateGradient : public OperatorFunction<Field> {
     
     Linop.HermOpAndNorm(psi, mmp, d, b);
 
-
     r = src - mmp;
     p = r;
 
@@ -96,38 +96,44 @@ class ConjugateGradient : public OperatorFunction<Field> {
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
     GridStopWatch LinalgTimer;
+    GridStopWatch InnerTimer;
+    GridStopWatch AxpyNormTimer;
+    GridStopWatch LinearCombTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
 
     SolverTimer.Start();
     int k;
-    for (k = 1; k <= MaxIterations; k++) {
+    for (k = 1; k <= MaxIterations*1000; k++) {
       c = cp;
 
       MatrixTimer.Start();
-      Linop.HermOpAndNorm(p, mmp, d, qq);
+      Linop.HermOp(p, mmp);
       MatrixTimer.Stop();
 
       LinalgTimer.Start();
-      //  RealD    qqck = norm2(mmp);
-      //  ComplexD dck  = innerProduct(p,mmp);
 
+      InnerTimer.Start();
+      ComplexD dc  = innerProduct(p,mmp);
+      InnerTimer.Stop();
+      d = dc.real();
       a = c / d;
-      b_pred = a * (a * qq - d) / c;
 
+      AxpyNormTimer.Start();
       cp = axpy_norm(r, -a, mmp, r);
+      AxpyNormTimer.Stop();
       b = cp / c;
 
-      // Fuse these loops ; should be really easy
-      psi = a * p + psi;
-      p = p * b + r;
-
+      LinearCombTimer.Start();
+      parallel_for(int ss=0;ss<src._grid->oSites();ss++){
+	vstream(psi[ss], a      *  p[ss] + psi[ss]);
+	vstream(p  [ss], b      *  p[ss] + r[ss]);
+      }
+      LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
       std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
                 << " residual " << cp << " target " << rsq << std::endl;
-      std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << "  b = "<< b << std::endl;
-      std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << "  c = "<< c << std::endl;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -148,6 +154,9 @@ class ConjugateGradient : public OperatorFunction<Field> {
 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

lib/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -43,6 +43,7 @@ namespace Grid {
 public:                                                
     RealD   Tolerance;
     Integer MaxIterations;
+    Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
     int verbose;
     MultiShiftFunction shifts;
 
@@ -164,6 +165,15 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
   }
  
+  ///////////////////////////////////////
+  // Timers
+  ///////////////////////////////////////
+  GridStopWatch AXPYTimer;
+  GridStopWatch ShiftTimer;
+  GridStopWatch QRTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch SolverTimer;
+  SolverTimer.Start();
   
   // Iteration loop
   int k;
@@ -171,7 +181,9 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   for (k=1;k<=MaxIterations;k++){
     
     a = c /cp;
+    AXPYTimer.Start();
     axpy(p,a,p,r);
+    AXPYTimer.Stop();
     
     // Note to self - direction ps is iterated seperately
     // for each shift. Does not appear to have any scope
@@ -180,6 +192,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     // However SAME r is used. Could load "r" and update
     // ALL ps[s]. 2/3 Bandwidth saving
     // New Kernel: Load r, vector of coeffs, vector of pointers ps
+    AXPYTimer.Start();
     for(int s=0;s<nshift;s++){
       if ( ! converged[s] ) { 
 	if (s==0){
@@ -190,22 +203,34 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	}
       }
     }
+    AXPYTimer.Stop();
     
     cp=c;
+    MatrixTimer.Start();  
+    //Linop.HermOpAndNorm(p,mmp,d,qq); // d is used
+    // The below is faster on KNL
+    Linop.HermOp(p,mmp); 
+    d=real(innerProduct(p,mmp));
     
-    Linop.HermOpAndNorm(p,mmp,d,qq);
+    MatrixTimer.Stop();  
+
+    AXPYTimer.Start();
     axpy(mmp,mass[0],p,mmp);
+    AXPYTimer.Stop();
     RealD rn = norm2(p);
     d += rn*mass[0];
     
     bp=b;
     b=-cp/d;
     
+    AXPYTimer.Start();
     c=axpy_norm(r,b,mmp,r);
+    AXPYTimer.Stop();
 
     // Toggle the recurrence history
     bs[0] = b;
     iz = 1-iz;
+    ShiftTimer.Start();
     for(int s=1;s<nshift;s++){
       if((!converged[s])){
 	RealD z0 = z[s][1-iz];
@@ -215,6 +240,7 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
       }
     }
+    ShiftTimer.Stop();
     
     for(int s=0;s<nshift;s++){
       int ss = s;
@@ -257,6 +283,9 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
     
     if ( all_converged ){
 
+    SolverTimer.Stop();
+
+
       std::cout<<GridLogMessage<< "CGMultiShift: All shifts have converged iteration "<<k<<std::endl;
       std::cout<<GridLogMessage<< "CGMultiShift: Checking solutions"<<std::endl;
       
@@ -269,8 +298,19 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	RealD cn = norm2(src);
 	std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
       }
+
+      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
+      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+
+      IterationsToComplete = k;	
+
       return;
     }
+
+   
   }
   // ugly hack
   std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;

lib/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -57,7 +57,8 @@ void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, i
       
   parallel_region
   {
-    std::vector < vobj > B(Nm); // Thread private
+
+    std::vector < vobj , commAllocator<vobj> > B(Nm); // Thread private
        
     parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
       for(int j=j0; j<j1; ++j) B[j]=0.;

lib/communicator/SharedMemoryMPI.cc

@@ -114,19 +114,151 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   assert(WorldNode!=-1);
   _ShmSetup=1;
 }
-
-void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+// Gray encode support 
+int BinaryToGray (int  binary) {
+  int gray = (binary>>1)^binary;
+  return gray;
+}
+int Log2Size(int TwoToPower,int MAXLOG2)
 {
-  ////////////////////////////////////////////////////////////////
-  // Assert power of two shm_size.
-  ////////////////////////////////////////////////////////////////
   int log2size = -1;
-  for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){  
-    if ( (0x1<<i) == WorldShmSize ) {
+  for(int i=0;i<=MAXLOG2;i++){
+    if ( (0x1<<i) == TwoToPower ) {
       log2size = i;
       break;
     }
   }
+  return log2size;
+}
+void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+#undef HYPERCUBE 
+#ifdef HYPERCUBE
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
+  assert(log2size != -1);
+
+  ////////////////////////////////////////////////////////////////
+  // Identify the hypercube coordinate of this node using hostname
+  ////////////////////////////////////////////////////////////////
+  // n runs 0...7 9...16 18...25 27...34     (8*4)  5 bits
+  // i runs 0..7                                    3 bits
+  // r runs 0..3                                    2 bits
+  // 2^10 = 1024 nodes
+  const int maxhdim = 10; 
+  std::vector<int> HyperCubeCoords(maxhdim,0);
+  std::vector<int> RootHyperCubeCoords(maxhdim,0);
+  int R;
+  int I;
+  int N;
+  const int namelen = _POSIX_HOST_NAME_MAX;
+  char name[namelen];
+
+  // Parse ICE-XA hostname to get hypercube location
+  gethostname(name,namelen);
+  int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
+  assert(nscan==3);
+
+  int nlo = N%9;
+  int nhi = N/9;
+  uint32_t hypercoor = (R<<8)|(I<<5)|(nhi<<3)|nlo ;
+  uint32_t rootcoor  = hypercoor;
+
+  //////////////////////////////////////////////////////////////////
+  // Print debug info
+  //////////////////////////////////////////////////////////////////
+  for(int d=0;d<maxhdim;d++){
+    HyperCubeCoords[d] = (hypercoor>>d)&0x1;
+  }
+
+  std::string hname(name);
+  std::cout << "hostname "<<hname<<std::endl;
+  std::cout << "R " << R << " I " << I << " N "<< N<<
+            << " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
+
+  //////////////////////////////////////////////////////////////////
+  // broadcast node 0's base coordinate for this partition.
+  //////////////////////////////////////////////////////////////////
+  MPI_Bcast(&rootcoor, sizeof(rootcoor), MPI_BYTE, 0, WorldComm); 
+  hypercoor=hypercoor-rootcoor;
+  assert(hypercoor<WorldSize);
+  assert(hypercoor>=0);
+
+  //////////////////////////////////////
+  // Printing
+  //////////////////////////////////////
+  for(int d=0;d<maxhdim;d++){
+    HyperCubeCoords[d] = (hypercoor>>d)&0x1;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Identify subblock of ranks on node spreading across dims
+  // in a maximally symmetrical way
+  ////////////////////////////////////////////////////////////////
+  int ndimension              = processors.size();
+  std::vector<int> processor_coor(ndimension);
+  std::vector<int> WorldDims = processors;   std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
+  std::vector<int> ShmCoor  (ndimension);    std::vector<int> NodeCoor (ndimension);    std::vector<int> WorldCoor(ndimension);
+  std::vector<int> HyperCoor(ndimension);
+  int dim = 0;
+  for(int l2=0;l2<log2size;l2++){
+    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
+    ShmDims[dim]*=2;
+    dim=(dim+1)%ndimension;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Establish torus of processes and nodes with sub-blockings
+  ////////////////////////////////////////////////////////////////
+  for(int d=0;d<ndimension;d++){
+    NodeDims[d] = WorldDims[d]/ShmDims[d];
+  }
+  ////////////////////////////////////////////////////////////////
+  // Map Hcube according to physical lattice 
+  // must partition. Loop over dims and find out who would join.
+  ////////////////////////////////////////////////////////////////
+  int hcoor = hypercoor;
+  for(int d=0;d<ndimension;d++){
+     int bits = Log2Size(NodeDims[d],MAXLOG2RANKSPERNODE);
+     int msk  = (0x1<<bits)-1;
+     HyperCoor[d]=hcoor & msk;  
+     HyperCoor[d]=BinaryToGray(HyperCoor[d]); // Space filling curve magic
+     hcoor = hcoor >> bits;
+  } 
+  ////////////////////////////////////////////////////////////////
+  // Check processor counts match
+  ////////////////////////////////////////////////////////////////
+  int Nprocessors=1;
+  for(int i=0;i<ndimension;i++){
+    Nprocessors*=processors[i];
+  }
+  assert(WorldSize==Nprocessors);
+
+  ////////////////////////////////////////////////////////////////
+  // Establish mapping between lexico physics coord and WorldRank
+  ////////////////////////////////////////////////////////////////
+  int rank;
+
+  Lexicographic::CoorFromIndexReversed(NodeCoor,WorldNode   ,NodeDims);
+
+  for(int d=0;d<ndimension;d++) NodeCoor[d]=HyperCoor[d];
+
+  Lexicographic::CoorFromIndexReversed(ShmCoor ,WorldShmRank,ShmDims);
+  for(int d=0;d<ndimension;d++) WorldCoor[d] = NodeCoor[d]*ShmDims[d]+ShmCoor[d];
+  Lexicographic::IndexFromCoorReversed(WorldCoor,rank,WorldDims);
+
+  /////////////////////////////////////////////////////////////////
+  // Build the new communicator
+  /////////////////////////////////////////////////////////////////
+  int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
+  assert(ierr==0);
+#else 
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
   assert(log2size != -1);
 
   ////////////////////////////////////////////////////////////////
@@ -175,7 +307,69 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
+#endif
 }
+////////////////////////////////////////////////////////////////////////////////////////////
+// SHMGET
+////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMGET
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  std::cout << "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+  std::vector<int> shmids(WorldShmSize);
+
+  if ( WorldShmRank == 0 ) {
+    for(int r=0;r<WorldShmSize;r++){
+      size_t size = bytes;
+      key_t key   = IPC_PRIVATE;
+      int flags = IPC_CREAT | SHM_R | SHM_W;
+#ifdef SHM_HUGETLB
+      if (Hugepages) flags|=SHM_HUGETLB;
+#endif
+      if ((shmids[r]= shmget(key,size, flags)) ==-1) {
+        int errsv = errno;
+        printf("Errno %d\n",errsv);
+        printf("key   %d\n",key);
+        printf("size  %lld\n",size);
+        printf("flags %d\n",flags);
+        perror("shmget");
+        exit(1);
+      }
+    }
+  }
+  MPI_Barrier(WorldShmComm);
+  MPI_Bcast(&shmids[0],WorldShmSize*sizeof(int),MPI_BYTE,0,WorldShmComm);
+  MPI_Barrier(WorldShmComm);
+
+  for(int r=0;r<WorldShmSize;r++){
+    WorldShmCommBufs[r] = (uint64_t *)shmat(shmids[r], NULL,0);
+    if (WorldShmCommBufs[r] == (uint64_t *)-1) {
+      perror("Shared memory attach failure");
+      shmctl(shmids[r], IPC_RMID, NULL);
+      exit(2);
+    }
+  }
+  MPI_Barrier(WorldShmComm);
+  ///////////////////////////////////
+  // Mark for clean up
+  ///////////////////////////////////
+  for(int r=0;r<WorldShmSize;r++){
+    shmctl(shmids[r], IPC_RMID,(struct shmid_ds *)NULL);
+  }
+  MPI_Barrier(WorldShmComm);
+
+  _ShmAlloc=1;
+  _ShmAllocBytes  = bytes;
+}
+#endif
  
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
@@ -183,7 +377,7 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<<std::endl;
+  std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -193,7 +387,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   WorldShmCommBufs.resize(WorldShmSize);
   
   ////////////////////////////////////////////////////////////////////////////////////////////
-  // Hugetlbf and others map filesystems as mappable huge pages
+  // Hugetlbfs and others map filesystems as mappable huge pages
   ////////////////////////////////////////////////////////////////////////////////////////////
   char shm_name [NAME_MAX];
   for(int r=0;r<WorldShmSize;r++){
@@ -344,6 +538,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 }
 #endif
 
+
+
+
   ////////////////////////////////////////////////////////
   // Global shared functionality finished
   // Now move to per communicator functionality

lib/lattice/Lattice_arith.h

@@ -244,19 +244,11 @@ namespace Grid {
 
   template<class sobj,class vobj> strong_inline
   RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
-    conformable(ret,x);
-    conformable(x,y);
-    axpy(ret,a,x,y);
-    return norm2(ret);
+    return axpy_norm_fast(ret,a,x,y);
   }
   template<class sobj,class vobj> strong_inline
   RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
-    ret.checkerboard = x.checkerboard;
-    conformable(ret,x);
-    conformable(x,y);
-    axpby(ret,a,b,x,y);
-    return norm2(ret); // FIXME implement parallel norm in ss loop
+    return axpby_norm_fast(ret,a,b,x,y);
   }
 
 }

lib/lattice/Lattice_reduction.h

@@ -33,7 +33,7 @@ namespace Grid {
   // Deterministic Reduction operations
   ////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
-  ComplexD nrm = innerProduct(arg,arg);
+  auto nrm = innerProduct(arg,arg);
   return std::real(nrm); 
 }
 
@@ -43,32 +43,85 @@ inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &righ
 {
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
-  scalar_type  nrm;
-  
   GridBase *grid = left._grid;
+  const int pad = 8;
 
-  std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
+  ComplexD  inner;
+  Vector<ComplexD> sumarray(grid->SumArraySize()*pad);
 
   parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
     int nwork, mywork, myoff;
     GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
     
-    decltype(innerProductD(left._odata[0],right._odata[0])) vnrm=zero; // private to thread; sub summation
+    decltype(innerProductD(left._odata[0],right._odata[0])) vinner=zero; // private to thread; sub summation
     for(int ss=myoff;ss<mywork+myoff; ss++){
-      vnrm = vnrm + innerProductD(left._odata[ss],right._odata[ss]);
+      vinner = vinner + innerProductD(left._odata[ss],right._odata[ss]);
     }
-    sumarray[thr]=TensorRemove(vnrm) ;
+    // All threads sum across SIMD; reduce serial work at end
+    // one write per cacheline with streaming store
+    ComplexD tmp = Reduce(TensorRemove(vinner)) ;
+    vstream(sumarray[thr*pad],tmp);
   }
   
-  vector_type vvnrm; vvnrm=zero;  // sum across threads
+  inner=0.0;
   for(int i=0;i<grid->SumArraySize();i++){
-    vvnrm = vvnrm+sumarray[i];
+    inner = inner+sumarray[i*pad];
   } 
-  nrm = Reduce(vvnrm);// sum across simd
-  right._grid->GlobalSum(nrm);
+  right._grid->GlobalSum(inner);
+  return inner;
+}
+
+/////////////////////////
+// Fast axpby_norm
+// z = a x + b y
+// return norm z
+/////////////////////////
+template<class sobj,class vobj> strong_inline RealD 
+axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y) 
+{
+  sobj one(1.0);
+  return axpby_norm_fast(z,a,one,x,y);
+}
+
+template<class sobj,class vobj> strong_inline RealD 
+axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) 
+{
+  const int pad = 8;
+  z.checkerboard = x.checkerboard;
+  conformable(z,x);
+  conformable(x,y);
+
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_typeD vector_type;
+  RealD  nrm;
+  
+  GridBase *grid = x._grid;
+  
+  Vector<RealD> sumarray(grid->SumArraySize()*pad);
+  
+  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
+    int nwork, mywork, myoff;
+    GridThread::GetWork(x._grid->oSites(),thr,mywork,myoff);
+    
+    // private to thread; sub summation
+    decltype(innerProductD(z._odata[0],z._odata[0])) vnrm=zero; 
+    for(int ss=myoff;ss<mywork+myoff; ss++){
+      vobj tmp = a*x._odata[ss]+b*y._odata[ss];
+      vnrm = vnrm + innerProductD(tmp,tmp);
+      vstream(z._odata[ss],tmp);
+    }
+    vstream(sumarray[thr*pad],real(Reduce(TensorRemove(vnrm)))) ;
+  }
+  
+  nrm = 0.0; // sum across threads; linear in thread count but fast
+  for(int i=0;i<grid->SumArraySize();i++){
+    nrm = nrm+sumarray[i*pad];
+  } 
+  z._grid->GlobalSum(nrm);
   return nrm; 
 }
 
+ 
 template<class Op,class T1>
 inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
   ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object

lib/lattice/Lattice_rng.h

@@ -158,10 +158,19 @@ namespace Grid {
       // tens of seconds per trajectory so this is clean in all reasonable cases,
       // and margin of safety is orders of magnitude.
       // We could hack Sitmo to skip in the higher order words of state if necessary
+      //
+      // Replace with 2^30 ; avoid problem on large volumes
+      //
       /////////////////////////////////////////////////////////////////////////////////////
       //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
+      const int shift = 30;
+
       uint64_t skip = site;
-      skip = skip<<40;
+
+      skip = skip<<shift;
+
+      assert((skip >> shift)==site); // check for overflow
+
       eng.discard(skip);
       //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
     } 

lib/parallelIO/BinaryIO.h

@@ -263,7 +263,7 @@ PARALLEL_CRITICAL
 			      GridBase *grid,
 			      std::vector<fobj> &iodata,
 			      std::string file,
-			      uint64_t offset,
+			      uint64_t& offset,
 			      const std::string &format, int control,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
@@ -431,14 +431,20 @@ PARALLEL_CRITICAL
           MPI_Abort(MPI_COMM_WORLD, 1); //assert(ierr == 0);
         }
 
-        std::cout << GridLogDebug << "MPI read I/O set view " << file << std::endl;
+        std::cout << GridLogDebug << "MPI write I/O set view " << file << std::endl;
         ierr = MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);
         assert(ierr == 0);
 
-        std::cout << GridLogDebug << "MPI read I/O write all " << file << std::endl;
+        std::cout << GridLogDebug << "MPI write I/O write all " << file << std::endl;
         ierr = MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);
         assert(ierr == 0);
 
+        MPI_Offset os;
+        MPI_File_get_position(fh, &os);
+        MPI_File_get_byte_offset(fh, os, &disp);
+        offset = disp;
+
+
         MPI_File_close(&fh);
         MPI_Type_free(&fileArray);
         MPI_Type_free(&localArray);
@@ -448,7 +454,7 @@ PARALLEL_CRITICAL
       } else { 
 
         std::cout << GridLogMessage << "IOobject: C++ write I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
+                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
         
 	std::ofstream fout; 
 	fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
@@ -495,6 +501,7 @@ PARALLEL_CRITICAL
 	  exit(1);
 #endif
 	}
+  offset  = fout.tellp();
 	fout.close();
       }
       timer.Stop();
@@ -699,7 +706,6 @@ PARALLEL_CRITICAL
 
     IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
 	     nersc_csum,scidac_csuma,scidac_csumb);
-
     iodata.resize(1);
     {
       std::vector<RngStateType> tmp(RngStateCount);

lib/parallelIO/IldgIO.h

@@ -182,6 +182,11 @@ class GridLimeReader : public BinaryIO {
    {
      filename= _filename;
      File = fopen(filename.c_str(), "r");
+     if (File == nullptr)
+     {
+       std::cerr << "cannot open file '" << filename << "'" << std::endl;
+       abort();
+     }
      LimeR = limeCreateReader(File);
    }
    /////////////////////////////////////////////

lib/perfmon/Timer.h

@@ -49,7 +49,8 @@ inline double usecond(void) {
 
 typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
-typedef  std::chrono::milliseconds          GridTime;
+typedef  std::chrono::milliseconds          GridMillisecs;
+typedef  std::chrono::microseconds          GridTime;
 typedef  std::chrono::microseconds          GridUsecs;
 
 inline std::ostream& operator<< (std::ostream & stream, const std::chrono::milliseconds & time)
@@ -57,6 +58,11 @@ inline std::ostream& operator<< (std::ostream & stream, const std::chrono::milli
   stream << time.count()<<" ms";
   return stream;
 }
+inline std::ostream& operator<< (std::ostream & stream, const std::chrono::microseconds & time)
+{
+  stream << time.count()<<" usec";
+  return stream;
+}
  
 class GridStopWatch {
 private:

lib/qcd/action/fermion/FermionOperator.h

@@ -63,9 +63,12 @@ namespace Grid {
       virtual RealD  M    (const FermionField &in, FermionField &out)=0;
       virtual RealD  Mdag (const FermionField &in, FermionField &out)=0;
 
-      // half checkerboard operaions
+      // Query the even even properties to make algorithmic decisions
       virtual int    ConstEE(void) { return 1; }; // clover returns zero as EE depends on gauge field
+      virtual int    isTrivialEE(void) { return 0; };
+      virtual RealD  Mass(void) {return 0.0;};
 
+      // half checkerboard operaions
       virtual void   Meooe       (const FermionField &in, FermionField &out)=0;
       virtual void   MeooeDag    (const FermionField &in, FermionField &out)=0;
       virtual void   Mooee       (const FermionField &in, FermionField &out)=0;

lib/qcd/action/fermion/FermionOperatorImpl.h

@@ -765,6 +765,11 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
       reg = memory;
     }
 
+    inline void InsertGaugeField(DoubledGaugeField &U_ds,
+				 const GaugeLinkField &U,int mu)
+    {
+      PokeIndex<LorentzIndex>(U_ds, U, mu);
+    }
     inline void DoubleStore(GridBase *GaugeGrid,
 			    DoubledGaugeField &UUUds, // for Naik term
 			    DoubledGaugeField &Uds,
@@ -803,8 +808,10 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
 	U    = U    *phases;
 	Udag = Udag *phases;
 
-	PokeIndex<LorentzIndex>(Uds, U, mu);
-	PokeIndex<LorentzIndex>(Uds, Udag, mu + 4);
+	InsertGaugeField(Uds,U,mu);
+	InsertGaugeField(Uds,Udag,mu+4);
+	//	PokeIndex<LorentzIndex>(Uds, U, mu);
+	//	PokeIndex<LorentzIndex>(Uds, Udag, mu + 4);
 
 	// 3 hop based on thin links. Crazy huh ?
 	U  = PeekIndex<LorentzIndex>(Uthin, mu);
@@ -816,8 +823,8 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
 	UUU    = UUU    *phases;
 	UUUdag = UUUdag *phases;
 
-	PokeIndex<LorentzIndex>(UUUds, UUU, mu);
-	PokeIndex<LorentzIndex>(UUUds, UUUdag, mu+4);
+	InsertGaugeField(UUUds,UUU,mu);
+	InsertGaugeField(UUUds,UUUdag,mu+4);
 
       }
     }
@@ -910,6 +917,23 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
       mac(&phi(), &UU(), &chi());
     }
       
+    inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
+    {
+      GridBase *GaugeGrid = U_ds._grid;
+      parallel_for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
+
+	SiteScalarGaugeLink   ScalarU;
+	SiteDoubledGaugeField ScalarUds;
+	
+	std::vector<int> lcoor;
+	GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
+	peekLocalSite(ScalarUds, U_ds, lcoor);
+	
+	peekLocalSite(ScalarU, U, lcoor);
+	ScalarUds(mu) = ScalarU();
+	
+      }
+    }
     inline void DoubleStore(GridBase *GaugeGrid,
 			    DoubledGaugeField &UUUds, // for Naik term
 			    DoubledGaugeField &Uds,
@@ -951,23 +975,8 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
 	U    = U    *phases;
 	Udag = Udag *phases;
 
-
-	for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
-	  SiteScalarGaugeLink   ScalarU;
-	  SiteDoubledGaugeField ScalarUds;
-	  
-	  std::vector<int> lcoor;
-	  GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
-	  peekLocalSite(ScalarUds, Uds, lcoor);
-
-	  peekLocalSite(ScalarU, U, lcoor);
-	  ScalarUds(mu) = ScalarU();
-
-	  peekLocalSite(ScalarU, Udag, lcoor);
-	  ScalarUds(mu + 4) = ScalarU();
-
-	  pokeLocalSite(ScalarUds, Uds, lcoor);
-	}
+	InsertGaugeField(Uds,U,mu);
+	InsertGaugeField(Uds,Udag,mu+4);
 
 	// 3 hop based on thin links. Crazy huh ?
 	U  = PeekIndex<LorentzIndex>(Uthin, mu);
@@ -979,24 +988,8 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
 	UUU    = UUU    *phases;
 	UUUdag = UUUdag *phases;
 
-	for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
-
-	  SiteScalarGaugeLink  ScalarU;
-	  SiteDoubledGaugeField ScalarUds;
-	  
-	  std::vector<int> lcoor;
-	  GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
-      
-	  peekLocalSite(ScalarUds, UUUds, lcoor);
-
-	  peekLocalSite(ScalarU, UUU, lcoor);
-	  ScalarUds(mu) = ScalarU();
-
-	  peekLocalSite(ScalarU, UUUdag, lcoor);
-	  ScalarUds(mu + 4) = ScalarU();
-	  
-	  pokeLocalSite(ScalarUds, UUUds, lcoor);
-	}
+	InsertGaugeField(UUUds,UUU,mu);
+	InsertGaugeField(UUUds,UUUdag,mu+4);
 
       }
     }

lib/qcd/action/fermion/ImprovedStaggeredFermion.cc

@@ -44,6 +44,7 @@ ImprovedStaggeredFermionStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1, 3, 3,
 template <class Impl>
 ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid, 
 							 RealD _mass,
+							 RealD _c1, RealD _c2,RealD _u0,
 							 const ImplParams &p)
     : Kernels(p),
       _grid(&Fgrid),
@@ -62,6 +63,16 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GridCartesian &Fgrid, G
       UUUmuOdd(&Hgrid) ,
       _tmp(&Hgrid)
 {
+  int vol4;
+  int LLs=1;
+  c1=_c1;
+  c2=_c2;
+  u0=_u0;
+  vol4= _grid->oSites();
+  Stencil.BuildSurfaceList(LLs,vol4);
+  vol4= _cbgrid->oSites();
+  StencilEven.BuildSurfaceList(LLs,vol4);
+  StencilOdd.BuildSurfaceList(LLs,vol4);
 }
 
 template <class Impl>
@@ -69,22 +80,10 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Uthin, Gau
 							 GridRedBlackCartesian &Hgrid, RealD _mass,
 							 RealD _c1, RealD _c2,RealD _u0,
 							 const ImplParams &p)
-  : ImprovedStaggeredFermion(Fgrid,Hgrid,_mass,p)
+  : ImprovedStaggeredFermion(Fgrid,Hgrid,_mass,_c1,_c2,_u0,p)
 {
-  c1=_c1;
-  c2=_c2;
-  u0=_u0;
   ImportGauge(_Uthin,_Ufat);
 }
-template <class Impl>
-ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Uthin,GaugeField &_Utriple, GaugeField &_Ufat, GridCartesian &Fgrid,
-							 GridRedBlackCartesian &Hgrid, RealD _mass,
-							 const ImplParams &p)
-  : ImprovedStaggeredFermion(Fgrid,Hgrid,_mass,p)
-{
-  ImportGaugeSimple(_Utriple,_Ufat);
-}
-
 
   ////////////////////////////////////////////////////////////
   // Momentum space propagator should be 
@@ -98,11 +97,6 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Uthin,Gaug
   // of above link to implmement fourier based solver.
   ////////////////////////////////////////////////////////////
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin) 
-{
-  ImportGauge(_Uthin,_Uthin);
-};
-template <class Impl>
 void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const GaugeField &_Utriple,const GaugeField &_Ufat) 
 {
   /////////////////////////////////////////////////////////////////
@@ -125,6 +119,20 @@ void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const GaugeField &_Utripl
     PokeIndex<LorentzIndex>(Umu, -U, mu+4);
 
   }
+  CopyGaugeCheckerboards();
+}
+template <class Impl>
+void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U) 
+{
+
+  Umu   = _U;
+  UUUmu = _UUU;
+  CopyGaugeCheckerboards();
+}
+
+template <class Impl>
+void ImprovedStaggeredFermion<Impl>::CopyGaugeCheckerboards(void)
+{
   pickCheckerboard(Even, UmuEven,  Umu);
   pickCheckerboard(Odd,  UmuOdd ,  Umu);
   pickCheckerboard(Even, UUUmuEven,UUUmu);
@@ -160,10 +168,7 @@ void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin,const
     PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
   }
 
-  pickCheckerboard(Even, UmuEven, Umu);
-  pickCheckerboard(Odd,  UmuOdd , Umu);
-  pickCheckerboard(Even, UUUmuEven, UUUmu);
-  pickCheckerboard(Odd,   UUUmuOdd, UUUmu);
+  CopyGaugeCheckerboards();
 }
 
 /////////////////////////////
@@ -322,6 +327,7 @@ void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionF
 
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) {
+  DhopCalls+=2;
   conformable(in._grid, _grid);  // verifies full grid
   conformable(in._grid, out._grid);
 
@@ -332,6 +338,7 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
 
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) {
+  DhopCalls+=1;
   conformable(in._grid, _cbgrid);    // verifies half grid
   conformable(in._grid, out._grid);  // drops the cb check
 
@@ -343,6 +350,7 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
 
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) {
+  DhopCalls+=1;
   conformable(in._grid, _cbgrid);    // verifies half grid
   conformable(in._grid, out._grid);  // drops the cb check
 
@@ -374,25 +382,193 @@ void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder
 						  DoubledGaugeField &U,
 						  DoubledGaugeField &UUU,
 						  const FermionField &in,
-						  FermionField &out, int dag) {
+						  FermionField &out, int dag) 
+{
+#ifdef GRID_OMP
+  if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
+    DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
+  else
+#endif
+    DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
+}
+template <class Impl>
+void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
+								 DoubledGaugeField &U,
+								 DoubledGaugeField &UUU,
+								 const FermionField &in,
+								 FermionField &out, int dag) 
+{
+#ifdef GRID_OMP
+  Compressor compressor; 
+  int len =  U._grid->oSites();
+  const int LLs =  1;
+
+  DhopTotalTime   -= usecond();
+
+  DhopFaceTime    -= usecond();
+  st.Prepare();
+  st.HaloGather(in,compressor);
+  st.CommsMergeSHM(compressor);
+  DhopFaceTime    += usecond();
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Ugly explicit thread mapping introduced for OPA reasons.
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime    -= usecond();
+#pragma omp parallel 
+  {
+    int tid = omp_get_thread_num();
+    int nthreads = omp_get_num_threads();
+    int ncomms = CartesianCommunicator::nCommThreads;
+    if (ncomms == -1) ncomms = 1;
+    assert(nthreads > ncomms);
+
+    if (tid >= ncomms) {
+      nthreads -= ncomms;
+      int ttid  = tid - ncomms;
+      int n     = len;
+      int chunk = n / nthreads;
+      int rem   = n % nthreads;
+      int myblock, myn;
+      if (ttid < rem) {
+        myblock = ttid * chunk + ttid;
+        myn = chunk+1;
+      } else {
+        myblock = ttid*chunk + rem;
+        myn = chunk;
+      }
+
+      // do the compute
+      if (dag == DaggerYes) {
+        for (int ss = myblock; ss < myblock+myn; ++ss) {
+          int sU = ss;
+	  // Interior = 1; Exterior = 0; must implement for staggered
+          Kernels::DhopSiteDag(st,lo,U,UUU,st.CommBuf(),1,sU,in,out,1,0); 
+        }
+      } else {
+        for (int ss = myblock; ss < myblock+myn; ++ss) {
+	  // Interior = 1; Exterior = 0;
+          int sU = ss;
+          Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),1,sU,in,out,1,0);
+        }
+      }
+    } else {
+      st.CommunicateThreaded();
+    }
+  }
+  DhopComputeTime    += usecond();
+
+  // First to enter, last to leave timing
+  DhopFaceTime    -= usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime    -= usecond();
+
+  DhopComputeTime2    -= usecond();
+  if (dag == DaggerYes) {
+    int sz=st.surface_list.size();
+    parallel_for (int ss = 0; ss < sz; ss++) {
+      int sU = st.surface_list[ss];
+      Kernels::DhopSiteDag(st,lo,U,UUU,st.CommBuf(),1,sU,in,out,0,1);
+    }
+  } else {
+    int sz=st.surface_list.size();
+    parallel_for (int ss = 0; ss < sz; ss++) {
+      int sU = st.surface_list[ss];
+      Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),1,sU,in,out,0,1);
+    }
+  }
+  DhopComputeTime2    += usecond();
+#else
+  assert(0);
+#endif
+}
+
+
+template <class Impl>
+void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
+							     DoubledGaugeField &U,
+							     DoubledGaugeField &UUU,
+							     const FermionField &in,
+							     FermionField &out, int dag) 
+{
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
+  DhopTotalTime   -= usecond();
+
+  DhopCommTime    -= usecond();
   Compressor compressor;
   st.HaloExchange(in, compressor);
+  DhopCommTime    += usecond();
 
+  DhopComputeTime -= usecond();
   if (dag == DaggerYes) {
-    PARALLEL_FOR_LOOP
-    for (int sss = 0; sss < in._grid->oSites(); sss++) {
+    parallel_for (int sss = 0; sss < in._grid->oSites(); sss++) {
       Kernels::DhopSiteDag(st, lo, U, UUU, st.CommBuf(), 1, sss, in, out);
     }
   } else {
-    PARALLEL_FOR_LOOP
-    for (int sss = 0; sss < in._grid->oSites(); sss++) {
+    parallel_for (int sss = 0; sss < in._grid->oSites(); sss++) {
       Kernels::DhopSite(st, lo, U, UUU, st.CommBuf(), 1, sss, in, out);
     }
   }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
 };
 
+  ////////////////////////////////////////////////////////////////
+  // Reporting
+  ////////////////////////////////////////////////////////////////
+template<class Impl>
+void ImprovedStaggeredFermion<Impl>::Report(void) 
+{
+  std::vector<int> latt = GridDefaultLatt();          
+  RealD volume = 1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _grid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void ImprovedStaggeredFermion<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime   = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
+
 //////////////////////////////////////////////////////// 
 // Conserved current - not yet implemented.
 ////////////////////////////////////////////////////////

lib/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -49,6 +49,18 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////
@@ -105,25 +117,34 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
 
   void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
+  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
+                    const FermionField &in, FermionField &out, int dag);
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
+                    const FermionField &in, FermionField &out, int dag);
 
-  // Constructor
+  //////////////////////////////////////////////////////////////////////////
+  // Grid own interface Constructor
+  //////////////////////////////////////////////////////////////////////////
   ImprovedStaggeredFermion(GaugeField &_Uthin, GaugeField &_Ufat, GridCartesian &Fgrid,
 			   GridRedBlackCartesian &Hgrid, RealD _mass,
-			   RealD _c1=9.0/8.0, RealD _c2=-1.0/24.0,RealD _u0=1.0,
-			   const ImplParams &p = ImplParams());
-
-  ImprovedStaggeredFermion(GaugeField &_Uthin, GaugeField &_Utriple, GaugeField &_Ufat, GridCartesian &Fgrid,
-			   GridRedBlackCartesian &Hgrid, RealD _mass,
+			   RealD _c1, RealD _c2,RealD _u0,
 			   const ImplParams &p = ImplParams());
 
+  //////////////////////////////////////////////////////////////////////////
+  // MILC constructor no gauge fields
+  //////////////////////////////////////////////////////////////////////////
   ImprovedStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid, RealD _mass,
+			   RealD _c1=1.0, RealD _c2=1.0,RealD _u0=1.0,
 			   const ImplParams &p = ImplParams());
 
-
   // DoubleStore impl dependent
-  void ImportGaugeSimple(const GaugeField &_Utriple, const GaugeField &_Ufat);
+  void ImportGauge      (const GaugeField &_Uthin ) { assert(0); }
   void ImportGauge      (const GaugeField &_Uthin  ,const GaugeField &_Ufat);
-  void ImportGauge(const GaugeField &_Uthin);
+  void ImportGaugeSimple(const GaugeField &_UUU    ,const GaugeField &_U);
+  void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
+  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  DoubledGaugeField &GetUUU(void) { return UUUmu; };
+  void CopyGaugeCheckerboards(void);
 
   ///////////////////////////////////////////////////////////////
   // Data members require to support the functionality
@@ -132,7 +153,8 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
   //    protected:
  public:
   // any other parameters of action ???
-
+  virtual int   isTrivialEE(void) { return 1; };
+  virtual RealD Mass(void) { return mass; }
   RealD mass;
   RealD u0;
   RealD c1;

lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc

@@ -41,8 +41,7 @@ ImprovedStaggeredFermion5DStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1, 3,
 
   // 5d lattice for DWF.
 template<class Impl>
-ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GaugeField &_Uthin,GaugeField &_Ufat,
-							     GridCartesian         &FiveDimGrid,
+ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GridCartesian         &FiveDimGrid,
 							     GridRedBlackCartesian &FiveDimRedBlackGrid,
 							     GridCartesian         &FourDimGrid,
 							     GridRedBlackCartesian &FourDimRedBlackGrid,
@@ -121,16 +120,74 @@ ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GaugeField &_Uthin,
     assert(FiveDimGrid._simd_layout[0]        ==1);
 
   }
+  int LLs = FiveDimGrid._rdimensions[0];
+  int vol4= FourDimGrid.oSites();
+  Stencil.BuildSurfaceList(LLs,vol4);
 
-  // Allocate the required comms buffer
+  vol4=FourDimRedBlackGrid.oSites();
+  StencilEven.BuildSurfaceList(LLs,vol4);
+  StencilOdd.BuildSurfaceList(LLs,vol4);
+}
+template <class Impl>
+void ImprovedStaggeredFermion5D<Impl>::CopyGaugeCheckerboards(void)
+{
+  pickCheckerboard(Even, UmuEven,  Umu);
+  pickCheckerboard(Odd,  UmuOdd ,  Umu);
+  pickCheckerboard(Even, UUUmuEven,UUUmu);
+  pickCheckerboard(Odd,  UUUmuOdd, UUUmu);
+}
+template<class Impl>
+ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GaugeField &_Uthin,GaugeField &_Ufat,
+							     GridCartesian         &FiveDimGrid,
+							     GridRedBlackCartesian &FiveDimRedBlackGrid,
+							     GridCartesian         &FourDimGrid,
+							     GridRedBlackCartesian &FourDimRedBlackGrid,
+							     RealD _mass,
+							     RealD _c1,RealD _c2, RealD _u0,
+							     const ImplParams &p) :
+  ImprovedStaggeredFermion5D(FiveDimGrid,FiveDimRedBlackGrid,
+			     FourDimGrid,FourDimRedBlackGrid,
+			     _mass,_c1,_c2,_u0,p)
+{
   ImportGauge(_Uthin,_Ufat);
 }
 
+///////////////////////////////////////////////////
+// For MILC use; pass three link U's and 1 link U
+///////////////////////////////////////////////////
 template <class Impl>
-void ImprovedStaggeredFermion5D<Impl>::ImportGauge(const GaugeField &_Uthin) 
+void ImprovedStaggeredFermion5D<Impl>::ImportGaugeSimple(const GaugeField &_Utriple,const GaugeField &_Ufat) 
 {
-  ImportGauge(_Uthin,_Uthin);
-};
+  /////////////////////////////////////////////////////////////////
+  // Trivial import; phases and fattening and such like preapplied
+  /////////////////////////////////////////////////////////////////
+  for (int mu = 0; mu < Nd; mu++) {
+
+    auto U = PeekIndex<LorentzIndex>(_Utriple, mu);
+    Impl::InsertGaugeField(UUUmu,U,mu);
+
+    U = adj( Cshift(U, mu, -3));
+    Impl::InsertGaugeField(UUUmu,-U,mu+4);
+
+    U = PeekIndex<LorentzIndex>(_Ufat, mu);
+    Impl::InsertGaugeField(Umu,U,mu);
+
+    U = adj( Cshift(U, mu, -1));
+    Impl::InsertGaugeField(Umu,-U,mu+4);
+
+  }
+  CopyGaugeCheckerboards();
+}
+template <class Impl>
+void ImprovedStaggeredFermion5D<Impl>::ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U) 
+{
+  /////////////////////////////////////////////////////////////////
+  // Trivial import; phases and fattening and such like preapplied
+  /////////////////////////////////////////////////////////////////
+  Umu   = _U;
+  UUUmu = _UUU;
+  CopyGaugeCheckerboards();
+}
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::ImportGauge(const GaugeField &_Uthin,const GaugeField &_Ufat)
 {
@@ -159,10 +216,7 @@ void ImprovedStaggeredFermion5D<Impl>::ImportGauge(const GaugeField &_Uthin,cons
     PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
   }
 
-  pickCheckerboard(Even, UmuEven, Umu);
-  pickCheckerboard(Odd,  UmuOdd , Umu);
-  pickCheckerboard(Even, UUUmuEven, UUUmu);
-  pickCheckerboard(Odd,  UUUmuOdd, UUUmu);
+  CopyGaugeCheckerboards();
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopDir(const FermionField &in, FermionField &out,int dir5,int disp)
@@ -223,6 +277,162 @@ void ImprovedStaggeredFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
   assert(0);
 }
 
+/*CHANGE */
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
+						    DoubledGaugeField & U,DoubledGaugeField & UUU,
+						    const FermionField &in, FermionField &out,int dag)
+{
+#ifdef GRID_OMP
+  if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
+    DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
+  else
+#endif
+    DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
+}
+
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
+								   DoubledGaugeField & U,DoubledGaugeField & UUU,
+								   const FermionField &in, FermionField &out,int dag)
+{
+#ifdef GRID_OMP
+  //  assert((dag==DaggerNo) ||(dag==DaggerYes));
+
+  Compressor compressor; 
+
+  int LLs = in._grid->_rdimensions[0];
+  int len =  U._grid->oSites();
+
+  DhopFaceTime-=usecond();
+  st.Prepare();
+  st.HaloGather(in,compressor);
+  //  st.HaloExchangeOptGather(in,compressor); // Wilson compressor
+  st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+  DhopFaceTime+=usecond();
+
+  double ctime=0;
+  double ptime=0;
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Ugly explicit thread mapping introduced for OPA reasons.
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
+  {
+    int tid = omp_get_thread_num();
+    int nthreads = omp_get_num_threads();
+    int ncomms = CartesianCommunicator::nCommThreads;
+    if (ncomms == -1) ncomms = 1;
+    assert(nthreads > ncomms);
+    if (tid >= ncomms) {
+      double start = usecond();
+      nthreads -= ncomms;
+      int ttid  = tid - ncomms;
+      int n     = U._grid->oSites(); // 4d vol
+      int chunk = n / nthreads;
+      int rem   = n % nthreads;
+      int myblock, myn;
+      if (ttid < rem) {
+        myblock = ttid * chunk + ttid;
+        myn = chunk+1;
+      } else {
+        myblock = ttid*chunk + rem;
+        myn = chunk;
+      }
+
+      // do the compute
+      if (dag == DaggerYes) {
+        for (int ss = myblock; ss < myblock+myn; ++ss) {
+          int sU = ss;
+	  // Interior = 1; Exterior = 0; must implement for staggered
+          Kernels::DhopSiteDag(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out,1,0); //<---------
+        }
+      } else {
+        for (int ss = myblock; ss < myblock+myn; ++ss) {
+	  // Interior = 1; Exterior = 0;
+          int sU = ss;
+          Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out,1,0); //<------------
+        }
+      }
+        ptime = usecond() - start;
+    } else {
+      double start = usecond();
+      st.CommunicateThreaded();
+      ctime = usecond() - start;
+    }
+  }
+  DhopCommTime += ctime;
+  DhopComputeTime+=ptime;
+
+  // First to enter, last to leave timing
+  st.CollateThreads();
+
+  DhopFaceTime-=usecond();
+  st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
+
+  DhopComputeTime2-=usecond();
+  if (dag == DaggerYes) {
+    int sz=st.surface_list.size();
+    parallel_for (int ss = 0; ss < sz; ss++) {
+      int sU = st.surface_list[ss];
+      Kernels::DhopSiteDag(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out,0,1); //<----------
+    }
+  } else {
+    int sz=st.surface_list.size();
+    parallel_for (int ss = 0; ss < sz; ss++) {
+      int sU = st.surface_list[ss];
+      Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out,0,1);//<----------
+    }
+  }
+  DhopComputeTime2+=usecond();
+#else
+  assert(0);
+#endif
+
+}
+
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
+						    DoubledGaugeField & U,DoubledGaugeField & UUU,
+						    const FermionField &in, FermionField &out,int dag)
+{
+  Compressor compressor;
+  int LLs = in._grid->_rdimensions[0];
+
+
+
+ //double t1=usecond();
+  DhopTotalTime -= usecond();
+  DhopCommTime -= usecond();
+  st.HaloExchange(in,compressor);
+  DhopCommTime += usecond();
+  
+  DhopComputeTime -= usecond();
+  // Dhop takes the 4d grid from U, and makes a 5d index for fermion
+  if (dag == DaggerYes) {
+    parallel_for (int ss = 0; ss < U._grid->oSites(); ss++) {
+      int sU=ss;
+      Kernels::DhopSiteDag(st, lo, U, UUU, st.CommBuf(), LLs, sU,in, out);
+    }
+  } else {
+    parallel_for (int ss = 0; ss < U._grid->oSites(); ss++) {
+      int sU=ss;
+      Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out);
+    }
+  }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
+ //double t2=usecond();
+ //std::cout << __FILE__ << " " << __func__  << " Total Time " << DhopTotalTime << std::endl;
+ //std::cout << __FILE__ << " " << __func__  << " Total Time Org " << t2-t1 << std::endl;
+ //std::cout << __FILE__ << " " << __func__  << " Comml Time " << DhopCommTime << std::endl;
+ //std::cout << __FILE__ << " " << __func__  << " Compute Time " << DhopComputeTime << std::endl;
+
+}
+/*CHANGE END*/
+
+/* ORG
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
@@ -254,6 +464,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
   DhopComputeTime += usecond();
   DhopTotalTime   += usecond();
 }
+*/
 
 
 template<class Impl>
@@ -336,6 +547,9 @@ void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void)
   DhopTotalTime    = 0;
   DhopCommTime    = 0;
   DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+
   Stencil.ZeroCounters();
   StencilEven.ZeroCounters();
   StencilOdd.ZeroCounters();

lib/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -64,6 +64,8 @@ namespace QCD {
       double DhopCalls;
       double DhopCommTime;
       double DhopComputeTime;
+      double DhopComputeTime2;
+      double DhopFaceTime;
 
       ///////////////////////////////////////////////////////////////
       // Implement the abstract base
@@ -119,7 +121,27 @@ namespace QCD {
 		      FermionField &out,
 		      int dag);
     
+    void DhopInternalOverlappedComms(StencilImpl & st,
+		      LebesgueOrder &lo,
+		      DoubledGaugeField &U,
+		      DoubledGaugeField &UUU,
+		      const FermionField &in, 
+		      FermionField &out,
+		      int dag);
+
+    void DhopInternalSerialComms(StencilImpl & st,
+		      LebesgueOrder &lo,
+		      DoubledGaugeField &U,
+		      DoubledGaugeField &UUU,
+		      const FermionField &in, 
+		      FermionField &out,
+		      int dag);
+    
+    
     // Constructors
+    ////////////////////////////////////////////////////////////////////////////////////////////////
+    // Grid internal interface -- Thin link and fat link, with coefficients
+    ////////////////////////////////////////////////////////////////////////////////////////////////
     ImprovedStaggeredFermion5D(GaugeField &_Uthin,
 			       GaugeField &_Ufat,
 			       GridCartesian         &FiveDimGrid,
@@ -127,18 +149,38 @@ namespace QCD {
 			       GridCartesian         &FourDimGrid,
 			       GridRedBlackCartesian &FourDimRedBlackGrid,
 			       double _mass,
-			       RealD _c1=9.0/8.0, RealD _c2=-1.0/24.0,RealD _u0=1.0,
+			       RealD _c1, RealD _c2,RealD _u0,
+			       const ImplParams &p= ImplParams());
+    ////////////////////////////////////////////////////////////////////////////////////////////////
+    // MILC constructor ; triple links, no rescale factors; must be externally pre multiplied
+    ////////////////////////////////////////////////////////////////////////////////////////////////
+    ImprovedStaggeredFermion5D(GridCartesian         &FiveDimGrid,
+			       GridRedBlackCartesian &FiveDimRedBlackGrid,
+			       GridCartesian         &FourDimGrid,
+			       GridRedBlackCartesian &FourDimRedBlackGrid,
+			       double _mass,
+			       RealD _c1=1.0, RealD _c2=1.0,RealD _u0=1.0,
 			       const ImplParams &p= ImplParams());
 
-    // DoubleStore
-    void ImportGauge(const GaugeField &_U);
+    // DoubleStore gauge field in operator
+    void ImportGauge      (const GaugeField &_Uthin ) { assert(0); }
     void ImportGauge      (const GaugeField &_Uthin  ,const GaugeField &_Ufat);
+    void ImportGaugeSimple(const GaugeField &_UUU,const GaugeField &_U);
+    void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
+    // Give a reference; can be used to do an assignment or copy back out after import
+    // if Carleton wants to cache them and not use the ImportSimple
+    DoubledGaugeField &GetU(void)   { return Umu ; } ;
+    DoubledGaugeField &GetUUU(void) { return UUUmu; };
+    void CopyGaugeCheckerboards(void);
     
     ///////////////////////////////////////////////////////////////
     // Data members require to support the functionality
     ///////////////////////////////////////////////////////////////
   public:
 
+    virtual int   isTrivialEE(void) { return 1; };
+    virtual RealD Mass(void) { return mass; }
+    
     GridBase *_FourDimGrid;
     GridBase *_FourDimRedBlackGrid;
     GridBase *_FiveDimGrid;

lib/qcd/action/fermion/StaggeredKernels.cc

@@ -32,223 +32,241 @@ namespace Grid {
 namespace QCD {
 
 int StaggeredKernelsStatic::Opt= StaggeredKernelsStatic::OptGeneric;
+int StaggeredKernelsStatic::Comms = StaggeredKernelsStatic::CommsAndCompute;
+
+#define GENERIC_STENCIL_LEG(U,Dir,skew,multLink)		\
+  SE = st.GetEntry(ptype, Dir+skew, sF);			\
+  if (SE->_is_local ) {						\
+    if (SE->_permute) {						\
+      chi_p = χ						\
+      permute(chi,  in._odata[SE->_offset], ptype);		\
+    } else {							\
+      chi_p = &in._odata[SE->_offset];				\
+    }								\
+  } else {							\
+    chi_p = &buf[SE->_offset];					\
+  }								\
+  multLink(Uchi, U._odata[sU], *chi_p, Dir);			
+
+#define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink)		\
+  SE = st.GetEntry(ptype, Dir+skew, sF);			\
+  if (SE->_is_local ) {						\
+    if (SE->_permute) {						\
+      chi_p = χ						\
+      permute(chi,  in._odata[SE->_offset], ptype);		\
+    } else {							\
+      chi_p = &in._odata[SE->_offset];				\
+    }								\
+  } else if ( st.same_node[Dir] ) {				\
+    chi_p = &buf[SE->_offset];					\
+  }								\
+  if (SE->_is_local || st.same_node[Dir] ) {			\
+    multLink(Uchi, U._odata[sU], *chi_p, Dir);			\
+  }
+
+#define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink)		\
+  SE = st.GetEntry(ptype, Dir+skew, sF);			\
+  if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
+    nmu++;							\
+    chi_p = &buf[SE->_offset];					\
+    multLink(Uchi, U._odata[sU], *chi_p, Dir);			\
+  }
 
 template <class Impl>
 StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
 
-////////////////////////////////////////////
+////////////////////////////////////////////////////////////////////////////////////
 // Generic implementation; move to different file?
-////////////////////////////////////////////
-
+// Int, Ext, Int+Ext cases for comms overlap
+////////////////////////////////////////////////////////////////////////////////////
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteDepth(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
-					   SiteSpinor *buf, int sF,
-					   int sU, const FermionField &in, SiteSpinor &out,int threeLink) {
+void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+					     DoubledGaugeField &U, DoubledGaugeField &UUU,
+					     SiteSpinor *buf, int LLs, int sU, 
+					     const FermionField &in, FermionField &out, int dag) {
   const SiteSpinor *chi_p;
   SiteSpinor chi;
   SiteSpinor Uchi;
   StencilEntry *SE;
   int ptype;
-  int skew = 0;
-  if (threeLink) skew=8;
-  ///////////////////////////
-  // Xp
-  ///////////////////////////
+  int skew;
 
-  SE = st.GetEntry(ptype, Xp+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
+  for(int s=0;s<LLs;s++){
+    int sF=LLs*sU+s;
+    skew = 0;
+    GENERIC_STENCIL_LEG(U,Xp,skew,Impl::multLink);
+    GENERIC_STENCIL_LEG(U,Yp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(U,Zp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(U,Tp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(U,Xm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(U,Ym,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(U,Zm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(U,Tm,skew,Impl::multLinkAdd);
+    skew=8;
+    GENERIC_STENCIL_LEG(UUU,Xp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Yp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Zp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Tp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Xm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Ym,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Zm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG(UUU,Tm,skew,Impl::multLinkAdd);
+    if ( dag ) { 
+      Uchi = - Uchi;
     } 
-  } else {
-    chi_p = &buf[SE->_offset];
+    vstream(out._odata[sF], Uchi);
   }
-  Impl::multLink(Uchi, U._odata[sU], *chi_p, Xp);
-
-  ///////////////////////////
-  // Yp
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Yp+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Yp);
-
-  ///////////////////////////
-  // Zp
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Zp+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Zp);
-
-  ///////////////////////////
-  // Tp
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Tp+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Tp);
-
-  ///////////////////////////
-  // Xm
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Xm+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Xm);
-
-  ///////////////////////////
-  // Ym
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Ym+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Ym);
-
-  ///////////////////////////
-  // Zm
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Zm+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Zm);
-
-  ///////////////////////////
-  // Tm
-  ///////////////////////////
-  SE = st.GetEntry(ptype, Tm+skew, sF);
-  if (SE->_is_local) {
-    if (SE->_permute) {
-      chi_p = &chi;
-      permute(chi,  in._odata[SE->_offset], ptype);
-    } else {
-      chi_p = &in._odata[SE->_offset];
-    }
-  } else {
-    chi_p = &buf[SE->_offset];
-  }
-  Impl::multLinkAdd(Uchi, U._odata[sU], *chi_p, Tm);
-
-  vstream(out, Uchi);
 };
 
+  ///////////////////////////////////////////////////
+  // Only contributions from interior of our node
+  ///////////////////////////////////////////////////
+template <class Impl>
+void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+						DoubledGaugeField &U, DoubledGaugeField &UUU,
+						SiteSpinor *buf, int LLs, int sU, 
+						const FermionField &in, FermionField &out,int dag) {
+  const SiteSpinor *chi_p;
+  SiteSpinor chi;
+  SiteSpinor Uchi;
+  StencilEntry *SE;
+  int ptype;
+  int skew ;
+
+  for(int s=0;s<LLs;s++){
+    int sF=LLs*sU+s;
+    skew = 0;
+    Uchi=zero;
+    GENERIC_STENCIL_LEG_INT(U,Xp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Yp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Zp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Tp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Xm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Ym,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Zm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(U,Tm,skew,Impl::multLinkAdd);
+    skew=8;
+    GENERIC_STENCIL_LEG_INT(UUU,Xp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Yp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Zp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Tp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Xm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Ym,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Zm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_INT(UUU,Tm,skew,Impl::multLinkAdd);
+    if ( dag ) {
+      Uchi = - Uchi;
+    }
+    vstream(out._odata[sF], Uchi);
+  }
+};
+
+
+  ///////////////////////////////////////////////////
+  // Only contributions from exterior of our node
+  ///////////////////////////////////////////////////
+template <class Impl>
+void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+						DoubledGaugeField &U, DoubledGaugeField &UUU,
+						SiteSpinor *buf, int LLs, int sU,
+						const FermionField &in, FermionField &out,int dag) {
+  const SiteSpinor *chi_p;
+  SiteSpinor chi;
+  SiteSpinor Uchi;
+  StencilEntry *SE;
+  int ptype;
+  int nmu=0;
+  int skew ;
+
+  for(int s=0;s<LLs;s++){
+    int sF=LLs*sU+s;
+    skew = 0;
+    Uchi=zero;
+    GENERIC_STENCIL_LEG_EXT(U,Xp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Yp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Zp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Tp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Xm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Ym,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Zm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(U,Tm,skew,Impl::multLinkAdd);
+    skew=8;
+    GENERIC_STENCIL_LEG_EXT(UUU,Xp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Yp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Zp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Tp,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Xm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Ym,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
+    GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
+
+    if ( nmu ) { 
+      if ( dag ) { 
+	out._odata[sF] = out._odata[sF] - Uchi;
+      } else { 
+	out._odata[sF] = out._odata[sF] + Uchi;
+      }
+    }
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////
+// Driving / wrapping routine to select right kernel
+////////////////////////////////////////////////////////////////////////////////////
+
 template <class Impl>
 void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,
 					 SiteSpinor *buf, int LLs, int sU,
-						  const FermionField &in, FermionField &out) {
-  SiteSpinor naik;
-  SiteSpinor naive;
-  int oneLink  =0;
-  int threeLink=1;
+					 const FermionField &in, FermionField &out,
+					 int interior,int exterior)
+{
   int dag=1;
-  switch(Opt) {
-#ifdef AVX512
-  //FIXME; move the sign into the Asm routine
-  case OptInlineAsm:
-    DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out);
-    for(int s=0;s<LLs;s++) {
-      int sF=s+LLs*sU;
-      out._odata[sF]=-out._odata[sF];
-    }
-    break;
-#endif
-  case OptHandUnroll:
-    DhopSiteHand(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
-    break;
-  case OptGeneric:
-    for(int s=0;s<LLs;s++){
-       int sF=s+LLs*sU;
-       DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
-       DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
-       out._odata[sF] =-naive-naik; 
-     }
-    break;
-  default:
-    std::cout<<"Oops Opt = "<<Opt<<std::endl;
-    assert(0);
-    break;
-  }
+  DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
+};
+
+template <class Impl>
+void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,
+				      SiteSpinor *buf, int LLs, int sU,
+				      const FermionField &in, FermionField &out,
+				      int interior,int exterior)
+{
+  int dag=0;
+  DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
 };
 
 template <class Impl>
 void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,
 				      SiteSpinor *buf, int LLs,
-				      int sU, const FermionField &in, FermionField &out) 
+				      int sU, const FermionField &in, FermionField &out,
+				      int dag,int interior,int exterior) 
 {
-  int oneLink  =0;
-  int threeLink=1;
-  SiteSpinor naik;
-  SiteSpinor naive;
-  int dag=0;
   switch(Opt) {
 #ifdef AVX512
   case OptInlineAsm:
-    DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out);
+    if ( interior && exterior ) {
+      DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
+    } else { 
+      std::cout << GridLogError << "Cannot overlap comms and compute with Staggered assembly"<<std::endl;
+      assert(0);
+    }
     break;
 #endif
   case OptHandUnroll:
+    if ( interior && exterior ) {
       DhopSiteHand   (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
+    } else if ( interior ) {
+      DhopSiteHandInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
+    } else if ( exterior ) {
+      DhopSiteHandExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
+    }
     break;
   case OptGeneric:
-    for(int s=0;s<LLs;s++){
-      int sF=LLs*sU+s;
-      //      assert(sF<in._odata.size());
-      //      assert(sU< U._odata.size());
-      //      assert(sF>=0);      assert(sU>=0);
-      DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
-      DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
-      out._odata[sF] =naive+naik;
+    if ( interior && exterior ) {
+      DhopSiteGeneric   (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
+    } else if ( interior ) {
+      DhopSiteGenericInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
+    } else if ( exterior ) {
+      DhopSiteGenericExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
     }
     break;
   default:

lib/qcd/action/fermion/StaggeredKernels.h

@@ -38,8 +38,9 @@ namespace QCD {
 class StaggeredKernelsStatic { 
  public:
   enum { OptGeneric, OptHandUnroll, OptInlineAsm };
-  // S-direction is INNERMOST and takes no part in the parity.
-  static int Opt;  // these are a temporary hack
+  enum { CommsAndCompute, CommsThenCompute };
+  static int Opt;
+  static int Comms;
 };
  
 template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , public StaggeredKernelsStatic { 
@@ -53,24 +54,62 @@ public:
    void DhopDir(StencilImpl &st, DoubledGaugeField &U, DoubledGaugeField &UUU, SiteSpinor * buf,
 		      int sF, int sU, const FermionField &in, FermionField &out, int dir,int disp);
 
-   void DhopSiteDepth(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteSpinor * buf,
-		     int sF, int sU, const FermionField &in, SiteSpinor &out,int threeLink);
+   ///////////////////////////////////////////////////////////////////////////////////////
+   // Generic Nc kernels
+   ///////////////////////////////////////////////////////////////////////////////////////
+   void DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo, 
+			DoubledGaugeField &U, DoubledGaugeField &UUU, 
+			SiteSpinor * buf, int LLs, int sU, 
+			const FermionField &in, FermionField &out,int dag);
+   void DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo, 
+			   DoubledGaugeField &U, DoubledGaugeField &UUU, 
+			   SiteSpinor * buf, int LLs, int sU, 
+			   const FermionField &in, FermionField &out,int dag);
+   void DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo, 
+			   DoubledGaugeField &U, DoubledGaugeField &UUU,
+			   SiteSpinor * buf, int LLs, int sU, 
+			   const FermionField &in, FermionField &out,int dag);
 
+   ///////////////////////////////////////////////////////////////////////////////////////
+   // Nc=3 specific kernels
+   ///////////////////////////////////////////////////////////////////////////////////////
+   void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+		     DoubledGaugeField &U,DoubledGaugeField &UUU, 
+		     SiteSpinor * buf, int LLs, int sU, 
+		     const FermionField &in, FermionField &out,int dag);
+   void DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+			DoubledGaugeField &U,DoubledGaugeField &UUU, 
+			SiteSpinor * buf, int LLs, int sU, 
+			const FermionField &in, FermionField &out,int dag);
+   void DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+			DoubledGaugeField &U,DoubledGaugeField &UUU, 
+			SiteSpinor * buf, int LLs, int sU, 
+			const FermionField &in, FermionField &out,int dag);
 
-   void DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteSpinor * buf,
-		     int sF, int sU, const FermionField &in, SiteSpinor&out,int threeLink);
+   ///////////////////////////////////////////////////////////////////////////////////////
+   // Asm Nc=3 specific kernels
+   ///////////////////////////////////////////////////////////////////////////////////////
+   void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
+		    DoubledGaugeField &U,DoubledGaugeField &UUU, 
+		    SiteSpinor * buf, int LLs, int sU, 
+		    const FermionField &in, FermionField &out,int dag);
+   ///////////////////////////////////////////////////////////////////////////////////////////////////
+   // Generic interface; fan out to right routine
+   ///////////////////////////////////////////////////////////////////////////////////////////////////
+   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
+		 DoubledGaugeField &U, DoubledGaugeField &UUU, 
+		 SiteSpinor * buf, int LLs, int sU,
+		 const FermionField &in, FermionField &out, int interior=1,int exterior=1);
 
-   void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,SiteSpinor * buf,
-		     int LLs, int sU, const FermionField &in, FermionField &out, int dag);
+   void DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, 
+		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
+		    SiteSpinor * buf, int LLs, int sU,
+		    const FermionField &in, FermionField &out, int interior=1,int exterior=1);
 
-   void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU, SiteSpinor * buf,
-			 int LLs, int sU, const FermionField &in, FermionField &out);
-      
-   void DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU, SiteSpinor * buf,
-		int sF, int sU, const FermionField &in, FermionField &out);
-
-   void DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU, SiteSpinor *buf, 
-                   int LLs, int sU, const FermionField &in, FermionField &out);
+   void DhopSite(StencilImpl &st, LebesgueOrder &lo, 
+		 DoubledGaugeField &U, DoubledGaugeField &UUU, 
+		 SiteSpinor * buf, int LLs, int sU,
+		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
   
 public:
 

lib/qcd/action/fermion/StaggeredKernelsAsm.cc

@@ -560,6 +560,27 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
        VSTORE(2,%0,pUChi_02)					\
        : : "r" (out) : "memory" );
 
+#define nREDUCE(out)							\
+  asm (									\
+       VADD(UChi_00,UChi_10,UChi_00)					\
+       VADD(UChi_01,UChi_11,UChi_01)					\
+       VADD(UChi_02,UChi_12,UChi_02)					\
+       VADD(UChi_30,UChi_20,UChi_30)					\
+       VADD(UChi_31,UChi_21,UChi_31)					\
+       VADD(UChi_32,UChi_22,UChi_32)					\
+       VADD(UChi_00,UChi_30,UChi_00)					\
+       VADD(UChi_01,UChi_31,UChi_01)					\
+       VADD(UChi_02,UChi_32,UChi_02)				);	\
+  asm (VZERO(Chi_00)							\
+       VSUB(UChi_00,Chi_00,UChi_00)					\
+       VSUB(UChi_01,Chi_00,UChi_01)					\
+       VSUB(UChi_02,Chi_00,UChi_02)				);	\
+  asm (								\
+       VSTORE(0,%0,pUChi_00)					\
+       VSTORE(1,%0,pUChi_01)					\
+       VSTORE(2,%0,pUChi_02)					\
+       : : "r" (out) : "memory" );
+
 #define REDUCEa(out)					\
   asm (							\
   VADD(UChi_00,UChi_10,UChi_00)				\
@@ -571,6 +592,22 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
        VSTORE(2,%0,pUChi_02)						\
        : : "r" (out) : "memory" );
 
+// FIXME is sign right in the VSUB ?
+#define nREDUCEa(out)					\
+  asm (							\
+  VADD(UChi_00,UChi_10,UChi_00)				\
+  VADD(UChi_01,UChi_11,UChi_01)				\
+  VADD(UChi_02,UChi_12,UChi_02)	);			\
+  asm (VZERO(Chi_00)							\
+       VSUB(UChi_00,Chi_00,UChi_00)					\
+       VSUB(UChi_01,Chi_00,UChi_01)					\
+       VSUB(UChi_02,Chi_00,UChi_02)				);	\
+  asm (									\
+       VSTORE(0,%0,pUChi_00)				\
+       VSTORE(1,%0,pUChi_01)				\
+       VSTORE(2,%0,pUChi_02)				\
+       : : "r" (out) : "memory" );
+
 #define PERMUTE_DIR(dir)			\
       permute##dir(Chi_0,Chi_0);\
       permute##dir(Chi_1,Chi_1);\
@@ -581,10 +618,9 @@ namespace QCD {
 
 template <class Impl>
 void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
-					 DoubledGaugeField &U,
-					 DoubledGaugeField &UUU,
-					 SiteSpinor *buf, int LLs,
-					 int sU, const FermionField &in, FermionField &out) 
+					 DoubledGaugeField &U, DoubledGaugeField &UUU,
+					 SiteSpinor *buf, int LLs, int sU, 
+					 const FermionField &in, FermionField &out,int dag) 
 {
   assert(0);
 };
@@ -645,10 +681,9 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
   // This is the single precision 5th direction vectorised kernel
 #include <simd/Intel512single.h>
 template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
-								    DoubledGaugeField &U,
-								    DoubledGaugeField &UUU,
-								    SiteSpinor *buf, int LLs,
-								    int sU, const FermionField &in, FermionField &out) 
+								    DoubledGaugeField &U, DoubledGaugeField &UUU,
+								    SiteSpinor *buf, int LLs, int sU, 
+								    const FermionField &in, FermionField &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -685,8 +720,12 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
     MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
 
     addr0 = (uint64_t) &out._odata[sF];
+    if ( dag ) {
+      nREDUCE(addr0);
+    } else { 
       REDUCE(addr0);
     }
+   }
 #else 
     assert(0);
 #endif
@@ -695,10 +734,9 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
 
 #include <simd/Intel512double.h>
 template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
-								    DoubledGaugeField &U,
-								    DoubledGaugeField &UUU,
-								    SiteSpinor *buf, int LLs,
-								    int sU, const FermionField &in, FermionField &out) 
+								    DoubledGaugeField &U, DoubledGaugeField &UUU,
+								    SiteSpinor *buf, int LLs, int sU, 
+								    const FermionField &in, FermionField &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -734,8 +772,12 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
     MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
 
     addr0 = (uint64_t) &out._odata[sF];
+    if ( dag ) {
+      nREDUCE(addr0);
+    } else { 
       REDUCE(addr0);
     }
+  }
 #else 
   assert(0);
 #endif
@@ -776,10 +818,9 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
 
 #include <simd/Intel512single.h>
 template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
-								    DoubledGaugeField &U,
-								    DoubledGaugeField &UUU,
-								    SiteSpinor *buf, int LLs,
-								    int sU, const FermionField &in, FermionField &out) 
+							       DoubledGaugeField &U, DoubledGaugeField &UUU,
+							       SiteSpinor *buf, int LLs, int sU, 
+							       const FermionField &in, FermionField &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -832,8 +873,12 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
     MULT_ADD_XYZT(gauge2,gauge3);  
 
     addr0 = (uint64_t) &out._odata[sF];
+    if ( dag ) { 
+      nREDUCEa(addr0);
+    } else { 
       REDUCEa(addr0);
     }
+  }
 #else 
   assert(0);
 #endif
@@ -841,10 +886,9 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
 
 #include <simd/Intel512double.h>
 template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, 
-								    DoubledGaugeField &U,
-								    DoubledGaugeField &UUU,
-								    SiteSpinor *buf, int LLs,
-								    int sU, const FermionField &in, FermionField &out) 
+							       DoubledGaugeField &U, DoubledGaugeField &UUU,
+							       SiteSpinor *buf, int LLs, int sU, 
+							       const FermionField &in, FermionField &out,int dag) 
 {
 #ifdef AVX512
   uint64_t gauge0,gauge1,gauge2,gauge3;
@@ -897,8 +941,12 @@ template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st,
     MULT_ADD_XYZT(gauge2,gauge3);  
     
     addr0 = (uint64_t) &out._odata[sF];
+    if ( dag ) {
+      nREDUCEa(addr0);
+    } else { 
       REDUCEa(addr0);
     }
+  }
 #else 
   assert(0);
 #endif
@@ -909,7 +957,7 @@ template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st,
 				  DoubledGaugeField &U,			\
 				  DoubledGaugeField &UUU,		\
 				  SiteSpinor *buf, int LLs,		\
-				  int sU, const FermionField &in, FermionField &out);
+				  int sU, const FermionField &in, FermionField &out,int dag);
 
 KERNEL_INSTANTIATE(StaggeredKernels,DhopSiteAsm,StaggeredImplD);
 KERNEL_INSTANTIATE(StaggeredKernels,DhopSiteAsm,StaggeredImplF);

lib/qcd/action/fermion/StaggeredKernelsHand.cc

@@ -28,7 +28,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     /*  END LEGAL */
 #include <Grid.h>
 
-#define REGISTER
 
 #define LOAD_CHI(b)		\
   const SiteSpinor & ref (b[offset]);	\
@@ -59,7 +58,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     UChi ## _1 += U_12*Chi_2;\
     UChi ## _2 += U_22*Chi_2;
 
-#define MULT_ADD(A,UChi)				\
+#define MULT_ADD(U,A,UChi)			\
   auto & ref(U._odata[sU](A));			\
    Impl::loadLinkElement(U_00,ref()(0,0));      \
    Impl::loadLinkElement(U_10,ref()(1,0));      \
@@ -86,237 +85,315 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   permute##dir(Chi_1,Chi_1);			\
   permute##dir(Chi_2,Chi_2);
 
+
+#define HAND_STENCIL_LEG_BASE(Dir,Perm,skew)	\
+  SE=st.GetEntry(ptype,Dir+skew,sF);	\
+  offset = SE->_offset;			\
+  local  = SE->_is_local;		\
+  perm   = SE->_permute;		\
+  if ( local ) {						\
+    LOAD_CHI(in._odata);					\
+    if ( perm) {						\
+      PERMUTE_DIR(Perm);					\
+    }								\
+  } else {							\
+    LOAD_CHI(buf);						\
+  }								
+
+#define HAND_STENCIL_LEG_BEGIN(Dir,Perm,skew,even)		\
+  HAND_STENCIL_LEG_BASE(Dir,Perm,skew)				\
+  {								\
+    MULT(Dir,even);						\
+  }
+
+#define HAND_STENCIL_LEG(U,Dir,Perm,skew,even)			\
+  HAND_STENCIL_LEG_BASE(Dir,Perm,skew)				\
+  {								\
+    MULT_ADD(U,Dir,even);					\
+  }
+
+
+
+#define HAND_STENCIL_LEG_INT(U,Dir,Perm,skew,even)	\
+  SE=st.GetEntry(ptype,Dir+skew,sF);			\
+  offset = SE->_offset;					\
+  local  = SE->_is_local;				\
+  perm   = SE->_permute;				\
+  if ( local ) {					\
+    LOAD_CHI(in._odata);				\
+    if ( perm) {					\
+      PERMUTE_DIR(Perm);				\
+    }							\
+  } else if ( st.same_node[Dir] ) {			\
+    LOAD_CHI(buf);					\
+  }							\
+  if (SE->_is_local || st.same_node[Dir] ) {		\
+    MULT_ADD(U,Dir,even);				\
+  }
+
+#define HAND_STENCIL_LEG_EXT(U,Dir,Perm,skew,even)	\
+  SE=st.GetEntry(ptype,Dir+skew,sF);			\
+  offset = SE->_offset;					\
+  local  = SE->_is_local;				\
+  perm   = SE->_permute;				\
+  if ((!SE->_is_local) && (!st.same_node[Dir]) ) {		\
+    nmu++;							\
+    { LOAD_CHI(buf);	  }					\
+    { MULT_ADD(U,Dir,even); }					\
+  }								
+
 namespace Grid {
 namespace QCD {
 
 
 template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
-					  SiteSpinor *buf, int LLs,
-					  int sU, const FermionField &in, FermionField &out, int dag) 
-{
-  SiteSpinor naik; 
-  SiteSpinor naive;
-  int oneLink  =0;
-  int threeLink=1;
-  int skew(0);
-  Real scale(1.0);
-  
-  if(dag) scale = -1.0;
-  
-  for(int s=0;s<LLs;s++){
-    int sF=s+LLs*sU;
-    DhopSiteDepthHand(st,lo,U,buf,sF,sU,in,naive,oneLink);
-    DhopSiteDepthHand(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
-    out._odata[sF] =scale*(naive+naik);
-  }
-}
-
-template <class Impl>
-void StaggeredKernels<Impl>::DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
-					       SiteSpinor *buf, int sF,
-					       int sU, const FermionField &in, SiteSpinor &out,int threeLink) 
+void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, 
+					  DoubledGaugeField &U,DoubledGaugeField &UUU,
+					  SiteSpinor *buf, int LLs, int sU, 
+					  const FermionField &in, FermionField &out,int dag) 
 {
   typedef typename Simd::scalar_type S;
   typedef typename Simd::vector_type V;
 
-  REGISTER Simd even_0; // 12 regs on knc
-  REGISTER Simd even_1;
-  REGISTER Simd even_2;
-  REGISTER Simd odd_0; // 12 regs on knc
-  REGISTER Simd odd_1;
-  REGISTER Simd odd_2;
+  Simd even_0; // 12 regs on knc
+  Simd even_1;
+  Simd even_2;
+  Simd odd_0; // 12 regs on knc
+  Simd odd_1;
+  Simd odd_2;
 
-  REGISTER Simd Chi_0;    // two spinor; 6 regs
-  REGISTER Simd Chi_1;
-  REGISTER Simd Chi_2;
+  Simd Chi_0;    // two spinor; 6 regs
+  Simd Chi_1;
+  Simd Chi_2;
   
-  REGISTER Simd U_00;  // two rows of U matrix
-  REGISTER Simd U_10;
-  REGISTER Simd U_20;  
-  REGISTER Simd U_01;
-  REGISTER Simd U_11;
-  REGISTER Simd U_21;  // 2 reg left.
-  REGISTER Simd U_02;
-  REGISTER Simd U_12;
-  REGISTER Simd U_22; 
-
-  int skew = 0;
-  if (threeLink) skew=8;
+  Simd U_00;  // two rows of U matrix
+  Simd U_10;
+  Simd U_20;  
+  Simd U_01;
+  Simd U_11;
+  Simd U_21;  // 2 reg left.
+  Simd U_02;
+  Simd U_12;
+  Simd U_22; 
 
+  SiteSpinor result;
   int offset,local,perm, ptype;
+
   StencilEntry *SE;
+  int skew;
 
-  // Xp
-  SE=st.GetEntry(ptype,Xp+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
+  for(int s=0;s<LLs;s++){
+    int sF=s+LLs*sU;
 
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
+    skew = 0;
+    HAND_STENCIL_LEG_BEGIN(Xp,3,skew,even);  
+    HAND_STENCIL_LEG_BEGIN(Yp,2,skew,odd);   
+    HAND_STENCIL_LEG      (U,Zp,1,skew,even);  
+    HAND_STENCIL_LEG      (U,Tp,0,skew,odd);  
+    HAND_STENCIL_LEG      (U,Xm,3,skew,even);  
+    HAND_STENCIL_LEG      (U,Ym,2,skew,odd);   
+    HAND_STENCIL_LEG      (U,Zm,1,skew,even);  
+    HAND_STENCIL_LEG      (U,Tm,0,skew,odd);  
+    skew = 8;
+    HAND_STENCIL_LEG(UUU,Xp,3,skew,even);  
+    HAND_STENCIL_LEG(UUU,Yp,2,skew,odd);   
+    HAND_STENCIL_LEG(UUU,Zp,1,skew,even);  
+    HAND_STENCIL_LEG(UUU,Tp,0,skew,odd);  
+    HAND_STENCIL_LEG(UUU,Xm,3,skew,even);  
+    HAND_STENCIL_LEG(UUU,Ym,2,skew,odd);   
+    HAND_STENCIL_LEG(UUU,Zm,1,skew,even);  
+    HAND_STENCIL_LEG(UUU,Tm,0,skew,odd);  
+    
+    if ( dag ) {
+      result()()(0) = - even_0 - odd_0;
+      result()()(1) = - even_1 - odd_1;
+      result()()(2) = - even_2 - odd_2;
     } else { 
-    LOAD_CHI(buf);
+      result()()(0) = even_0 + odd_0;
+      result()()(1) = even_1 + odd_1;
+      result()()(2) = even_2 + odd_2;
     }
+    vstream(out._odata[sF],result);
+  }
+}
+
+
+template <class Impl>
+void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, 
+					     DoubledGaugeField &U, DoubledGaugeField &UUU,
+					     SiteSpinor *buf, int LLs, int sU, 
+					     const FermionField &in, FermionField &out,int dag) 
 {
-    MULT(Xp,even);
-  }
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
 
-  // Yp
-  SE=st.GetEntry(ptype,Yp+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
+  Simd even_0; // 12 regs on knc
+  Simd even_1;
+  Simd even_2;
+  Simd odd_0; // 12 regs on knc
+  Simd odd_1;
+  Simd odd_2;
 
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
+  Simd Chi_0;    // two spinor; 6 regs
+  Simd Chi_1;
+  Simd Chi_2;
+  
+  Simd U_00;  // two rows of U matrix
+  Simd U_10;
+  Simd U_20;  
+  Simd U_01;
+  Simd U_11;
+  Simd U_21;  // 2 reg left.
+  Simd U_02;
+  Simd U_12;
+  Simd U_22; 
+
+  SiteSpinor result;
+  int offset,local,perm, ptype;
+
+  StencilEntry *SE;
+  int skew;
+
+  for(int s=0;s<LLs;s++){
+    int sF=s+LLs*sU;
+
+    even_0 = zero;    even_1 = zero;    even_2 = zero;
+     odd_0 = zero;     odd_1 = zero;     odd_2 = zero;
+
+    skew = 0;
+    HAND_STENCIL_LEG_INT(U,Xp,3,skew,even);  
+    HAND_STENCIL_LEG_INT(U,Yp,2,skew,odd);   
+    HAND_STENCIL_LEG_INT(U,Zp,1,skew,even);  
+    HAND_STENCIL_LEG_INT(U,Tp,0,skew,odd);  
+    HAND_STENCIL_LEG_INT(U,Xm,3,skew,even);  
+    HAND_STENCIL_LEG_INT(U,Ym,2,skew,odd);   
+    HAND_STENCIL_LEG_INT(U,Zm,1,skew,even);  
+    HAND_STENCIL_LEG_INT(U,Tm,0,skew,odd);  
+    skew = 8;
+    HAND_STENCIL_LEG_INT(UUU,Xp,3,skew,even);  
+    HAND_STENCIL_LEG_INT(UUU,Yp,2,skew,odd);   
+    HAND_STENCIL_LEG_INT(UUU,Zp,1,skew,even);  
+    HAND_STENCIL_LEG_INT(UUU,Tp,0,skew,odd);  
+    HAND_STENCIL_LEG_INT(UUU,Xm,3,skew,even);  
+    HAND_STENCIL_LEG_INT(UUU,Ym,2,skew,odd);   
+    HAND_STENCIL_LEG_INT(UUU,Zm,1,skew,even);  
+    HAND_STENCIL_LEG_INT(UUU,Tm,0,skew,odd);  
+
+    // Assume every site must be connected to at least one interior point. No 1^4 subvols.
+    if ( dag ) {
+      result()()(0) = - even_0 - odd_0;
+      result()()(1) = - even_1 - odd_1;
+      result()()(2) = - even_2 - odd_2;
     } else { 
-    LOAD_CHI(buf);
+      result()()(0) = even_0 + odd_0;
+      result()()(1) = even_1 + odd_1;
+      result()()(2) = even_2 + odd_2;
     }
+    vstream(out._odata[sF],result);
+  }
+}
+
+
+template <class Impl>
+void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, 
+					     DoubledGaugeField &U, DoubledGaugeField &UUU,
+					     SiteSpinor *buf, int LLs, int sU, 
+					     const FermionField &in, FermionField &out,int dag) 
 {
-    MULT(Yp,odd);
-  }
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
 
+  Simd even_0; // 12 regs on knc
+  Simd even_1;
+  Simd even_2;
+  Simd odd_0; // 12 regs on knc
+  Simd odd_1;
+  Simd odd_2;
 
-  // Zp
-  SE=st.GetEntry(ptype,Zp+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
+  Simd Chi_0;    // two spinor; 6 regs
+  Simd Chi_1;
+  Simd Chi_2;
   
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
+  Simd U_00;  // two rows of U matrix
+  Simd U_10;
+  Simd U_20;  
+  Simd U_01;
+  Simd U_11;
+  Simd U_21;  // 2 reg left.
+  Simd U_02;
+  Simd U_12;
+  Simd U_22; 
+
+  SiteSpinor result;
+  int offset,local,perm, ptype;
+
+  StencilEntry *SE;
+  int skew;
+
+  for(int s=0;s<LLs;s++){
+    int sF=s+LLs*sU;
+
+    even_0 = zero;    even_1 = zero;    even_2 = zero;
+     odd_0 = zero;     odd_1 = zero;     odd_2 = zero;
+    int nmu=0;
+    skew = 0;
+    HAND_STENCIL_LEG_EXT(U,Xp,3,skew,even);  
+    HAND_STENCIL_LEG_EXT(U,Yp,2,skew,odd);   
+    HAND_STENCIL_LEG_EXT(U,Zp,1,skew,even);  
+    HAND_STENCIL_LEG_EXT(U,Tp,0,skew,odd);  
+    HAND_STENCIL_LEG_EXT(U,Xm,3,skew,even);  
+    HAND_STENCIL_LEG_EXT(U,Ym,2,skew,odd);   
+    HAND_STENCIL_LEG_EXT(U,Zm,1,skew,even);  
+    HAND_STENCIL_LEG_EXT(U,Tm,0,skew,odd);  
+    skew = 8;
+    HAND_STENCIL_LEG_EXT(UUU,Xp,3,skew,even);  
+    HAND_STENCIL_LEG_EXT(UUU,Yp,2,skew,odd);   
+    HAND_STENCIL_LEG_EXT(UUU,Zp,1,skew,even);  
+    HAND_STENCIL_LEG_EXT(UUU,Tp,0,skew,odd);  
+    HAND_STENCIL_LEG_EXT(UUU,Xm,3,skew,even);  
+    HAND_STENCIL_LEG_EXT(UUU,Ym,2,skew,odd);   
+    HAND_STENCIL_LEG_EXT(UUU,Zm,1,skew,even);  
+    HAND_STENCIL_LEG_EXT(UUU,Tm,0,skew,odd);  
+
+    // Add sum of all exterior connected stencil legs
+    if ( nmu ) { 
+      if ( dag ) {
+	result()()(0) = - even_0 - odd_0;
+	result()()(1) = - even_1 - odd_1;
+	result()()(2) = - even_2 - odd_2;
       } else { 
-    LOAD_CHI(buf);
+	result()()(0) = even_0 + odd_0;
+	result()()(1) = even_1 + odd_1;
+	result()()(2) = even_2 + odd_2;
+      }
+      out._odata[sF] = out._odata[sF] + result;
+    }
   }
-  {
-    MULT_ADD(Zp,even);
 }
 
-  // Tp
-  SE=st.GetEntry(ptype,Tp+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
-  
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
-  } else { 
-    LOAD_CHI(buf);
-  }
-  {
-    MULT_ADD(Tp,odd);
-  }
-  
-  // Xm
-  SE=st.GetEntry(ptype,Xm+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
-  
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
-  } else { 
-    LOAD_CHI(buf);
-  }
-  {
-    MULT_ADD(Xm,even);
-  }
-  
-  
-  // Ym
-  SE=st.GetEntry(ptype,Ym+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
-  
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
-  } else { 
-    LOAD_CHI(buf);
-  }
-  {
-    MULT_ADD(Ym,odd);
-  }
-
-  // Zm
-  SE=st.GetEntry(ptype,Zm+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
-  
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
-  } else { 
-    LOAD_CHI(buf);
-  }
-  {
-    MULT_ADD(Zm,even);
-  }
-
-  // Tm
-  SE=st.GetEntry(ptype,Tm+skew,sF);
-  offset = SE->_offset;
-  local  = SE->_is_local;
-  perm   = SE->_permute;
-  
-  if ( local ) {
-    LOAD_CHI(in._odata);
-    if ( perm) {
-      PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
-    }
-  } else { 
-    LOAD_CHI(buf);
-  }
-  {
-    MULT_ADD(Tm,odd);
-  }
-
-  vstream(out()()(0),even_0+odd_0);
-  vstream(out()()(1),even_1+odd_1);
-  vstream(out()()(2),even_2+odd_2);
-
-}
 
 #define DHOP_SITE_HAND_INSTANTIATE(IMPL)				\
   template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
 						     DoubledGaugeField &U,DoubledGaugeField &UUU, \
-						     SiteSpinor *buf, int LLs, \
-						     int sU, const FermionField &in, FermionField &out, int dag);
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionField &in, FermionField &out, int dag); \
+									\
+  template void StaggeredKernels<IMPL>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeField &U,DoubledGaugeField &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionField &in, FermionField &out, int dag); \
+									\
+  template void StaggeredKernels<IMPL>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeField &U,DoubledGaugeField &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionField &in, FermionField &out, int dag); \
 
-#define DHOP_SITE_DEPTH_HAND_INSTANTIATE(IMPL)				\
-  template void StaggeredKernels<IMPL>::DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, \
-							  SiteSpinor *buf, int sF, \
-							  int sU, const FermionField &in, SiteSpinor &out,int threeLink) ;
 DHOP_SITE_HAND_INSTANTIATE(StaggeredImplD);
 DHOP_SITE_HAND_INSTANTIATE(StaggeredImplF);
 DHOP_SITE_HAND_INSTANTIATE(StaggeredVec5dImplD);
 DHOP_SITE_HAND_INSTANTIATE(StaggeredVec5dImplF);
 
-DHOP_SITE_DEPTH_HAND_INSTANTIATE(StaggeredImplD);
-DHOP_SITE_DEPTH_HAND_INSTANTIATE(StaggeredImplF);
-DHOP_SITE_DEPTH_HAND_INSTANTIATE(StaggeredVec5dImplD);
-DHOP_SITE_DEPTH_HAND_INSTANTIATE(StaggeredVec5dImplF);
 
-}}
+}
+}
+

lib/qcd/action/fermion/WilsonCompressor.h

@@ -69,39 +69,47 @@ class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
   /*****************************************************/
   /* Compress includes precision change if mpi data is not same */
   /*****************************************************/
-  inline void Compress(SiteHalfSpinor *buf,Integer o,const SiteSpinor &in) {
-    projector::Proj(buf[o],in,mu,dag);
+  inline void Compress(SiteHalfSpinor * __restrict__ buf,Integer o,const SiteSpinor &in) {
+    SiteHalfSpinor tmp;
+    projector::Proj(tmp,in,mu,dag);
+    vstream(buf[o],tmp);
   }
 
   /*****************************************************/
   /* Exchange includes precision change if mpi data is not same */
   /*****************************************************/
-  inline void Exchange(SiteHalfSpinor *mp,
-                       SiteHalfSpinor *vp0,
-                       SiteHalfSpinor *vp1,
+  inline void Exchange(SiteHalfSpinor * __restrict__ mp,
+                       const SiteHalfSpinor * __restrict__ vp0,
+                       const SiteHalfSpinor * __restrict__ vp1,
 		       Integer type,Integer o){
-    exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+    SiteHalfSpinor tmp1;
+    SiteHalfSpinor tmp2;
+    exchange(tmp1,tmp2,vp0[o],vp1[o],type);
+    vstream(mp[2*o  ],tmp1);
+    vstream(mp[2*o+1],tmp2);
   }
 
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  inline void Decompress(SiteHalfSpinor *out,
-			 SiteHalfSpinor *in, Integer o) {    
+  inline void Decompress(SiteHalfSpinor * __restrict__ out,
+			 SiteHalfSpinor * __restrict__ in, Integer o) {    
     assert(0);
   }
 
   /*****************************************************/
   /* Compress Exchange                                 */
   /*****************************************************/
-  inline void CompressExchange(SiteHalfSpinor *out0,
-			       SiteHalfSpinor *out1,
-			       const SiteSpinor *in,
+  inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
+			       SiteHalfSpinor * __restrict__ out1,
+			       const SiteSpinor * __restrict__ in,
 			       Integer j,Integer k, Integer m,Integer type){
     SiteHalfSpinor temp1, temp2,temp3,temp4;
     projector::Proj(temp1,in[k],mu,dag);
     projector::Proj(temp2,in[m],mu,dag);
-    exchange(out0[j],out1[j],temp1,temp2,type);
+    exchange(temp3,temp4,temp1,temp2,type);
+    vstream(out0[j],temp3);
+    vstream(out1[j],temp4);
   }
 
   /*****************************************************/
@@ -266,41 +274,16 @@ public:
     if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
   }
 
-  std::vector<int> same_node;
-  std::vector<int> surface_list;
-
   WilsonStencil(GridBase *grid,
 		int npoints,
 		int checkerboard,
 		const std::vector<int> &directions,
 		const std::vector<int> &distances)  
-    : CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
-    same_node(npoints)
+    : CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) 
   { 
     ZeroCountersi();
-    surface_list.resize(0);
   };
 
-  void BuildSurfaceList(int Ls,int vol4){
-
-    // find same node for SHM
-    // Here we know the distance is 1 for WilsonStencil
-    for(int point=0;point<this->_npoints;point++){
-      same_node[point] = this->SameNode(point);
-    }
-    
-    for(int site = 0 ;site< vol4;site++){
-      int local = 1;
-      for(int point=0;point<this->_npoints;point++){
-	if( (!this->GetNodeLocal(site*Ls,point)) && (!same_node[point]) ){ 
-	  local = 0;
-	}
-      }
-      if(local == 0) { 
-	surface_list.push_back(site);
-      }
-    }
-  }
 
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
@@ -361,23 +344,23 @@ public:
     int dag = compress.dag;
     int face_idx=0;
     if ( dag ) { 
-      assert(same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
-      assert(same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
-      assert(same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
-      assert(same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
-      assert(same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
-      assert(same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
-      assert(same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
-      assert(same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
+      assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
+      assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
+      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
+      assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
+      assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
+      assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
+      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
+      assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
     } else {
-      assert(same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
-      assert(same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
-      assert(same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
-      assert(same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
-      assert(same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
-      assert(same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
-      assert(same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
-      assert(same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
+      assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
+      assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
+      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
+      assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
+      assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
+      assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
+      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
+      assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);

lib/qcd/action/fermion/WilsonFermion.cc

@@ -348,15 +348,98 @@ void WilsonFermion<Impl>::DhopDirDisp(const FermionField &in, FermionField &out,
   parallel_for (int sss = 0; sss < in._grid->oSites(); sss++) {
     Kernels::DhopDir(Stencil, Umu, Stencil.CommBuf(), sss, sss, in, out, dirdisp, gamma);
   }
-};
-
+} 
+/*Change starts*/
 template <class Impl>
 void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        DoubledGaugeField &U,
                                        const FermionField &in,
                                        FermionField &out, int dag) {
-  assert((dag == DaggerNo) || (dag == DaggerYes));
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
+  else
+#endif 
+    DhopInternalSerial(st,lo,U,in,out,dag);
 
+}
+
+template <class Impl>
+void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
+                                       DoubledGaugeField &U,
+                                       const FermionField &in,
+                                       FermionField &out, int dag) {
+  assert((dag == DaggerNo) || (dag == DaggerYes));
+#ifdef GRID_OMP
+  Compressor compressor;
+  int len =  U._grid->oSites();
+  const int LLs =  1;
+
+  st.Prepare();
+  st.HaloGather(in,compressor);
+  st.CommsMergeSHM(compressor);
+#pragma omp parallel
+  {
+    int tid = omp_get_thread_num();
+    int nthreads = omp_get_num_threads();
+    int ncomms = CartesianCommunicator::nCommThreads;
+    if (ncomms == -1) ncomms = 1;
+    assert(nthreads > ncomms);
+    if (tid >= ncomms) {
+      nthreads -= ncomms;
+      int ttid  = tid - ncomms;
+      int n     = len;
+      int chunk = n / nthreads;
+      int rem   = n % nthreads;
+      int myblock, myn;
+      if (ttid < rem) {
+        myblock = ttid * chunk + ttid;
+        myn = chunk+1;
+      } else {
+        myblock = ttid*chunk + rem;
+        myn = chunk;
+      }
+      // do the compute
+     if (dag == DaggerYes) {
+
+        for (int sss = myblock; sss < myblock+myn; ++sss) {
+         Kernels::DhopSiteDag(st, lo, U, st.CommBuf(), sss, sss, 1, 1, in, out);
+       }
+     } else {
+        for (int sss = myblock; sss < myblock+myn; ++sss) {
+         Kernels::DhopSite(st, lo, U, st.CommBuf(), sss, sss, 1, 1, in, out);
+       }
+    } //else
+
+    } else {
+      st.CommunicateThreaded();
+    }
+
+  Compressor compressor(dag);
+
+  if (dag == DaggerYes) {
+    parallel_for (int sss = 0; sss < in._grid->oSites(); sss++) {
+      Kernels::DhopSiteDag(st, lo, U, st.CommBuf(), sss, sss, 1, 1, in, out);
+    }
+  } else {
+    parallel_for (int sss = 0; sss < in._grid->oSites(); sss++) {
+      Kernels::DhopSite(st, lo, U, st.CommBuf(), sss, sss, 1, 1, in, out);
+    }
+  }
+
+  }  //pragma
+#else
+  assert(0);
+#endif
+};
+
+
+template <class Impl>
+void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
+                                       DoubledGaugeField &U,
+                                       const FermionField &in,
+                                       FermionField &out, int dag) {
+  assert((dag == DaggerNo) || (dag == DaggerYes));
   Compressor compressor(dag);
   st.HaloExchange(in, compressor);
 
@@ -370,6 +453,7 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
     }
   }
 };
+/*Change ends */
 
 /*******************************************************************************
  * Conserved current utilities for Wilson fermions, for contracting propagators

lib/qcd/action/fermion/WilsonFermion.h

@@ -130,6 +130,12 @@ class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic {
   void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
                     const FermionField &in, FermionField &out, int dag);
 
+  void DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
+
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
+
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
                 GridRedBlackCartesian &Hgrid, RealD _mass, 
@@ -145,6 +151,8 @@ class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic {
 
   //    protected:
  public:
+  virtual RealD Mass(void) { return mass; }
+  virtual int   isTrivialEE(void) { return 1; };
   RealD mass;
   RealD diag_mass;
 

lib/qcd/action/fermion/WilsonFermion5D.cc

@@ -445,8 +445,7 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
 	}
       }
 	ptime = usecond() - start;
-    }
-    {
+    } else {
       double start = usecond();
       st.CommunicateThreaded();
       ctime = usecond() - start;

lib/qcd/action/fermion/WilsonKernels.h

@@ -232,6 +232,7 @@ private:
   void GenericDhopSiteDagExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
 			     int sF, int sU, const FermionField &in, FermionField &out);
 
+
   void AsmDhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
 		   int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
 

lib/qcd/action/fermion/WilsonKernelsHand.cc

@@ -30,147 +30,33 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 #define REGISTER
 
-#define LOAD_CHIMU_BODY(F)			\
-  Chimu_00=ref(F)(0)(0);			\
-  Chimu_01=ref(F)(0)(1);			\
-  Chimu_02=ref(F)(0)(2);			\
-  Chimu_10=ref(F)(1)(0);			\
-  Chimu_11=ref(F)(1)(1);			\
-  Chimu_12=ref(F)(1)(2);			\
-  Chimu_20=ref(F)(2)(0);			\
-  Chimu_21=ref(F)(2)(1);			\
-  Chimu_22=ref(F)(2)(2);			\
-  Chimu_30=ref(F)(3)(0);			\
-  Chimu_31=ref(F)(3)(1);			\
-  Chimu_32=ref(F)(3)(2)
-
-#define LOAD_CHIMU(DIR,F,PERM)						\
-  { const SiteSpinor & ref (in._odata[offset]); LOAD_CHIMU_BODY(F); }
-
-#define LOAD_CHI_BODY(F)				\
-    Chi_00 = ref(F)(0)(0);\
-    Chi_01 = ref(F)(0)(1);\
-    Chi_02 = ref(F)(0)(2);\
-    Chi_10 = ref(F)(1)(0);\
-    Chi_11 = ref(F)(1)(1);\
-    Chi_12 = ref(F)(1)(2)
-
-#define LOAD_CHI(DIR,F,PERM)					\
-  {const SiteHalfSpinor &ref(buf[offset]); LOAD_CHI_BODY(F); }
-
-
-//G-parity implementations using in-place intrinsic ops
-
-//1l 1h -> 1h 1l
-//0l 0h , 1h 1l -> 0l 1h 0h,1l
-//0h,1l -> 1l,0h
-//if( (distance == 1 && !perm_will_occur) || (distance == -1 && perm_will_occur) )
-//Pulled fermion through forwards face, GPBC on upper component
-//Need 0= 0l 1h   1= 1l 0h
-//else if( (distance == -1 && !perm) || (distance == 1 && perm) )
-//Pulled fermion through backwards face, GPBC on lower component
-//Need 0= 1l 0h   1= 0l 1h
-
-//1l 1h -> 1h 1l
-//0l 0h , 1h 1l -> 0l 1h 0h,1l
-#define DO_TWIST_0L_1H(INTO,S,C,F, PERM, tmp1, tmp2, tmp3)			\
-  permute##PERM(tmp1, ref(1)(S)(C));				\
-  exchange##PERM(tmp2,tmp3, ref(0)(S)(C), tmp1);		\
-  INTO = tmp2;
-
-//0l 0h -> 0h 0l
-//1l 1h, 0h 0l -> 1l 0h, 1h 0l
-#define DO_TWIST_1L_0H(INTO,S,C,F, PERM, tmp1, tmp2, tmp3)			\
-  permute##PERM(tmp1, ref(0)(S)(C));				\
-  exchange##PERM(tmp2,tmp3, ref(1)(S)(C), tmp1);		\
-  INTO = tmp2;
-
-
-
-
-#define LOAD_CHI_SETUP(DIR,F)						\
-  g = F;								\
-  direction = st._directions[DIR];				\
-  distance = st._distances[DIR];				\
-  sl = st._grid->_simd_layout[direction];			\
-  inplace_twist = 0;						\
-  if(SE->_around_the_world && this->Params.twists[DIR % 4]){		\
-    if(sl == 1){							\
-      g = (F+1) % 2;							\
-    }else{								\
-      inplace_twist = 1;						\
-    }									\
-  }  
-
-#define LOAD_CHIMU_GPARITY_INPLACE_TWIST(DIR,F,PERM)			\
+#define LOAD_CHIMU \
   {const SiteSpinor & ref (in._odata[offset]);	\
-    LOAD_CHI_SETUP(DIR,F);						\
-    if(!inplace_twist){							\
-      LOAD_CHIMU_BODY(g);						\
-    }else{								\
-      if(  ( F==0 && ((distance == 1 && !perm) || (distance == -1 && perm)) ) || \
-	   ( F==1 && ((distance == -1 && !perm) || (distance == 1 && perm)) ) ){ \
-	DO_TWIST_0L_1H(Chimu_00,0,0,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_0L_1H(Chimu_01,0,1,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_0L_1H(Chimu_02,0,2,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_0L_1H(Chimu_10,1,0,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_0L_1H(Chimu_11,1,1,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_0L_1H(Chimu_12,1,2,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_0L_1H(Chimu_20,2,0,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_0L_1H(Chimu_21,2,1,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_0L_1H(Chimu_22,2,2,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_0L_1H(Chimu_30,3,0,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_0L_1H(Chimu_31,3,1,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_0L_1H(Chimu_32,3,2,F,PERM,  U_11,U_20,U_21);		\
-      }else{								\
-	DO_TWIST_1L_0H(Chimu_00,0,0,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_1L_0H(Chimu_01,0,1,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_1L_0H(Chimu_02,0,2,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_1L_0H(Chimu_10,1,0,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_1L_0H(Chimu_11,1,1,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_1L_0H(Chimu_12,1,2,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_1L_0H(Chimu_20,2,0,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_1L_0H(Chimu_21,2,1,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_1L_0H(Chimu_22,2,2,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_1L_0H(Chimu_30,3,0,F,PERM,  U_11,U_20,U_21);		\
-	DO_TWIST_1L_0H(Chimu_31,3,1,F,PERM,  U_00,U_01,U_10);		\
-	DO_TWIST_1L_0H(Chimu_32,3,2,F,PERM,  U_11,U_20,U_21);		\
-      } \
-    } \
-  }
+    Chimu_00=ref()(0)(0);\
+    Chimu_01=ref()(0)(1);\
+    Chimu_02=ref()(0)(2);\
+    Chimu_10=ref()(1)(0);\
+    Chimu_11=ref()(1)(1);\
+    Chimu_12=ref()(1)(2);\
+    Chimu_20=ref()(2)(0);\
+    Chimu_21=ref()(2)(1);\
+    Chimu_22=ref()(2)(2);\
+    Chimu_30=ref()(3)(0);\
+    Chimu_31=ref()(3)(1);\
+    Chimu_32=ref()(3)(2);}
 
-
-#define LOAD_CHI_GPARITY_INPLACE_TWIST(DIR,F,PERM)				\
+#define LOAD_CHI\
   {const SiteHalfSpinor &ref(buf[offset]);	\
-    LOAD_CHI_SETUP(DIR,F);						\
-    if(!inplace_twist){							\
-      LOAD_CHI_BODY(g);							\
-    }else{								\
-      if(  ( F==0 && ((distance == 1 && !perm) || (distance == -1 && perm)) ) || \
-	   ( F==1 && ((distance == -1 && !perm) || (distance == 1 && perm)) ) ){ \
-	DO_TWIST_0L_1H(Chi_00,0,0,F,PERM,  U_00,U_01,U_10);			\
-	DO_TWIST_0L_1H(Chi_01,0,1,F,PERM,  U_11,U_20,U_21);			\
-	DO_TWIST_0L_1H(Chi_02,0,2,F,PERM,  UChi_00,UChi_01,UChi_02);		\
-	DO_TWIST_0L_1H(Chi_10,1,0,F,PERM,  UChi_10,UChi_11,UChi_12);		\
-	DO_TWIST_0L_1H(Chi_11,1,1,F,PERM,  U_00,U_01,U_10);			\
-	DO_TWIST_0L_1H(Chi_12,1,2,F,PERM,  U_11,U_20,U_21);			\
-      }else{								\
-	DO_TWIST_1L_0H(Chi_00,0,0,F,PERM,  U_00,U_01,U_10);			\
-	DO_TWIST_1L_0H(Chi_01,0,1,F,PERM,  U_11,U_20,U_21);			\
-	DO_TWIST_1L_0H(Chi_02,0,2,F,PERM,  UChi_00,UChi_01,UChi_02);		\
-	DO_TWIST_1L_0H(Chi_10,1,0,F,PERM,  UChi_10,UChi_11,UChi_12);		\
-	DO_TWIST_1L_0H(Chi_11,1,1,F,PERM,  U_00,U_01,U_10);			\
-	DO_TWIST_1L_0H(Chi_12,1,2,F,PERM,  U_11,U_20,U_21);			\
-      }									\
-    }									\
-  }
-
-
-#define LOAD_CHI_GPARITY(DIR,F,PERM) LOAD_CHI_GPARITY_INPLACE_TWIST(DIR,F,PERM)
-#define LOAD_CHIMU_GPARITY(DIR,F,PERM) LOAD_CHIMU_GPARITY_INPLACE_TWIST(DIR,F,PERM)
+    Chi_00 = ref()(0)(0);\
+    Chi_01 = ref()(0)(1);\
+    Chi_02 = ref()(0)(2);\
+    Chi_10 = ref()(1)(0);\
+    Chi_11 = ref()(1)(1);\
+    Chi_12 = ref()(1)(2);}
 
 // To splat or not to splat depends on the implementation
-#define MULT_2SPIN_BODY \
+#define MULT_2SPIN(A)\
+  {auto & ref(U._odata[sU](A));			\
    Impl::loadLinkElement(U_00,ref()(0,0));	\
    Impl::loadLinkElement(U_10,ref()(1,0));	\
    Impl::loadLinkElement(U_20,ref()(2,0));	\
@@ -197,14 +83,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     UChi_01+= U_10*Chi_02;\
     UChi_11+= U_10*Chi_12;\
     UChi_02+= U_20*Chi_02;\
-  UChi_12+= U_20*Chi_12
-
-
-#define MULT_2SPIN(A,F)					\
-  {auto & ref(U._odata[sU](A)); MULT_2SPIN_BODY; }
-
-#define MULT_2SPIN_GPARITY(A,F)				\
-  {auto & ref(U._odata[sU](F)(A)); MULT_2SPIN_BODY; }
+    UChi_12+= U_20*Chi_12;}
 
 
 #define PERMUTE_DIR(dir)			\
@@ -428,87 +307,84 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
   result_31-= UChi_11;	\
   result_32-= UChi_12;
 
-#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
   offset = SE->_offset;				\
   local  = SE->_is_local;			\
   perm   = SE->_permute;			\
   if ( local ) {				\
-    LOAD_CHIMU_IMPL(DIR,F,PERM);			\
+    LOAD_CHIMU;					\
     PROJ;					\
     if ( perm) {				\
       PERMUTE_DIR(PERM);			\
     }						\
   } else {					\
-    LOAD_CHI_IMPL(DIR,F,PERM);			\
+    LOAD_CHI;					\
   }						\
-  MULT_2SPIN_IMPL(DIR,F);			\
+  MULT_2SPIN(DIR);				\
   RECON;					
 
-
-#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL)	\
+#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
   offset = SE->_offset;				\
   local  = SE->_is_local;			\
   perm   = SE->_permute;			\
   if ( local ) {				\
-    LOAD_CHIMU_IMPL(DIR,F,PERM);			\
+    LOAD_CHIMU;					\
     PROJ;					\
     if ( perm) {				\
       PERMUTE_DIR(PERM);			\
     }						\
   } else if ( st.same_node[DIR] ) {		\
-    LOAD_CHI_IMPL(DIR,F,PERM);			\
+    LOAD_CHI;					\
   }						\
   if (local || st.same_node[DIR] ) {		\
-    MULT_2SPIN_IMPL(DIR,F);			\
+    MULT_2SPIN(DIR);				\
     RECON;					\
   }
 
-#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL)	\
+#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON)	\
   SE=st.GetEntry(ptype,DIR,ss);			\
   offset = SE->_offset;				\
-  local  = SE->_is_local;			\
-  perm   = SE->_permute;			\
   if((!SE->_is_local)&&(!st.same_node[DIR]) ) {	\
-    LOAD_CHI_IMPL(DIR,F,PERM);			\
-    MULT_2SPIN_IMPL(DIR,F);			\
+    LOAD_CHI;					\
+    MULT_2SPIN(DIR);				\
     RECON;					\
     nmu++;					\
   }
 
-#define HAND_RESULT(ss,F)			\
+#define HAND_RESULT(ss)				\
   {						\
     SiteSpinor & ref (out._odata[ss]);		\
-    vstream(ref(F)(0)(0),result_00);		\
-    vstream(ref(F)(0)(1),result_01);		\
-    vstream(ref(F)(0)(2),result_02);		\
-    vstream(ref(F)(1)(0),result_10);		\
-    vstream(ref(F)(1)(1),result_11);		\
-    vstream(ref(F)(1)(2),result_12);		\
-    vstream(ref(F)(2)(0),result_20);		\
-    vstream(ref(F)(2)(1),result_21);		\
-    vstream(ref(F)(2)(2),result_22);		\
-    vstream(ref(F)(3)(0),result_30);		\
-    vstream(ref(F)(3)(1),result_31);		\
-    vstream(ref(F)(3)(2),result_32);		\
+    vstream(ref()(0)(0),result_00);		\
+    vstream(ref()(0)(1),result_01);		\
+    vstream(ref()(0)(2),result_02);		\
+    vstream(ref()(1)(0),result_10);		\
+    vstream(ref()(1)(1),result_11);		\
+    vstream(ref()(1)(2),result_12);		\
+    vstream(ref()(2)(0),result_20);		\
+    vstream(ref()(2)(1),result_21);		\
+    vstream(ref()(2)(2),result_22);		\
+    vstream(ref()(3)(0),result_30);		\
+    vstream(ref()(3)(1),result_31);		\
+    vstream(ref()(3)(2),result_32);		\
   }
 
-#define HAND_RESULT_EXT(ss,F)			\
+#define HAND_RESULT_EXT(ss)			\
   if (nmu){					\
     SiteSpinor & ref (out._odata[ss]);		\
-    ref(F)(0)(0)+=result_00;		\
-    ref(F)(0)(1)+=result_01;		\
-    ref(F)(0)(2)+=result_02;		\
-    ref(F)(1)(0)+=result_10;		\
-    ref(F)(1)(1)+=result_11;		\
-    ref(F)(1)(2)+=result_12;		\
-    ref(F)(2)(0)+=result_20;		\
-    ref(F)(2)(1)+=result_21;		\
-    ref(F)(2)(2)+=result_22;		\
-    ref(F)(3)(0)+=result_30;		\
-    ref(F)(3)(1)+=result_31;		\
-    ref(F)(3)(2)+=result_32;		\
+    ref()(0)(0)+=result_00;		\
+    ref()(0)(1)+=result_01;		\
+    ref()(0)(2)+=result_02;		\
+    ref()(1)(0)+=result_10;		\
+    ref()(1)(1)+=result_11;		\
+    ref()(1)(2)+=result_12;		\
+    ref()(2)(0)+=result_20;		\
+    ref()(2)(1)+=result_21;		\
+    ref()(2)(2)+=result_22;		\
+    ref()(3)(0)+=result_30;		\
+    ref()(3)(1)+=result_31;		\
+    ref()(3)(2)+=result_32;		\
   }
 
 
@@ -587,18 +463,15 @@ WilsonKernels<Impl>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGauge
   int offset,local,perm, ptype;
   StencilEntry *SE;
 
-#define HAND_DOP_SITE(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
-  HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);	\
-  HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_RESULT(ss,F)
-
-  HAND_DOP_SITE(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+  HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
+  HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
 }
 
 template<class Impl>
@@ -613,18 +486,15 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,Doub
   StencilEntry *SE;
   int offset,local,perm, ptype;
   
-#define HAND_DOP_SITE_DAG(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
-  HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_RESULT(ss,F)
-
-  HAND_DOP_SITE_DAG(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+  HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
+  HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT(ss);
 }
 
 template<class Impl> void 
@@ -639,20 +509,16 @@ WilsonKernels<Impl>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGa
 
   int offset,local,perm, ptype;
   StencilEntry *SE;
-
-#define HAND_DOP_SITE_INT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
-  ZERO_RESULT; \
-  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_RESULT(ss,F)
-
-  HAND_DOP_SITE_INT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT(ss);
 }
 
 template<class Impl>
@@ -666,20 +532,16 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,D
 
   StencilEntry *SE;
   int offset,local,perm, ptype;
-
-#define HAND_DOP_SITE_DAG_INT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL)				\
-  ZERO_RESULT;							\
-  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
-  HAND_RESULT(ss,F)
-  
-  HAND_DOP_SITE_DAG_INT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT(ss);
 }
 
 template<class Impl> void 
@@ -695,20 +557,16 @@ WilsonKernels<Impl>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGa
   int offset,local,perm, ptype;
   StencilEntry *SE;
   int nmu=0;
-
-#define HAND_DOP_SITE_EXT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
-  ZERO_RESULT; \
-  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_RESULT_EXT(ss,F)
-
-  HAND_DOP_SITE_EXT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
+  HAND_RESULT_EXT(ss);
 }
 
 template<class Impl>
@@ -723,193 +581,18 @@ void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,D
   StencilEntry *SE;
   int offset,local,perm, ptype;
   int nmu=0;
-
-#define HAND_DOP_SITE_DAG_EXT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
-  ZERO_RESULT; \
-  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
-  HAND_RESULT_EXT(ss,F)
-
-  HAND_DOP_SITE_DAG_EXT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+  ZERO_RESULT;
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
+  HAND_RESULT_EXT(ss);
 }
 
-  ////////////////////////////////////////////////
-  // Specialise Gparity to simple implementation
-  ////////////////////////////////////////////////
-#define HAND_SPECIALISE_EMPTY(IMPL)					\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSite(StencilImpl &st,			\
-				    LebesgueOrder &lo,			\
-				    DoubledGaugeField &U,		\
-				    SiteHalfSpinor *buf,		\
-				    int sF,int sU,			\
-				    const FermionField &in,		\
-				    FermionField &out){ assert(0); }	\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteDag(StencilImpl &st,			\
-				    LebesgueOrder &lo,			\
-				    DoubledGaugeField &U,		\
-				    SiteHalfSpinor *buf,		\
-				    int sF,int sU,			\
-				    const FermionField &in,		\
-				    FermionField &out){ assert(0); }	\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteInt(StencilImpl &st,			\
-				    LebesgueOrder &lo,			\
-				    DoubledGaugeField &U,		\
-				    SiteHalfSpinor *buf,		\
-				    int sF,int sU,			\
-				    const FermionField &in,		\
-				    FermionField &out){ assert(0); }	\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteExt(StencilImpl &st,			\
-				    LebesgueOrder &lo,			\
-				    DoubledGaugeField &U,		\
-				    SiteHalfSpinor *buf,		\
-				    int sF,int sU,			\
-				    const FermionField &in,		\
-				    FermionField &out){ assert(0); }	\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilImpl &st,	       	\
-				    LebesgueOrder &lo,			\
-				    DoubledGaugeField &U,		\
-				    SiteHalfSpinor *buf,		\
-				    int sF,int sU,			\
-				    const FermionField &in,		\
-				    FermionField &out){ assert(0); }	\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilImpl &st,	       	\
-				    LebesgueOrder &lo,			\
-				    DoubledGaugeField &U,		\
-				    SiteHalfSpinor *buf,		\
-				    int sF,int sU,			\
-				    const FermionField &in,		\
-				    FermionField &out){ assert(0); }	\
-
-
-
-#define HAND_SPECIALISE_GPARITY(IMPL)					\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf, \
-				    int ss,int sU,const FermionField &in, FermionField &out) \
-  {									\
-    typedef IMPL Impl;							\
-    typedef typename Simd::scalar_type S;				\
-    typedef typename Simd::vector_type V;				\
-									\
-    HAND_DECLARATIONS(ignore);						\
-									\
-    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
-    StencilEntry *SE;							\
-    HAND_DOP_SITE(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-    HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-  }									\
-									\
-  template<>								\
-  void WilsonKernels<IMPL>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
-					    int ss,int sU,const FermionField &in, FermionField &out) \
-  {									\
-    typedef IMPL Impl;							\
-    typedef typename Simd::scalar_type S;				\
-    typedef typename Simd::vector_type V;				\
-									\
-    HAND_DECLARATIONS(ignore);						\
-									\
-    StencilEntry *SE;							\
-    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist;					\
-    HAND_DOP_SITE_DAG(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-    HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-  }									\
-									\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf, \
-						     int ss,int sU,const FermionField &in, FermionField &out) \
-  {									\
-    typedef IMPL Impl;							\
-    typedef typename Simd::scalar_type S;				\
-    typedef typename Simd::vector_type V;				\
-									\
-    HAND_DECLARATIONS(ignore);						\
-									\
-    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist;					\
-    StencilEntry *SE;							\
-    HAND_DOP_SITE_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-    HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-  }									\
-									\
-  template<>								\
-  void WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
-							     int ss,int sU,const FermionField &in, FermionField &out) \
-  {									\
-    typedef IMPL Impl;							\
-    typedef typename Simd::scalar_type S;				\
-    typedef typename Simd::vector_type V;				\
-									\
-    HAND_DECLARATIONS(ignore);						\
-									\
-    StencilEntry *SE;							\
-    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
-    HAND_DOP_SITE_DAG_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-    HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-  }									\
-									\
-  template<> void							\
-  WilsonKernels<IMPL>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf, \
-						     int ss,int sU,const FermionField &in, FermionField &out) \
-  {									\
-    typedef IMPL Impl;							\
-    typedef typename Simd::scalar_type S;				\
-    typedef typename Simd::vector_type V;				\
-									\
-    HAND_DECLARATIONS(ignore);						\
-									\
-    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
-    StencilEntry *SE;							\
-    int nmu=0;								\
-    HAND_DOP_SITE_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-    nmu = 0;								\
-    HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-  }									\
-  template<>								\
-  void WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
-							     int ss,int sU,const FermionField &in, FermionField &out) \
-  {									\
-    typedef IMPL Impl;							\
-    typedef typename Simd::scalar_type S;				\
-    typedef typename Simd::vector_type V;				\
-									\
-    HAND_DECLARATIONS(ignore);						\
-									\
-    StencilEntry *SE;							\
-    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
-    int nmu=0;								\
-    HAND_DOP_SITE_DAG_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-    nmu = 0;								\
-    HAND_DOP_SITE_DAG_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
-  }
-
-
-HAND_SPECIALISE_GPARITY(GparityWilsonImplF);
-HAND_SPECIALISE_GPARITY(GparityWilsonImplD);
-HAND_SPECIALISE_GPARITY(GparityWilsonImplFH);
-HAND_SPECIALISE_GPARITY(GparityWilsonImplDF);
-
-
-
-
-
-
-
-
-
-
-  
 ////////////// Wilson ; uses this implementation /////////////////////
 
 #define INSTANTIATE_THEM(A) \
@@ -930,8 +613,6 @@ INSTANTIATE_THEM(WilsonImplF);
 INSTANTIATE_THEM(WilsonImplD);
 INSTANTIATE_THEM(ZWilsonImplF);
 INSTANTIATE_THEM(ZWilsonImplD);
-INSTANTIATE_THEM(GparityWilsonImplF);
-INSTANTIATE_THEM(GparityWilsonImplD);
 INSTANTIATE_THEM(DomainWallVec5dImplF);
 INSTANTIATE_THEM(DomainWallVec5dImplD);
 INSTANTIATE_THEM(ZDomainWallVec5dImplF);
@@ -940,12 +621,11 @@ INSTANTIATE_THEM(WilsonImplFH);
 INSTANTIATE_THEM(WilsonImplDF);
 INSTANTIATE_THEM(ZWilsonImplFH);
 INSTANTIATE_THEM(ZWilsonImplDF);
-INSTANTIATE_THEM(GparityWilsonImplFH);
-INSTANTIATE_THEM(GparityWilsonImplDF);
 INSTANTIATE_THEM(DomainWallVec5dImplFH);
 INSTANTIATE_THEM(DomainWallVec5dImplDF);
 INSTANTIATE_THEM(ZDomainWallVec5dImplFH);
 INSTANTIATE_THEM(ZDomainWallVec5dImplDF);
 INSTANTIATE_THEM(WilsonTwoIndexAntiSymmetricImplF);
 INSTANTIATE_THEM(WilsonTwoIndexAntiSymmetricImplD);
+
 }}

lib/qcd/action/fermion/WilsonKernelsHandGparity.cc

@@ -0,0 +1,878 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/WilsonKernelsHand.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/qcd/action/fermion/FermionCore.h>
+
+#define REGISTER
+
+#define LOAD_CHIMU_BODY(F)			\
+  Chimu_00=ref(F)(0)(0);			\
+  Chimu_01=ref(F)(0)(1);			\
+  Chimu_02=ref(F)(0)(2);			\
+  Chimu_10=ref(F)(1)(0);			\
+  Chimu_11=ref(F)(1)(1);			\
+  Chimu_12=ref(F)(1)(2);			\
+  Chimu_20=ref(F)(2)(0);			\
+  Chimu_21=ref(F)(2)(1);			\
+  Chimu_22=ref(F)(2)(2);			\
+  Chimu_30=ref(F)(3)(0);			\
+  Chimu_31=ref(F)(3)(1);			\
+  Chimu_32=ref(F)(3)(2)
+
+#define LOAD_CHIMU(DIR,F,PERM)						\
+  { const SiteSpinor & ref (in._odata[offset]); LOAD_CHIMU_BODY(F); }
+
+#define LOAD_CHI_BODY(F)				\
+    Chi_00 = ref(F)(0)(0);\
+    Chi_01 = ref(F)(0)(1);\
+    Chi_02 = ref(F)(0)(2);\
+    Chi_10 = ref(F)(1)(0);\
+    Chi_11 = ref(F)(1)(1);\
+    Chi_12 = ref(F)(1)(2)
+
+#define LOAD_CHI(DIR,F,PERM)					\
+  {const SiteHalfSpinor &ref(buf[offset]); LOAD_CHI_BODY(F); }
+
+
+//G-parity implementations using in-place intrinsic ops
+
+//1l 1h -> 1h 1l
+//0l 0h , 1h 1l -> 0l 1h 0h,1l
+//0h,1l -> 1l,0h
+//if( (distance == 1 && !perm_will_occur) || (distance == -1 && perm_will_occur) )
+//Pulled fermion through forwards face, GPBC on upper component
+//Need 0= 0l 1h   1= 1l 0h
+//else if( (distance == -1 && !perm) || (distance == 1 && perm) )
+//Pulled fermion through backwards face, GPBC on lower component
+//Need 0= 1l 0h   1= 0l 1h
+
+//1l 1h -> 1h 1l
+//0l 0h , 1h 1l -> 0l 1h 0h,1l
+#define DO_TWIST_0L_1H(INTO,S,C,F, PERM, tmp1, tmp2, tmp3)			\
+  permute##PERM(tmp1, ref(1)(S)(C));				\
+  exchange##PERM(tmp2,tmp3, ref(0)(S)(C), tmp1);		\
+  INTO = tmp2;
+
+//0l 0h -> 0h 0l
+//1l 1h, 0h 0l -> 1l 0h, 1h 0l
+#define DO_TWIST_1L_0H(INTO,S,C,F, PERM, tmp1, tmp2, tmp3)			\
+  permute##PERM(tmp1, ref(0)(S)(C));				\
+  exchange##PERM(tmp2,tmp3, ref(1)(S)(C), tmp1);		\
+  INTO = tmp2;
+
+
+
+
+#define LOAD_CHI_SETUP(DIR,F)						\
+  g = F;								\
+  direction = st._directions[DIR];				\
+  distance = st._distances[DIR];				\
+  sl = st._grid->_simd_layout[direction];			\
+  inplace_twist = 0;						\
+  if(SE->_around_the_world && this->Params.twists[DIR % 4]){		\
+    if(sl == 1){							\
+      g = (F+1) % 2;							\
+    }else{								\
+      inplace_twist = 1;						\
+    }									\
+  }  
+
+#define LOAD_CHIMU_GPARITY_INPLACE_TWIST(DIR,F,PERM)			\
+  { const SiteSpinor &ref(in._odata[offset]);				\
+    LOAD_CHI_SETUP(DIR,F);						\
+    if(!inplace_twist){							\
+      LOAD_CHIMU_BODY(g);						\
+    }else{								\
+      if(  ( F==0 && ((distance == 1 && !perm) || (distance == -1 && perm)) ) || \
+	   ( F==1 && ((distance == -1 && !perm) || (distance == 1 && perm)) ) ){ \
+	DO_TWIST_0L_1H(Chimu_00,0,0,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_0L_1H(Chimu_01,0,1,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_0L_1H(Chimu_02,0,2,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_0L_1H(Chimu_10,1,0,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_0L_1H(Chimu_11,1,1,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_0L_1H(Chimu_12,1,2,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_0L_1H(Chimu_20,2,0,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_0L_1H(Chimu_21,2,1,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_0L_1H(Chimu_22,2,2,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_0L_1H(Chimu_30,3,0,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_0L_1H(Chimu_31,3,1,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_0L_1H(Chimu_32,3,2,F,PERM,  U_11,U_20,U_21);		\
+      }else{								\
+	DO_TWIST_1L_0H(Chimu_00,0,0,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_1L_0H(Chimu_01,0,1,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_1L_0H(Chimu_02,0,2,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_1L_0H(Chimu_10,1,0,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_1L_0H(Chimu_11,1,1,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_1L_0H(Chimu_12,1,2,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_1L_0H(Chimu_20,2,0,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_1L_0H(Chimu_21,2,1,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_1L_0H(Chimu_22,2,2,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_1L_0H(Chimu_30,3,0,F,PERM,  U_11,U_20,U_21);		\
+	DO_TWIST_1L_0H(Chimu_31,3,1,F,PERM,  U_00,U_01,U_10);		\
+	DO_TWIST_1L_0H(Chimu_32,3,2,F,PERM,  U_11,U_20,U_21);		\
+      } \
+    } \
+  }
+
+
+#define LOAD_CHI_GPARITY_INPLACE_TWIST(DIR,F,PERM)				\
+  { const SiteHalfSpinor &ref(buf[offset]);				\
+    LOAD_CHI_SETUP(DIR,F);						\
+    if(!inplace_twist){							\
+      LOAD_CHI_BODY(g);							\
+    }else{								\
+      if(  ( F==0 && ((distance == 1 && !perm) || (distance == -1 && perm)) ) || \
+	   ( F==1 && ((distance == -1 && !perm) || (distance == 1 && perm)) ) ){ \
+	DO_TWIST_0L_1H(Chi_00,0,0,F,PERM,  U_00,U_01,U_10);			\
+	DO_TWIST_0L_1H(Chi_01,0,1,F,PERM,  U_11,U_20,U_21);			\
+	DO_TWIST_0L_1H(Chi_02,0,2,F,PERM,  UChi_00,UChi_01,UChi_02);		\
+	DO_TWIST_0L_1H(Chi_10,1,0,F,PERM,  UChi_10,UChi_11,UChi_12);		\
+	DO_TWIST_0L_1H(Chi_11,1,1,F,PERM,  U_00,U_01,U_10);			\
+	DO_TWIST_0L_1H(Chi_12,1,2,F,PERM,  U_11,U_20,U_21);			\
+      }else{								\
+	DO_TWIST_1L_0H(Chi_00,0,0,F,PERM,  U_00,U_01,U_10);			\
+	DO_TWIST_1L_0H(Chi_01,0,1,F,PERM,  U_11,U_20,U_21);			\
+	DO_TWIST_1L_0H(Chi_02,0,2,F,PERM,  UChi_00,UChi_01,UChi_02);		\
+	DO_TWIST_1L_0H(Chi_10,1,0,F,PERM,  UChi_10,UChi_11,UChi_12);		\
+	DO_TWIST_1L_0H(Chi_11,1,1,F,PERM,  U_00,U_01,U_10);			\
+	DO_TWIST_1L_0H(Chi_12,1,2,F,PERM,  U_11,U_20,U_21);			\
+      }									\
+    }									\
+  }
+
+
+#define LOAD_CHI_GPARITY(DIR,F,PERM) LOAD_CHI_GPARITY_INPLACE_TWIST(DIR,F,PERM)
+#define LOAD_CHIMU_GPARITY(DIR,F,PERM) LOAD_CHIMU_GPARITY_INPLACE_TWIST(DIR,F,PERM)
+
+// To splat or not to splat depends on the implementation
+#define MULT_2SPIN_BODY \
+  Impl::loadLinkElement(U_00,ref()(0,0));	\
+  Impl::loadLinkElement(U_10,ref()(1,0));	\
+  Impl::loadLinkElement(U_20,ref()(2,0));	\
+  Impl::loadLinkElement(U_01,ref()(0,1));	\
+  Impl::loadLinkElement(U_11,ref()(1,1));	\
+  Impl::loadLinkElement(U_21,ref()(2,1));	\
+  UChi_00 = U_00*Chi_00;			\
+  UChi_10 = U_00*Chi_10;			\
+  UChi_01 = U_10*Chi_00;			\
+  UChi_11 = U_10*Chi_10;			\
+  UChi_02 = U_20*Chi_00;			\
+  UChi_12 = U_20*Chi_10;			\
+  UChi_00+= U_01*Chi_01;			\
+  UChi_10+= U_01*Chi_11;			\
+  UChi_01+= U_11*Chi_01;			\
+  UChi_11+= U_11*Chi_11;			\
+  UChi_02+= U_21*Chi_01;			\
+  UChi_12+= U_21*Chi_11;			\
+  Impl::loadLinkElement(U_00,ref()(0,2));	\
+  Impl::loadLinkElement(U_10,ref()(1,2));	\
+  Impl::loadLinkElement(U_20,ref()(2,2));	\
+  UChi_00+= U_00*Chi_02;			\
+  UChi_10+= U_00*Chi_12;			\
+  UChi_01+= U_10*Chi_02;			\
+  UChi_11+= U_10*Chi_12;			\
+  UChi_02+= U_20*Chi_02;			\
+  UChi_12+= U_20*Chi_12
+
+
+#define MULT_2SPIN(A,F)					\
+  {auto & ref(U._odata[sU](A)); MULT_2SPIN_BODY; }
+
+#define MULT_2SPIN_GPARITY(A,F)				\
+  {auto & ref(U._odata[sU](F)(A)); MULT_2SPIN_BODY; }
+
+
+#define PERMUTE_DIR(dir)			\
+      permute##dir(Chi_00,Chi_00);\
+      permute##dir(Chi_01,Chi_01);\
+      permute##dir(Chi_02,Chi_02);\
+      permute##dir(Chi_10,Chi_10);\
+      permute##dir(Chi_11,Chi_11);\
+      permute##dir(Chi_12,Chi_12);
+
+//      hspin(0)=fspin(0)+timesI(fspin(3));
+//      hspin(1)=fspin(1)+timesI(fspin(2));
+#define XP_PROJ \
+    Chi_00 = Chimu_00+timesI(Chimu_30);\
+    Chi_01 = Chimu_01+timesI(Chimu_31);\
+    Chi_02 = Chimu_02+timesI(Chimu_32);\
+    Chi_10 = Chimu_10+timesI(Chimu_20);\
+    Chi_11 = Chimu_11+timesI(Chimu_21);\
+    Chi_12 = Chimu_12+timesI(Chimu_22);
+
+#define YP_PROJ \
+    Chi_00 = Chimu_00-Chimu_30;\
+    Chi_01 = Chimu_01-Chimu_31;\
+    Chi_02 = Chimu_02-Chimu_32;\
+    Chi_10 = Chimu_10+Chimu_20;\
+    Chi_11 = Chimu_11+Chimu_21;\
+    Chi_12 = Chimu_12+Chimu_22;
+
+#define ZP_PROJ \
+  Chi_00 = Chimu_00+timesI(Chimu_20);		\
+  Chi_01 = Chimu_01+timesI(Chimu_21);		\
+  Chi_02 = Chimu_02+timesI(Chimu_22);		\
+  Chi_10 = Chimu_10-timesI(Chimu_30);		\
+  Chi_11 = Chimu_11-timesI(Chimu_31);		\
+  Chi_12 = Chimu_12-timesI(Chimu_32);
+
+#define TP_PROJ \
+  Chi_00 = Chimu_00+Chimu_20;		\
+  Chi_01 = Chimu_01+Chimu_21;		\
+  Chi_02 = Chimu_02+Chimu_22;		\
+  Chi_10 = Chimu_10+Chimu_30;		\
+  Chi_11 = Chimu_11+Chimu_31;		\
+  Chi_12 = Chimu_12+Chimu_32;
+
+
+//      hspin(0)=fspin(0)-timesI(fspin(3));
+//      hspin(1)=fspin(1)-timesI(fspin(2));
+#define XM_PROJ \
+    Chi_00 = Chimu_00-timesI(Chimu_30);\
+    Chi_01 = Chimu_01-timesI(Chimu_31);\
+    Chi_02 = Chimu_02-timesI(Chimu_32);\
+    Chi_10 = Chimu_10-timesI(Chimu_20);\
+    Chi_11 = Chimu_11-timesI(Chimu_21);\
+    Chi_12 = Chimu_12-timesI(Chimu_22);
+
+#define YM_PROJ \
+    Chi_00 = Chimu_00+Chimu_30;\
+    Chi_01 = Chimu_01+Chimu_31;\
+    Chi_02 = Chimu_02+Chimu_32;\
+    Chi_10 = Chimu_10-Chimu_20;\
+    Chi_11 = Chimu_11-Chimu_21;\
+    Chi_12 = Chimu_12-Chimu_22;
+
+#define ZM_PROJ \
+  Chi_00 = Chimu_00-timesI(Chimu_20);		\
+  Chi_01 = Chimu_01-timesI(Chimu_21);		\
+  Chi_02 = Chimu_02-timesI(Chimu_22);		\
+  Chi_10 = Chimu_10+timesI(Chimu_30);		\
+  Chi_11 = Chimu_11+timesI(Chimu_31);		\
+  Chi_12 = Chimu_12+timesI(Chimu_32);
+
+#define TM_PROJ \
+  Chi_00 = Chimu_00-Chimu_20;		\
+  Chi_01 = Chimu_01-Chimu_21;		\
+  Chi_02 = Chimu_02-Chimu_22;		\
+  Chi_10 = Chimu_10-Chimu_30;		\
+  Chi_11 = Chimu_11-Chimu_31;		\
+  Chi_12 = Chimu_12-Chimu_32;
+
+//      fspin(0)=hspin(0);
+//      fspin(1)=hspin(1);
+//      fspin(2)=timesMinusI(hspin(1));
+//      fspin(3)=timesMinusI(hspin(0));
+#define XP_RECON\
+  result_00 = UChi_00;\
+  result_01 = UChi_01;\
+  result_02 = UChi_02;\
+  result_10 = UChi_10;\
+  result_11 = UChi_11;\
+  result_12 = UChi_12;\
+  result_20 = timesMinusI(UChi_10);\
+  result_21 = timesMinusI(UChi_11);\
+  result_22 = timesMinusI(UChi_12);\
+  result_30 = timesMinusI(UChi_00);\
+  result_31 = timesMinusI(UChi_01);\
+  result_32 = timesMinusI(UChi_02);
+
+#define XP_RECON_ACCUM\
+  result_00+=UChi_00;\
+  result_01+=UChi_01;\
+  result_02+=UChi_02;\
+  result_10+=UChi_10;\
+  result_11+=UChi_11;\
+  result_12+=UChi_12;\
+  result_20-=timesI(UChi_10);\
+  result_21-=timesI(UChi_11);\
+  result_22-=timesI(UChi_12);\
+  result_30-=timesI(UChi_00);\
+  result_31-=timesI(UChi_01);\
+  result_32-=timesI(UChi_02);
+
+#define XM_RECON\
+  result_00 = UChi_00;\
+  result_01 = UChi_01;\
+  result_02 = UChi_02;\
+  result_10 = UChi_10;\
+  result_11 = UChi_11;\
+  result_12 = UChi_12;\
+  result_20 = timesI(UChi_10);\
+  result_21 = timesI(UChi_11);\
+  result_22 = timesI(UChi_12);\
+  result_30 = timesI(UChi_00);\
+  result_31 = timesI(UChi_01);\
+  result_32 = timesI(UChi_02);
+
+#define XM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= timesI(UChi_10);\
+  result_21+= timesI(UChi_11);\
+  result_22+= timesI(UChi_12);\
+  result_30+= timesI(UChi_00);\
+  result_31+= timesI(UChi_01);\
+  result_32+= timesI(UChi_02);
+
+#define YP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= UChi_10;\
+  result_21+= UChi_11;\
+  result_22+= UChi_12;\
+  result_30-= UChi_00;\
+  result_31-= UChi_01;\
+  result_32-= UChi_02;
+
+#define YM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= UChi_10;\
+  result_21-= UChi_11;\
+  result_22-= UChi_12;\
+  result_30+= UChi_00;\
+  result_31+= UChi_01;\
+  result_32+= UChi_02;
+
+#define ZP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= timesI(UChi_00);			\
+  result_21-= timesI(UChi_01);			\
+  result_22-= timesI(UChi_02);			\
+  result_30+= timesI(UChi_10);			\
+  result_31+= timesI(UChi_11);			\
+  result_32+= timesI(UChi_12);
+
+#define ZM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= timesI(UChi_00);			\
+  result_21+= timesI(UChi_01);			\
+  result_22+= timesI(UChi_02);			\
+  result_30-= timesI(UChi_10);			\
+  result_31-= timesI(UChi_11);			\
+  result_32-= timesI(UChi_12);
+
+#define TP_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20+= UChi_00;			\
+  result_21+= UChi_01;			\
+  result_22+= UChi_02;			\
+  result_30+= UChi_10;			\
+  result_31+= UChi_11;			\
+  result_32+= UChi_12;
+
+#define TM_RECON_ACCUM\
+  result_00+= UChi_00;\
+  result_01+= UChi_01;\
+  result_02+= UChi_02;\
+  result_10+= UChi_10;\
+  result_11+= UChi_11;\
+  result_12+= UChi_12;\
+  result_20-= UChi_00;	\
+  result_21-= UChi_01;	\
+  result_22-= UChi_02;	\
+  result_30-= UChi_10;	\
+  result_31-= UChi_11;	\
+  result_32-= UChi_12;
+
+#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU_IMPL(DIR,F,PERM);			\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else {					\
+    LOAD_CHI_IMPL(DIR,F,PERM);			\
+  }						\
+  MULT_2SPIN_IMPL(DIR,F);			\
+  RECON;					
+
+
+#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL)	\
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if ( local ) {				\
+    LOAD_CHIMU_IMPL(DIR,F,PERM);			\
+    PROJ;					\
+    if ( perm) {				\
+      PERMUTE_DIR(PERM);			\
+    }						\
+  } else if ( st.same_node[DIR] ) {		\
+    LOAD_CHI_IMPL(DIR,F,PERM);			\
+  }						\
+  if (local || st.same_node[DIR] ) {		\
+    MULT_2SPIN_IMPL(DIR,F);			\
+    RECON;					\
+  }
+
+#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL)	\
+  SE=st.GetEntry(ptype,DIR,ss);			\
+  offset = SE->_offset;				\
+  local  = SE->_is_local;			\
+  perm   = SE->_permute;			\
+  if((!SE->_is_local)&&(!st.same_node[DIR]) ) {	\
+    LOAD_CHI_IMPL(DIR,F,PERM);			\
+    MULT_2SPIN_IMPL(DIR,F);			\
+    RECON;					\
+    nmu++;					\
+  }
+
+#define HAND_RESULT(ss,F)			\
+  {						\
+    SiteSpinor & ref (out._odata[ss]);		\
+    vstream(ref(F)(0)(0),result_00);		\
+    vstream(ref(F)(0)(1),result_01);		\
+    vstream(ref(F)(0)(2),result_02);		\
+    vstream(ref(F)(1)(0),result_10);		\
+    vstream(ref(F)(1)(1),result_11);		\
+    vstream(ref(F)(1)(2),result_12);		\
+    vstream(ref(F)(2)(0),result_20);		\
+    vstream(ref(F)(2)(1),result_21);		\
+    vstream(ref(F)(2)(2),result_22);		\
+    vstream(ref(F)(3)(0),result_30);		\
+    vstream(ref(F)(3)(1),result_31);		\
+    vstream(ref(F)(3)(2),result_32);		\
+  }
+
+#define HAND_RESULT_EXT(ss,F)			\
+  if (nmu){					\
+    SiteSpinor & ref (out._odata[ss]);		\
+    ref(F)(0)(0)+=result_00;		\
+    ref(F)(0)(1)+=result_01;		\
+    ref(F)(0)(2)+=result_02;		\
+    ref(F)(1)(0)+=result_10;		\
+    ref(F)(1)(1)+=result_11;		\
+    ref(F)(1)(2)+=result_12;		\
+    ref(F)(2)(0)+=result_20;		\
+    ref(F)(2)(1)+=result_21;		\
+    ref(F)(2)(2)+=result_22;		\
+    ref(F)(3)(0)+=result_30;		\
+    ref(F)(3)(1)+=result_31;		\
+    ref(F)(3)(2)+=result_32;		\
+  }
+
+
+#define HAND_DECLARATIONS(a)			\
+  Simd result_00;				\
+  Simd result_01;				\
+  Simd result_02;				\
+  Simd result_10;				\
+  Simd result_11;				\
+  Simd result_12;				\
+  Simd result_20;				\
+  Simd result_21;				\
+  Simd result_22;				\
+  Simd result_30;				\
+  Simd result_31;				\
+  Simd result_32;				\
+  Simd Chi_00;					\
+  Simd Chi_01;					\
+  Simd Chi_02;					\
+  Simd Chi_10;					\
+  Simd Chi_11;					\
+  Simd Chi_12;					\
+  Simd UChi_00;					\
+  Simd UChi_01;					\
+  Simd UChi_02;					\
+  Simd UChi_10;					\
+  Simd UChi_11;					\
+  Simd UChi_12;					\
+  Simd U_00;					\
+  Simd U_10;					\
+  Simd U_20;					\
+  Simd U_01;					\
+  Simd U_11;					\
+  Simd U_21;
+
+#define ZERO_RESULT				\
+  result_00=zero;				\
+  result_01=zero;				\
+  result_02=zero;				\
+  result_10=zero;				\
+  result_11=zero;				\
+  result_12=zero;				\
+  result_20=zero;				\
+  result_21=zero;				\
+  result_22=zero;				\
+  result_30=zero;				\
+  result_31=zero;				\
+  result_32=zero;			
+
+#define Chimu_00 Chi_00
+#define Chimu_01 Chi_01
+#define Chimu_02 Chi_02
+#define Chimu_10 Chi_10
+#define Chimu_11 Chi_11
+#define Chimu_12 Chi_12
+#define Chimu_20 UChi_00
+#define Chimu_21 UChi_01
+#define Chimu_22 UChi_02
+#define Chimu_30 UChi_10
+#define Chimu_31 UChi_11
+#define Chimu_32 UChi_12
+
+namespace Grid {
+namespace QCD {
+
+template<class Impl> void 
+WilsonKernels<Impl>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionField &in, FermionField &out)
+{
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+
+#define HAND_DOP_SITE(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+  HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);	\
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_RESULT(ss,F)
+
+  HAND_DOP_SITE(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionField &in, FermionField &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset,local,perm, ptype;
+
+#define HAND_DOP_SITE_DAG(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+  HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_RESULT(ss,F)
+
+  HAND_DOP_SITE_DAG(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+}
+
+template<class Impl> void 
+WilsonKernels<Impl>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionField &in, FermionField &out)
+{
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+
+#define HAND_DOP_SITE_INT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+  ZERO_RESULT; \
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_RESULT(ss,F)
+
+  HAND_DOP_SITE_INT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionField &in, FermionField &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset,local,perm, ptype;
+
+#define HAND_DOP_SITE_DAG_INT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL)				\
+  ZERO_RESULT;							\
+  HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL);		\
+  HAND_RESULT(ss,F)
+  
+  HAND_DOP_SITE_DAG_INT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+}
+
+template<class Impl> void 
+WilsonKernels<Impl>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf,
+					  int ss,int sU,const FermionField &in, FermionField &out)
+{
+// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  int offset,local,perm, ptype;
+  StencilEntry *SE;
+  int nmu=0;
+
+#define HAND_DOP_SITE_EXT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+  ZERO_RESULT; \
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_RESULT_EXT(ss,F)
+
+  HAND_DOP_SITE_EXT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+}
+
+template<class Impl>
+void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
+						  int ss,int sU,const FermionField &in, FermionField &out)
+{
+  typedef typename Simd::scalar_type S;
+  typedef typename Simd::vector_type V;
+
+  HAND_DECLARATIONS(ignore);
+
+  StencilEntry *SE;
+  int offset,local,perm, ptype;
+  int nmu=0;
+
+#define HAND_DOP_SITE_DAG_EXT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
+  ZERO_RESULT; \
+  HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
+  HAND_RESULT_EXT(ss,F)
+
+  HAND_DOP_SITE_DAG_EXT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
+}
+
+#define HAND_SPECIALISE_GPARITY(IMPL)					\
+  template<> void							\
+  WilsonKernels<IMPL>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf, \
+				    int ss,int sU,const FermionField &in, FermionField &out) \
+  {									\
+    typedef IMPL Impl;							\
+    typedef typename Simd::scalar_type S;				\
+    typedef typename Simd::vector_type V;				\
+									\
+    HAND_DECLARATIONS(ignore);						\
+									\
+    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
+    StencilEntry *SE;							\
+    HAND_DOP_SITE(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+    HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+  }									\
+									\
+  template<>								\
+  void WilsonKernels<IMPL>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
+					    int ss,int sU,const FermionField &in, FermionField &out) \
+  {									\
+    typedef IMPL Impl;							\
+    typedef typename Simd::scalar_type S;				\
+    typedef typename Simd::vector_type V;				\
+									\
+    HAND_DECLARATIONS(ignore);						\
+									\
+    StencilEntry *SE;							\
+    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist;					\
+    HAND_DOP_SITE_DAG(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+    HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+  }									\
+									\
+  template<> void							\
+  WilsonKernels<IMPL>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf, \
+						     int ss,int sU,const FermionField &in, FermionField &out) \
+  {									\
+    typedef IMPL Impl;							\
+    typedef typename Simd::scalar_type S;				\
+    typedef typename Simd::vector_type V;				\
+									\
+    HAND_DECLARATIONS(ignore);						\
+									\
+    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist;					\
+    StencilEntry *SE;							\
+    HAND_DOP_SITE_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+    HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+  }									\
+									\
+  template<>								\
+  void WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
+							     int ss,int sU,const FermionField &in, FermionField &out) \
+  {									\
+    typedef IMPL Impl;							\
+    typedef typename Simd::scalar_type S;				\
+    typedef typename Simd::vector_type V;				\
+									\
+    HAND_DECLARATIONS(ignore);						\
+									\
+    StencilEntry *SE;							\
+    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
+    HAND_DOP_SITE_DAG_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+    HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+  }									\
+									\
+  template<> void							\
+  WilsonKernels<IMPL>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor  *buf, \
+						     int ss,int sU,const FermionField &in, FermionField &out) \
+  {									\
+    typedef IMPL Impl;							\
+    typedef typename Simd::scalar_type S;				\
+    typedef typename Simd::vector_type V;				\
+									\
+    HAND_DECLARATIONS(ignore);						\
+									\
+    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
+    StencilEntry *SE;							\
+    int nmu=0;								\
+    HAND_DOP_SITE_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+    nmu = 0;								\
+    HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+  }									\
+  template<>								\
+  void WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
+							     int ss,int sU,const FermionField &in, FermionField &out) \
+  {									\
+    typedef IMPL Impl;							\
+    typedef typename Simd::scalar_type S;				\
+    typedef typename Simd::vector_type V;				\
+									\
+    HAND_DECLARATIONS(ignore);						\
+									\
+    StencilEntry *SE;							\
+    int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
+    int nmu=0;								\
+    HAND_DOP_SITE_DAG_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+    nmu = 0;								\
+    HAND_DOP_SITE_DAG_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
+  }
+
+
+HAND_SPECIALISE_GPARITY(GparityWilsonImplF);
+HAND_SPECIALISE_GPARITY(GparityWilsonImplD);
+HAND_SPECIALISE_GPARITY(GparityWilsonImplFH);
+HAND_SPECIALISE_GPARITY(GparityWilsonImplDF);
+
+
+
+
+
+
+
+
+
+
+  
+////////////// Wilson ; uses this implementation /////////////////////
+
+#define INSTANTIATE_THEM(A) \
+template void WilsonKernels<A>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,\
+					     int ss,int sU,const FermionField &in, FermionField &out); \
+template void WilsonKernels<A>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
+						int ss,int sU,const FermionField &in, FermionField &out);\
+template void WilsonKernels<A>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,\
+						int ss,int sU,const FermionField &in, FermionField &out); \
+template void WilsonKernels<A>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
+						   int ss,int sU,const FermionField &in, FermionField &out); \
+template void WilsonKernels<A>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,\
+						int ss,int sU,const FermionField &in, FermionField &out); \
+template void WilsonKernels<A>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
+						   int ss,int sU,const FermionField &in, FermionField &out); 
+
+INSTANTIATE_THEM(GparityWilsonImplF);
+INSTANTIATE_THEM(GparityWilsonImplD);
+INSTANTIATE_THEM(GparityWilsonImplFH);
+INSTANTIATE_THEM(GparityWilsonImplDF);
+}}

lib/qcd/hmc/HMCResourceManager.h

@@ -48,6 +48,22 @@ with this program; if not, write to the Free Software Foundation, Inc.,
     }                                                                    \
   }
 
+#define RegisterLoadCheckPointerMetadataFunction(NAME)                   \
+  template < class Metadata >                                            \
+  void Load##NAME##Checkpointer(const CheckpointerParameters& Params_, const Metadata& M_) { \
+    if (!have_CheckPointer) {                                            \
+      std::cout << GridLogDebug << "Loading Metadata Checkpointer " << #NAME      \
+                << std::endl;                                            \
+      CP = std::unique_ptr<CheckpointerBaseModule>(                      \
+        new NAME##CPModule<ImplementationPolicy, Metadata >(Params_, M_));   \
+      have_CheckPointer = true;                                          \
+    } else {                                                             \
+      std::cout << GridLogError << "Checkpointer already loaded "        \
+                << std::endl;                                            \
+      exit(1);                                                           \
+    }                                                                    \
+  }
+
 namespace Grid {
 namespace QCD {
 
@@ -77,7 +93,7 @@ class HMCResourceManager {
   bool have_CheckPointer;
 
   // NOTE: operator << is not overloaded for std::vector<string> 
-  // so thsi function is necessary
+  // so this function is necessary
   void output_vector_string(const std::vector<std::string> &vs){
     for (auto &i: vs)
       std::cout << i << " ";
@@ -254,6 +270,7 @@ class HMCResourceManager {
   RegisterLoadCheckPointerFunction(Nersc);
   #ifdef HAVE_LIME
   RegisterLoadCheckPointerFunction(ILDG);
+  RegisterLoadCheckPointerMetadataFunction(Scidac);
   #endif
 
   ////////////////////////////////////////////////////////

lib/qcd/hmc/checkpointers/BaseCheckpointer.h

@@ -76,6 +76,14 @@ class BaseHmcCheckpointer : public HmcObservable<typename Impl::Field> {
     }
  	} 
 
+  void check_filename(const std::string &filename){
+    std::ifstream f(filename.c_str());
+    if(!f.good()){
+      std::cout << GridLogError << "Filename " << filename << " not found. Aborting. " << std::endl;
+      abort();
+    };
+  }
+
   virtual void initialize(const CheckpointerParameters &Params) = 0;
 
   virtual void CheckpointRestore(int traj, typename Impl::Field &U,

lib/qcd/hmc/checkpointers/BinaryCheckpointer.h

@@ -93,6 +93,9 @@ class BinaryHmcCheckpointer : public BaseHmcCheckpointer<Impl> {
   void CheckpointRestore(int traj, Field &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
     std::string config, rng;
     this->build_filenames(traj, Params, config, rng);
+    this->check_filename(rng);
+    this->check_filename(config);
+
 
     BinarySimpleMunger<sobj_double, sobj> munge;
 

lib/qcd/hmc/checkpointers/CheckPointerModules.h

@@ -136,6 +136,20 @@ class ILDGCPModule: public CheckPointerModule< ImplementationPolicy> {
 
 };
 
+template<class ImplementationPolicy, class Metadata>
+class ScidacCPModule: public CheckPointerModule< ImplementationPolicy> {
+  typedef CheckPointerModule< ImplementationPolicy> CPBase;
+  Metadata M;
+
+  // acquire resource
+  virtual void initialize(){
+     this->CheckPointPtr.reset(new ScidacHmcCheckpointer<ImplementationPolicy, Metadata>(this->Par_, M));
+  }
+public:
+  ScidacCPModule(typename CPBase::APar Par, Metadata M_):M(M_), CPBase(Par) {}
+  template <class ReaderClass>
+  ScidacCPModule(Reader<ReaderClass>& Reader) : Parametrized<typename CPBase::APar>(Reader), M(Reader){};
+};
 #endif
 
 

lib/qcd/hmc/checkpointers/CheckPointers.h

@@ -34,6 +34,7 @@ directory
 #include <Grid/qcd/hmc/checkpointers/NerscCheckpointer.h>
 #include <Grid/qcd/hmc/checkpointers/BinaryCheckpointer.h>
 #include <Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h>
+#include <Grid/qcd/hmc/checkpointers/ScidacCheckpointer.h>
 //#include <Grid/qcd/hmc/checkpointers/CheckPointerModules.h>
 
 

lib/qcd/hmc/checkpointers/ILDGCheckpointer.h

@@ -95,6 +95,10 @@ class ILDGHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
                          GridParallelRNG &pRNG) {
     std::string config, rng;
     this->build_filenames(traj, Params, config, rng);
+    this->check_filename(rng);
+    this->check_filename(config);
+
+    
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     BinaryIO::readRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);

lib/qcd/hmc/checkpointers/NerscCheckpointer.h

@@ -69,6 +69,9 @@ class NerscHmcCheckpointer : public BaseHmcCheckpointer<Gimpl> {
                          GridParallelRNG &pRNG) {
     std::string config, rng;
     this->build_filenames(traj, Params, config, rng);
+    this->check_filename(rng);
+    this->check_filename(config);
+
 
     FieldMetaData header;
     NerscIO::readRNGState(sRNG, pRNG, header, rng);

lib/qcd/hmc/checkpointers/ScidacCheckpointer.h

@@ -0,0 +1,125 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/ScidacCheckpointer.h
+
+Copyright (C) 2018
+
+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 */
+#ifndef SCIDAC_CHECKPOINTER
+#define SCIDAC_CHECKPOINTER
+
+#ifdef HAVE_LIME
+
+#include <iostream>
+#include <sstream>
+#include <string>
+
+namespace Grid {
+namespace QCD {
+
+// For generic fields
+template <class Implementation, class Metadata>
+class ScidacHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
+ private:
+  CheckpointerParameters Params;
+  Metadata MData;
+
+  typedef typename Implementation::Field Field;
+
+ public:
+  //INHERIT_GIMPL_TYPES(Implementation);
+
+  ScidacHmcCheckpointer(const CheckpointerParameters &Params_) { initialize(Params_); }
+  ScidacHmcCheckpointer(const CheckpointerParameters &Params_, const Metadata& M_):MData(M_) { initialize(Params_); }
+
+  void initialize(const CheckpointerParameters &Params_) {
+    Params = Params_;
+
+    // check here that the format is valid
+    int ieee32big = (Params.format == std::string("IEEE32BIG"));
+    int ieee32    = (Params.format == std::string("IEEE32"));
+    int ieee64big = (Params.format == std::string("IEEE64BIG"));
+    int ieee64    = (Params.format == std::string("IEEE64"));
+
+    if (!(ieee64big || ieee32 || ieee32big || ieee64)) {
+      std::cout << GridLogError << "Unrecognized file format " << Params.format
+                << std::endl;
+      std::cout << GridLogError
+                << "Allowed: IEEE32BIG | IEEE32 | IEEE64BIG | IEEE64"
+                << std::endl;
+
+      exit(1);
+    }
+  }
+
+  void TrajectoryComplete(int traj, Field &U, GridSerialRNG &sRNG,
+                          GridParallelRNG &pRNG) {
+    if ((traj % Params.saveInterval) == 0) {
+      std::string config, rng;
+      this->build_filenames(traj, Params, config, rng);
+      GridBase *grid = U._grid;
+      uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+      BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
+      ScidacWriter _ScidacWriter(grid->IsBoss());
+      _ScidacWriter.open(config);
+      _ScidacWriter.writeScidacFieldRecord(U, MData);
+      _ScidacWriter.close();
+
+      std::cout << GridLogMessage << "Written Scidac Configuration on " << config
+                << " checksum " << std::hex << nersc_csum<<"/"
+		            << scidac_csuma<<"/" << scidac_csumb
+		            << std::dec << std::endl;
+    }
+  };
+
+  void CheckpointRestore(int traj, Field &U, GridSerialRNG &sRNG,
+                         GridParallelRNG &pRNG) {
+    std::string config, rng;
+    this->build_filenames(traj, Params, config, rng);
+    this->check_filename(rng);
+    this->check_filename(config);
+
+
+    uint32_t nersc_csum,scidac_csuma,scidac_csumb;
+    BinaryIO::readRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
+
+    Metadata md_content;
+    ScidacReader _ScidacReader;
+    _ScidacReader.open(config);
+    _ScidacReader.readScidacFieldRecord(U,md_content);  // format from the header
+    _ScidacReader.close();
+
+    std::cout << GridLogMessage << "Read Scidac Configuration from " << config
+              << " checksum " << std::hex 
+	      << nersc_csum<<"/"
+	      << scidac_csuma<<"/"
+	      << scidac_csumb
+	      << std::dec << std::endl;
+  };
+};
+}
+}
+
+#endif  // HAVE_LIME
+#endif  // ILDG_CHECKPOINTER

lib/stencil/Stencil.h

@@ -66,6 +66,8 @@ void Gather_plane_simple_table (std::vector<std::pair<int,int> >& table,const La
   parallel_for(int i=0;i<num;i++){
     compress.Compress(&buffer[off],table[i].first,rhs._odata[so+table[i].second]);
   }
+// Further optimisatoin: i) streaming store the result
+//                       ii) software prefetch the first element of the next table entry
 }
 
 ///////////////////////////////////////////////////////////////////
@@ -148,6 +150,8 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
   std::vector<int>                  _distances;
   std::vector<int>                  _comm_buf_size;
   std::vector<int>                  _permute_type;
+  std::vector<int> same_node;
+  std::vector<int> surface_list;
 
   Vector<StencilEntry>  _entries;
   std::vector<Packet> Packets;
@@ -199,7 +203,7 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
 
     int dimension    = _directions[point];
     int displacement = _distances[point];
-    assert( (displacement==1) || (displacement==-1));
+
 
     int pd              = _grid->_processors[dimension];
     int fd              = _grid->_fdimensions[dimension];
@@ -214,9 +218,12 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
     if ( ! comm_dim ) return 1;
 
     int nbr_proc;
-    if (displacement==1) nbr_proc = 1;
+    if (displacement>0) nbr_proc = 1;
     else                nbr_proc = pd-1;
 
+    // FIXME  this logic needs to be sorted for three link term
+    //    assert( (displacement==1) || (displacement==-1));
+    // Present hack only works for >= 4^4 subvol per node
     _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
     void *shm = (void *) _grid->ShmBufferTranslate(recv_from_rank,u_recv_buf_p);
@@ -505,25 +512,24 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
   template<class decompressor>
   void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) { 
 
-    for(int i=0;i<mm.size();i++){	
     mergetime-=usecond();
+    for(int i=0;i<mm.size();i++){	
       parallel_for(int o=0;o<mm[i].buffer_size/2;o++){
 	decompress.Exchange(mm[i].mpointer,
 			    mm[i].vpointers[0],
 			    mm[i].vpointers[1],
 			    mm[i].type,o);
       }
-      mergetime+=usecond();
     }
+    mergetime+=usecond();
 
-    for(int i=0;i<dd.size();i++){	
     decompresstime-=usecond();
+    for(int i=0;i<dd.size();i++){	
       parallel_for(int o=0;o<dd[i].buffer_size;o++){
 	decompress.Decompress(dd[i].kernel_p,dd[i].mpi_p,o);
       }      
-      decompresstime+=usecond();
     }
-
+    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -538,6 +544,29 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
     }
   };
 
+  // Move interior/exterior split into the generic stencil
+  // FIXME Explicit Ls in interface is a pain. Should just use a vol
+  void BuildSurfaceList(int Ls,int vol4){
+
+    // find same node for SHM
+    // Here we know the distance is 1 for WilsonStencil
+    for(int point=0;point<this->_npoints;point++){
+      same_node[point] = this->SameNode(point);
+    }
+    
+    for(int site = 0 ;site< vol4;site++){
+      int local = 1;
+      for(int point=0;point<this->_npoints;point++){
+	if( (!this->GetNodeLocal(site*Ls,point)) && (!same_node[point]) ){ 
+	  local = 0;
+	}
+      }
+      if(local == 0) { 
+	surface_list.push_back(site);
+      }
+    }
+  }
+
  CartesianStencil(GridBase *grid,
 		  int npoints,
 		  int checkerboard,
@@ -548,7 +577,8 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
     comm_bytes_thr(npoints), 
     comm_enter_thr(npoints),
     comm_leave_thr(npoints), 
-       comm_time_thr(npoints)
+    comm_time_thr(npoints),
+    same_node(npoints)
   {
     face_table_computed=0;
     _npoints = npoints;
@@ -556,6 +586,7 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
     _directions = directions;
     _distances  = distances;
     _unified_buffer_size=0;
+    surface_list.resize(0);
 
     int osites  = _grid->oSites();
     

lib/threads/Threads.h

@@ -40,7 +40,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 #define PARALLEL_FOR_LOOP        _Pragma("omp parallel for schedule(static)")
 #define PARALLEL_FOR_LOOP_INTERN _Pragma("omp for schedule(static)")
-#define PARALLEL_NESTED_LOOP2 _Pragma("omp parallel for schedule(static) collapse(2)")
+#define PARALLEL_NESTED_LOOP2 _Pragma("omp parallel for collapse(2)")
 #define PARALLEL_REGION       _Pragma("omp parallel")
 #define PARALLEL_CRITICAL     _Pragma("omp critical")
 #else

lib/util/Init.cc

@@ -368,8 +368,10 @@ void Grid_init(int *argc,char ***argv)
   }
   if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-overlap") ){
     QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsAndCompute;
+    QCD::StaggeredKernelsStatic::Comms = QCD::StaggeredKernelsStatic::CommsAndCompute;
   } else {
     QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
+    QCD::StaggeredKernelsStatic::Comms = QCD::StaggeredKernelsStatic::CommsThenCompute;
   }
   if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-concurrent") ){
     CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicyConcurrent);
@@ -385,6 +387,7 @@ void Grid_init(int *argc,char ***argv)
   if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-threads") ){
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--comms-threads");
     GridCmdOptionInt(arg,CartesianCommunicator::nCommThreads);
+    assert(CartesianCommunicator::nCommThreads > 0);
   }
   if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");

tests/Test_compressed_lanczos_hot_start.cc

@@ -167,7 +167,7 @@ int main (int argc, char ** argv) {
   RealD mass = Params.mass;
   RealD M5   = Params.M5;
   std::vector<int> blockSize = Params.blockSize;
-  std::vector<int> latt({16,16,16,16});
+  std::vector<int> latt({32,32,32,32});
   uint64_t     vol = Ls*latt[0]*latt[1]*latt[2]*latt[3];
   double   mat_flop= 2.0*1320.0*vol;    
   // Grids

tests/core/Test_staggered5Dvec.cc

@@ -141,6 +141,7 @@ int main (int argc, char ** argv)
   t1=usecond();
  
   std::cout<<GridLogMessage << "Called Ds ASM"<<std::endl;
+  std::cout<<GridLogMessage << "norm src "<< norm2(src)<<std::endl;
   std::cout<<GridLogMessage << "norm result "<< norm2(tmp)<<std::endl;
   std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
 
@@ -160,7 +161,8 @@ int main (int argc, char ** argv)
   localConvert(sresult,tmp);
  
   std::cout<<GridLogMessage << "Called sDs unroll"<<std::endl;
-  std::cout<<GridLogMessage << "norm result "<< norm2(sresult)<<std::endl;
+  std::cout<<GridLogMessage << "norm ssrc "<< norm2(ssrc)<<std::endl;
+  std::cout<<GridLogMessage << "norm sresult "<< norm2(sresult)<<std::endl;
   std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
 
 
@@ -181,6 +183,7 @@ int main (int argc, char ** argv)
   localConvert(sresult,tmp);
  
   std::cout<<GridLogMessage << "Called sDs asm"<<std::endl;
+  std::cout<<GridLogMessage << "norm ssrc   "<< norm2(ssrc)<<std::endl;
   std::cout<<GridLogMessage << "norm result "<< norm2(sresult)<<std::endl;
   std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)*extra<<std::endl;
 

tests/core/Test_staggered5DvecF.cc

@@ -0,0 +1,196 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_wilson.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>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+
+  const int Ls=16;
+  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);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  int threads = GridThread::GetThreads();
+
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  std::vector<int> seeds({1,2,3,4});
+
+  GridParallelRNG          pRNG4(UGrid);
+  GridParallelRNG          pRNG5(FGrid);
+  pRNG4.SeedFixedIntegers(seeds);
+  pRNG5.SeedFixedIntegers(seeds);
+
+  typedef typename ImprovedStaggeredFermion5DF::FermionField FermionField; 
+  typedef typename ImprovedStaggeredFermion5DF::ComplexField ComplexField; 
+  typename ImprovedStaggeredFermion5DF::ImplParams params; 
+
+  FermionField src   (FGrid);
+  random(pRNG5,src);
+  /*
+  std::vector<int> site({0,1,2,0,0});
+  ColourVector cv = zero;
+  cv()()(0)=1.0;
+  src = zero;
+  pokeSite(cv,src,site);
+  */
+  FermionField result(FGrid); result=zero;
+  FermionField    tmp(FGrid);    tmp=zero;
+  FermionField    err(FGrid);    tmp=zero;
+  FermionField phi   (FGrid); random(pRNG5,phi);
+  FermionField chi   (FGrid); random(pRNG5,chi);
+
+  LatticeGaugeFieldF Umu(UGrid);
+  SU3::HotConfiguration(pRNG4,Umu);
+
+  /*
+  for(int mu=1;mu<4;mu++){
+    auto tmp = PeekIndex<LorentzIndex>(Umu,mu);
+        tmp = zero;
+    PokeIndex<LorentzIndex>(Umu,tmp,mu);
+  }
+  */
+  double volume=Ls;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
+  RealD mass=0.1;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+
+  ImprovedStaggeredFermion5DF     Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,c1,c2,u0,params);
+  ImprovedStaggeredFermionVec5dF sDs(Umu,Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,mass,c1,c2,u0,params);
+
+  std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
+  std::cout<<GridLogMessage<<"= Testing Dhop against cshift implementation         "<<std::endl;
+  std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
+
+  int ncall=1000;
+  int ncall1=1000;
+  double t0(0),t1(0);
+  double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 +  == 1146
+
+  std::cout<<GridLogMessage << "Calling staggered operator"<<std::endl;
+  t0=usecond();
+  for(int i=0;i<ncall1;i++){
+    Ds.Dhop(src,result,0);
+  }
+  t1=usecond();
+
+  
+  std::cout<<GridLogMessage << "Called Ds"<<std::endl;
+  std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+
+  std::cout<<GridLogMessage << "Calling vectorised staggered operator"<<std::endl;
+
+#ifdef AVX512
+  QCD::StaggeredKernelsStatic::Opt=QCD::StaggeredKernelsStatic::OptInlineAsm;
+#else
+  QCD::StaggeredKernelsStatic::Opt=QCD::StaggeredKernelsStatic::OptGeneric;
+#endif
+
+  t0=usecond();
+  for(int i=0;i<ncall1;i++){
+    Ds.Dhop(src,tmp,0);
+  }
+  t1=usecond();
+ 
+  std::cout<<GridLogMessage << "Called Ds ASM"<<std::endl;
+  std::cout<<GridLogMessage << "norm src "<< norm2(src)<<std::endl;
+  std::cout<<GridLogMessage << "norm result "<< norm2(tmp)<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+
+  err = tmp-result; 
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+  
+  FermionField ssrc  (sFGrid);  localConvert(src,ssrc);
+  FermionField sresult(sFGrid); sresult=zero;
+
+  QCD::StaggeredKernelsStatic::Opt=QCD::StaggeredKernelsStatic::OptHandUnroll;
+  t0=usecond();
+  for(int i=0;i<ncall1;i++){
+    sDs.Dhop(ssrc,sresult,0);
+  }
+  t1=usecond();
+  localConvert(sresult,tmp);
+ 
+  std::cout<<GridLogMessage << "Called sDs unroll"<<std::endl;
+  std::cout<<GridLogMessage << "norm ssrc "<< norm2(ssrc)<<std::endl;
+  std::cout<<GridLogMessage << "norm sresult "<< norm2(sresult)<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+
+
+#ifdef AVX512
+  QCD::StaggeredKernelsStatic::Opt=QCD::StaggeredKernelsStatic::OptInlineAsm;
+#else
+  QCD::StaggeredKernelsStatic::Opt=QCD::StaggeredKernelsStatic::OptGeneric;
+#endif
+
+  err = tmp-result; 
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+  int extra=1;
+  t0=usecond();
+  for(int i=0;i<ncall1*extra;i++){
+    sDs.Dhop(ssrc,sresult,0);
+  }
+  t1=usecond();
+  localConvert(sresult,tmp);
+ 
+  std::cout<<GridLogMessage << "Called sDs asm"<<std::endl;
+  std::cout<<GridLogMessage << "norm ssrc   "<< norm2(ssrc)<<std::endl;
+  std::cout<<GridLogMessage << "norm result "<< norm2(sresult)<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)*extra<<std::endl;
+
+  err = tmp-result; 
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+
+
+  Grid_finalize();
+}

tests/hmc/Test_hmc_ScalarActionNxN.cc

@@ -34,6 +34,8 @@ class ScalarActionParameters : Serializable {
     double, lambda,
     double, g);
 
+  ScalarActionParameters() = default;
+
     template <class ReaderClass >
   ScalarActionParameters(Reader<ReaderClass>& Reader){
     read(Reader, "ScalarAction", *this);
@@ -125,9 +127,12 @@ int main(int argc, char **argv) {
   TheHMC.Resources.AddGrid("scalar", ScalarGrid);
   std::cout << "Lattice size : " << GridDefaultLatt() << std::endl;
   
+  ScalarActionParameters SPar(Reader);
+
   // Checkpointer definition
   CheckpointerParameters CPparams(Reader);
-  TheHMC.Resources.LoadBinaryCheckpointer(CPparams);
+  //TheHMC.Resources.LoadBinaryCheckpointer(CPparams);
+  TheHMC.Resources.LoadScidacCheckpointer(CPparams, SPar);
 
   RNGModuleParameters RNGpar(Reader);
   TheHMC.Resources.SetRNGSeeds(RNGpar);
@@ -140,7 +145,6 @@ int main(int argc, char **argv) {
   // Collect actions, here use more encapsulation
 
   // Scalar action in adjoint representation
-  ScalarActionParameters SPar(Reader);
   ScalarAction Saction(SPar.mass_squared, SPar.lambda, SPar.g);
 
   // Collect actions

tests/hmc/Test_hmc_WG_Production.cc

@@ -33,6 +33,7 @@ namespace Grid{
     GRID_SERIALIZABLE_CLASS_MEMBERS(ActionParameters,
 				    double, beta)
 
+    ActionParameters() = default;
 
     template <class ReaderClass >
     ActionParameters(Reader<ReaderClass>& Reader){
@@ -68,11 +69,15 @@ int main(int argc, char **argv) {
   }
   Serialiser Reader(TheHMC.ParameterFile);
 
-
+  // Read parameters from input file
+  ActionParameters WilsonPar(Reader);
 
   // Checkpointer definition
   CheckpointerParameters CPparams(Reader);
-  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+  //TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  // Store metadata in the Scidac checkpointer
+  TheHMC.Resources.LoadScidacCheckpointer(CPparams, WilsonPar);
 
   RNGModuleParameters RNGpar(Reader);
   TheHMC.Resources.SetRNGSeeds(RNGpar);
@@ -91,8 +96,6 @@ int main(int argc, char **argv) {
   // need wrappers of the fermionic classes
   // that have a complex construction
   // standard
-  ActionParameters WilsonPar(Reader);
-  //RealD beta = 6.4 ;
   WilsonGaugeActionR Waction(WilsonPar.beta);
 
   ActionLevel<HMCWrapper::Field> Level1(1);

tests/solver/Test_staggered_block_cg_prec.cc

@@ -74,8 +74,16 @@ int main (int argc, char ** argv)
 
   LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(pRNG,Umu);
 
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
   RealD mass=0.003;
-  ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass); 
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,c1,c2,u0); 
   SchurStaggeredOperator<ImprovedStaggeredFermion5DR,FermionField> HermOp(Ds);
 
   ConjugateGradient<FermionField> CG(1.0e-8,10000);
@@ -87,14 +95,26 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "****************************************************************** "<<std::endl;
   std::cout << GridLogMessage << " Calling 4d CG "<<std::endl;
   std::cout << GridLogMessage << "****************************************************************** "<<std::endl;
-  ImprovedStaggeredFermionR Ds4d(Umu,Umu,*UGrid,*UrbGrid,mass);
+  ImprovedStaggeredFermionR Ds4d(Umu,Umu,*UGrid,*UrbGrid,mass,c1,c2,u0);
   SchurStaggeredOperator<ImprovedStaggeredFermionR,FermionField> HermOp4d(Ds4d);
   FermionField src4d(UGrid); random(pRNG,src4d);
   FermionField src4d_o(UrbGrid);   pickCheckerboard(Odd,src4d_o,src4d);
   FermionField result4d_o(UrbGrid); 
 
+  double deodoe_flops=(16*(3*(6+8+8)) + 15*3*2)*volume; // == 66*16 +  == 1146
   result4d_o=zero;
+  {
+    double t1=usecond();
     CG(HermOp4d,src4d_o,result4d_o);
+    double t2=usecond();
+    double ncall=CG.IterationsToComplete;
+    double flops = deodoe_flops * ncall;
+    std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+    std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+    HermOp4d.Report();
+  }
+  Ds4d.Report();
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 
 
@@ -103,7 +123,17 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
   Ds.ZeroCounters();
   result_o=zero;
+  {
+    double t1=usecond();
     CG(HermOp,src_o,result_o);
+    double t2=usecond();
+    double ncall=CG.IterationsToComplete*Ls;
+    double flops = deodoe_flops * ncall;
+    std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+    std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+    HermOp.Report();
+  }
   Ds.Report();
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 
@@ -112,7 +142,18 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
   Ds.ZeroCounters();
   result_o=zero;
+  {
+    double t1=usecond();
     mCG(HermOp,src_o,result_o);
+    double t2=usecond();
+    double ncall=mCG.IterationsToComplete*Ls;
+    double flops = deodoe_flops * ncall;
+    std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+    std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+    HermOp.Report();
+  }
+
   Ds.Report();
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 
@@ -121,7 +162,17 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
   Ds.ZeroCounters();
   result_o=zero;
+  {
+    double t1=usecond();
     BCGrQ(HermOp,src_o,result_o);
+    double t2=usecond();
+    double ncall=BCGrQ.IterationsToComplete*Ls;
+    double flops = deodoe_flops * ncall;
+    std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+    std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+    HermOp.Report();
+  }
   Ds.Report();
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 

tests/solver/Test_staggered_block_cg_unprec.cc

@@ -74,7 +74,16 @@ int main (int argc, char ** argv)
   LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(pRNG,Umu);
 
   RealD mass=0.003;
-  ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass); 
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,c1,c2,u0); 
   MdagMLinearOperator<ImprovedStaggeredFermion5DR,FermionField> HermOp(Ds);
 
   ConjugateGradient<FermionField> CG(1.0e-8,10000);
@@ -86,11 +95,23 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "****************************************************************** "<<std::endl;
   std::cout << GridLogMessage << " Calling 4d CG "<<std::endl;
   std::cout << GridLogMessage << "****************************************************************** "<<std::endl;
-  ImprovedStaggeredFermionR Ds4d(Umu,Umu,*UGrid,*UrbGrid,mass);
+  ImprovedStaggeredFermionR Ds4d(Umu,Umu,*UGrid,*UrbGrid,mass,c1,c2,u0);
   MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp4d(Ds4d);
   FermionField src4d(UGrid); random(pRNG,src4d);
   FermionField result4d(UGrid); result4d=zero;
+
+  double deodoe_flops=(16*(3*(6+8+8)) + 15*3*2)*volume; // == 66*16 +  == 1146
+  {
+    double t1=usecond();
     CG(HermOp4d,src4d,result4d);
+    double t2=usecond();
+    double ncall=CG.IterationsToComplete;
+    double flops = deodoe_flops * ncall;
+    std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+    std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+   }
+
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 
 
@@ -98,9 +119,18 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << " Calling 5d CG for "<<Ls <<" right hand sides" <<std::endl;
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
   result=zero;
+{
   Ds.ZeroCounters();
+  double t1=usecond();
   CG(HermOp,src,result);
+  double t2=usecond();
+  double ncall=CG.IterationsToComplete;
+  double flops = deodoe_flops * ncall;
+  std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+    std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
   Ds.Report();
+}
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
@@ -108,7 +138,16 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
   result=zero;
   Ds.ZeroCounters();
+{
+  double t1=usecond();
   mCG(HermOp,src,result);
+  double t2=usecond();
+  double ncall=CG.IterationsToComplete;
+  double flops = deodoe_flops * ncall;
+  std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+  std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+}
   Ds.Report();
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 
@@ -117,7 +156,16 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
   result=zero;
   Ds.ZeroCounters();
+{
+  double t1=usecond();
   BCGrQ(HermOp,src,result);
+  double t2=usecond();
+  double ncall=CG.IterationsToComplete;
+  double flops = deodoe_flops * ncall;
+  std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+  std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+}
   Ds.Report();
   std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
 

tests/solver/Test_staggered_cg_prec.cc

@@ -71,7 +71,10 @@ int main (int argc, char ** argv)
   }  
   
   RealD mass=0.003;
-  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass);
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
 
   FermionField res_o(&RBGrid); 
   FermionField src_o(&RBGrid); 
@@ -80,7 +83,19 @@ int main (int argc, char ** argv)
 
   SchurStaggeredOperator<ImprovedStaggeredFermionR,FermionField> HermOpEO(Ds);
   ConjugateGradient<FermionField> CG(1.0e-8,10000);
+  double t1=usecond();
   CG(HermOpEO,src_o,res_o);
+  double t2=usecond();
+
+  // Schur solver: uses DeoDoe => volume * 1146
+  double ncall=CG.IterationsToComplete;
+  double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 +  == 1146
+
+  std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+  std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+
+
 
   FermionField tmp(&RBGrid);
 

tests/solver/Test_staggered_cg_schur.cc

@@ -65,7 +65,10 @@ int main (int argc, char ** argv)
   FermionField  resid(&Grid); 
 
   RealD mass=0.1;
-  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass);
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
 
   ConjugateGradient<FermionField> CG(1.0e-8,10000);
   SchurRedBlackStaggeredSolve<FermionField> SchurSolver(CG);

tests/solver/Test_staggered_cg_unprec.cc

@@ -73,7 +73,10 @@ int main (int argc, char ** argv)
   }  
   
   RealD mass=0.1;
-  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass);
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
 
   MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
   ConjugateGradient<FermionField> CG(1.0e-6,10000);

tests/solver/Test_staggered_multishift.cc

@@ -0,0 +1,121 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_wilson_cg_unprec.cc
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+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 */
+#include <Grid/Grid.h>
+#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+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)
+{
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField; 
+  typename ImprovedStaggeredFermionR::ImplParams params; 
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
+  ////////////////////////////////////////
+  // sqrt 
+  ////////////////////////////////////////
+  double     lo=0.001;
+  double     hi=1.0;
+  int precision=64;
+  int    degree=10;
+  AlgRemez remez(lo,hi,precision);
+  remez.generateApprox(degree,1,2);
+  MultiShiftFunction Sqrt(remez,1.0e-6,false);
+  std::cout<<GridLogMessage << "Generating degree "<<degree<<" for x^(1/2)"<<std::endl;
+
+
+  ////////////////////////////////////////////
+  // Setup staggered
+  ////////////////////////////////////////////
+  RealD mass=0.003;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
+  SchurStaggeredOperator<ImprovedStaggeredFermionR,FermionField> HermOpEO(Ds);
+
+  FermionField src(&Grid); random(pRNG,src);
+  FermionField src_o(&RBGrid); 
+  pickCheckerboard(Odd,src_o,src);
+
+
+  /////////////////////////////////
+  //Multishift CG
+  /////////////////////////////////
+  std::vector<FermionField> result(degree,&RBGrid);
+  ConjugateGradientMultiShift<FermionField> MSCG(10000,Sqrt);
+
+  double deodoe_flops=(1205+15*degree)*volume; // == 66*16 +  == 1146
+
+  double t1=usecond();
+  MSCG(HermOpEO,src_o,result);
+  double t2=usecond();
+  double ncall=MSCG.IterationsToComplete;
+  double flops = deodoe_flops * ncall;
+  std::cout<<GridLogMessage << "usec    =   "<< (t2-t1)<<std::endl;
+  std::cout<<GridLogMessage << "flops   =   "<< flops<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t2-t1)<<std::endl;
+  //  HermOpEO.Report();
+
+  Grid_finalize();
+}