Scripts

Feature/wilson clover 5d

Grid/algorithms/FFT.h

@@ -191,7 +191,7 @@ public:
       
     Lattice<sobj> pgbuf(&pencil_g);
     autoView(pgbuf_v , pgbuf, CpuWrite);
-    std::cout << "CPU view" << std::endl;
+    //std::cout << "CPU view" << std::endl;
     
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -215,7 +215,7 @@ public:
     else if ( sign == forward ) div = 1.0;
     else assert(0);
       
-    std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
+    //std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
     FFTW_plan p;
     {
       FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
@@ -229,7 +229,7 @@ public:
     }
       
     // Barrel shift and collect global pencil
-    std::cout << GridLogPerformance<<"Making pencil" << std::endl;
+    //std::cout << GridLogPerformance<<"Making pencil" << std::endl;
     Coordinate lcoor(Nd), gcoor(Nd);
     result = source;
     int pc = processor_coor[dim];
@@ -251,7 +251,7 @@ public:
       }
     }
       
-    std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
+    //std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
     // Loop over orthog coords
     int NN=pencil_g.lSites();
     GridStopWatch timer;
@@ -274,7 +274,7 @@ public:
     usec += timer.useconds();
     flops+= flops_call*NN;
       
-    std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
+    //std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
     // writing out result
     {
       autoView(pgbuf_v,pgbuf,CpuRead);
@@ -291,7 +291,7 @@ public:
     }
     result = result*div;
       
-    std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
+    //std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
     // destroying plan
     FFTW<scalar>::fftw_destroy_plan(p);
 #endif

Grid/algorithms/LinearOperator.h

@@ -277,6 +277,38 @@ public:
     assert(0);
   }
 };
+template<class Matrix,class Field>
+class ShiftedNonHermitianLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD shift;
+public:
+  ShiftedNonHermitianLinearOperator(Matrix &Mat,RealD shft): _Mat(Mat),shift(shft){};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    _Mat.Mdiag(in,out);
+    out = out + shift*in;
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    _Mat.M(in,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    _Mat.Mdag(in,out);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    assert(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    assert(0);
+  }
+};
 
 //////////////////////////////////////////////////////////
 // Even Odd Schur decomp operators; there are several

Grid/algorithms/approx/Chebyshev.h

@@ -269,7 +269,9 @@ public:
     RealD xscale = 2.0/(hi-lo);
     RealD mscale = -(hi+lo)/(hi-lo);
     Linop.HermOp(T0,y);
+    grid->Barrier();
     axpby(T1,xscale,mscale,y,in);
+    grid->Barrier();
 
     // sum = .5 c[0] T0 + c[1] T1
     //    out = ()*T0 + Coeffs[1]*T1;

Grid/algorithms/blas/BatchedBlas.h

@@ -208,8 +208,8 @@ public:
     assert(Bkn.size()==batchCount);
     assert(Cmn.size()==batchCount);
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    //assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    //assert(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -367,28 +367,67 @@ public:
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    eCmn = alpha * eAmk * eBkn ;
         });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.adjoint() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
 	  });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  } );
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  } );
       } else { 
 	assert(0);
@@ -414,8 +453,8 @@ public:
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    //assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    //assert(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -514,28 +553,70 @@ public:
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    eCmn = alpha * eAmk * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.adjoint() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.transpose() ;
 	  });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  } );
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  } );
       } else { 
 	assert(0);
@@ -661,29 +742,41 @@ public:
 	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    eCmn = alpha * eAmk * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.transpose() ;	  
 	  });
       } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
-	  } );
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
+	  });
       } else { 
 	assert(0);
       }
@@ -809,28 +902,40 @@ public:
 	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    eCmn = alpha * eAmk * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.transpose() ;
 	  });
       } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
-	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  });
       } else { 
 	assert(0);

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -245,9 +245,10 @@ until convergence
 	_HermOp(src_n,tmp);
 	//	std::cout << GridLogMessage<< tmp<<std::endl; exit(0);
 	//	std::cout << GridLogIRL << " _HermOp " << norm2(tmp) << std::endl;
-	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+//	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vnum = real(innerProduct(tmp,tmp)); // HermOp^2.
 	RealD vden = norm2(src_n);
-	RealD na = vnum/vden;
+	RealD na = std::sqrt(vnum/vden);
 	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
 	  i=_MAX_ITER_IRL_MEVAPP_;
 	evalMaxApprox = na;
@@ -255,6 +256,7 @@ until convergence
 	src_n = tmp;
       }
     }
+    std::cout << GridLogIRL << " Final evalMaxApprox  " << evalMaxApprox << std::endl;
 	
     std::vector<RealD> lme(Nm);  
     std::vector<RealD> lme2(Nm);

Grid/algorithms/multigrid/Aggregates.h

@@ -97,7 +97,7 @@ public:
 
     RealD scale;
 
-    ConjugateGradient<FineField> CG(1.0e-2,100,false);
+    ConjugateGradient<FineField> CG(1.0e-3,400,false);
     FineField noise(FineGrid);
     FineField Mn(FineGrid);
 
@@ -110,7 +110,7 @@ public:
       
       hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
 
-      for(int i=0;i<1;i++){
+      for(int i=0;i<4;i++){
 
 	CG(hermop,noise,subspace[b]);
 
@@ -146,7 +146,7 @@ public:
       
       DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
 
-      for(int i=0;i<3;i++){
+      for(int i=0;i<2;i++){
 	//  void operator() (const Field &src, Field &psi){
 #if 1
 	std::cout << GridLogMessage << " inverting on noise "<<std::endl;

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -441,8 +441,20 @@ public:
     std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
   }
 #else
+  //////////////////////////////////////////////////////////////////////
+  // Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
   void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
 		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    CoarsenOperator(linop,Subspace,Subspace);
+  }
+  //////////////////////////////////////////////////////////////////////
+  // Petrov - Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & U,
+		       Aggregation<Fobj,CComplex,nbasis> & V)
   {
     std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
     GridBase *grid = FineGrid();
@@ -458,11 +470,9 @@ public:
     // Orthogonalise the subblocks over the basis
     /////////////////////////////////////////////////////////////
     CoarseScalar InnerProd(CoarseGrid()); 
-    blockOrthogonalise(InnerProd,Subspace.subspace);
+    blockOrthogonalise(InnerProd,V.subspace);
+    blockOrthogonalise(InnerProd,U.subspace);
 
-    //    for(int s=0;s<Subspace.subspace.size();s++){
-      //      std::cout << " subspace norm "<<norm2(Subspace.subspace[s])<<std::endl;
-    //    }
     const int npoint = geom.npoint;
       
     Coordinate clatt = CoarseGrid()->GlobalDimensions();
@@ -542,7 +552,7 @@ public:
       std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
       for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
 	tphaseBZ-=usecond();
-	phaV = phaF[p]*Subspace.subspace[i];
+	phaV = phaF[p]*V.subspace[i];
 	tphaseBZ+=usecond();
 
 	/////////////////////////////////////////////////////////////////////
@@ -555,7 +565,7 @@ public:
 	//	std::cout << i << " " <<p << " MphaV "<<norm2(MphaV)<<" "<<norm2(phaV)<<std::endl;
 
 	tproj-=usecond();
-	blockProject(coarseInner,MphaV,Subspace.subspace);
+	blockProject(coarseInner,MphaV,U.subspace);
 	coarseInner = conjugate(pha[p]) * coarseInner;
 
 	ComputeProj[p] = coarseInner;

Grid/allocator/AlignedAllocator.h

@@ -69,7 +69,7 @@ public:
   }
 
   // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
-  void construct(pointer __p, const _Tp& __val) { assert(0);};
+  void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
@@ -175,10 +175,11 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
 template<class T> using hostVector          = std::vector<T,alignedAllocator<T> >;           // Needs autoview
-template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // 
+template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // Really want to deprecate
 template<class T> using uvmVector           = std::vector<T,uvmAllocator<T> >;               // auto migrating page
 template<class T> using deviceVector        = std::vector<T,devAllocator<T> >;               // device vector
 
+/*
 template<class T> class vecView
 {
  protected:
@@ -214,6 +215,7 @@ template<class T> vecView<T> VectorView(Vector<T> &vec,ViewMode _mode)
 #define autoVecView(v_v,v,mode)					\
   auto v_v = VectorView(v,mode);				\
   ViewCloser<decltype(v_v)> _autoView##v_v(v_v);
+*/
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryManagerCache.cc

@@ -9,6 +9,7 @@ static char print_buffer [ MAXLINE ];
 #define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer << std::endl;
 #define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug  << print_buffer << std::endl;
 //#define dprintf(...) 
+//#define mprintf(...) 
 
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
@@ -109,7 +110,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Discard(%lx) %lx",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -119,7 +120,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
     AccCache.AccPtr=(uint64_t) NULL;
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -139,7 +140,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld\n",
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld",
 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
   if (AccCache.accLock!=0) return;
@@ -153,7 +154,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)NULL;
     AccCache.state=CpuDirty; // CPU primary now
     DeviceBytes   -=AccCache.bytes;
-    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld ",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   //  uint64_t CpuPtr = AccCache.CpuPtr;
   DeviceEvictions++;
@@ -167,7 +168,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -182,7 +183,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx\n",
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx",
 	  (uint64_t)AccCache.bytes,
 	  (uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
@@ -210,7 +211,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
+    dprintf("AcceleratorViewClose %lx",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -222,7 +223,7 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
+    dprintf("AcceleratorViewOpen %lx",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -233,6 +234,9 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  if(bytes>=DeviceMaxBytes) {
+    printf("EvictVictims bytes %ld DeviceMaxBytes %ld\n",bytes,DeviceMaxBytes);
+  }
   assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
@@ -265,7 +269,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld\n",
+    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
@@ -305,7 +309,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
-    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
+    dprintf("Copied Empty entry into device accLock= %d",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -318,21 +322,21 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d",AccCache.accLock);
   } else {
     assert(0);
   }
@@ -341,7 +345,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
-    dprintf("AccCache entry removed from LRU \n");
+    dprintf("AccCache entry removed from LRU ");
     LRUremove(AccCache);
   }
 
@@ -364,10 +368,10 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   AccCache.accLock--;
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
   } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)

Grid/communicator/Communicator_base.h

@@ -33,6 +33,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
+#define NVLINK_GET
+
 NAMESPACE_BEGIN(Grid);
 
 extern bool Stencil_force_mpi ;
@@ -147,7 +149,8 @@ public:
 			    sizeof(obj),d*100+p);
 
       }
-      CommsComplete(list);
+      if (!list.empty()) // avoid triggering assert in comms == none
+	CommsComplete(list);
       for(int p=1;p<_processors[d];p++){
 	accum = accum + column[p];
       }
@@ -192,6 +195,11 @@ public:
 				      void *recv,
 				      int recv_from_rank,int do_recv,
 				      int xbytes,int rbytes,int dir);
+
+  // Could do a PollHtoD and have a CommsMerge dependence
+  void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
+  void StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list);
+
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
 				    int xmit_to_rank,int do_xmit,

Grid/communicator/Communicator_mpi3.cc

@@ -30,6 +30,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
 ////////////////////////////////////////////
@@ -259,32 +260,39 @@ CartesianCommunicator::~CartesianCommunicator()
 }
 #ifdef USE_GRID_REDUCTION
 void CartesianCommunicator::GlobalSum(float &f){
+  FlightRecorder::StepLog("GlobalSumP2P");
   CartesianCommunicator::GlobalSumP2P(f);
 }
 void CartesianCommunicator::GlobalSum(double &d)
 {
+  FlightRecorder::StepLog("GlobalSumP2P");
   CartesianCommunicator::GlobalSumP2P(d);
 }
 #else
 void CartesianCommunicator::GlobalSum(float &f){
+  FlightRecorder::StepLog("AllReduce");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSum(double &d)
 {
+  FlightRecorder::StepLog("AllReduce");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
   assert(ierr==0);
 }
 #endif
 void CartesianCommunicator::GlobalSum(uint32_t &u){
+  FlightRecorder::StepLog("AllReduce");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSum(uint64_t &u){
+  FlightRecorder::StepLog("AllReduce");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
+  FlightRecorder::StepLog("AllReduceVector");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
@@ -362,8 +370,6 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int bytes)
 {
   std::vector<MpiCommsRequest_t> reqs(0);
-  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
 
   int myrank = _processor;
   int ierr;
@@ -379,9 +385,6 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 		    communicator,MPI_STATUS_IGNORE);
   assert(ierr==0);
 
-  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
@@ -399,6 +402,8 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 
 
 #ifdef ACCELERATOR_AWARE_MPI
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list) {};
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list) {};
 double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
 							   void *xmit,
 							   int dest,int dox,
@@ -440,8 +445,15 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(rrq);
       off_node_bytes+=rbytes;
     }
+#ifdef NVLINK_GET
+    else { 
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+    }
+#endif
   }
-  
+  // This is a NVLINK PUT  
   if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
@@ -450,9 +462,11 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(xrq);
       off_node_bytes+=xbytes;
     } else {
+#ifndef NVLINK_GET
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
       assert(shm!=NULL);
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+#endif
     }
   }
   return off_node_bytes;
@@ -461,7 +475,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
   int nreq=list.size();
-
+  /*finishes Get/Put*/
   acceleratorCopySynchronise();
 
   if (nreq==0) return;
@@ -561,53 +575,105 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
   
   if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-#undef DEVICE_TO_HOST_CONCURRENT // pipeline
-#ifdef DEVICE_TO_HOST_CONCURRENT
+
       tag= dir+_processor*32;
 
       host_xmit = this->HostBufferMalloc(xbytes);
-      acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes); // Make this Asynch
+      CommsRequest_t srq;
+
+      srq.ev = acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes); // Make this Asynch
       
       //      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
       //      assert(ierr==0);
       //      off_node_bytes+=xbytes;
 
-      CommsRequest_t srq;
       srq.PacketType = InterNodeXmit;
       srq.bytes      = xbytes;
       //      srq.req        = xrq;
       srq.host_buf   = host_xmit;
       srq.device_buf = xmit;
+      srq.tag        = tag;
+      srq.dest       = dest;
+      srq.commdir    = commdir;
       list.push_back(srq);
-#else
-      tag= dir+_processor*32;
-
-      host_xmit = this->HostBufferMalloc(xbytes);
-      const int chunks=1;
-      for(int n=0;n<chunks;n++){
-	void * host_xmitc = (void *)( (uint64_t) host_xmit + n*xbytes/chunks);
-	void * xmitc      = (void *)( (uint64_t) xmit      + n*xbytes/chunks);
-	acceleratorCopyFromDeviceAsynch(xmitc, host_xmitc,xbytes/chunks); // Make this Asynch
-      }
-      acceleratorCopySynchronise(); // Complete all pending copy transfers
-      
-      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
-      off_node_bytes+=xbytes;
-
-      CommsRequest_t srq;
-      srq.PacketType = InterNodeXmit;
-      srq.bytes      = xbytes;
-      srq.req        = xrq;
-      srq.host_buf   = host_xmit;
-      srq.device_buf = xmit;
-      list.push_back(srq);
-#endif
     }
   }
 
   return off_node_bytes;
 }
+/*
+ * In the interest of better pipelining, poll for completion on each DtoH and 
+ * start MPI_ISend in the meantime
+ */
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list)
+{
+  int pending = 0;
+  do {
+
+    pending = 0;
+
+    for(int idx = 0; idx<list.size();idx++){
+
+      if ( list[idx].PacketType==InterNodeRecv ) {
+
+	int flag = 0;
+	MPI_Status status;
+	int ierr = MPI_Test(&list[idx].req,&flag,&status);
+	assert(ierr==0);
+
+	if ( flag ) {
+	  //	  std::cout << " PollIrecv "<<idx<<" flag "<<flag<<std::endl;
+	  acceleratorCopyToDeviceAsynch(list[idx].host_buf,list[idx].device_buf,list[idx].bytes);
+	  list[idx].PacketType=InterNodeReceiveHtoD;
+	} else {
+	  pending ++;
+	}
+      }
+    }
+    //    std::cout << " PollIrecv "<<pending<<" pending requests"<<std::endl;
+  } while ( pending );
+  
+}
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list)
+{
+  int pending = 0;
+  do {
+
+    pending = 0;
+
+    for(int idx = 0; idx<list.size();idx++){
+
+      if ( list[idx].PacketType==InterNodeXmit ) {
+
+	if ( acceleratorEventIsComplete(list[idx].ev) ) {
+
+	  void *host_xmit = list[idx].host_buf;
+	  uint32_t xbytes = list[idx].bytes;
+	  int dest        = list[idx].dest;
+	  int tag         = list[idx].tag;
+	  int commdir     = list[idx].commdir;
+	  ///////////////////
+	  // Send packet
+	  ///////////////////
+
+	  //	  std::cout << " DtoH is complete for index "<<idx<<" calling MPI_Isend "<<std::endl;
+	  
+	  MPI_Request xrq;
+	  int ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+	  assert(ierr==0);
+
+	  list[idx].req        = xrq; // Update the MPI request in the list
+
+	  list[idx].PacketType=InterNodeXmitISend;
+
+	} else {
+	  // not done, so return to polling loop
+	  pending++;
+	}
+      }
+    }
+  } while (pending);
+}  
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
@@ -644,69 +710,89 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
    * - complete all copies
    * - post MPI send asynch
    */
+#ifdef NVLINK_GET
+  if ( dor ) {
 
-  //  static int printed;
-  //  if((printed<8) && this->IsBoss() ) {
-  //    printf("dir %d doX %d doR %d Face size %ld %ld\n",dir,dox,dor,xbytes,rbytes);
-  //    printed++;
-  //  }
-  
+    if ( ! ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      assert(shm!=NULL);
+
+      CommsRequest_t srq;
+
+      srq.ev = acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+
+      srq.PacketType = IntraNodeRecv;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = NULL;
+      srq.device_buf = xmit;
+      srq.tag        = -1;
+      srq.dest       = dest;
+      srq.commdir    = dir;
+      list.push_back(srq);
+    }
+  }  
+#else
   if (dox) {
 
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-#ifdef DEVICE_TO_HOST_CONCURRENT
-      tag= dir+_processor*32;
-      // Find the send in the prepared list
-      int list_idx=-1;
-      for(int idx = 0; idx<list.size();idx++){
-
-	if ( (list[idx].device_buf==xmit)
-	   &&(list[idx].PacketType==InterNodeXmit)
-	   &&(list[idx].bytes==xbytes) ) {
-
-	  list_idx = idx;
-	  host_xmit = list[idx].host_buf;
-	}
-      }
-      assert(list_idx != -1); // found it
-      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
-      list[list_idx].req        = xrq; // Update the MPI request in the list
-      off_node_bytes+=xbytes;
-#endif      
-    } else {
+    if ( !( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
       assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+
+      CommsRequest_t srq;
+      
+      srq.ev = acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+
+      srq.PacketType = IntraNodeXmit;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = NULL;
+      srq.device_buf = xmit;
+      srq.tag        = -1;
+      srq.dest       = dest;
+      srq.commdir    = dir;
+      list.push_back(srq);
+      
     }
   }
+#endif
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
-  int nreq=list.size();
+  acceleratorCopySynchronise(); // Complete all pending copy transfers D2D
 
-  if (nreq==0) return;
-  std::vector<MPI_Status> status(nreq);
-  std::vector<MPI_Request> MpiRequests(nreq);
+  std::vector<MPI_Status> status;
+  std::vector<MPI_Request> MpiRequests;
+    
+  for(int r=0;r<list.size();r++){
+    // Must check each Send buf is clear to reuse
+    if ( list[r].PacketType == InterNodeXmitISend ) MpiRequests.push_back(list[r].req);
+    //    if ( list[r].PacketType == InterNodeRecv ) MpiRequests.push_back(list[r].req); // Already "Test" passed
+  }
 
-  for(int r=0;r<nreq;r++){
-    MpiRequests[r] = list[r].req;
+  int nreq=MpiRequests.size();
+
+  if (nreq>0) {
+    status.resize(MpiRequests.size());
+    int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.
+    assert(ierr==0);
   }
   
-  int ierr = MPI_Waitall(nreq,&MpiRequests[0],&status[0]);
-  assert(ierr==0);
-
-  for(int r=0;r<nreq;r++){
-    if ( list[r].PacketType==InterNodeRecv ) {
-      acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
-    }
-  }
+  //  for(int r=0;r<nreq;r++){
+  //    if ( list[r].PacketType==InterNodeRecv ) {
+  //      acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
+  //    }
+  //  }
+  
   
-  acceleratorCopySynchronise(); // Complete all pending copy transfers
   list.resize(0);               // Delete the list
   this->HostBufferFreeAll();    // Clean up the buffer allocs
-  this->StencilBarrier(); 
+#ifndef NVLINK_GET
+  this->StencilBarrier(); // if PUT must check our nbrs have filled our receive buffers.
+#endif   
 }
 #endif
 ////////////////////////////////////////////
@@ -715,6 +801,7 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
 
 void CartesianCommunicator::StencilBarrier(void)
 {
+  FlightRecorder::StepLog("NodeBarrier");
   MPI_Barrier  (ShmComm);
 }
 //void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
@@ -722,11 +809,13 @@ void CartesianCommunicator::StencilBarrier(void)
 //}
 void CartesianCommunicator::Barrier(void)
 {
+  FlightRecorder::StepLog("GridBarrier");
   int ierr = MPI_Barrier(communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 {
+  FlightRecorder::StepLog("Broadcast");
   int ierr=MPI_Bcast(data,
 		     bytes,
 		     MPI_BYTE,
@@ -745,6 +834,7 @@ void CartesianCommunicator::BarrierWorld(void){
 }
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
+  FlightRecorder::StepLog("BroadcastWorld");
   int ierr= MPI_Bcast(data,
 		      bytes,
 		      MPI_BYTE,
@@ -767,6 +857,7 @@ void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,
 }
 void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
 {
+  FlightRecorder::StepLog("AllToAll");
   // MPI is a pain and uses "int" arguments
   // 64*64*64*128*16 == 500Million elements of data.
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.

Grid/communicator/Communicator_none.cc

@@ -91,7 +91,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(list.size()==0);}
 void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
 						int dest,
@@ -132,6 +132,8 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 {
   return 2.0*bytes;
 }
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list) {};
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list) {};
 double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
 							   void *xmit,
 							   int xmit_to_rank,int dox,
@@ -139,7 +141,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
 							   int recv_from_rank,int dor,
 							   int xbytes,int rbytes, int dir)
 {
-  return xbytes+rbytes;
+  return 0.0;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,

Grid/communicator/SharedMemory.h

@@ -50,12 +50,30 @@ typedef MPI_Request MpiCommsRequest_t;
 #ifdef ACCELERATOR_AWARE_MPI
 typedef MPI_Request CommsRequest_t;
 #else
-enum PacketType_t { InterNodeXmit, InterNodeRecv, IntraNodeXmit, IntraNodeRecv };
+/*
+ * Enable state transitions as each packet flows.
+ */
+enum PacketType_t {
+  FaceGather,
+  InterNodeXmit,
+  InterNodeRecv,
+  IntraNodeXmit,
+  IntraNodeRecv,
+  InterNodeXmitISend,
+  InterNodeReceiveHtoD
+};
+/*
+ *Package arguments needed for various actions along packet flow
+ */
 typedef struct {
   PacketType_t PacketType;
   void *host_buf;
   void *device_buf;
+  int dest;
+  int tag;
+  int commdir;
   unsigned long bytes;
+  acceleratorEvent_t ev;
   MpiCommsRequest_t req;
 } CommsRequest_t;
 #endif
@@ -119,7 +137,7 @@ public:
   ///////////////////////////////////////////////////
   static void SharedMemoryAllocate(uint64_t bytes, int flags);
   static void SharedMemoryFree(void);
-  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
+  //  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
   static void SharedMemoryZero(void *dest,size_t bytes);
 
 };

Grid/communicator/SharedMemoryMPI.cc

@@ -542,12 +542,12 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
 #ifndef ACCELERATOR_AWARE_MPI
-  printf("Host buffer allocate for GPU non-aware MPI\n");
+  // printf("Host buffer allocate for GPU non-aware MPI\n");
 #if 0
   HostCommBuf= acceleratorAllocHost(bytes);
 #else 
   HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
-#ifdef HAVE_NUMAIF_H
+#if 0
   #warning "Moving host buffers to specific NUMA domain"
   int numa;
   char *numa_name=(char *)getenv("MPI_BUF_NUMA");
@@ -916,14 +916,14 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
-#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  acceleratorCopyToDevice(src,dest,bytes);
-#else   
-  bcopy(src,dest,bytes);
-#endif
-}
+//void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+//{
+//#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+//  acceleratorCopyToDevice(src,dest,bytes);
+//#else   
+//  bcopy(src,dest,bytes);
+//#endif
+//}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
@@ -959,6 +959,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
     MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
 
     ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
+    //    std::cerr << " SetCommunicator rank "<<r<<" comm "<<ShmCommBufs[r] <<std::endl;
   }
   ShmBufferFreeAll();
 
@@ -989,7 +990,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
 #endif
 
-  //SharedMemoryTest();
+  //  SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier
@@ -1011,19 +1012,18 @@ void SharedMemory::SharedMemoryTest(void)
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
        check[2]=magic;
-       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
+       acceleratorCopyToDevice(check,ShmCommBufs[r],3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
   for(uint64_t r=0;r<ShmSize;r++){
-    ShmBarrier();
-    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
-    ShmBarrier();
+    acceleratorCopyFromDevice(ShmCommBufs[r],check,3*sizeof(uint64_t));
     assert(check[0]==GlobalSharedMemory::WorldNode);
     assert(check[1]==r);
     assert(check[2]==magic);
-    ShmBarrier();
   }
+  ShmBarrier();
+  std::cout << GridLogDebug << " SharedMemoryTest has passed "<<std::endl;
 }
 
 void *SharedMemory::ShmBuffer(int rank)

Grid/communicator/SharedMemoryNone.cc

@@ -122,10 +122,10 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
   acceleratorMemSet(dest,0,bytes);
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
-  acceleratorCopyToDevice(src,dest,bytes);
-}
+//void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+//{
+//  acceleratorCopyToDevice(src,dest,bytes);
+//}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality

Grid/cshift/Cshift_mpi.h

@@ -68,7 +68,7 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   if(Cshift_verbose) std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
-#if 1
+
 template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
 {
   int sshift[2];
@@ -125,7 +125,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
   static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
   static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
-    
+#ifndef ACCELERATOR_AWARE_MPI
+  static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size);
+  static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size);
+#endif
+  
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
   RealD tcopy=0.0;
@@ -156,16 +160,29 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
+
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
       
       tcomms-=usecond();
       grid->Barrier();
 
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+#else
+      // bouncy bouncy
+      acceleratorCopyFromDevice(&send_buf[0],&hsend_buf[0],bytes);
+      grid->SendToRecvFrom((void *)&hsend_buf[0],
+			   xmit_to_rank,
+			   (void *)&hrecv_buf[0],
+			   recv_from_rank,
+			   bytes);
+      acceleratorCopyToDevice(&hrecv_buf[0],&recv_buf[0],bytes);
+#endif
+
       xbytes+=bytes;
       grid->Barrier();
       tcomms+=usecond();
@@ -226,12 +243,16 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   static std::vector<deviceVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
- 
+
   for(int s=0;s<Nsimd;s++){
     send_buf_extract[s].resize(buffer_size);
     recv_buf_extract[s].resize(buffer_size);
   }
-
+#ifndef ACCELERATOR_AWARE_MPI
+  hostVector<scalar_object> hsend_buf; hsend_buf.resize(buffer_size);
+  hostVector<scalar_object> hrecv_buf; hrecv_buf.resize(buffer_size);
+#endif
+  
   int bytes = buffer_size*sizeof(scalar_object);
 
   ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
@@ -283,11 +304,22 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
+#ifdef ACCELERATOR_AWARE_MPI
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
 			     xmit_to_rank,
 			     (void *)recv_buf_extract_mpi,
 			     recv_from_rank,
 			     bytes);
+#else
+      // bouncy bouncy
+	acceleratorCopyFromDevice((void *)send_buf_extract_mpi,(void *)&hsend_buf[0],bytes);
+	grid->SendToRecvFrom((void *)&hsend_buf[0],
+			     xmit_to_rank,
+			     (void *)&hrecv_buf[0],
+			     recv_from_rank,
+			     bytes);
+	acceleratorCopyToDevice((void *)&hrecv_buf[0],(void *)recv_buf_extract_mpi,bytes);
+#endif
 
 	xbytes+=bytes;
 	grid->Barrier();
@@ -311,234 +343,6 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
   }
 }
-#else 
-template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  GridBase *grid=rhs.Grid();
-  Lattice<vobj> temp(rhs.Grid());
-
-  int fd              = rhs.Grid()->_fdimensions[dimension];
-  int rd              = rhs.Grid()->_rdimensions[dimension];
-  int pd              = rhs.Grid()->_processors[dimension];
-  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  assert(simd_layout==1);
-  assert(comm_dim==1);
-  assert(shift>=0);
-  assert(shift<fd);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-  
-  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf_v; recv_buf_v.resize(buffer_size);
-  vobj *send_buf;
-  vobj *recv_buf;
-  {
-    grid->ShmBufferFreeAll();
-    size_t bytes = buffer_size*sizeof(vobj);
-    send_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-    recv_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-  }
-    
-  int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  for(int x=0;x<rd;x++){       
-
-    int sx        =  (x+sshift)%rd;
-    int comm_proc = ((x+sshift)/rd)%pd;
-    
-    if (comm_proc==0) {
-
-      tcopy-=usecond();
-      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-      tcopy+=usecond();
-
-    } else {
-
-      int words = buffer_size;
-      if (cbmask != 0x3) words=words>>1;
-
-      int bytes = words * sizeof(vobj);
-
-      tgather-=usecond();
-      Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
-      tgather+=usecond();
-
-      //      int rank           = grid->_processor;
-      int recv_from_rank;
-      int xmit_to_rank;
-      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-
-
-      tcomms-=usecond();
-      //      grid->Barrier();
-
-      acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
-      grid->SendToRecvFrom((void *)&send_buf[0],
-			   xmit_to_rank,
-			   (void *)&recv_buf[0],
-			   recv_from_rank,
-			   bytes);
-      xbytes+=bytes;
-      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
-
-      //      grid->Barrier();
-      tcomms+=usecond();
-
-      tscatter-=usecond();
-      Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
-      tscatter+=usecond();
-    }
-  }
-  if(Cshift_verbose){
-    std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  }
-}
-
-template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  GridBase *grid=rhs.Grid();
-  const int Nsimd = grid->Nsimd();
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-   
-  int fd = grid->_fdimensions[dimension];
-  int rd = grid->_rdimensions[dimension];
-  int ld = grid->_ldimensions[dimension];
-  int pd = grid->_processors[dimension];
-  int simd_layout     = grid->_simd_layout[dimension];
-  int comm_dim        = grid->_processors[dimension] >1 ;
-
-  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
-  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
-  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
-
-  assert(comm_dim==1);
-  assert(simd_layout==2);
-  assert(shift>=0);
-  assert(shift<fd);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-
-  int permute_type=grid->PermuteType(dimension);
-
-  ///////////////////////////////////////////////
-  // Simd direction uses an extract/merge pair
-  ///////////////////////////////////////////////
-  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
-  //  int words = sizeof(vobj)/sizeof(vector_type);
-
-  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
-  scalar_object *  recv_buf_extract_mpi;
-  scalar_object *  send_buf_extract_mpi;
-  {
-    size_t bytes = sizeof(scalar_object)*buffer_size;
-    grid->ShmBufferFreeAll();
-    send_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-    recv_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-  }
-  for(int s=0;s<Nsimd;s++){
-    send_buf_extract[s].resize(buffer_size);
-    recv_buf_extract[s].resize(buffer_size);
-  }
-
-  int bytes = buffer_size*sizeof(scalar_object);
-
-  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
-  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
-
-  ///////////////////////////////////////////
-  // Work out what to send where
-  ///////////////////////////////////////////
-  int cb    = (cbmask==0x2)? Odd : Even;
-  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  // loop over outer coord planes orthog to dim
-  for(int x=0;x<rd;x++){       
-
-    // FIXME call local permute copy if none are offnode.
-    for(int i=0;i<Nsimd;i++){       
-      pointers[i] = &send_buf_extract[i][0];
-    }
-    tgather-=usecond();
-    int sx   = (x+sshift)%rd;
-    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
-    tgather+=usecond();
-
-    for(int i=0;i<Nsimd;i++){
-      
-      int inner_bit = (Nsimd>>(permute_type+1));
-      int ic= (i&inner_bit)? 1:0;
-
-      int my_coor          = rd*ic + x;
-      int nbr_coor         = my_coor+sshift;
-      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-
-      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
-      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
-      int nbr_lane = (i&(~inner_bit));
-
-      int recv_from_rank;
-      int xmit_to_rank;
-
-      if (nbr_ic) nbr_lane|=inner_bit;
-
-      assert (sx == nbr_ox);
-
-      if(nbr_proc){
-	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
-
-	tcomms-=usecond();
-	//	grid->Barrier();
-
-	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
-	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
-			     xmit_to_rank,
-			     (void *)recv_buf_extract_mpi,
-			     recv_from_rank,
-			     bytes);
-	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
-	xbytes+=bytes;
-
-	//	grid->Barrier();
-	tcomms+=usecond();
-	rpointers[i] = &recv_buf_extract[i][0];
-      } else { 
-	rpointers[i] = &send_buf_extract[nbr_lane][0];
-      }
-
-    }
-    tscatter-=usecond();
-    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-    tscatter+=usecond();
-
-  }
-  if(Cshift_verbose){
-    std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
-    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
-  }
-}
-#endif
 
 NAMESPACE_END(Grid); 
 

Grid/lattice/Lattice_base.h

@@ -236,7 +236,7 @@ public:
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
     vobj vtmp;
     vtmp = r;
-#if 0
+#if 1
     deviceVector<vobj> vvtmp(1);
     acceleratorPut(vvtmp[0],vtmp);
     vobj *vvtmp_p = & vvtmp[0];

Grid/lattice/Lattice_slicesum_core.h

@@ -55,7 +55,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
   d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
   
   //copy offsets to device
-  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  acceleratorCopyToDeviceAsynch(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
   
   
   gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
@@ -88,7 +88,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
     exit(EXIT_FAILURE);
   }
   
-  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  acceleratorCopyFromDeviceAsynch(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
   
   //sync after copy
   accelerator_barrier();

Grid/lattice/PaddedCell.h

@@ -466,6 +466,12 @@ public:
     static deviceVector<vobj> recv_buf;
     send_buf.resize(buffer_size*2*depth);    
     recv_buf.resize(buffer_size*2*depth);
+#ifndef ACCELERATOR_AWARE_MPI
+    static hostVector<vobj> hsend_buf; 
+    static hostVector<vobj> hrecv_buf;
+    hsend_buf.resize(buffer_size*2*depth);    
+    hrecv_buf.resize(buffer_size*2*depth);
+#endif    
 
     std::vector<MpiCommsRequest_t> fwd_req;   
     std::vector<MpiCommsRequest_t> bwd_req;   
@@ -495,9 +501,16 @@ public:
       t_gather+=usecond()-t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(fwd_req,
 				(void *)&send_buf[d*buffer_size], xmit_to_rank,
 				(void *)&recv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[d*buffer_size],&hsend_buf[d*buffer_size],bytes);
+      grid->SendToRecvFromBegin(fwd_req,
+				(void *)&hsend_buf[d*buffer_size], xmit_to_rank,
+				(void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#endif
       t_comms+=usecond()-t;
      }
     for ( int d=0;d < depth ; d ++ ) {
@@ -508,9 +521,16 @@ public:
       t_gather+= usecond() - t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(bwd_req,
 				(void *)&send_buf[(d+depth)*buffer_size], recv_from_rank,
 				(void *)&recv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[(d+depth)*buffer_size],&hsend_buf[(d+depth)*buffer_size],bytes);
+      grid->SendToRecvFromBegin(bwd_req,
+				(void *)&hsend_buf[(d+depth)*buffer_size], recv_from_rank,
+				(void *)&hrecv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#endif      
       t_comms+=usecond()-t;
     }
 
@@ -533,8 +553,13 @@ public:
 
     t=usecond();
     grid->CommsComplete(fwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
-
+    
     t=usecond();
     for ( int d=0;d < depth ; d ++ ) {
       ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++;
@@ -543,6 +568,11 @@ public:
 
     t=usecond();
     grid->CommsComplete(bwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
     
     t=usecond();

Grid/qcd/action/fermion/CompactWilsonCloverFermion5D.h

@@ -0,0 +1,196 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion5D.h
+
+    Copyright (C) 2020 - 2025
+
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Nils Meyer <nils.meyer@ur.de>
+    Author: Christoph Lehner <christoph@lhnr.de>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#pragma once
+
+#include <Grid/qcd/action/fermion/WilsonFermion5D.h>
+#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
+#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
+
+NAMESPACE_BEGIN(Grid);
+
+// see Grid/qcd/action/fermion/CompactWilsonCloverFermion.h for description
+
+template<class Impl, class CloverHelpers>
+class CompactWilsonCloverFermion5D : public WilsonFermion5D<Impl>,
+				     public WilsonCloverHelpers<Impl>,
+				     public CompactWilsonCloverHelpers<Impl> {
+  /////////////////////////////////////////////
+  // Sizes
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_COMPACT_CLOVER_SIZES(Impl);
+
+  /////////////////////////////////////////////
+  // Type definitions
+  /////////////////////////////////////////////
+
+public:
+
+  INHERIT_IMPL_TYPES(Impl);
+  INHERIT_CLOVER_TYPES(Impl);
+  INHERIT_COMPACT_CLOVER_TYPES(Impl);
+
+  typedef WilsonFermion5D<Impl>            WilsonBase;
+  typedef WilsonCloverHelpers<Impl>        Helpers;
+  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
+
+  /////////////////////////////////////////////
+  // Constructors
+  /////////////////////////////////////////////
+
+public:
+
+  CompactWilsonCloverFermion5D(GaugeField& _Umu,
+			       GridCartesian         &FiveDimGrid,
+			       GridRedBlackCartesian &FiveDimRedBlackGrid,
+			       GridCartesian         &FourDimGrid,
+			       GridRedBlackCartesian &FourDimRedBlackGrid,
+			       const RealD _mass,
+			       const RealD _csw_r = 0.0,
+			       const RealD _csw_t = 0.0,
+			       const RealD _cF = 1.0,
+			       const ImplParams& impl_p = ImplParams());
+
+  /////////////////////////////////////////////
+  // Member functions (implementing interface)
+  /////////////////////////////////////////////
+
+public:
+
+  virtual void Instantiatable() {};
+  int          ConstEE()     override { return 0; };
+  int          isTrivialEE() override { return 0; };
+
+  void Dhop(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopOE(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopEO(const FermionField& in, FermionField& out, int dag) override;
+
+  void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
+
+  void M(const FermionField& in, FermionField& out) override;
+
+  void Mdag(const FermionField& in, FermionField& out) override;
+
+  void Meooe(const FermionField& in, FermionField& out) override;
+
+  void MeooeDag(const FermionField& in, FermionField& out) override;
+
+  void Mooee(const FermionField& in, FermionField& out) override;
+
+  void MooeeDag(const FermionField& in, FermionField& out) override;
+
+  void MooeeInv(const FermionField& in, FermionField& out) override;
+
+  void MooeeInvDag(const FermionField& in, FermionField& out) override;
+
+  void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
+
+  void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
+
+  void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
+
+  void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
+
+  /////////////////////////////////////////////
+  // Member functions (internals)
+  /////////////////////////////////////////////
+
+  void MooeeInternal(const FermionField&        in,
+                     FermionField&              out,
+                     const CloverDiagonalField& diagonal,
+                     const CloverTriangleField& triangle);
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+  void ImportGauge(const GaugeField& _Umu) override;
+
+  /////////////////////////////////////////////
+  // Helpers
+  /////////////////////////////////////////////
+
+private:
+
+  template<class Field>
+  const MaskField* getCorrectMaskField(const Field &in) const {
+    if(in.Grid()->_isCheckerBoarded) {
+      if(in.Checkerboard() == Odd) {
+        return &this->BoundaryMaskOdd;
+      } else {
+        return &this->BoundaryMaskEven;
+      }
+    } else {
+      return &this->BoundaryMask;
+    }
+  }
+
+  template<class Field>
+  void ApplyBoundaryMask(Field& f) {
+    const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
+    assert(m != nullptr);
+    CompactHelpers::ApplyBoundaryMask(f, *m);
+  }
+
+  /////////////////////////////////////////////
+  // Member Data
+  /////////////////////////////////////////////
+
+public:
+
+  RealD csw_r;
+  RealD csw_t;
+  RealD cF;
+  int n_rhs;
+  
+  bool fixedBoundaries;
+
+  CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
+  CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
+
+  CloverTriangleField Triangle,    TriangleEven,    TriangleOdd;
+  CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
+
+  FermionField Tmp;
+
+  MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
+};
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -55,6 +55,7 @@ NAMESPACE_CHECK(Wilson);
 NAMESPACE_CHECK(WilsonTM);
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion5D.h> // 5d compact wilson clover fermions
 NAMESPACE_CHECK(WilsonClover);
 #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 NAMESPACE_CHECK(Wilson5D);
@@ -164,12 +165,17 @@ typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiS
 
 // Compact Clover fermions
 template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
+template <typename WImpl> using CompactWilsonClover5D = CompactWilsonCloverFermion5D<WImpl, CompactCloverHelpers<WImpl>>;
 template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
 
 typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
 typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
 typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
 
+typedef CompactWilsonClover5D<WilsonImplD2> CompactWilsonCloverFermion5DD2;
+typedef CompactWilsonClover5D<WilsonImplF> CompactWilsonCloverFermion5DF;
+typedef CompactWilsonClover5D<WilsonImplD> CompactWilsonCloverFermion5DD;
+
 typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
 typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
 typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -484,6 +484,11 @@ public:
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
     accelerator_barrier();
+#ifdef NVLINK_GET
+    this->_grid->StencilBarrier(); // He can now get mu local gather, I can get his
+    // Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
+    // Or issue barrier AFTER the DMA is running
+#endif    
   }
 
 };

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -91,13 +91,13 @@ public:
   virtual void   Mdag (const FermionField &in, FermionField &out){assert(0);};
 
   // half checkerboard operations; leave unimplemented as abstract for now
-  virtual void   Meooe       (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   Mooee       (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   MooeeInv    (const FermionField &in, FermionField &out){assert(0);};
+  virtual void   Meooe       (const FermionField &in, FermionField &out);
+  virtual void   Mooee       (const FermionField &in, FermionField &out);
+  virtual void   MooeeInv    (const FermionField &in, FermionField &out);
 
-  virtual void   MeooeDag    (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   MooeeDag    (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   MooeeInvDag (const FermionField &in, FermionField &out){assert(0);};
+  virtual void   MeooeDag    (const FermionField &in, FermionField &out);
+  virtual void   MooeeDag    (const FermionField &in, FermionField &out);
+  virtual void   MooeeInvDag (const FermionField &in, FermionField &out);
   virtual void   Mdir   (const FermionField &in, FermionField &out,int dir,int disp){assert(0);};   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
   virtual void   MdirAll(const FermionField &in, std::vector<FermionField> &out){assert(0);};   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
 

Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermion5DImplementation.h

@@ -0,0 +1,376 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion5DImplementation.h
+
+    Copyright (C) 2017 - 2025
+
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Christoph Lehner <christoph@lhnr.de>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Grid.h>
+#include <Grid/qcd/spin/Dirac.h>
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion5D.h>
+
+
+NAMESPACE_BEGIN(Grid);
+template<class Impl, class CloverHelpers>
+CompactWilsonCloverFermion5D<Impl, CloverHelpers>::CompactWilsonCloverFermion5D(GaugeField& _Umu,
+										GridCartesian         &FiveDimGrid,
+										GridRedBlackCartesian &FiveDimRedBlackGrid,
+										GridCartesian         &FourDimGrid,
+										GridRedBlackCartesian &FourDimRedBlackGrid,
+										const RealD _mass,
+										const RealD _csw_r,
+										const RealD _csw_t,
+										const RealD _cF,
+										const ImplParams& impl_p)
+  : WilsonBase(_Umu, FiveDimGrid, FiveDimRedBlackGrid, FourDimGrid, FourDimRedBlackGrid, _mass, impl_p)
+  , csw_r(_csw_r)
+  , csw_t(_csw_t)
+  , cF(_cF)
+  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , Diagonal(&FourDimGrid),        Triangle(&FourDimGrid)
+  , DiagonalEven(&FourDimRedBlackGrid),    TriangleEven(&FourDimRedBlackGrid)
+  , DiagonalOdd(&FourDimRedBlackGrid),     TriangleOdd(&FourDimRedBlackGrid)
+  , DiagonalInv(&FourDimGrid),     TriangleInv(&FourDimGrid)
+  , DiagonalInvEven(&FourDimRedBlackGrid), TriangleInvEven(&FourDimRedBlackGrid)
+  , DiagonalInvOdd(&FourDimRedBlackGrid),  TriangleInvOdd(&FourDimRedBlackGrid)
+  , Tmp(&FiveDimGrid)
+  , BoundaryMask(&FiveDimGrid)
+  , BoundaryMaskEven(&FiveDimRedBlackGrid), BoundaryMaskOdd(&FiveDimRedBlackGrid)
+{
+  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
+
+  csw_r *= 0.5;
+  csw_t *= 0.5;
+  //if (clover_anisotropy.isAnisotropic)
+  //  csw_r /= clover_anisotropy.xi_0;
+
+  ImportGauge(_Umu);
+  if (fixedBoundaries) {
+    this->BoundaryMaskEven.Checkerboard() = Even;
+    this->BoundaryMaskOdd.Checkerboard() = Odd;
+    CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
+  }
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::Dhop(in, out, dag);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::DhopOE(in, out, dag);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+  WilsonBase::DhopEO(in, out, dag);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+  WilsonBase::DhopDir(in, out, dir, disp);
+  if(this->fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+  WilsonBase::DhopDirAll(in, out);
+  if(this->fixedBoundaries) {
+    for(auto& o : out) ApplyBoundaryMask(o);
+  }
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
+  out.Checkerboard() = in.Checkerboard();
+  WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
+  Mooee(in, Tmp);
+  axpy(out, 1.0, out, Tmp);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
+  out.Checkerboard() = in.Checkerboard();
+  WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
+  MooeeDag(in, Tmp);
+  axpy(out, 1.0, out, Tmp);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
+  WilsonBase::Meooe(in, out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
+  WilsonBase::MeooeDag(in, out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
+  if(in.Grid()->_isCheckerBoarded) {
+    if(in.Checkerboard() == Odd) {
+      MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
+    } else {
+      MooeeInternal(in, out, DiagonalEven, TriangleEven);
+    }
+  } else {
+    MooeeInternal(in, out, Diagonal, Triangle);
+  }
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
+  Mooee(in, out); // blocks are hermitian
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
+  if(in.Grid()->_isCheckerBoarded) {
+    if(in.Checkerboard() == Odd) {
+      MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
+    } else {
+      MooeeInternal(in, out, DiagonalInvEven, TriangleInvEven);
+    }
+  } else {
+    MooeeInternal(in, out, DiagonalInv, TriangleInv);
+  }
+  if(fixedBoundaries) ApplyBoundaryMask(out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
+  MooeeInv(in, out); // blocks are hermitian
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+  DhopDir(in, out, dir, disp);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+  DhopDirAll(in, out);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+  assert(!fixedBoundaries); // TODO check for changes required for open bc
+
+  // NOTE: code copied from original clover term
+  conformable(X.Grid(), Y.Grid());
+  conformable(X.Grid(), force.Grid());
+  GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
+  GaugeField clover_force(force.Grid());
+  PropagatorField Lambda(force.Grid());
+
+  // Guido: Here we are hitting some performance issues:
+  // need to extract the components of the DoubledGaugeField
+  // for each call
+  // Possible solution
+  // Create a vector object to store them? (cons: wasting space)
+  std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
+
+  Impl::extractLinkField(U, this->Umu);
+
+  force = Zero();
+  // Derivative of the Wilson hopping term
+  this->DhopDeriv(force, X, Y, dag);
+
+  ///////////////////////////////////////////////////////////
+  // Clover term derivative
+  ///////////////////////////////////////////////////////////
+  Impl::outerProductImpl(Lambda, X, Y);
+  //std::cout << "Lambda:" << Lambda << std::endl;
+
+  Gamma::Algebra sigma[] = {
+      Gamma::Algebra::SigmaXY,
+      Gamma::Algebra::SigmaXZ,
+      Gamma::Algebra::SigmaXT,
+      Gamma::Algebra::MinusSigmaXY,
+      Gamma::Algebra::SigmaYZ,
+      Gamma::Algebra::SigmaYT,
+      Gamma::Algebra::MinusSigmaXZ,
+      Gamma::Algebra::MinusSigmaYZ,
+      Gamma::Algebra::SigmaZT,
+      Gamma::Algebra::MinusSigmaXT,
+      Gamma::Algebra::MinusSigmaYT,
+      Gamma::Algebra::MinusSigmaZT};
+
+  /*
+    sigma_{\mu \nu}=
+    | 0         sigma[0]  sigma[1]  sigma[2] |
+    | sigma[3]    0       sigma[4]  sigma[5] |
+    | sigma[6]  sigma[7]     0      sigma[8] |
+    | sigma[9]  sigma[10] sigma[11]   0      |
+  */
+
+  int count = 0;
+  clover_force = Zero();
+  for (int mu = 0; mu < 4; mu++)
+  {
+    force_mu = Zero();
+    for (int nu = 0; nu < 4; nu++)
+    {
+      if (mu == nu)
+        continue;
+
+      RealD factor;
+      if (nu == 4 || mu == 4)
+      {
+        factor = 2.0 * csw_t;
+      }
+      else
+      {
+        factor = 2.0 * csw_r;
+      }
+      PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
+      Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
+      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
+      count++;
+    }
+
+    pokeLorentz(clover_force, U[mu] * force_mu, mu);
+  }
+  //clover_force *= csw;
+  force += clover_force;
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+  assert(0);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+  assert(0);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
+								      FermionField&              out,
+								      const CloverDiagonalField& diagonal,
+								      const CloverTriangleField& triangle) {
+  assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
+  out.Checkerboard() = in.Checkerboard();
+  conformable(in, out);
+  CompactHelpers::MooeeKernel(diagonal.oSites(), this->Ls, in, out, diagonal, triangle);
+}
+
+template<class Impl, class CloverHelpers>
+void CompactWilsonCloverFermion5D<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
+  // NOTE: parts copied from original implementation
+
+  // Import gauge into base class
+  double t0 = usecond();
+  WilsonBase::ImportGauge(_Umu); // NOTE: called here and in wilson constructor -> performed twice, but can't avoid that
+
+  // Initialize temporary variables
+  double t1 = usecond();
+  conformable(_Umu.Grid(), this->GaugeGrid());
+  GridBase* grid = _Umu.Grid();
+  typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
+  CloverField TmpOriginal(grid);
+  CloverField TmpInverse(grid);
+
+  // Compute the field strength terms mu>nu
+  double t2 = usecond();
+  WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
+  WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
+  WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
+  WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
+
+  // Compute the Clover Operator acting on Colour and Spin
+  // multiply here by the clover coefficients for the anisotropy
+  double t3 = usecond();
+  TmpOriginal  = Helpers::fillCloverYZ(Bx) * csw_r;
+  TmpOriginal += Helpers::fillCloverXZ(By) * csw_r;
+  TmpOriginal += Helpers::fillCloverXY(Bz) * csw_r;
+  TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
+  TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
+  TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
+
+  // Instantiate the clover term
+  // - In case of the standard clover the mass term is added
+  // - In case of the exponential clover the clover term is exponentiated
+  double t4 = usecond();
+  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, 4.0 + this->M5 /*this->diag_mass*/);
+
+  // Convert the data layout of the clover term
+  double t5 = usecond();
+  CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
+
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
+  double t6 = usecond();
+  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, 4.0 + this->M5 /*this->diag_mass*/);
+
+  // Invert the Clover term
+  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
+  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
+  // TODO: For now this inversion is explictly done on the CPU
+  double t7 = usecond();
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
+
+  // Fill the remaining clover fields
+  double t8 = usecond();
+  pickCheckerboard(Even, DiagonalEven,    Diagonal);
+  pickCheckerboard(Even, TriangleEven,    Triangle);
+  pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
+  pickCheckerboard(Odd,  TriangleOdd,     Triangle);
+  pickCheckerboard(Even, DiagonalInvEven, DiagonalInv);
+  pickCheckerboard(Even, TriangleInvEven, TriangleInv);
+  pickCheckerboard(Odd,  DiagonalInvOdd,  DiagonalInv);
+  pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
+
+  // Report timings
+  double t9 = usecond();
+
+  std::cout << GridLogDebug << "CompactWilsonCloverFermion5D::ImportGauge timings:" << std::endl;
+  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
+}
+
+NAMESPACE_END(Grid);

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

@@ -14,6 +14,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 Author: Guido Cossu <guido.cossu@ed.ac.uk>
 Author: Andrew Lawson <andrew.lawson1991@gmail.com>
 Author: Vera Guelpers <V.M.Guelpers@soton.ac.uk>
+Author: Christoph Lehner <christoph@lhnr.de>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -484,6 +485,54 @@ void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag
   Dhop(in,out,dag); // -0.5 is included
   axpy(out,4.0-M5,in,out);
 }
+template <class Impl>
+void WilsonFermion5D<Impl>::Meooe(const FermionField &in, FermionField &out)
+{
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerNo);
+  } else {
+    DhopOE(in, out, DaggerNo);
+  }
+}
+
+template <class Impl>
+void WilsonFermion5D<Impl>::MeooeDag(const FermionField &in, FermionField &out)
+{
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerYes);
+  } else {
+    DhopOE(in, out, DaggerYes);
+  }
+}
+
+template <class Impl>
+void WilsonFermion5D<Impl>::Mooee(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  typename FermionField::scalar_type scal(4.0 + M5);
+  out = scal * in;
+}
+
+template <class Impl>
+void WilsonFermion5D<Impl>::MooeeDag(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  Mooee(in, out);
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::MooeeInv(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  out = (1.0/(4.0 + M5))*in;
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  MooeeInv(in,out);
+}
 
 template<class Impl>
 void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const FermionField &in, RealD mass,std::vector<double> twist)

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

@@ -63,7 +63,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
   } else {							\
     chi = coalescedRead(buf[SE->_offset],lane);			\
   }								\
-  acceleratorSynchronise();						\
+  acceleratorSynchronise();					\
   Impl::multLink(Uchi, U[sU], chi, Dir, SE, st);		\
   Recon(result, Uchi);
 
@@ -517,7 +517,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
-     // dependent on result of merge
+     //     // dependent on result of merge
      acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt);    return;}

Grid/qcd/action/fermion/instantiation/CompactWilsonCloverFermion5DInstantiation.cc.master

@@ -0,0 +1,45 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/ qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation5D.cc.master
+
+    Copyright (C) 2017 - 2025
+
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@ed.ac.uk>
+    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
+    Author: Mattia Bruno <mattia.bruno@cern.ch>
+    Author: Christoph Lehner <christoph@lhnr.de>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Grid.h>
+#include <Grid/qcd/spin/Dirac.h>
+#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion5D.h>
+#include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermion5DImplementation.h>
+#include <Grid/qcd/action/fermion/CloverHelpers.h>
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class CompactWilsonCloverFermion5D<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>; 
+template class CompactWilsonCloverFermion5D<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>; 
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplD/CompactWilsonCloverFermion5DInstantiationWilsonImplD.cc

@@ -0,0 +1 @@
+../CompactWilsonCloverFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplF/CompactWilsonCloverFermion5DInstantiationWilsonImplF.cc

@@ -0,0 +1 @@
+../CompactWilsonCloverFermion5DInstantiation.cc.master

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

@@ -62,7 +62,7 @@ do
 done
 done
 
-CC_LIST="CompactWilsonCloverFermionInstantiation"
+CC_LIST="CompactWilsonCloverFermionInstantiation CompactWilsonCloverFermion5DInstantiation"
 
 for impl in $COMPACT_WILSON_IMPL_LIST
 do

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

@@ -76,27 +76,27 @@ public:
     return action;
   };
 
-  virtual void deriv(const GaugeField &Umu,GaugeField & dSdU) {
+  virtual void deriv(const GaugeField &U, GaugeField &dSdU) {
     //extend Ta to include Lorentz indexes
     RealD factor_p = c_plaq/RealD(Nc)*0.5;
     RealD factor_r = c_rect/RealD(Nc)*0.5;
 
-    GridBase *grid = Umu.Grid();
+    GridBase *grid = U.Grid();
 
-    std::vector<GaugeLinkField> U (Nd,grid);
+    std::vector<GaugeLinkField> Umu (Nd,grid);
     for(int mu=0;mu<Nd;mu++){
-      U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+      Umu[mu] = PeekIndex<LorentzIndex>(U,mu);
     }
     std::vector<GaugeLinkField> RectStaple(Nd,grid), Staple(Nd,grid);
-    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, U, workspace);
+    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, Umu, workspace);
 
     GaugeLinkField dSdU_mu(grid);
     GaugeLinkField staple(grid);
 
     for (int mu=0; mu < Nd; mu++){
-      dSdU_mu = Ta(U[mu]*Staple[mu])*factor_p;
-      dSdU_mu = dSdU_mu + Ta(U[mu]*RectStaple[mu])*factor_r;
-	  
+      dSdU_mu = Ta(Umu[mu]*Staple[mu])*factor_p;
+      dSdU_mu = dSdU_mu + Ta(Umu[mu]*RectStaple[mu])*factor_r;
+
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);
     }
 

Grid/qcd/action/gauge/WilsonGaugeAction.h

@@ -73,20 +73,23 @@ public:
     // extend Ta to include Lorentz indexes
 
     RealD factor = 0.5 * beta / RealD(Nc);
+    GridBase *grid = U.Grid();
 
-    GaugeLinkField Umu(U.Grid());
-    GaugeLinkField dSdU_mu(U.Grid());
+    GaugeLinkField dSdU_mu(grid);
+    std::vector<GaugeLinkField> Umu(Nd, grid);
     for (int mu = 0; mu < Nd; mu++) {
+      Umu[mu] = PeekIndex<LorentzIndex>(U, mu);
+    }
 
-      Umu = PeekIndex<LorentzIndex>(U, mu);
-      
+    for (int mu = 0; mu < Nd; mu++) {
       // Staple in direction mu
-      WilsonLoops<Gimpl>::Staple(dSdU_mu, U, mu);
-      dSdU_mu = Ta(Umu * dSdU_mu) * factor;
-      
+      WilsonLoops<Gimpl>::Staple(dSdU_mu, Umu, mu);
+      dSdU_mu = Ta(Umu[mu] * dSdU_mu) * factor;
+
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);
     }
   }
+
 private:
   RealD beta;  
  };

Grid/qcd/hmc/checkpointers/BinaryCheckpointer.h

@@ -111,8 +111,8 @@ public:
   };
 
   void CheckpointRestore(int traj, Field &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
-    std::string config, rng;
-    this->build_filenames(traj, Params, config, rng);
+    std::string config, rng, smr;
+    this->build_filenames(traj, Params, config, smr, rng);
     this->check_filename(rng);
     this->check_filename(config);
 

Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h

@@ -75,7 +75,7 @@ public:
                           GridParallelRNG &pRNG) {
     if ((traj % Params.saveInterval) == 0) {
       std::string config, rng, smr;
-      this->build_filenames(traj, Params, config, rng);
+      this->build_filenames(traj, Params, config, smr, rng);
       GridBase *grid = SmartConfig.get_U(false).Grid();
       uint32_t nersc_csum,scidac_csuma,scidac_csumb;
       BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
@@ -102,7 +102,7 @@ public:
       if ( Params.saveSmeared ) { 
 	IldgWriter _IldgWriter(grid->IsBoss());
 	_IldgWriter.open(smr);
-	_IldgWriter.writeConfiguration<GaugeStats>(SmartConfig.get_U(true), traj, config, config);
+	_IldgWriter.writeConfiguration<GaugeStats>(SmartConfig.get_U(true), traj, smr, smr);
 	_IldgWriter.close();
 
 	std::cout << GridLogMessage << "Written ILDG Configuration on " << smr
@@ -118,8 +118,8 @@ public:
 
   void CheckpointRestore(int traj, GaugeField &U, GridSerialRNG &sRNG,
                          GridParallelRNG &pRNG) {
-    std::string config, rng;
-    this->build_filenames(traj, Params, config, rng);
+    std::string config, rng, smr;
+    this->build_filenames(traj, Params, config, smr, rng);
     this->check_filename(rng);
     this->check_filename(config);
 

Grid/qcd/hmc/checkpointers/ScidacCheckpointer.h

@@ -107,8 +107,8 @@ class ScidacHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
 
   void CheckpointRestore(int traj, Field &U, GridSerialRNG &sRNG,
                          GridParallelRNG &pRNG) {
-    std::string config, rng;
-    this->build_filenames(traj, Params, config, rng);
+    std::string config, rng, smr;
+    this->build_filenames(traj, Params, config, smr, rng);
     this->check_filename(rng);
     this->check_filename(config);
 

Grid/qcd/smearing/HISQSmearing.h

@@ -62,15 +62,15 @@ accelerator_inline int stencilIndex(int mu, int nu) {
 
 
 /*!  @brief structure holding the link treatment */
-struct SmearingParameters{
-    SmearingParameters(){}
+struct HISQSmearingParameters{
+    HISQSmearingParameters(){}
     Real c_1;               // 1 link
     Real c_naik;            // Naik term
     Real c_3;               // 3 link
     Real c_5;               // 5 link
     Real c_7;               // 7 link
     Real c_lp;              // 5 link Lepage
-    SmearingParameters(Real c1, Real cnaik, Real c3, Real c5, Real c7, Real clp) 
+    HISQSmearingParameters(Real c1, Real cnaik, Real c3, Real c5, Real c7, Real clp) 
         : c_1(c1),
           c_naik(cnaik),
           c_3(c3),
@@ -86,7 +86,7 @@ class Smear_HISQ : public Gimpl {
 
 private:
     GridCartesian* const _grid;
-    SmearingParameters _linkTreatment;
+    HISQSmearingParameters _linkTreatment;
 
 public:
 
@@ -117,7 +117,7 @@ public:
     //          IN--u_thin
     void smear(GF& u_smr, GF& u_naik, GF& u_thin) const {
 
-        SmearingParameters lt = this->_linkTreatment;
+        HISQSmearingParameters lt = this->_linkTreatment;
         auto grid = this->_grid;
 
         // Create a padded cell of extra padding depth=1 and fill the padding.

Grid/qcd/smearing/WilsonFlow.h

@@ -252,6 +252,11 @@ void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const{
 
   out = in;
   RealD taus = 0.;
+
+  // Perform initial t=0 measurements
+  for(auto const &meas : this->functions)
+    meas.second(0,taus,out);
+  
   for (unsigned int step = 1; step <= Nstep; step++) { //step indicates the number of smearing steps applied at the time of measurement
     auto start = std::chrono::high_resolution_clock::now();
     evolve_step(out, taus);
@@ -336,6 +341,11 @@ void WilsonFlowAdaptive<Gimpl>::smear(GaugeField& out, const GaugeField& in) con
   RealD taus = 0.;
   RealD eps = init_epsilon;
   unsigned int step = 0;
+
+  // Perform initial t=0 measurements
+  for(auto const &meas : this->functions)
+    meas.second(step,taus,out);
+  
   do{
     int step_success = evolve_step_adaptive(out, taus, eps); 
     step += step_success; //step will not be incremented if the integration step fails

Grid/qcd/utils/WilsonLoops.h

@@ -292,19 +292,21 @@ public:
   //////////////////////////////////////////////////
   // the sum over all nu-oriented staples for nu != mu on each site
   //////////////////////////////////////////////////
-  static void Staple(GaugeMat &staple, const GaugeLorentz &Umu, int mu) {
+  static void Staple(GaugeMat &staple, const GaugeLorentz &U, int mu) {
 
-    GridBase *grid = Umu.Grid();
-
-    std::vector<GaugeMat> U(Nd, grid);
+    std::vector<GaugeMat> Umu(Nd, U.Grid());
     for (int d = 0; d < Nd; d++) {
-      U[d] = PeekIndex<LorentzIndex>(Umu, d);
+      Umu[d] = PeekIndex<LorentzIndex>(U, d);
     }
-    Staple(staple, U, mu);
+    Staple(staple, Umu, mu);
   }
 
-  static void Staple(GaugeMat &staple, const std::vector<GaugeMat> &U, int mu) {
-    staple = Zero();
+  static void Staple(GaugeMat &staple, const std::vector<GaugeMat> &Umu, int mu) {
+
+    autoView(staple_v, staple, AcceleratorWrite);
+    accelerator_for(i, staple.Grid()->oSites(), Simd::Nsimd(), {
+        staple_v[i] = Zero();
+    });
 
     for (int nu = 0; nu < Nd; nu++) {
 
@@ -318,12 +320,12 @@ public:
         //      |
         //    __|
         //
-     
+
         staple += Gimpl::ShiftStaple(
 				     Gimpl::CovShiftForward(
-							    U[nu], nu,
+							    Umu[nu], nu,
 							    Gimpl::CovShiftBackward(
-										    U[mu], mu, Gimpl::CovShiftIdentityBackward(U[nu], nu))),
+										    Umu[mu], mu, Gimpl::CovShiftIdentityBackward(Umu[nu], nu))),
 				     mu);
 
         //  __
@@ -333,8 +335,8 @@ public:
         //
 
         staple += Gimpl::ShiftStaple(
-				     Gimpl::CovShiftBackward(U[nu], nu,
-							     Gimpl::CovShiftBackward(U[mu], mu, U[nu])), mu);
+				     Gimpl::CovShiftBackward(Umu[nu], nu,
+							     Gimpl::CovShiftBackward(Umu[mu], mu, Umu[nu])), mu);
       }
     }
   }

Grid/stencil/Stencil.h

@@ -363,12 +363,16 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
+    //    std::cout << "Communicate Begin "<<std::endl;
+    //    _grid->Barrier();
     FlightRecorder::StepLog("Communicate begin");
     // All GPU kernel tasks must complete
     //    accelerator_barrier();     // All kernels should ALREADY be complete
     //    _grid->StencilBarrier();   // Everyone is here, so noone running slow and still using receive buffer
                                // But the HaloGather had a barrier too.
     for(int i=0;i<Packets.size();i++){
+      //      std::cout << "Communicate prepare "<<i<<std::endl;
+      //      _grid->Barrier();
       _grid->StencilSendToRecvFromPrepare(MpiReqs,
 					  Packets[i].send_buf,
 					  Packets[i].to_rank,Packets[i].do_send,
@@ -376,8 +380,15 @@ public:
 					  Packets[i].from_rank,Packets[i].do_recv,
 					  Packets[i].xbytes,Packets[i].rbytes,i);
     }
+    //    std::cout << "Communicate PollDtoH "<<std::endl;
+    //    _grid->Barrier();
+    _grid->StencilSendToRecvFromPollDtoH (MpiReqs); /* Starts MPI*/
+    //    std::cout << "Communicate CopySynch "<<std::endl;
+    //    _grid->Barrier();
     acceleratorCopySynchronise();
+    // Starts intranode
     for(int i=0;i<Packets.size();i++){
+      //      std::cout << "Communicate Begin "<<i<<std::endl;
       _grid->StencilSendToRecvFromBegin(MpiReqs,
 					Packets[i].send_buf,
 					Packets[i].to_rank,Packets[i].do_send,
@@ -385,6 +396,7 @@ public:
 					Packets[i].from_rank,Packets[i].do_recv,
 					Packets[i].xbytes,Packets[i].rbytes,i);
     }
+    FlightRecorder::StepLog("Communicate begin has finished");
     // Get comms started then run checksums
     // Having this PRIOR to the dslash seems to make Sunspot work... (!)
     for(int i=0;i<Packets.size();i++){
@@ -395,7 +407,14 @@ public:
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
+    //    std::cout << "Communicate Complete "<<std::endl;
+    //    _grid->Barrier();
     FlightRecorder::StepLog("Start communicate complete");
+    //    std::cout << "Communicate Complete PollIRecv "<<std::endl;
+    //    _grid->Barrier();
+    _grid->StencilSendToRecvFromPollIRecv(MpiReqs);
+    //    std::cout << "Communicate Complete Complete "<<std::endl;
+    //    _grid->Barrier();
     _grid->StencilSendToRecvFromComplete(MpiReqs,0); // MPI is done
     if   ( this->partialDirichlet ) DslashLogPartial();
     else if ( this->fullDirichlet ) DslashLogDirichlet();
@@ -428,6 +447,7 @@ public:
     Communicate();
     CommsMergeSHM(compress);
     CommsMerge(compress);
+    accelerator_barrier();
   }
 
   template<class compressor> int HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point,int & face_idx)
@@ -483,6 +503,9 @@ public:
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
     //    accelerator_barrier();
+    //////////////////////////////////
+    // I will overwrite my send buffers
+    //////////////////////////////////
     _grid->StencilBarrier();// Synch shared memory on a single nodes
 
     assert(source.Grid()==_grid);
@@ -496,7 +519,11 @@ public:
       HaloGatherDir(source,compress,point,face_idx);
     }
     accelerator_barrier(); // All my local gathers are complete
-    //    _grid->StencilBarrier();// Synch shared memory on a single nodes
+#ifdef NVLINK_GET
+    _grid->StencilBarrier(); // He can now get mu local gather, I can get his
+    // Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
+    // Or issue barrier AFTER the DMA is running
+#endif    
     face_table_computed=1;
     assert(u_comm_offset==_unified_buffer_size);
   }
@@ -535,6 +562,7 @@ public:
 	  coalescedWrite(to[j] ,coalescedRead(from [j]));
       });
       acceleratorFenceComputeStream();
+      // Also fenced in WilsonKernels
     }
   }
   
@@ -663,7 +691,7 @@ public:
 	}
       }
     }
-    std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
+    //    std::cout << "BuildSurfaceList size is "<<surface_list_size<<std::endl;
     surface_list.resize(surface_list_size);
     std::vector<int> surface_list_host(surface_list_size);
     int32_t ss=0;
@@ -683,6 +711,7 @@ public:
       }
     }
     acceleratorCopyToDevice(&surface_list_host[0],&surface_list[0],surface_list_size*sizeof(int));
+    //    std::cout << GridLogMessage<<"BuildSurfaceList size is "<<surface_list_size<<std::endl;
   }
   /// Introduce a block structure and switch off comms on boundaries
   void DirichletBlock(const Coordinate &dirichlet_block)
@@ -774,8 +803,8 @@ public:
     this->_entries_host_p = &_entries[0];
     this->_entries_p = &_entries_device[0];
 
-    std::cout << GridLogMessage << " Stencil object allocated for "<<std::dec<<this->_osites
-	      <<" sites table "<<std::hex<<this->_entries_p<< " GridPtr "<<_grid<<std::dec<<std::endl;
+    //    std::cout << GridLogMessage << " Stencil object allocated for "<<std::dec<<this->_osites
+    //	      <<" sites table "<<std::hex<<this->_entries_p<< " GridPtr "<<_grid<<std::dec<<std::endl;
     
     for(int ii=0;ii<npoints;ii++){
 

Grid/threads/Accelerator.h

@@ -242,19 +242,33 @@ inline void *acceleratorAllocDevice(size_t bytes)
   return ptr;
 };
 
+typedef int acceleratorEvent_t;
+
 inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
 inline void acceleratorFreeHost(void *ptr){ cudaFree(ptr);};
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
-inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyHostToDevice, stream);}
-inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToHost, stream);}
+inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
+inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
+inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) {
+  acceleratorCopyToDevice(from,to,bytes);
+  return 0;
+}
+inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) {
+  acceleratorCopyFromDevice(from,to,bytes);
+  return 0;
+}
+inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
   cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
+  return 0;
 }
 inline void acceleratorCopySynchronise(void) { cudaStreamSynchronize(copyStream); };
+inline void acceleratorEventWait(acceleratorEvent_t ev)
+{
+  //auto discard=cudaStreamSynchronize(ev);
+}
+inline int acceleratorEventIsComplete(acceleratorEvent_t ev){ acceleratorEventWait(ev) ; return 1;}
 
 
 inline int  acceleratorIsCommunicable(void *ptr)
@@ -323,7 +337,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
     cgh.parallel_for(							\
 		     sycl::nd_range<3>(global,local),			\
 		     [=] (sycl::nd_item<3> item) /*mutable*/		\
-		     [[intel::reqd_sub_group_size(16)]]			\
+		     [[sycl::reqd_sub_group_size(16)]]			\
 		     {							\
 		       auto iter1    = item.get_global_id(0);		\
 		       auto iter2    = item.get_global_id(1);		\
@@ -343,11 +357,28 @@ inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
 
 inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
 
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {  theCopyAccelerator->memcpy(to,from,bytes);}
-inline void acceleratorCopyToDeviceAsynch(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); }
-inline void acceleratorCopyFromDeviceAsynch(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); }
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+
+///////
+// Asynch event interface
+///////
+typedef sycl::event acceleratorEvent_t;
+
+inline void acceleratorEventWait(acceleratorEvent_t ev)
+{
+  ev.wait();
+}
+
+inline int acceleratorEventIsComplete(acceleratorEvent_t ev)
+{
+  return (ev.get_info<sycl::info::event::command_execution_status>() == sycl::info::event_command_status::complete);
+}
+
+inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { return theCopyAccelerator->memcpy(to,from,bytes);}
+inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from,void *to,size_t bytes)        { return theCopyAccelerator->memcpy(to,from,bytes); }
+inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from,void *to,size_t bytes)      { return theCopyAccelerator->memcpy(to,from,bytes); }
+
+inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
 
 inline int  acceleratorIsCommunicable(void *ptr)
@@ -358,8 +389,10 @@ inline int  acceleratorIsCommunicable(void *ptr)
   else return 0;
 #endif
   return 1;
+
 }
 
+
 #endif
 
 //////////////////////////////////////////////
@@ -459,7 +492,7 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 inline void *acceleratorAllocHost(size_t bytes)
 {
   void *ptr=NULL;
-  auto err = hipMallocHost((void **)&ptr,bytes);
+  auto err = hipHostMalloc((void **)&ptr,bytes);
   if( err != hipSuccess ) {
     ptr = (void *) NULL;
     fprintf(stderr," hipMallocManaged failed for %ld %s \n",bytes,hipGetErrorString(err)); fflush(stderr);
@@ -492,23 +525,35 @@ inline void *acceleratorAllocDevice(size_t bytes)
 inline void acceleratorFreeHost(void *ptr){ auto discard=hipFree(ptr);};
 inline void acceleratorFreeShared(void *ptr){ auto discard=hipFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ auto discard=hipFree(ptr);};
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
+inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes)  { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
+inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
 
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto discard=hipMemset(base,value,bytes);}
 
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
+typedef int acceleratorEvent_t;
+
+inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
   auto discard=hipMemcpyDtoDAsync(to,from,bytes, copyStream);
+  return 0;
 }
-inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
-  auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyHostToDevice, stream);
+inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
+  acceleratorCopyToDevice(from,to,bytes);
+  return 0;
 }
-inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
-  auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToHost, stream);
+inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
+  acceleratorCopyFromDevice(from,to,bytes);
+  return 0;
 }
 inline void acceleratorCopySynchronise(void) { auto discard=hipStreamSynchronize(copyStream); };
 
+inline void acceleratorEventWait(acceleratorEvent_t ev)
+{
+  //  auto discard=hipStreamSynchronize(ev);
+}
+inline int acceleratorEventIsComplete(acceleratorEvent_t ev){ acceleratorEventWait(ev) ; return 1;}
+
+
 #endif
 
 inline void acceleratorPin(void *ptr,unsigned long bytes)
@@ -545,6 +590,8 @@ inline void acceleratorPin(void *ptr,unsigned long bytes)
 
 #undef GRID_SIMT
 
+typedef int acceleratorEvent_t;
+
 inline void acceleratorMem(void)
 {
   /*
@@ -565,8 +612,13 @@ inline void acceleratorMem(void)
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
 
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes);}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
+inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from,void *to,size_t bytes)        { acceleratorCopyToDevice(from,to,bytes); return 0; }
+inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from,void *to,size_t bytes)      { acceleratorCopyFromDevice(from,to,bytes); return 0; }
+inline void acceleratorEventWait(acceleratorEvent_t ev){}
+inline int acceleratorEventIsComplete(acceleratorEvent_t ev){ acceleratorEventWait(ev); return 1;}
+inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); return 0;}
+
 inline void acceleratorCopySynchronise(void) {};
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
@@ -655,9 +707,9 @@ inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)
   acceleratorCopySynchronise();
 }
 
-template<class T> void acceleratorPut(T& dev,T&host)
+template<class T> void acceleratorPut(T& dev,const T&host)
 {
-  acceleratorCopyToDevice(&host,&dev,sizeof(T));
+  acceleratorCopyToDevice((void *)&host,&dev,sizeof(T));
 }
 template<class T> T acceleratorGet(T& dev)
 {

Grid/threads/Threads.h

@@ -73,9 +73,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define thread_critical                                     DO_PRAGMA(omp critical)
 
 #ifdef GRID_OMP
-inline void thread_bcopy(void *from, void *to,size_t bytes)
+inline void thread_bcopy(const void *from, void *to,size_t bytes)
 {
-  uint64_t *ufrom = (uint64_t *)from;
+  const uint64_t *ufrom = (const uint64_t *)from;
   uint64_t *uto   = (uint64_t *)to;
   assert(bytes%8==0);
   uint64_t words=bytes/8;
@@ -84,7 +84,7 @@ inline void thread_bcopy(void *from, void *to,size_t bytes)
   });
 }
 #else
-inline void thread_bcopy(void *from, void *to,size_t bytes)
+inline void thread_bcopy(const void *from, void *to,size_t bytes)
 {
   bcopy(from,to,bytes);
 }

Grid/util/Init.cc

@@ -509,7 +509,14 @@ void Grid_init(int *argc,char ***argv)
 		  Grid_default_latt,
 		  Grid_default_mpi);
 
-
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--flightrecorder") ){
+    std::cout << GridLogMessage <<" Enabling flight recorder " <<std::endl;
+    FlightRecorder::SetLoggingMode(FlightRecorder::LoggingModeRecord);
+    FlightRecorder::PrintEntireLog = 1;
+    FlightRecorder::ChecksumComms  = 1;
+    FlightRecorder::ChecksumCommsSend=1;
+  }
+  
   if( GridCmdOptionExists(*argv,*argv+*argc,"--decomposition") ){
     std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
     std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
@@ -631,12 +638,11 @@ void Grid_debug_handler_init(void)
   sa.sa_flags    = SA_SIGINFO;
   //  sigaction(SIGSEGV,&sa,NULL);
   sigaction(SIGTRAP,&sa,NULL);
-  sigaction(SIGBUS,&sa,NULL);
+  //  sigaction(SIGBUS,&sa,NULL);
   //  sigaction(SIGUSR2,&sa,NULL);
 
-  feenableexcept( FE_INVALID|FE_OVERFLOW|FE_DIVBYZERO);
-
-  sigaction(SIGFPE,&sa,NULL);
+  //  feenableexcept( FE_INVALID|FE_OVERFLOW|FE_DIVBYZERO);
+  //  sigaction(SIGFPE,&sa,NULL);
   sigaction(SIGKILL,&sa,NULL);
   sigaction(SIGILL,&sa,NULL);
 
@@ -651,3 +657,4 @@ void Grid_debug_handler_init(void)
 }
 
 NAMESPACE_END(Grid);
+

Grid/util/Lexicographic.h

@@ -50,7 +50,7 @@ namespace Grid{
       int64_t index64;
       IndexFromCoorReversed(coor,index64,dims);
       if ( index64>=2*1024*1024*1024LL ){
-	std::cout << " IndexFromCoorReversed " << coor<<" index " << index64<< " dims "<<dims<<std::endl;
+	//	std::cout << " IndexFromCoorReversed " << coor<<" index " << index64<< " dims "<<dims<<std::endl;
       }
       assert(index64<2*1024*1024*1024LL);
       index = (int) index64;

HMC/ComputeWilsonFlow.cc

@@ -66,6 +66,7 @@ namespace Grid{
   };
 }
 
+
 template <class T> void writeFile(T& in, std::string const fname){  
 #ifdef HAVE_LIME
   // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
@@ -73,7 +74,7 @@ template <class T> void writeFile(T& in, std::string const fname){
   Grid::emptyUserRecord record;
   Grid::ScidacWriter WR(in.Grid()->IsBoss());
   WR.open(fname);
-  WR.writeScidacFieldRecord(in,record,0);
+  WR.writeScidacFieldRecord(in,record,0); // Lexico
   WR.close();
 #endif
   // What is the appropriate way to throw error?
@@ -107,8 +108,18 @@ int main(int argc, char **argv) {
 
   for (int conf = CPar.StartConfiguration; conf <= CPar.EndConfiguration; conf+= CPar.Skip){
 
+#if 0    
   CPNersc.CheckpointRestore(conf, Umu, sRNG, pRNG);
+#else
+  // Don't require Grid format RNGs
+  FieldMetaData header;
+  std::string file, filesmr;
+  file    = CPar.conf_path + "/" + CPar.conf_prefix      + "." + std::to_string(conf);
+  filesmr = CPar.conf_path + "/" + CPar.conf_smr_prefix  + "." + std::to_string(conf);
 
+  NerscIO::readConfiguration(Umu,header,file);
+#endif
+  
   std::cout << std::setprecision(15);
   std::cout << GridLogMessage << "Initial plaquette: "<< WilsonLoops<PeriodicGimplR>::avgPlaquette(Umu) << std::endl;
   
@@ -116,6 +127,7 @@ int main(int argc, char **argv) {
   std::string file_post = CPar.conf_prefix + "." + std::to_string(conf);
 
   WilsonFlow<PeriodicGimplR> WF(WFPar.step_size,WFPar.steps,WFPar.meas_interval);
+  
   WF.addMeasurement(WFPar.meas_interval_density, [&file_pre,&file_post,&conf](int step, RealD t, const typename PeriodicGimplR::GaugeField &U){
     
     typedef typename PeriodicGimplR::GaugeLinkField GaugeMat;
@@ -165,33 +177,48 @@ int main(int argc, char **argv) {
     //double coeff = 2.0 / (1.0 * Nd * (Nd - 1)) / 3.0;
     //Plq = coeff * Plq;
 
-    int tau = std::round(t);
-    std::string efile = file_pre + "E_dnsty_" + std::to_string(tau) + "_" + file_post;
-    writeFile(R,efile);
-    std::string tfile = file_pre + "Top_dnsty_" + std::to_string(tau) + "_" + file_post;
-    writeFile(qfield,tfile);
 
+    RealD WFlow_TC5Li   = WilsonLoops<PeriodicGimplR>::TopologicalCharge5Li(U);
+
+    int tau = std::round(t);
+
+    std::string efile = file_pre + "E_dnsty_" + std::to_string(tau) + "_" + file_post;
+    //    writeFile(R,efile);
+
+    std::string tfile = file_pre + "Top_dnsty_" + std::to_string(tau) + "_" + file_post;
+    //    writeFile(qfield,tfile);
+
+    std::string ufile = file_pre + "U_" + std::to_string(tau) + "_" + file_post;
+    {
+      //      PeriodicGimplR::GaugeField Ucopy = U;
+      //      NerscIO::writeConfiguration(Ucopy,ufile);
+    }
+    
     RealD E = real(sum(R))/ RealD(U.Grid()->gSites());
     RealD T = real( sum(qfield) );
     Coordinate scoor; for (int mu=0; mu < Nd; mu++) scoor[mu] = 0;
     RealD E0 = real(peekSite(R,scoor));
     RealD T0 = real(peekSite(qfield,scoor));
     std::cout << GridLogMessage << "[WilsonFlow] Saved energy density (clover) & topo. charge density: "  << conf << " " << step << "  " << tau << "  "
-	      << "(E_avg,T_sum) " << E << " " << T << " (E, T at origin) " << E0 << " " << T0 << std::endl;
+	      << "(E_avg,T_sum) " << E << " " << T << " (E, T at origin) " << E0 << " " << T0 << " Q5Li "<< WFlow_TC5Li << std::endl;
     
   });
   
   int t=WFPar.maxTau;
   WF.smear(Uflow, Umu);
-
+  //  NerscIO::writeConfiguration(Uflow,filesmr);
+  
+  
   RealD WFlow_plaq = WilsonLoops<PeriodicGimplR>::avgPlaquette(Uflow);
   RealD WFlow_TC   = WilsonLoops<PeriodicGimplR>::TopologicalCharge(Uflow);
+  RealD WFlow_TC5Li   = WilsonLoops<PeriodicGimplR>::TopologicalCharge5Li(Uflow);
   RealD WFlow_T0   = WF.energyDensityPlaquette(t,Uflow); // t
   RealD WFlow_EC   = WF.energyDensityCloverleaf(t,Uflow);
-  std::cout << GridLogMessage << "Plaquette          "<< conf << "   " << WFlow_plaq << std::endl;
-  std::cout << GridLogMessage << "T0                 "<< conf << "   " << WFlow_T0 << std::endl;
-  std::cout << GridLogMessage << "TC0                 "<< conf << "   " << WFlow_EC << std::endl;
-  std::cout << GridLogMessage << "TopologicalCharge  "<< conf << "   " << WFlow_TC   << std::endl;
+  std::cout << GridLogMessage << "Plaquette            "<< conf << "   " << WFlow_plaq << std::endl;
+  std::cout << GridLogMessage << "T0                   "<< conf << "   " << WFlow_T0 << std::endl;
+  std::cout << GridLogMessage << "TC0                  "<< conf << "   " << WFlow_EC << std::endl;
+  std::cout << GridLogMessage << "TopologicalCharge    "<< conf << "   " << WFlow_TC   << std::endl;
+  std::cout << GridLogMessage << "TopologicalCharge5Li "<< conf << "   " << WFlow_TC5Li<< std::endl;
 
   std::cout<< GridLogMessage << " Admissibility check:\n";
   const double sp_adm = 0.067;                // admissible threshold

HMC/FTHMC2p1f.cc

@@ -25,13 +25,20 @@ directory
 *************************************************************************************/
 /*  END LEGAL */
 #include <Grid/Grid.h>
+
+#if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
 #include <Grid/qcd/smearing/JacobianAction.h>
+#endif
 
 using namespace Grid;
 
 int main(int argc, char **argv)
 {
+#if Nc != 3
+#warning FTHMC2p1f will not work for Nc != 3
+  std::cout << "This program will currently only work for Nc == 3." << std::endl;
+#else
   std::cout << std::setprecision(12);
   
   Grid_init(&argc, &argv);
@@ -220,7 +227,6 @@ int main(int argc, char **argv)
   TheHMC.Run(SmearingPolicy); // for smearing
 
   Grid_finalize();
+#endif
 } // main
 
-
-

HMC/FTHMC2p1f_3GeV.cc

@@ -24,14 +24,22 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
+
 #include <Grid/Grid.h>
+
+#if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
 #include <Grid/qcd/smearing/JacobianAction.h>
+#endif
 
 using namespace Grid;
 
 int main(int argc, char **argv)
 {
+#if Nc != 3
+#warning FTHMC2p1f_3GeV will not work for Nc != 3
+  std::cout << "This program will currently only work for Nc == 3." << std::endl;
+#else
   std::cout << std::setprecision(12);
   
   Grid_init(&argc, &argv);
@@ -220,6 +228,7 @@ int main(int argc, char **argv)
   TheHMC.Run(SmearingPolicy); // for smearing
 
   Grid_finalize();
+#endif
 } // main
 
 

HMC/HMC2p1f_3GeV.cc

@@ -25,13 +25,20 @@ directory
 *************************************************************************************/
 /*  END LEGAL */
 #include <Grid/Grid.h>
+
+#if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
 #include <Grid/qcd/smearing/JacobianAction.h>
+#endif
 
 using namespace Grid;
 
 int main(int argc, char **argv)
 {
+#if Nc != 3
+#warning HMC2p1f_3GeV will not work for Nc != 3
+  std::cout << "This program will currently only work for Nc == 3." << std::endl;
+#else
   std::cout << std::setprecision(12);
   
   Grid_init(&argc, &argv);
@@ -220,6 +227,7 @@ int main(int argc, char **argv)
   TheHMC.Run(SmearingPolicy); // for smearing
 
   Grid_finalize();
+#endif
 } // main
 
 

benchmarks/Benchmark_dwf_fp32.cc

@@ -52,7 +52,7 @@ int main (int argc, char ** argv)
 
   int threads = GridThread::GetThreads();
 
-  int Ls=8;
+  int Ls=16;
   for(int i=0;i<argc;i++) {
     if(std::string(argv[i]) == "-Ls"){
       std::stringstream ss(argv[i+1]); ss >> Ls;

benchmarks/Benchmark_usqcd.cc

@@ -492,17 +492,18 @@ public:
 	}
 	FGrid->Barrier();
 	double t1=usecond();
-	uint64_t ncall = 500;
-
-	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+	uint64_t no    = 50;
+	uint64_t ni    = 100;
 
 	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
 
 	time_statistics timestat;
-	std::vector<double> t_time(ncall);
-	for(uint64_t i=0;i<ncall;i++){
+	std::vector<double> t_time(no);
+	for(uint64_t i=0;i<no;i++){
 	  t0=usecond();
-	  Dw.DhopEO(src_o,r_e,DaggerNo);
+	  for(uint64_t j=0;j<ni;j++){
+	    Dw.DhopEO(src_o,r_e,DaggerNo);
+	  }
 	  t1=usecond();
 	  t_time[i] = t1-t0;
 	}
@@ -520,11 +521,11 @@ public:
 	double mf_hi, mf_lo, mf_err;
 
 	timestat.statistics(t_time);
-	mf_hi = flops/timestat.min;
-	mf_lo = flops/timestat.max;
+	mf_hi = flops/timestat.min*ni;
+	mf_lo = flops/timestat.max*ni;
 	mf_err= flops/timestat.min * timestat.err/timestat.mean;
 
-	mflops = flops/timestat.mean;
+	mflops = flops/timestat.mean*ni;
 	mflops_all.push_back(mflops);
 	if ( mflops_best == 0   ) mflops_best = mflops;
 	if ( mflops_worst== 0   ) mflops_worst= mflops;
@@ -535,6 +536,7 @@ public:
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank   "<< mflops/NP<<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node   "<< mflops/NN<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo us per call   "<< timestat.mean/ni<<std::endl;
 
       }
 
@@ -654,17 +656,19 @@ public:
 	}
 	FGrid->Barrier();
 	double t1=usecond();
-	uint64_t ncall = 500;
 
-	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+	uint64_t no    = 50;
+	uint64_t ni    = 100;
 
 	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
 
 	time_statistics timestat;
-	std::vector<double> t_time(ncall);
-	for(uint64_t i=0;i<ncall;i++){
+	std::vector<double> t_time(no);
+	for(uint64_t i=0;i<no;i++){
 	  t0=usecond();
-	  Ds.DhopEO(src_o,r_e,DaggerNo);
+	  for(uint64_t j=0;j<ni;j++){
+	    Ds.DhopEO(src_o,r_e,DaggerNo);
+	  }
 	  t1=usecond();
 	  t_time[i] = t1-t0;
 	}
@@ -675,11 +679,11 @@ public:
 	double mf_hi, mf_lo, mf_err;
 	
 	timestat.statistics(t_time);
-	mf_hi = flops/timestat.min;
-	mf_lo = flops/timestat.max;
+	mf_hi = flops/timestat.min*ni;
+	mf_lo = flops/timestat.max*ni;
 	mf_err= flops/timestat.min * timestat.err/timestat.mean;
 
-	mflops = flops/timestat.mean;
+	mflops = flops/timestat.mean*ni;
 	mflops_all.push_back(mflops);
 	if ( mflops_best == 0   ) mflops_best = mflops;
 	if ( mflops_worst== 0   ) mflops_worst= mflops;
@@ -689,6 +693,7 @@ public:
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank   "<< mflops/NP<<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node   "<< mflops/NN<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo us per call   "<< timestat.mean/ni<<std::endl;
       
       }
 
@@ -792,19 +797,18 @@ public:
 	  Dc.M(src,r);
 	}
 	FGrid->Barrier();
-	double t1=usecond();
-	uint64_t ncall = 500;
-
-	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+	uint64_t ni = 100;
+	uint64_t no = 50;
 
 	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
-
 	time_statistics timestat;
-	std::vector<double> t_time(ncall);
-	for(uint64_t i=0;i<ncall;i++){
-	  t0=usecond();
-	  Dc.M(src,r);
-	  t1=usecond();
+	std::vector<double> t_time(no);
+	for(uint64_t i=0;i<no;i++){
+	  double t0=usecond();
+	  for(uint64_t j=0;j<ni;j++){
+	    Dc.M(src,r);
+	  }
+	  double t1=usecond();
 	  t_time[i] = t1-t0;
 	}
 	FGrid->Barrier();
@@ -814,20 +818,21 @@ public:
 	double mf_hi, mf_lo, mf_err;
 	
 	timestat.statistics(t_time);
-	mf_hi = flops/timestat.min;
-	mf_lo = flops/timestat.max;
+	mf_hi = flops/timestat.min*ni;
+	mf_lo = flops/timestat.max*ni;
 	mf_err= flops/timestat.min * timestat.err/timestat.mean;
 
-	mflops = flops/timestat.mean;
+	mflops = flops/timestat.mean*ni;
 	mflops_all.push_back(mflops);
 	if ( mflops_best == 0   ) mflops_best = mflops;
 	if ( mflops_worst== 0   ) mflops_worst= mflops;
 	if ( mflops>mflops_best ) mflops_best = mflops;
 	if ( mflops<mflops_worst) mflops_worst= mflops;
 	
-	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<" "<<timestat.mean<<" us"<<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s per rank   "<< mflops/NP<<std::endl;
 	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s per node   "<< mflops/NN<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov us per call   "<< timestat.mean/ni<<std::endl;
       
       }
 
@@ -872,7 +877,7 @@ int main (int argc, char ** argv)
   int do_dslash=1;
 
   int sel=4;
-  std::vector<int> L_list({8,12,16,24});
+  std::vector<int> L_list({8,12,16,24,32});
   int selm1=sel-1;
 
   std::vector<double> clover;

configure.ac

@@ -151,7 +151,7 @@ AC_ARG_ENABLE([tracing],
 case ${ac_TRACING} in
     nvtx)
         AC_DEFINE([GRID_TRACING_NVTX],[1],[use NVTX])
-	LIBS="${LIBS} -lnvToolsExt64_1"
+	LIBS="${LIBS} -lnvToolsExt"
 	;;
     roctx)
         AC_DEFINE([GRID_TRACING_ROCTX],[1],[use ROCTX])

examples/Example_Laplacian_smearing.cc

@@ -93,10 +93,13 @@ int main(int argc, char ** argv)
   Real coeff = (width*width) / Real(4*Iterations);
 
   chi=kronecker;
+
   //  chi = (1-p^2/2N)^N kronecker
   for(int n = 0; n < Iterations; ++n) {
     Laplacian.M(chi,psi);
     chi = chi - coeff*psi;
+    RealD n2 = norm2(chi);
+    chi = chi * (1.0/std::sqrt(n2));
   }
 
   std::cout << " Wuppertal smeared operator is chi = \n" << chi <<std::endl;

systems/Aurora/benchmarks/bench16.pbs

@@ -32,15 +32,9 @@ export MPICH_OFI_NIC_POLICY=GPU
 # Local vol 16.16.16.32
 #
 
-VOL 128.64.128.96
-MPI 4.4.4.3
-NPROC 192
-mpiexec -np 192 -ppn 12 -envall ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 4.4.4.3 --grid 128.64.128.96 --shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap
-
-
-LX=32
+LX=16
 LY=16
-LZ=32
+LZ=16
 LT=32
 
 NX=2

systems/Aurora/benchmarks/gpu_tile.sh

@@ -19,7 +19,7 @@ export ONEAPI_DEVICE_FILTER=gpu,level_zero
 
 export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=0
 export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
-export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:3
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:4
 export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
 #export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:2
 #export SYCL_PI_LEVEL_ZERO_USM_RESIDENT=1

systems/Aurora/config-command

@@ -1,18 +1,19 @@
 #Ahead of time compile for PVC
 
-export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs -device -Xs pvc -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl -lnuma -L/opt/aurora/24.180.3/spack/unified/0.8.0/install/linux-sles15-x86_64/oneapi-2024.07.30.002/numactl-2.0.14-7v6edad/lib" 
-export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions -I/opt/aurora/24.180.3/spack/unified/0.8.0/install/linux-sles15-x86_64/oneapi-2024.07.30.002/numactl-2.0.14-7v6edad/include/"
+export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs -device -Xs pvc -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl -lnuma -L/opt/aurora/24.180.3/spack/unified/0.8.0/install/linux-sles15-x86_64/oneapi-2024.07.30.002/numactl-2.0.14-7v6edad/lib -fPIC -fsycl-max-parallel-link-jobs=16 -fno-sycl-rdc" 
+export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions -I/opt/aurora/24.180.3/spack/unified/0.8.0/install/linux-sles15-x86_64/oneapi-2024.07.30.002/numactl-2.0.14-7v6edad/include/ -fPIC"
 
 #JIT compile 
 #export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel  -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl " 
 #export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions "
 
-../../configure \
+../configure \
 	--enable-simd=GPU \
 	--enable-reduction=grid \
 	--enable-gen-simd-width=64 \
 	--enable-comms=mpi-auto \
 	--enable-debug \
+	--prefix $HOME/gpt-install \
 	--disable-gparity \
 	--disable-fermion-reps \
 	--with-lime=$CLIME \

systems/Frontier-rocm631/config-command

@@ -0,0 +1,22 @@
+CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
+../../configure --enable-comms=mpi-auto \
+--with-lime=$CLIME \
+--enable-unified=no \
+--enable-shm=nvlink \
+--enable-tracing=none \
+--enable-accelerator=hip \
+--enable-gen-simd-width=64 \
+--disable-gparity \
+--disable-fermion-reps \
+--enable-simd=GPU \
+--with-gmp=$OLCF_GMP_ROOT \
+--with-fftw=$FFTW_DIR/.. \
+--with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
+--disable-fermion-reps \
+CXX=hipcc MPICXX=mpicxx \
+CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 " \
+ LDFLAGS="-L/lib64 -L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lhipblas -lrocblas"
+
+
+
+

systems/Frontier-rocm631/sourceme631.sh

@@ -0,0 +1,16 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/Crusher/Grid/spack/share/spack/setup-env.sh
+
+#module load cce/15.0.1
+
+module load rocm/6.3.1
+module load cray-fftw
+module load craype-accel-amd-gfx90a
+export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
+
+#Ugly hacks to get down level software working on current system
+#export LD_LIBRARY_PATH=/opt/cray/libfabric/1.20.1/lib64/:$LD_LIBRARY_PATH
+#export LD_LIBRARY_PATH=`pwd`/:$LD_LIBRARY_PATH
+#ln -s /opt/rocm-6.0.0/lib/libamdhip64.so.6 .
+

systems/Frontier/benchmarks/bench2.slurm

@@ -30,14 +30,10 @@ source ${root}/sourceme.sh
 
 export OMP_NUM_THREADS=7
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-
-for vol in 32.32.32.64
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#64.64.32.96
+for vol in 64.64.32.64
 do
-srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-overlap --shm 2048 --shm-mpi 0 --grid $vol  > log.shm0.ov.$vol
-srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-overlap --shm 2048 --shm-mpi 1 --grid $vol  > log.shm1.ov.$vol
-
-srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-sequential --shm 2048 --shm-mpi 0 --grid $vol  > log.shm0.seq.$vol
-srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-sequential --shm 2048 --shm-mpi 1 --grid $vol > log.shm1.seq.$vol
+srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-overlap --shm 2048 --shm-mpi 0 --grid $vol -Ls 16
 done
 

systems/Frontier/config-command

@@ -3,20 +3,19 @@ CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
 --with-lime=$CLIME \
 --enable-unified=no \
 --enable-shm=nvlink \
---enable-tracing=timer \
+--enable-tracing=none \
 --enable-accelerator=hip \
 --enable-gen-simd-width=64 \
 --disable-gparity \
 --disable-fermion-reps \
 --enable-simd=GPU \
---enable-accelerator-cshift \
 --with-gmp=$OLCF_GMP_ROOT \
 --with-fftw=$FFTW_DIR/.. \
 --with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 --disable-fermion-reps \
 CXX=hipcc MPICXX=mpicxx \
-CXXFLAGS="-fPIC -I{$ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 " \
- LDFLAGS="-L/lib64 -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64 -lhipblas -lrocblas"
+CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 " \
+ LDFLAGS="-L/lib64 -L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lhipblas -lrocblas"
 
 
 

systems/Frontier/sourceme.sh

@@ -1,12 +1,25 @@
+
+echo spack
 . /autofs/nccs-svm1_home1/paboyle/Crusher/Grid/spack/share/spack/setup-env.sh
-spack load c-lime
-module load emacs 
-module load PrgEnv-gnu
-module load rocm/6.0.0
-module load cray-mpich
-module load gmp
+
+module load cce/15.0.1
+module load rocm/5.3.0
 module load cray-fftw
 module load craype-accel-amd-gfx90a
+
+#Ugly hacks to get down level software working on current system
+export LD_LIBRARY_PATH=/opt/cray/libfabric/1.20.1/lib64/:$LD_LIBRARY_PATH
 export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=`pwd`/:$LD_LIBRARY_PATH
+ln -s /opt/rocm-6.0.0/lib/libamdhip64.so.6 .
+
+#echo spack load c-lime
+#spack load c-lime
+#module load emacs 
+##module load PrgEnv-gnu
+##module load cray-mpich
+##module load cray-fftw
+##module load craype-accel-amd-gfx90a
+##export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
 #Hack for lib
-#export LD_LIBRARY_PATH=`pwd`:$LD_LIBRARY_PATH
+##export LD_LIBRARY_PATH=`pwd`/:$LD_LIBRARY_PATH

systems/Jupiter/benchmarks/dwf.1node.perf

@@ -0,0 +1,273 @@
+RANK 1 using NUMA 1 GPU 1 NIC mlx5_1:1
+RANK 3 using NUMA 3 GPU 3 NIC mlx5_3:1
+RANK 0 using NUMA 0 GPU 0 NIC mlx5_0:1
+RANK 2 using NUMA 2 GPU 2 NIC mlx5_2:1
+SLURM detected
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device Number    : 0
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device identifier: NVIDIA GH200 120GB
+AcceleratorCudaInit[0]:   totalGlobalMem: 102005473280 
+AcceleratorCudaInit[0]:   managedMemory: 1 
+AcceleratorCudaInit[0]:   isMultiGpuBoard: 0 
+AcceleratorCudaInit[0]:   warpSize: 32 
+AcceleratorCudaInit[0]:   pciBusID: 1 
+AcceleratorCudaInit[0]:   pciDeviceID: 0 
+AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+local rank 0 device 0 bus id: 0009:01:00.0
+AcceleratorCudaInit: ================================================
+SharedMemoryMpi:  World communicator of size 4
+SharedMemoryMpi:  Node  communicator of size 4
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x4002c0000000 - 40033fffffff for comms buffers 
+Setting up IPC
+
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|_ |  |  |  |  |  |  |  |  |  |  |  | _|__
+__|_                                    _|__
+__|_   GGGG    RRRR    III    DDDD      _|__
+__|_  G        R   R    I     D   D     _|__
+__|_  G        R   R    I     D    D    _|__
+__|_  G  GG    RRRR     I     D    D    _|__
+__|_  G   G    R  R     I     D   D     _|__
+__|_   GGGG    R   R   III    DDDD      _|__
+__|_                                    _|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+  |  |  |  |  |  |  |  |  |  |  |  |  |  |  
+
+
+Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+Current Grid git commit hash=3737a24096282ea179607fc879814710860a0de6: (HEAD -> develop, origin/develop, origin/HEAD) clean
+
+Grid : Message : ================================================ 
+Grid : Message : MPI is initialised and logging filters activated 
+Grid : Message : ================================================ 
+Grid : Message : This rank is running on host jpbo-119-30.jupiter.internal
+Grid : Message : Requested 2147483648 byte stencil comms buffers 
+Grid : Message : MemoryManager Cache 81604378624 bytes 
+Grid : Message : MemoryManager::Init() setting up
+Grid : Message : MemoryManager::Init() cache pool for recent host   allocations: SMALL 8 LARGE 2 HUGE 0
+Grid : Message : MemoryManager::Init() cache pool for recent device allocations: SMALL 16 LARGE 8 Huge 0
+Grid : Message : MemoryManager::Init() cache pool for recent shared allocations: SMALL 16 LARGE 8 Huge 0
+Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
+Grid : Message : MemoryManager::Init() Using cudaMalloc
+
+
+
+
+
+
+
+Grid : Message : 0.303000 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 0.309000 s :  Testing with full communication 
+Grid : Message : 0.312000 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 0.313000 s : Grid Layout
+Grid : Message : 0.313000 s : 	Global lattice size  : 32 32 64 64 
+Grid : Message : 0.319000 s : 	OpenMP threads       : 4
+Grid : Message : 0.320000 s : 	MPI tasks            : 1 1 2 2 
+Grid : Message : 0.129590 s : Initialising 4d RNG
+Grid : Message : 0.764790 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 0.764920 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 0.942440 s : Initialising 5d RNG
+Grid : Message : 1.149388 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 1.149404 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+local rank 1 device 0 bus id: 0019:01:00.0
+local rank 2 device 0 bus id: 0029:01:00.0
+local rank 3 device 0 bus id: 0039:01:00.0
+Grid : Message : 43.893114 s : Drawing gauge field
+Grid : Message : 54.574150 s : Random gauge initialised 
+Grid : Message : 54.574170 s : Applying BCs for Dirichlet Block5 [0 0 0 0 0]
+Grid : Message : 54.574172 s : Applying BCs for Dirichlet Block4 [0 0 0 0]
+Grid : Message : 54.580032 s : Setting up Cshift based reference 
+Grid : Message : 60.407451 s : *****************************************************************
+Grid : Message : 60.407469 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 60.407470 s : *****************************************************************
+Grid : Message : 60.407471 s : *****************************************************************
+Grid : Message : 60.407472 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 60.407473 s : * Vectorising space-time by 8
+Grid : Message : 60.407475 s : * VComplex size is 64 B
+Grid : Message : 60.407477 s : * Using Overlapped Comms/Compute
+Grid : Message : 60.407479 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 60.407480 s : *****************************************************************
+Grid : Message : 61.102178 s : Called warmup
+Grid : Message : 62.177160 s : Called Dw 300 times in 1074958 us
+Grid : Message : 62.177198 s : mflop/s =   24721998.6
+Grid : Message : 62.177201 s : mflop/s per rank =  6180499.64
+Grid : Message : 62.177204 s : mflop/s per node =  24721998.6
+Grid : Message : 62.182696 s : norm diff   5.8108784e-14  Line 306
+Grid : Message : 71.328862 s : ----------------------------------------------------------------
+Grid : Message : 71.328884 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 71.328885 s : ----------------------------------------------------------------
+Grid : Message : 71.328886 s : Called DwDag
+Grid : Message : 71.328887 s : norm dag result 4.12810493
+Grid : Message : 71.329493 s : norm dag ref    4.12810493
+Grid : Message : 71.331967 s : norm dag diff   3.40632318e-14  Line 377
+Grid : Message : 71.394727 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 71.803650 s : src_e0.500003185
+Grid : Message : 71.819727 s : src_o0.499996882
+Grid : Message : 71.821991 s : *********************************************************
+Grid : Message : 71.821993 s : * Benchmarking DomainWallFermion::DhopEO                
+Grid : Message : 71.821995 s : * Vectorising space-time by 8
+Grid : Message : 71.821998 s : * Using Overlapped Comms/Compute
+Grid : Message : 71.822002 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 71.822003 s : *********************************************************
+Grid : Message : 72.377054 s : Deo mflop/s =   24065467
+Grid : Message : 72.377071 s : Deo mflop/s per rank   6016366.75
+Grid : Message : 72.377074 s : Deo mflop/s per node   24065467
+Grid : Message : 72.624877 s : r_e2.06377678
+Grid : Message : 72.625198 s : r_o2.06381058
+Grid : Message : 72.625507 s : res4.12758736
+Grid : Message : 73.759140 s : norm diff   0
+Grid : Message : 73.868204 s : norm diff even  0
+Grid : Message : 73.907201 s : norm diff odd   0
+
+
+
+
+
+
+
+Grid : Message : 74.414580 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 74.414582 s :  Testing without internode communication 
+Grid : Message : 74.414584 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 74.414586 s : Grid Layout
+Grid : Message : 74.414586 s : 	Global lattice size  : 32 32 64 64 
+Grid : Message : 74.414594 s : 	OpenMP threads       : 4
+Grid : Message : 74.414595 s : 	MPI tasks            : 1 1 2 2 
+Grid : Message : 74.679364 s : Initialising 4d RNG
+Grid : Message : 74.742332 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 74.742343 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 74.759525 s : Initialising 5d RNG
+Grid : Message : 75.812412 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 75.812429 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 119.252016 s : Drawing gauge field
+Grid : Message : 129.919846 s : Random gauge initialised 
+Grid : Message : 129.919863 s : Applying BCs for Dirichlet Block5 [0 0 0 0 0]
+Grid : Message : 129.919865 s : Applying BCs for Dirichlet Block4 [0 0 0 0]
+Grid : Message : 129.923611 s : Setting up Cshift based reference 
+Grid : Message : 135.522878 s : *****************************************************************
+Grid : Message : 135.522897 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 135.522899 s : *****************************************************************
+Grid : Message : 135.522899 s : *****************************************************************
+Grid : Message : 135.522900 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 135.522901 s : * Vectorising space-time by 8
+Grid : Message : 135.522903 s : * VComplex size is 64 B
+Grid : Message : 135.522905 s : * Using Overlapped Comms/Compute
+Grid : Message : 135.522907 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 135.522908 s : *****************************************************************
+Grid : Message : 136.151202 s : Called warmup
+Grid : Message : 137.224721 s : Called Dw 300 times in 1073490 us
+Grid : Message : 137.224748 s : mflop/s =   24755806
+Grid : Message : 137.224751 s : mflop/s per rank =  6188951.49
+Grid : Message : 137.224753 s : mflop/s per node =  24755806
+Grid : Message : 137.235239 s : norm diff   5.8108784e-14  Line 306
+Grid : Message : 146.451686 s : ----------------------------------------------------------------
+Grid : Message : 146.451708 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 146.451710 s : ----------------------------------------------------------------
+Grid : Message : 146.451712 s : Called DwDag
+Grid : Message : 146.451714 s : norm dag result 4.12810493
+Grid : Message : 146.452323 s : norm dag ref    4.12810493
+Grid : Message : 146.454799 s : norm dag diff   3.40632318e-14  Line 377
+Grid : Message : 146.498557 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 146.940894 s : src_e0.500003185
+Grid : Message : 146.953676 s : src_o0.499996882
+Grid : Message : 146.955927 s : *********************************************************
+Grid : Message : 146.955929 s : * Benchmarking DomainWallFermion::DhopEO                
+Grid : Message : 146.955932 s : * Vectorising space-time by 8
+Grid : Message : 146.955936 s : * Using Overlapped Comms/Compute
+Grid : Message : 146.955938 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 146.955941 s : *********************************************************
+Grid : Message : 147.511975 s : Deo mflop/s =   24036256.5
+Grid : Message : 147.511989 s : Deo mflop/s per rank   6009064.13
+Grid : Message : 147.511991 s : Deo mflop/s per node   24036256.5
+Grid : Message : 147.522100 s : r_e2.06377678
+Grid : Message : 147.522433 s : r_o2.06381058
+Grid : Message : 147.522745 s : res4.12758736
+Grid : Message : 148.229848 s : norm diff   0
+Grid : Message : 149.233474 s : norm diff even  0
+Grid : Message : 149.235815 s : norm diff odd   0
+
+
+
+
+
+
+
+Grid : Message : 149.960985 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 149.960990 s :  Testing without intranode communication 
+Grid : Message : 149.960991 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 149.960995 s : Grid Layout
+Grid : Message : 149.960995 s : 	Global lattice size  : 32 32 64 64 
+Grid : Message : 149.961003 s : 	OpenMP threads       : 4
+Grid : Message : 149.961004 s : 	MPI tasks            : 1 1 2 2 
+Grid : Message : 150.155810 s : Initialising 4d RNG
+Grid : Message : 150.800200 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 150.800340 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 150.973420 s : Initialising 5d RNG
+Grid : Message : 151.131117 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 151.131136 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 193.933765 s : Drawing gauge field
+Grid : Message : 204.611551 s : Random gauge initialised 
+Grid : Message : 204.611574 s : Applying BCs for Dirichlet Block5 [0 0 0 0 0]
+Grid : Message : 204.611576 s : Applying BCs for Dirichlet Block4 [0 0 0 0]
+Grid : Message : 204.615265 s : Setting up Cshift based reference 
+Grid : Message : 210.117788 s : *****************************************************************
+Grid : Message : 210.117807 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 210.117809 s : *****************************************************************
+Grid : Message : 210.117810 s : *****************************************************************
+Grid : Message : 210.117812 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 210.117813 s : * Vectorising space-time by 8
+Grid : Message : 210.117814 s : * VComplex size is 64 B
+Grid : Message : 210.117817 s : * Using Overlapped Comms/Compute
+Grid : Message : 210.117818 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 210.117819 s : *****************************************************************
+Grid : Message : 210.714641 s : Called warmup
+Grid : Message : 211.892227 s : Called Dw 300 times in 1177557 us
+Grid : Message : 211.892252 s : mflop/s =   22568003.2
+Grid : Message : 211.892255 s : mflop/s per rank =  5642000.8
+Grid : Message : 211.892257 s : mflop/s per node =  22568003.2
+Grid : Message : 211.896037 s : norm diff   5.8108784e-14  Line 306
+Grid : Message : 220.751375 s : ----------------------------------------------------------------
+Grid : Message : 220.751406 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 220.751409 s : ----------------------------------------------------------------
+Grid : Message : 220.751411 s : Called DwDag
+Grid : Message : 220.751412 s : norm dag result 4.12810493
+Grid : Message : 220.753307 s : norm dag ref    4.12810493
+Grid : Message : 220.755796 s : norm dag diff   3.40632318e-14  Line 377
+Grid : Message : 220.813226 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 221.697800 s : src_e0.500003185
+Grid : Message : 221.890920 s : src_o0.499996882
+Grid : Message : 221.913430 s : *********************************************************
+Grid : Message : 221.913450 s : * Benchmarking DomainWallFermion::DhopEO                
+Grid : Message : 221.913480 s : * Vectorising space-time by 8
+Grid : Message : 221.913500 s : * Using Overlapped Comms/Compute
+Grid : Message : 221.913530 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 221.913550 s : *********************************************************
+Grid : Message : 221.645213 s : Deo mflop/s =   24114032
+Grid : Message : 221.645228 s : Deo mflop/s per rank   6028508.01
+Grid : Message : 221.645231 s : Deo mflop/s per node   24114032
+Grid : Message : 221.656021 s : r_e2.06377678
+Grid : Message : 221.656389 s : r_o2.06381058
+Grid : Message : 221.656698 s : res4.12758736
+Grid : Message : 222.110075 s : norm diff   0
+Grid : Message : 222.857692 s : norm diff even  0
+Grid : Message : 222.875763 s : norm diff odd   0
+Grid : Message : 223.598127 s : *******************************************
+Grid : Message : 223.598145 s : ******* Grid Finalize                ******
+Grid : Message : 223.598146 s : *******************************************

systems/Jupiter/benchmarks/dwf.4node.perf

@@ -0,0 +1,286 @@
+RANK 2 using NUMA 2 GPU 2 NIC mlx5_2:1
+RANK 3 using NUMA 3 GPU 3 NIC mlx5_3:1
+RANK 0 using NUMA 0 GPU 0 NIC mlx5_0:1
+RANK 1 using NUMA 1 GPU 1 NIC mlx5_1:1
+RANK 0 using NUMA 0 GPU 0 NIC mlx5_0:1
+RANK 2 using NUMA 2 GPU 2 NIC mlx5_2:1
+RANK 1 using NUMA 1 GPU 1 NIC mlx5_1:1
+RANK 3 using NUMA 3 GPU 3 NIC mlx5_3:1
+RANK 3 using NUMA 3 GPU 3 NIC mlx5_3:1
+RANK 0 using NUMA 0 GPU 0 NIC mlx5_0:1
+RANK 1 using NUMA 1 GPU 1 NIC mlx5_1:1
+RANK 2 using NUMA 2 GPU 2 NIC mlx5_2:1
+RANK 1 using NUMA 1 GPU 1 NIC mlx5_1:1
+RANK 3 using NUMA 3 GPU 3 NIC mlx5_3:1
+RANK 0 using NUMA 0 GPU 0 NIC mlx5_0:1
+RANK 2 using NUMA 2 GPU 2 NIC mlx5_2:1
+SLURM detected
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device Number    : 0
+AcceleratorCudaInit[0]: ========================
+AcceleratorCudaInit[0]: Device identifier: NVIDIA GH200 120GB
+AcceleratorCudaInit[0]:   totalGlobalMem: 102005473280 
+AcceleratorCudaInit[0]:   managedMemory: 1 
+AcceleratorCudaInit[0]:   isMultiGpuBoard: 0 
+AcceleratorCudaInit[0]:   warpSize: 32 
+AcceleratorCudaInit[0]:   pciBusID: 1 
+AcceleratorCudaInit[0]:   pciDeviceID: 0 
+AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+local rank 0 device 0 bus id: 0009:01:00.0
+AcceleratorCudaInit: ================================================
+SharedMemoryMpi:  World communicator of size 16
+SharedMemoryMpi:  Node  communicator of size 4
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x4002a0000000 - 40031fffffff for comms buffers 
+Setting up IPC
+
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|_ |  |  |  |  |  |  |  |  |  |  |  | _|__
+__|_                                    _|__
+__|_   GGGG    RRRR    III    DDDD      _|__
+__|_  G        R   R    I     D   D     _|__
+__|_  G        R   R    I     D    D    _|__
+__|_  G  GG    RRRR     I     D    D    _|__
+__|_  G   G    R  R     I     D   D     _|__
+__|_   GGGG    R   R   III    DDDD      _|__
+__|_                                    _|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
+  |  |  |  |  |  |  |  |  |  |  |  |  |  |  
+
+
+Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+Current Grid git commit hash=3737a24096282ea179607fc879814710860a0de6: (HEAD -> develop, origin/develop, origin/HEAD) clean
+
+Grid : Message : ================================================ 
+Grid : Message : MPI is initialised and logging filters activated 
+Grid : Message : ================================================ 
+Grid : Message : This rank is running on host jpbo-012-11.jupiter.internal
+Grid : Message : Requested 2147483648 byte stencil comms buffers 
+Grid : Message : MemoryManager Cache 81604378624 bytes 
+Grid : Message : MemoryManager::Init() setting up
+Grid : Message : MemoryManager::Init() cache pool for recent host   allocations: SMALL 8 LARGE 2 HUGE 0
+Grid : Message : MemoryManager::Init() cache pool for recent device allocations: SMALL 16 LARGE 8 Huge 0
+Grid : Message : MemoryManager::Init() cache pool for recent shared allocations: SMALL 16 LARGE 8 Huge 0
+Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
+Grid : Message : MemoryManager::Init() Using cudaMalloc
+
+
+
+
+
+
+
+Grid : Message : 0.834000 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 0.838000 s :  Testing with full communication 
+Grid : Message : 0.839000 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 0.840000 s : Grid Layout
+Grid : Message : 0.840000 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 0.846000 s : 	OpenMP threads       : 4
+Grid : Message : 0.846000 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 0.165970 s : Initialising 4d RNG
+Grid : Message : 0.787270 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 0.787340 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 0.960410 s : Initialising 5d RNG
+Grid : Message : 1.142344 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 1.142352 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+local rank 2 device 0 bus id: 0029:01:00.0
+local rank 3 device 0 bus id: 0039:01:00.0
+local rank 1 device 0 bus id: 0019:01:00.0
+Grid : Message : 44.657270 s : Drawing gauge field
+Grid : Message : 55.247733 s : Random gauge initialised 
+Grid : Message : 55.247745 s : Applying BCs for Dirichlet Block5 [0 0 0 0 0]
+Grid : Message : 55.247747 s : Applying BCs for Dirichlet Block4 [0 0 0 0]
+Grid : Message : 55.253053 s : Setting up Cshift based reference 
+Grid : Message : 62.191747 s : *****************************************************************
+Grid : Message : 62.191767 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 62.191768 s : *****************************************************************
+Grid : Message : 62.191769 s : *****************************************************************
+Grid : Message : 62.191769 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 62.191769 s : * Vectorising space-time by 8
+Grid : Message : 62.191770 s : * VComplex size is 64 B
+Grid : Message : 62.191771 s : * Using Overlapped Comms/Compute
+Grid : Message : 62.191771 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 62.191772 s : *****************************************************************
+Grid : Message : 62.857568 s : Called warmup
+Grid : Message : 65.581790 s : Called Dw 300 times in 2200540 us
+Grid : Message : 65.582120 s : mflop/s =   48306525
+Grid : Message : 65.582140 s : mflop/s per rank =  3019157.81
+Grid : Message : 65.582150 s : mflop/s per node =  12076631.3
+Grid : Message : 65.637550 s : norm diff   5.80156793e-14  Line 306
+Grid : Message : 75.122153 s : ----------------------------------------------------------------
+Grid : Message : 75.122166 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 75.122167 s : ----------------------------------------------------------------
+Grid : Message : 75.122167 s : Called DwDag
+Grid : Message : 75.122167 s : norm dag result 4.12801829
+Grid : Message : 75.123295 s : norm dag ref    4.12801829
+Grid : Message : 75.125890 s : norm dag diff   3.42093991e-14  Line 377
+Grid : Message : 75.188462 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 75.605683 s : src_e0.500004005
+Grid : Message : 75.617824 s : src_o0.499996067
+Grid : Message : 75.620089 s : *********************************************************
+Grid : Message : 75.620091 s : * Benchmarking DomainWallFermion::DhopEO                
+Grid : Message : 75.620093 s : * Vectorising space-time by 8
+Grid : Message : 75.620094 s : * Using Overlapped Comms/Compute
+Grid : Message : 75.620095 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 75.620096 s : *********************************************************
+Grid : Message : 76.732272 s : Deo mflop/s =   48068252.4
+Grid : Message : 76.732283 s : Deo mflop/s per rank   3004265.77
+Grid : Message : 76.732285 s : Deo mflop/s per node   12017063.1
+Grid : Message : 76.749317 s : r_e2.06443136
+Grid : Message : 76.749652 s : r_o2.06378451
+Grid : Message : 76.749955 s : res4.12821587
+Grid : Message : 77.198827 s : norm diff   0
+Grid : Message : 77.981760 s : norm diff even  0
+Grid : Message : 78.455900 s : norm diff odd   0
+
+
+
+
+
+
+
+Grid : Message : 78.539333 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 78.539337 s :  Testing without internode communication 
+Grid : Message : 78.539338 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 78.539339 s : Grid Layout
+Grid : Message : 78.539339 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 78.539347 s : 	OpenMP threads       : 4
+Grid : Message : 78.539348 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 78.798501 s : Initialising 4d RNG
+Grid : Message : 78.862916 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 78.862925 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 78.879916 s : Initialising 5d RNG
+Grid : Message : 79.941271 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 79.941280 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 124.586264 s : Drawing gauge field
+Grid : Message : 135.338090 s : Random gauge initialised 
+Grid : Message : 135.338102 s : Applying BCs for Dirichlet Block5 [0 0 0 0 0]
+Grid : Message : 135.338103 s : Applying BCs for Dirichlet Block4 [0 0 0 0]
+Grid : Message : 135.341266 s : Setting up Cshift based reference 
+Grid : Message : 142.604280 s : *****************************************************************
+Grid : Message : 142.604450 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 142.604460 s : *****************************************************************
+Grid : Message : 142.604470 s : *****************************************************************
+Grid : Message : 142.604480 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 142.604480 s : * Vectorising space-time by 8
+Grid : Message : 142.604500 s : * VComplex size is 64 B
+Grid : Message : 142.604510 s : * Using Overlapped Comms/Compute
+Grid : Message : 142.604510 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 142.604520 s : *****************************************************************
+Grid : Message : 142.686034 s : Called warmup
+Grid : Message : 144.868543 s : Called Dw 300 times in 2182483 us
+Grid : Message : 144.868559 s : mflop/s =   48706194.1
+Grid : Message : 144.868561 s : mflop/s per rank =  3044137.13
+Grid : Message : 144.868562 s : mflop/s per node =  12176548.5
+Grid : Message : 144.887595 s : norm diff   5.80156793e-14  Line 306
+Grid : Message : 153.622978 s : ----------------------------------------------------------------
+Grid : Message : 153.622994 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 153.622995 s : ----------------------------------------------------------------
+Grid : Message : 153.622995 s : Called DwDag
+Grid : Message : 153.622996 s : norm dag result 4.12801829
+Grid : Message : 153.623604 s : norm dag ref    4.12801829
+Grid : Message : 153.626098 s : norm dag diff   3.42093991e-14  Line 377
+Grid : Message : 153.691426 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 154.148319 s : src_e0.500004005
+Grid : Message : 154.151454 s : src_o0.499996067
+Grid : Message : 154.153722 s : *********************************************************
+Grid : Message : 154.153724 s : * Benchmarking DomainWallFermion::DhopEO                
+Grid : Message : 154.153725 s : * Vectorising space-time by 8
+Grid : Message : 154.153726 s : * Using Overlapped Comms/Compute
+Grid : Message : 154.153727 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 154.153728 s : *********************************************************
+Grid : Message : 155.200671 s : Deo mflop/s =   51121022.4
+Grid : Message : 155.200682 s : Deo mflop/s per rank   3195063.9
+Grid : Message : 155.200684 s : Deo mflop/s per node   12780255.6
+Grid : Message : 155.217204 s : r_e2.06443136
+Grid : Message : 155.217550 s : r_o2.06378451
+Grid : Message : 155.217869 s : res4.12821587
+Grid : Message : 155.673744 s : norm diff   0
+Grid : Message : 156.463329 s : norm diff even  0
+Grid : Message : 156.878866 s : norm diff odd   0
+
+
+
+
+
+
+
+Grid : Message : 157.620761 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 157.620764 s :  Testing without intranode communication 
+Grid : Message : 157.620765 s : ++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 157.620766 s : Grid Layout
+Grid : Message : 157.620766 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 157.620773 s : 	OpenMP threads       : 4
+Grid : Message : 157.620774 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 157.671479 s : Initialising 4d RNG
+Grid : Message : 157.738691 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 157.738698 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 157.755651 s : Initialising 5d RNG
+Grid : Message : 158.848676 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 158.848685 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 202.465158 s : Drawing gauge field
+Grid : Message : 213.214546 s : Random gauge initialised 
+Grid : Message : 213.214561 s : Applying BCs for Dirichlet Block5 [0 0 0 0 0]
+Grid : Message : 213.214563 s : Applying BCs for Dirichlet Block4 [0 0 0 0]
+Grid : Message : 213.217711 s : Setting up Cshift based reference 
+Grid : Message : 219.662772 s : *****************************************************************
+Grid : Message : 219.662786 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 219.662787 s : *****************************************************************
+Grid : Message : 219.662788 s : *****************************************************************
+Grid : Message : 219.662788 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 219.662789 s : * Vectorising space-time by 8
+Grid : Message : 219.662790 s : * VComplex size is 64 B
+Grid : Message : 219.662791 s : * Using Overlapped Comms/Compute
+Grid : Message : 219.662791 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 219.662791 s : *****************************************************************
+Grid : Message : 220.425592 s : Called warmup
+Grid : Message : 222.536249 s : Called Dw 300 times in 2110597 us
+Grid : Message : 222.536267 s : mflop/s =   50365105.5
+Grid : Message : 222.536269 s : mflop/s per rank =  3147819.09
+Grid : Message : 222.536270 s : mflop/s per node =  12591276.4
+Grid : Message : 222.541053 s : norm diff   5.80156793e-14  Line 306
+Grid : Message : 232.135901 s : ----------------------------------------------------------------
+Grid : Message : 232.135915 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 232.135916 s : ----------------------------------------------------------------
+Grid : Message : 232.135917 s : Called DwDag
+Grid : Message : 232.135918 s : norm dag result 4.12801829
+Grid : Message : 232.151938 s : norm dag ref    4.12801829
+Grid : Message : 232.154451 s : norm dag diff   3.42093991e-14  Line 377
+Grid : Message : 232.216117 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 232.630529 s : src_e0.500004005
+Grid : Message : 232.643197 s : src_o0.499996067
+Grid : Message : 232.645527 s : *********************************************************
+Grid : Message : 232.645529 s : * Benchmarking DomainWallFermion::DhopEO                
+Grid : Message : 232.645532 s : * Vectorising space-time by 8
+Grid : Message : 232.645533 s : * Using Overlapped Comms/Compute
+Grid : Message : 232.645534 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 232.645535 s : *********************************************************
+Grid : Message : 233.774184 s : Deo mflop/s =   47432091.9
+Grid : Message : 233.774194 s : Deo mflop/s per rank   2964505.74
+Grid : Message : 233.774196 s : Deo mflop/s per node   11858023
+Grid : Message : 233.791552 s : r_e2.06443136
+Grid : Message : 233.791899 s : r_o2.06378451
+Grid : Message : 233.792204 s : res4.12821587
+Grid : Message : 234.230783 s : norm diff   0
+Grid : Message : 235.162780 s : norm diff even  0
+Grid : Message : 235.291950 s : norm diff odd   0
+Grid : Message : 235.765411 s : *******************************************
+Grid : Message : 235.765424 s : ******* Grid Finalize                ******
+Grid : Message : 235.765425 s : *******************************************
+

systems/Jupiter/benchmarks/dwf1.slurm

@@ -0,0 +1,57 @@
+#!/bin/sh
+#SBATCH --account=jureap14
+#SBATCH --nodes=1
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=4
+#SBATCH --cpus-per-task=64
+#SBATCH --time=2:00:00
+#SBATCH --partition=booster
+#SBATCH --gres=gpu:4
+
+export OMP_NUM_THREADS=4
+export OMPI_MCA_btl=^uct,openib
+export UCX_TLS=gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
+export UCX_RNDV_SCHEME=put_zcopy
+export UCX_RNDV_THRESH=16384
+export UCX_IB_GPU_DIRECT_RDMA=yes
+export UCX_MEMTYPE_CACHE=n
+
+OPT="--comms-overlap"
+
+source ../sourceme.sh
+
+cat << EOF > bind_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=(0 1 2 3)
+export NIC_MAP=(0 1 2 3)
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export NIC=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+export UCX_NET_DEVICES=mlx5_\${NIC}:1
+
+echo RANK \$SLURM_LOCALID using NUMA \$NUMA  GPU \$GPU NIC \$UCX_NET_DEVICES
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./bind_gpu
+
+srun --cpu-bind=no -N 1 -n $SLURM_NTASKS \
+        ./bind_gpu ./Benchmark_dwf_fp32 \
+	$OPT \
+	--mpi 1.1.2.2 \
+	--accelerator-threads 8 \
+	--grid 32.32.64.64 \
+	--shm 2048 > dwf.1node.perf
+
+srun --cpu-bind=no -N 1  -n $SLURM_NTASKS \
+	./bind_gpu ./Benchmark_comms_host_device \
+	--mpi 1.1.2.2 \
+	--accelerator-threads 8 \
+	--grid 32.32.64.64 \
+	--shm 2048 > comms.1node.perf
+
+
+
+

systems/Jupiter/benchmarks/dwf4.slurm

@@ -0,0 +1,57 @@
+#!/bin/sh
+#SBATCH --account=jureap14
+#SBATCH --nodes=4
+#SBATCH --ntasks=16
+#SBATCH --ntasks-per-node=4
+#SBATCH --cpus-per-task=64
+#SBATCH --time=2:00:00
+#SBATCH --partition=booster
+#SBATCH --gres=gpu:4
+
+export OMP_NUM_THREADS=4
+export OMPI_MCA_btl=^uct,openib
+export UCX_TLS=gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
+export UCX_RNDV_SCHEME=put_zcopy
+export UCX_RNDV_THRESH=16384
+export UCX_IB_GPU_DIRECT_RDMA=yes
+export UCX_MEMTYPE_CACHE=n
+
+OPT="--comms-overlap"
+
+source ../sourceme.sh
+
+cat << EOF > bind_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=(0 1 2 3)
+export NIC_MAP=(0 1 2 3)
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export NIC=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+export UCX_NET_DEVICES=mlx5_\${NIC}:1
+
+echo RANK \$SLURM_LOCALID using NUMA \$NUMA  GPU \$GPU NIC \$UCX_NET_DEVICES
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./bind_gpu
+
+srun --cpu-bind=no -N 4 -n $SLURM_NTASKS \
+        ./bind_gpu ./Benchmark_dwf_fp32 \
+	$OPT \
+	--mpi 2.2.2.2 \
+	--accelerator-threads 8 \
+	--grid 64.64.64.64 \
+	--shm 2048 > dwf.4node.perf
+
+srun --cpu-bind=no -N 4  -n $SLURM_NTASKS \
+	./bind_gpu ./Benchmark_comms_host_device \
+	--mpi 2.2.2.2 \
+	--accelerator-threads 8 \
+	--grid 32.32.64.64 \
+	--shm 2048 > comms.4node.perf
+
+
+
+

systems/Jupiter/config-command

@@ -0,0 +1,16 @@
+export CXX=nvcc
+export OPENMPI=/p/software/default/stages/2025/software/OpenMPI/5.0.5-NVHPC-24.9-CUDA-12/
+export LDFLAGS="-cudart shared -L${OPENMPI}/lib" 
+export CXXFLAGS="-ccbin clang++ -gencode arch=compute_90,code=sm_90 -std=c++17 -cudart shared -lcublas -lmpi -I${OPENMPI}/include"
+
+../../configure \
+    --enable-comms=mpi \
+    --enable-simd=GPU \
+    --enable-gen-simd-width=64 \
+    --enable-shm=nvlink \
+    --enable-accelerator=cuda \
+    --with-lime=$CLIME \
+    --disable-gparity \
+    --disable-fermion-reps \
+    --disable-unified 
+

systems/Jupiter/sourceme.sh

@@ -0,0 +1,9 @@
+CLIME=$HOME/install/
+module load Clang
+module load CUDA
+module load FFTW
+module load OpenSSL
+module load MPFR
+module load NVHPC
+module load UCX
+module load OpenMPI

systems/WorkArounds.txt

@@ -0,0 +1,206 @@
+The purpose of this file is to collate all non-obvious known magic shell variables
+and compiler flags required for either correctness or performance on various systems.
+
+A repository of work-arounds.
+
+Contents:
+1. Interconnect + MPI
+2. Compilation
+3. Profiling
+
+************************
+* 1. INTERCONNECT + MPI
+************************
+
+--------------------------------------------------------------------
+MPI2-IO correctness: force OpenMPI to use the MPICH romio implementation for parallel I/O 
+--------------------------------------------------------------------
+export OMPI_MCA_io=romio321
+
+--------------------------------------
+ROMIO fail with > 2GB per node read (32 bit issue)
+--------------------------------------
+
+Use later MPICH
+
+https://github.com/paboyle/Grid/issues/381
+
+https://github.com/pmodels/mpich/commit/3a479ab0
+
+--------------------------------------------------------------------
+Slingshot: Frontier and Perlmutter libfabric slow down 
+and physical memory fragmentation 
+--------------------------------------------------------------------
+export FI_MR_CACHE_MONITOR=disabled
+or
+export FI_MR_CACHE_MONITOR=kdreg2
+
+--------------------------------------------------------------------
+Perlmutter
+--------------------------------------------------------------------
+
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+
+--------------------------------------------------------------------
+Frontier/LumiG
+--------------------------------------------------------------------
+
+Hiding ROCR_VISIBLE_DEVICES triggers SDMA engines to be used for GPU-GPU
+
+cat << EOF > select_gpu
+#!/bin/bash
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export GPU_MAP=(0 1 2 3 7 6 5 4)
+export NUMA_MAP=(3 3 1 1 2 2 0 0)
+export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
+export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
+export HIP_VISIBLE_DEVICES=\$GPU
+unset ROCR_VISIBLE_DEVICES
+echo RANK \$SLURM_LOCALID using GPU \$GPU    
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+chmod +x ./select_gpu
+
+srun ./select_gpu BINARY
+
+
+--------------------------------------------------------------------
+Mellanox performance with A100 GPU (Tursa, Booster, Leonardo)
+--------------------------------------------------------------------
+export OMPI_MCA_btl=^uct,openib
+export UCX_TLS=gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
+export UCX_RNDV_SCHEME=put_zcopy
+export UCX_RNDV_THRESH=16384
+export UCX_IB_GPU_DIRECT_RDMA=yes
+
+--------------------------------------------------------------------
+Mellanox + A100 correctness (Tursa, Booster, Leonardo)
+--------------------------------------------------------------------
+export UCX_MEMTYPE_CACHE=n
+
+--------------------------------------------------------------------
+MPICH/Aurora/PVC correctness and performance 
+--------------------------------------------------------------------
+
+https://github.com/pmodels/mpich/issues/7302
+
+--enable-cuda-aware-mpi=no  
+--enable-unified=no
+
+Grid's internal D-H-H-D pipeline mode, avoid device memory in MPI
+Do not use SVM
+
+Ideally use MPICH with fix to issue 7302:
+
+https://github.com/pmodels/mpich/pull/7312
+
+Ideally:
+MPIR_CVAR_CH4_IPC_GPU_HANDLE_CACHE=generic
+
+Alternatives:
+export MPIR_CVAR_NOLOCAL=1
+export MPIR_CVAR_CH4_IPC_GPU_P2P_THRESHOLD=1000000000
+
+--------------------------------------------------------------------
+MPICH/Aurora/PVC correctness and performance 
+--------------------------------------------------------------------
+
+Broken:
+export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+
+This gives good peformance without requiring 
+--enable-cuda-aware-mpi=no  
+
+But is an open issue reported by James Osborn
+https://github.com/pmodels/mpich/issues/7139
+
+Possibly resolved but unclear if in the installed software yet.
+
+************************
+* 2. COMPILATION
+************************
+
+--------------------------------------------------------------------
+G++ compiler breakage / graveyard
+--------------------------------------------------------------------
+
+9.3.0, 10.3.1, 
+https://github.com/paboyle/Grid/issues/290
+https://github.com/paboyle/Grid/issues/264
+
+Working (-) Broken (X):
+
+4.9.0 -
+4.9.1 -
+5.1.0 X
+5.2.0 X
+5.3.0 X
+5.4.0 X
+6.1.0 X
+6.2.0 X
+6.3.0 -
+7.1.0 -
+8.0.0 (HEAD) -
+
+https://github.com/paboyle/Grid/issues/100
+
+--------------------------------------------------------------------
+AMD GPU nodes :
+--------------------------------------------------------------------
+
+multiple ROCM versions broken; use 5.3.0
+manifests itself as wrong results in fp32 
+
+https://github.com/paboyle/Grid/issues/464
+
+--------------------------------------------------------------------
+Aurora/PVC
+--------------------------------------------------------------------
+
+SYCL ahead of time compilation (fixes rare runtime JIT errors and faster runtime, PB)
+SYCL slow link and relocatable code issues (Christoph Lehner)
+Opt large register file required for good performance in fp64
+
+
+export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
+export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs -device -Xs pvc -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl -fPIC -fsycl-max-parallel-link-jobs=16 -fno-sycl-rdc" 
+export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions -fPIC"
+
+--------------------------------------------------------------------
+Aurora/PVC useful extra options
+--------------------------------------------------------------------
+
+Host only sanitizer:
+-Xarch_host -fsanitize=leak
+-Xarch_host -fsanitize=address
+
+Deterministic MPI reduction:
+export MPIR_CVAR_ALLREDUCE_DEVICE_COLLECTIVE=0
+export MPIR_CVAR_REDUCE_DEVICE_COLLECTIVE=0
+export MPIR_CVAR_ALLREDUCE_INTRA_ALGORITHM=recursive_doubling
+unset MPIR_CVAR_CH4_COLL_SELECTION_TUNING_JSON_FILE
+unset MPIR_CVAR_COLL_SELECTION_TUNING_JSON_FILE
+unset MPIR_CVAR_CH4_POSIX_COLL_SELECTION_TUNING_JSON_FILE
+
+
+
+************************
+* 3. Visual profile tools
+************************
+
+--------------------------------------------------------------------
+Frontier/rocprof
+--------------------------------------------------------------------
+
+--------------------------------------------------------------------
+Aurora/unitrace
+--------------------------------------------------------------------
+
+
+--------------------------------------------------------------------
+Tursa/nsight-sys
+--------------------------------------------------------------------

systems/mac-arm/config-command-mpi

@@ -1,2 +1,14 @@
-CXXFLAGS=-I/opt/local/include LDFLAGS=-L/opt/local/lib/ CXX=c++-13 MPICXX=mpicxx ../../configure --enable-simd=GEN --enable-comms=mpi-auto --enable-unified=yes --prefix $HOME/QCD/GridInstall --with-lime=/Users/peterboyle/QCD/SciDAC/install/ --with-openssl=$BREW --disable-fermion-reps --disable-gparity --disable-debug 
+
+
+CXX=mpicxx ../../configure \
+	   --enable-simd=GEN \
+	   --enable-comms=mpi-auto \
+	   --enable-Sp=yes \
+	   --enable-unified=yes \
+	   --prefix /Users/peterboyle/QCD/vtk/Grid/install \
+	   --with-lime=$CLIME \
+	   --with-openssl=$OPENSSL \
+	   --with-gmp=$GMP \
+	   --with-mpfr=$MPFR \
+	   --disable-debug 
 

systems/sdcc-genoa/bench.slurm

@@ -0,0 +1,32 @@
+#!/bin/bash
+#SBATCH --partition lqcd
+#SBATCH --time=00:50:00
+#SBATCH -A lqcdtest
+#SBATCH -q lqcd
+#SBATCH --exclusive
+#SBATCH --nodes=1
+#SBATCH -w genoahost001,genoahost003,genoahost050,genoahost054
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=64
+#SBATCH --qos lqcd
+
+source sourceme.sh
+
+export PLACES=(1:16:4 1:32:2 0:64:1);
+export THR=(16 32 64)
+
+for t in  2 
+do
+   
+export OMP_NUM_THREADS=${THR[$t]}
+export OMP_PLACES=${PLACES[$t]}
+export thr=${THR[$t]}
+
+#for vol in 24.24.24.24 32.32.32.32 48.48.48.96
+for vol in 48.48.48.96
+do
+srun -N1 -n1 ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid $vol --dslash-asm --shm 8192 > $vol.1node.thr$thr
+done
+#srun -N1 -n1 ./benchmarks/Benchmark_usqcd --mpi 1.1.1.1 --grid $vol > usqcd.1node.thr$thr
+done
+

systems/sdcc-genoa/bench2.slurm

@@ -0,0 +1,36 @@
+#!/bin/bash
+#SBATCH --partition lqcd
+#SBATCH --time=00:50:00
+#SBATCH -A lqcdtest
+#SBATCH -q lqcd
+#SBATCH --exclusive
+#SBATCH --nodes=2
+#SBATCH -w genoahost001,genoahost003,genoahost050,genoahost054
+#SBATCH --ntasks=2
+#SBATCH --cpus-per-task=64
+#SBATCH --qos lqcd
+
+source sourceme.sh
+
+export PLACES=(1:16:4 1:32:2 0:64:1);
+export THR=(16 32 64)
+
+nodes=2
+mpi=1.1.1.2
+
+for t in 2 
+do
+   
+export OMP_NUM_THREADS=${THR[$t]}
+export OMP_PLACES=${PLACES[$t]}
+export thr=${THR[$t]}
+
+#srun -N$nodes -n$nodes ./benchmarks/Benchmark_usqcd --mpi $mpi --grid 32.32.32.32 > usqcd.n$nodes.thr$thr
+
+for vol in 64.64.64.128
+do
+srun -N$nodes -n$nodes ./benchmarks/Benchmark_dwf_fp32 --mpi $mpi --grid $vol --dslash-asm --comms-overlap --shm 8192 > $vol.n$nodes.overlap.thr$thr
+done
+
+done
+

systems/sdcc-genoa/config-command

@@ -0,0 +1,29 @@
+
+spack load c-lime
+spack load fftw
+spack load hdf5+cxx
+
+export FFTW=`spack find --paths fftw       | grep ^fftw   | awk '{print $2}' `
+export HDF5=`spack find --paths hdf5+cxx   | grep ^hdf5   | awk '{print $2}' `
+export CLIME=`spack find --paths c-lime    | grep ^c-lime | awk '{print $2}' `
+
+../../configure \
+--enable-comms=mpi-auto \
+--enable-unified=yes \
+--enable-shm=shmopen \
+--enable-shm-fast-path=shmopen \
+--enable-accelerator=none \
+--enable-simd=AVX512 \
+--with-lime=$CLIME \
+--with-hdf5=$HDF5 \
+--with-fftw=$FFTW \
+--disable-fermion-reps \
+--disable-gparity \
+CXX=clang++ \
+MPICXX=mpicxx \
+LIBS=-llime \
+LDFLAGS=-L$CLIME/lib/ \
+CXXFLAGS="-std=c++17 -fPIE"
+
+
+

systems/sdcc-genoa/sourceme.sh

@@ -0,0 +1,5 @@
+source $HOME/spack/share/spack/setup-env.sh
+spack load llvm@17.0.4
+export LD_LIBRARY_PATH=/direct/sdcc+u/paboyle/spack/opt/spack/linux-almalinux8-icelake/gcc-8.5.0/llvm-17.0.4-laufdrcip63ivkadmtgoepwmj3dtztdu/lib:$LD_LIBRARY_PATH
+module load openmpi/4.1.8
+spack load c-lime

systems/spack-linux/config-command

@@ -0,0 +1,17 @@
+../../src/Grid/configure \
+    --prefix /home/pab/NPR/install \
+    --enable-comms=mpi-auto \
+    --enable-simd=AVX2 \
+    --enable-shm=none \
+    --enable-debug \
+    --with-lime=$CLIME \
+    --with-hdf5=$HDF5 \
+    --with-fftw=$FFTW \
+    --with-gmp=$GMP \
+    --with-mpfr=$MPFR \
+    --disable-gparity \
+    --disable-fermion-reps \
+    CXX=clang++ \
+    MPICXX=mpicxx \
+    CXXFLAGS="-std=c++17 "
+

systems/spack-linux/sourceme.sh

@@ -0,0 +1,28 @@
+source $HOME/spack/share/spack/setup-env.sh
+spack load llvm@12
+spack load autoconf%clang@12.0.1
+spack load automake%clang@12.0.1
+spack load c-lime%clang@12.0.1
+spack load fftw%clang@12.0.1
+spack load gmp%clang@12.0.1
+spack load mpfr%clang@12.0.1
+spack load openmpi%clang@12.0.1
+spack load openssl%clang@12.0.1
+spack load hdf5+cxx%clang@12.0.1
+spack load cmake%clang@12.0.1
+export FFTW=`spack find --paths fftw%clang@12.0.1    | grep ^fftw   | awk '{print $2}' `
+export HDF5=`spack find --paths hdf5+cxx%clang@12.0.1   | grep ^hdf5   | awk '{print $2}' `
+export CLIME=`spack find --paths c-lime%clang@12.0.1 | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr%clang@12.0.1    | grep ^mpfr  | awk '{print $2}' `
+export LLVM=`spack find --paths llvm@12    | grep ^llvm  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl%clang@12.0.1 | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp%clang@12.0.1      | grep ^gmp | awk '{print $2}' `
+export TCLAP=`spack find --paths tclap%clang@12.0.1  | grep ^tclap | awk '{print $2}' `
+export LD_LIBRARY_PATH=${TCLAP}/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$MPFR/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$GMP/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$FFTW/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$LLVM/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$LLVM/lib/x86_64-unknown-linux-gnu/:$LD_LIBRARY_PATH
+
+ulimit -s 81920

systems/spack-linux/spack-install

@@ -0,0 +1,19 @@
+cd
+git clone https://github.com/spack/spack.git
+source $HOME/spack/share/spack/setup-env.sh
+
+spack install llvm@12
+
+spack install autoconf%clang@12.0.1
+spack install automake%clang@12.0.1
+spack install c-lime%clang@12.0.1
+spack install fftw%clang@12.0.1
+spack install gmp%clang@12.0.1
+spack install mpfr%clang@12.0.1
+spack install openmpi%clang@12.0.1
+spack install openssl%clang@12.0.1
+spack install hdf5+cxx%clang@12.0.1
+spack install cmake%clang@12.0.1
+spack install tclap%clang@12.0.1
+spack install emacs%clang@12.0.1
+

tests/Test_dwf_dslash_repro.cc

@@ -62,7 +62,7 @@ int VerifyOnDevice(const FermionField &res, FermionField &ref)
     if (((random()&0xF)==0)&&injection) {
       uint64_t sF = random()%(NN);
       int lane=0;
-      printf("Error injection site %ld on rank %d\n",sF,res.Grid()->ThisRank());
+      printf("Error injection site %ld on rank %d\n",(long)sF,res.Grid()->ThisRank());
       auto vv = acceleratorGet(res_v[sF]);
       double *dd = (double *)&vv;
       *dd=M_PI;

tests/debug/Test_general_coarse_hdcg.cc

@@ -195,8 +195,8 @@ int main (int argc, char ** argv)
 
   int Nk=nrhs;
   int Nm=Nk*3;
-  int Nk=36;
-  int Nm=144;
+  //  int Nk=36;
+  //  int Nm=144;
   int Nstop=Nk;
   int Nconv_test_interval=1;
   

tests/debug/Test_general_coarse_pvdagm.cc

@@ -47,20 +47,20 @@ public:
   void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
   void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
   void Op     (const Field &in, Field &out){
-    std::cout << "Op: PVdag M "<<std::endl;
+    //    std::cout << "Op: PVdag M "<<std::endl;
     Field tmp(in.Grid());
     _Mat.M(in,tmp);
     _PV.Mdag(tmp,out);
   }
   void AdjOp     (const Field &in, Field &out){
-    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    //    std::cout << "AdjOp: Mdag PV "<<std::endl;
     Field tmp(in.Grid());
     _PV.M(in,tmp);
     _Mat.Mdag(tmp,out);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
   void HermOp(const Field &in, Field &out){
-    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    //    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
     Field tmp(in.Grid());
     //    _Mat.M(in,tmp);
     //    _PV.Mdag(tmp,out);
@@ -83,14 +83,14 @@ public:
   void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
   void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
   void Op     (const Field &in, Field &out){
-    std::cout << "Op: PVdag M "<<std::endl;
+    //    std::cout << "Op: PVdag M "<<std::endl;
     Field tmp(in.Grid());
     _Mat.M(in,tmp);
     _PV.Mdag(tmp,out);
     out = out + shift * in;
   }
   void AdjOp     (const Field &in, Field &out){
-    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    //    std::cout << "AdjOp: Mdag PV "<<std::endl;
     Field tmp(in.Grid());
     _PV.M(tmp,out);
     _Mat.Mdag(in,tmp);
@@ -98,7 +98,7 @@ public:
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
   void HermOp(const Field &in, Field &out){
-    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    //    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
     Field tmp(in.Grid());
     Op(in,tmp);
     AdjOp(tmp,out);

tests/debug/Test_heatbath_dwf_eofa_gparity.cc

@@ -54,6 +54,7 @@ const RealD            M5       = 1.8;
 
 int main(int argc, char** argv)
 {
+#ifdef ENABLE_GPARITY
   Grid_init(&argc, &argv);
 
   int threads = GridThread::GetThreads();
@@ -106,6 +107,6 @@ int main(int argc, char** argv)
     Meofa.refresh(Umu,sRNG, RNG5);
     printf("<Phi|Meofa|Phi> = %1.15e\n", Meofa.S(Umu));
   }
-
+#endif
   return 0;
 }

tests/debug/Test_heatbath_mobius_eofa_gparity.cc

@@ -56,6 +56,7 @@ const RealD            M5       = 1.8;
 
 int main(int argc, char** argv)
 {
+#ifdef ENABLE_GPARITY  
   Grid_init(&argc, &argv);
 
   int threads = GridThread::GetThreads();
@@ -106,6 +107,6 @@ int main(int argc, char** argv)
     Meofa.refresh(Umu, sRNG, RNG5);
     printf("<Phi|Meofa|Phi> = %1.15e\n", Meofa.S(Umu));
   }
-
+#endif
   return 0;
 }

tests/debug/Test_padded_cell.cc

@@ -33,6 +33,7 @@ using namespace std;
 using namespace Grid;
 
 // This is to optimize the SIMD
+/*
 template<class vobj> void gpermute(vobj & inout,int perm){
   vobj tmp=inout;
   if (perm & 0x1 ) { permute(inout,tmp,0); tmp=inout;}
@@ -40,7 +41,7 @@ template<class vobj> void gpermute(vobj & inout,int perm){
   if (perm & 0x4 ) { permute(inout,tmp,2); tmp=inout;}
   if (perm & 0x8 ) { permute(inout,tmp,3); tmp=inout;}
 }
-
+*/
 
 int main (int argc, char ** argv)
 {

tests/debug/Test_padded_cell_staple.cc

@@ -153,7 +153,7 @@ public:
     t=usecond();
     {
       autoView( gStaple_v , gStaple, AcceleratorWrite);
-      auto gStencil_v = gStencil.View();
+      auto gStencil_v = gStencil.View(AcceleratorRead);
       autoView( Ug_mu_v , Ug_mu, AcceleratorRead);
       autoView( Ug_nu_v , Ug_nu, AcceleratorRead);
   
@@ -389,7 +389,7 @@ public:
     GeneralLocalStencil gStencil(ggrid,shifts);
     {
       autoView( gStaple_v , gStaple, AcceleratorWrite);
-      auto gStencil_v = gStencil.View();
+      auto gStencil_v = gStencil.View(AcceleratorRead);
 
       typedef LatticeView<typename GaugeMat::vector_object> GaugeViewType;
       size_t vsize = Nd*sizeof(GaugeViewType);

tests/debug/Test_reweight_dwf_eofa_gparity.cc

@@ -83,6 +83,7 @@ std::vector<RealD> jack_stats(const std::vector<RealD>& data)
 
 int main(int argc, char **argv)
 {
+#ifdef ENABLE_GPARITY  
   Grid_init(&argc, &argv);
 
   // Initialize spacetime grid
@@ -206,4 +207,5 @@ int main(int argc, char **argv)
   std::cout << std::endl << "EOFA: rw = " << eofa_result[0] << " +/- " << eofa_result[1] << std::endl;
 
   Grid_finalize();
+#endif
 }

tests/debug/Test_reweight_mobius_eofa_gparity.cc

@@ -85,6 +85,7 @@ std::vector<RealD> jack_stats(const std::vector<RealD>& data)
 
 int main(int argc, char **argv)
 {
+#ifdef ENABLE_GPARITY  
   Grid_init(&argc, &argv);
 
   // Initialize spacetime grid
@@ -215,4 +216,5 @@ int main(int argc, char **argv)
   std::cout << std::endl << "EOFA: rw = " << eofa_result[0] << " +/- " << eofa_result[1] << std::endl;
 
   Grid_finalize();
+#endif
 }

tests/forces/Test_dwf_gpforce.cc

@@ -35,6 +35,7 @@ using namespace Grid;
 
 int main (int argc, char ** argv)
 {
+#ifdef ENABLE_GPARITY  
   Grid_init(&argc,&argv);
 
   Coordinate latt_size   = GridDefaultLatt();
@@ -244,4 +245,5 @@ int main (int argc, char ** argv)
 
   std::cout<< GridLogMessage << "Done" <<std::endl;
   Grid_finalize();
+#endif  
 }

tests/forces/Test_dwf_gpforce_eofa.cc

@@ -38,6 +38,7 @@ typedef typename FermionAction::FermionField FermionField;
 
 int main (int argc, char** argv)
 {
+#ifdef ENABLE_GPARITY  
   Grid_init(&argc, &argv);
 
   Coordinate latt_size   = GridDefaultLatt();
@@ -173,4 +174,5 @@ int main (int argc, char** argv)
 
   std::cout << GridLogMessage << "Done" << std::endl;
   Grid_finalize();
+#endif
 }

tests/forces/Test_gpdwf_force.cc

@@ -35,6 +35,7 @@ using namespace Grid;
 
 int main (int argc, char ** argv)
 {
+#ifdef ENABLE_GPARITY
   Grid_init(&argc,&argv);
 
   Coordinate latt_size   = GridDefaultLatt();
@@ -204,4 +205,5 @@ int main (int argc, char ** argv)
   assert( fabs(real(Sprime-S-dSpred)) < 1.0 ) ;
   std::cout<< GridLogMessage << "Done" <<std::endl;
   Grid_finalize();
+#endif
 }

tests/forces/Test_gpdwf_force_1f_2f.cc

@@ -32,6 +32,7 @@ using namespace std;
 using namespace Grid;
 
 //Here we test the G-parity action and force between the 1f (doubled-lattice) and 2f approaches 
+#ifdef ENABLE_GPARITY
 
 
 void copyConjGauge(LatticeGaugeFieldD &Umu_1f, const LatticeGaugeFieldD &Umu_2f, const int nu){
@@ -444,3 +445,7 @@ int main (int argc, char ** argv)
     assert(0);
   }
 }
+
+#else
+int main (int argc, char ** argv){};
+#endif

tests/forces/Test_gpwilson_force.cc

@@ -32,6 +32,7 @@ using namespace Grid;
 
 int main (int argc, char ** argv)
 {
+#ifdef ENABLE_GPARITY
   Grid_init(&argc,&argv);
 
   Coordinate latt_size   = GridDefaultLatt();
@@ -155,4 +156,5 @@ int main (int argc, char ** argv)
 
   std::cout<< GridLogMessage << "Done" <<std::endl;
   Grid_finalize();
+#endif
 }

tests/forces/Test_mobius_gparity_eofa_mixed.cc

@@ -30,9 +30,10 @@ See the full license in the file "LICENSE" in the top level distribution directo
 
 #include <Grid/Grid.h>
 
+#ifdef ENABLE_GPARITY
+
 using namespace std;
 using namespace Grid;
- ;
 
 typedef GparityWilsonImplD FermionImplPolicyD;
 typedef GparityMobiusEOFAFermionD FermionActionD;
@@ -231,3 +232,7 @@ int main (int argc, char** argv)
   std::cout << GridLogMessage << "Done" << std::endl;
   Grid_finalize();
 }
+#else
+int main(int argc,char ** argv) { return 0;};
+
+#endif

tests/forces/Test_mobius_gpforce_eofa.cc

@@ -31,14 +31,14 @@ See the full license in the file "LICENSE" in the top level distribution directo
 
 using namespace std;
 using namespace Grid;
- ;
+
 
 typedef GparityWilsonImplR FermionImplPolicy;
 typedef GparityMobiusEOFAFermionD FermionAction;
 typedef typename FermionAction::FermionField FermionField;
-
 int main (int argc, char** argv)
 {
+#ifdef ENABLE_GPARITY
   Grid_init(&argc, &argv);
 
   Coordinate latt_size   = GridDefaultLatt();
@@ -171,4 +171,5 @@ int main (int argc, char** argv)
 
   std::cout << GridLogMessage << "Done" << std::endl;
   Grid_finalize();
+#endif
 }

tests/hmc/Test_action_dwf_gparity2fvs1f.cc

@@ -30,7 +30,7 @@
 
 using namespace Grid;
 
-
+#ifdef ENABLE_GPARITY
 
 template<typename FermionField2f, typename FermionField1f>
 void copy2fTo1fFermionField(FermionField1f &out, const FermionField2f &in, int gpdir){
@@ -255,3 +255,6 @@ int main(int argc, char **argv) {
 } // main
 
 
+#else
+int main(int argc, char **argv){};
+#endif

tests/hmc/Test_hmc_EODWFRatio_Gparity.cc

@@ -30,6 +30,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 
 int main(int argc, char **argv) {
+#ifdef ENABLE_GPARITY
   using namespace Grid;
    ;
 
@@ -139,7 +140,7 @@ int main(int argc, char **argv) {
   Grid_finalize();
 
 
-
+#endif
 
 } // main
 

tests/hmc/Test_hmc_EOMobiusRatio.cc

@@ -55,13 +55,13 @@ namespace Grid{
 
 };
 
-  struct SmearingParameters: Serializable {
-    GRID_SERIALIZABLE_CLASS_MEMBERS(SmearingParameters,
+  struct HmcSmearingParameters: Serializable {
+    GRID_SERIALIZABLE_CLASS_MEMBERS(HmcSmearingParameters,
 				    double, rho,
 				    Integer, Nsmear)
 
     template <class ReaderClass >
-    SmearingParameters(Reader<ReaderClass>& Reader){
+    HmcSmearingParameters(Reader<ReaderClass>& Reader){
       read(Reader, "StoutSmearing", *this);
     }
 
@@ -213,7 +213,7 @@ int main(int argc, char **argv) {
   // Reset performance counters 
 
   if (ApplySmearing){
-    SmearingParameters SmPar(Reader);
+    HmcSmearingParameters SmPar(Reader);
     //double rho = 0.1;  // smearing parameter
     //int Nsmear = 3;    // number of smearing levels
     Smear_Stout<HMCWrapper::ImplPolicy> Stout(SmPar.rho);

tests/lanczos/LanParams.xml

@@ -0,0 +1,14 @@
+<?xml version="1.0"?>
+<grid>
+  <LanczosParameters>
+    <mass>0.00107</mass>
+    <M5>1.8</M5>
+    <Ls>48</Ls>
+    <Nstop>10</Nstop>
+    <Nk>15</Nk>
+    <Np>85</Np>
+    <ChebyLow>0.003</ChebyLow>
+    <ChebyHigh>60</ChebyHigh>
+    <ChebyOrder>201</ChebyOrder>
+  </LanczosParameters>
+</grid>

tests/lanczos/Test_compressed_lanczos_gparity.cc

@@ -35,6 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/iterative/LocalCoherenceLanczos.h>
 
+#ifdef ENABLE_GPARITY
+
 using namespace std;
 using namespace Grid;
 
@@ -378,7 +380,8 @@ void runTest(const Options &opt){
 
 
 //Note:  because we rely upon physical properties we must use a "real" gauge configuration
-int main (int argc, char ** argv) {
+int main (int argc, char ** argv)
+{
   Grid_init(&argc,&argv);
   GridLogIRL.TimingMode(1);
 
@@ -482,4 +485,8 @@ int main (int argc, char ** argv) {
   
   Grid_finalize();
 }
+#else
+int main(int argc, char **argv){};
+
+#endif
 

tests/lanczos/Test_dwf_G5R5.cc

@@ -0,0 +1,428 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_dwf_G5R5.cc
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+From Duo and Bob's Chirality study
+
+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;
+
+//typedef WilsonFermionD FermionOp;
+typedef DomainWallFermionD FermionOp;
+typedef typename DomainWallFermionD::FermionField FermionField;
+
+template <class T> void writeFile(T& in, std::string const fname){  
+#ifdef HAVE_LIME
+  // Ref: https://github.com/paboyle/Grid/blob/feature/scidac-wp1/tests/debug/Test_general_coarse_hdcg_phys48.cc#L111
+  std::cout << Grid::GridLogMessage << "Writes to: " << fname << std::endl;
+  Grid::emptyUserRecord record;
+  Grid::ScidacWriter WR(in.Grid()->IsBoss());
+  WR.open(fname);
+  WR.writeScidacFieldRecord(in,record,0);
+  WR.close();
+#endif
+  // What is the appropriate way to throw error?
+}
+
+RealD AllZero(RealD x) { return 0.; }
+
+namespace Grid {
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass , 
+		  		RealD, M5 , 
+	  			Integer, Ls,
+	  			Integer, Nstop,
+	  			Integer, Nk,
+	  			Integer, Np,
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+//                                  Integer, StartTrajectory,
+//                                  Integer, Trajectories, /* @brief Number of sweeps in this run */
+//                                  bool, MetropolisTest,
+//                                  Integer, NoMetropolisUntil,
+//                                  std::string, StartingType,
+//                                  Integer, SW,
+//				  RealD, Kappa,
+//                                  IntegratorParameters, MD)
+
+  LanczosParameters() {
+    ////////////////////////////// Default values
+      mass = 0;
+//    MetropolisTest    = true;
+//    NoMetropolisUntil = 10;
+//    StartTrajectory   = 0;
+//    SW                = 2;
+//    Trajectories      = 10;
+//    StartingType      = "HotStart";
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass > 
+  void initialize(Reader<ReaderClass> &TheReader){
+//    std::cout << GridLogMessage << "Reading HMC\n";
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+  
+};
+
+}
+
+int main(int argc, char** argv) {
+  Grid_init(&argc, &argv);
+
+  LanczosParameters LanParams;
+#if 1
+  {
+    XmlReader  HMCrd("LanParams.xml");
+    read(HMCrd,"LanczosParameters",LanParams);
+  }
+#else
+  {
+    LanParams.mass = mass;
+  }
+#endif
+  std::cout << GridLogMessage<< LanParams <<std::endl;
+  { 
+    XmlWriter HMCwr("LanParams.xml.out");
+    write(HMCwr,"LanczosParameters",LanParams);
+  }
+
+  int Ls=16;
+  RealD M5=1.8;
+  RealD mass = 0.01;
+
+  mass=LanParams.mass;
+  Ls=LanParams.Ls;
+  M5=LanParams.M5;
+
+  GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(
+      GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()),
+      GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid =
+      SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+//  GridCartesian* FGrid = UGrid;
+//  GridRedBlackCartesian* FrbGrid = UrbGrid;
+  GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls, UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, UGrid);
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n", UGrid, UrbGrid, FGrid, FrbGrid);
+
+  std::vector<int> seeds4({1, 2, 3, 4});
+  std::vector<int> seeds5({5, 6, 7, 8});
+  GridParallelRNG RNG5(FGrid); RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG RNG5rb(FrbGrid); RNG5.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("./config");
+
+  int precision32 = 0;
+  int tworow      = 0;
+  NerscIO::readConfiguration(Umu,header,file);
+
+/*
+  std::vector<LatticeColourMatrix> U(4, UGrid);
+  for (int mu = 0; mu < Nd; mu++) {
+    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+  }
+*/
+  int Nk = 20;
+  int Nstop = Nk;
+  int Np = 80;
+
+  Nstop=LanParams.Nstop;
+  Nk=LanParams.Nk;
+  Np=LanParams.Np;
+
+  int Nm = Nk + Np;
+  int MaxIt = 10000;
+  RealD resid = 1.0e-5;
+
+
+//while ( mass > - 5.0){
+  FermionOp Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  MdagMLinearOperator<FermionOp,FermionField> HermOp(Ddwf); /// <-----
+//  Gamma5HermitianLinearOperator <FermionOp,LatticeFermion> HermOp2(WilsonOperator); /// <-----
+  Gamma5R5HermitianLinearOperator<FermionOp, LatticeFermion> G5R5Herm(Ddwf);
+//  Gamma5R5HermitianLinearOperator
+  std::vector<double> Coeffs{0, 1.};
+  Polynomial<FermionField> PolyX(Coeffs);
+
+  Chebyshev<FermionField> Cheby(LanParams.ChebyLow,LanParams.ChebyHigh,LanParams.ChebyOrder);
+
+  FunctionHermOp<FermionField> OpCheby(Cheby,HermOp);
+  PlainHermOp<FermionField> Op     (HermOp);
+  PlainHermOp<FermionField> Op2     (G5R5Herm);
+
+  ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby, Op, Nstop, Nk, Nm, resid, MaxIt);
+
+  std::vector<RealD> eval(Nm);
+  FermionField src(FGrid);
+  gaussian(RNG5, src);
+  std::vector<FermionField> evec(Nm, FGrid);
+  for (int i = 0; i < 1; i++) {
+    std::cout << i << " / " << Nm << " grid pointer " << evec[i].Grid()
+              << std::endl;
+  };
+
+  int Nconv;
+  IRL.calc(eval, evec, src, Nconv);
+
+  std::cout << mass <<" : " << eval        << std::endl;
+  std::cout << " #evecs "   << evec.size() << std::endl;
+  std::cout << " Nconv  "   << Nconv       << std::endl;
+  std::cout << " Nm     "   << Nm          << std::endl;
+  if ( Nconv > evec.size() ) Nconv = evec.size();
+  
+#if 0
+  Gamma g5(Gamma::Algebra::Gamma5) ;
+  ComplexD dot;
+  FermionField tmp(FGrid);
+//  RealD eMe,eMMe;
+  for (int i = 0; i < Nstop ; i++) {
+//    tmp = g5*evec[i];
+    dot = innerProduct(evec[i],evec[i]);
+//    G5R5(tmp,evec[i]);
+    G5R5Herm.HermOpAndNorm(evec[i],tmp,eMe,eMMe);
+    std::cout <<"Norm "<<M5<<" "<< mass << " : " << i << " " << real(dot) << " " << imag(dot)  << " "<< eMe << " " <<eMMe<< std::endl ;
+    for (int j = 0; j < Nstop ; j++) {
+      dot = innerProduct(tmp,evec[j]);
+      std::cout <<"G5R5 "<<M5<<" "<< mass << " : " << i << " " <<j<<" " << real(dot) << " " << imag(dot)  << std::endl ;
+    }
+  }
+//  src  = evec[0]+evec[1]+evec[2];
+//  mass += -0.1;
+#endif
+
+  //**********************************************************************
+  //orthogonalization
+  //calculat the matrix
+  cout << "Start orthogonalization " << endl;
+  cout << "calculate the matrix element" << endl;
+  vector<LatticeFermion> G5R5Mevec(Nconv, FGrid);
+  vector<LatticeFermion> finalevec(Nconv, FGrid);
+  vector<RealD> eMe(Nconv), eMMe(Nconv);
+  for(int i = 0; i < Nconv; i++){
+    cout << "calculate the matrix element["<<i<<"]" << endl;
+    G5R5Herm.HermOpAndNorm(evec[i], G5R5Mevec[i], eMe[i], eMMe[i]);
+  }
+  cout << "Re<evec, G5R5M(evec)>: " << endl;
+  cout << eMe << endl;
+  cout << "<G5R5M(evec), G5R5M(evec)>" << endl;
+  cout << eMMe << endl;
+  vector<vector<ComplexD>> VevecG5R5Mevec(Nconv);
+  Eigen::MatrixXcd evecG5R5Mevec = Eigen::MatrixXcd::Zero(Nconv, Nconv);
+  for(int i = 0; i < Nconv; i++){
+    VevecG5R5Mevec[i].resize(Nconv);
+    for(int j = 0; j < Nconv; j++){
+      VevecG5R5Mevec[i][j] = innerProduct(evec[i], G5R5Mevec[j]);
+      evecG5R5Mevec(i, j) = VevecG5R5Mevec[i][j];
+    }
+  }
+  //calculate eigenvector
+  cout << "Eigen solver" << endl;
+  Eigen::SelfAdjointEigenSolver<Eigen::MatrixXcd> eigensolver(evecG5R5Mevec);
+  vector<RealD> eigeneval(Nconv);
+  vector<vector<ComplexD>> eigenevec(Nconv);
+  for(int i = 0; i < Nconv; i++){
+    eigeneval[i] = eigensolver.eigenvalues()[i];
+    eigenevec[i].resize(Nconv);
+    for(int j = 0; j < Nconv; j++){
+      eigenevec[i][j] = eigensolver.eigenvectors()(i, j);
+    }
+  }
+  //rotation
+  cout << "Do rotation" << endl;
+  for(int i = 0; i < Nconv; i++){
+    finalevec[i] = finalevec[i] - finalevec[i];
+    for(int j = 0; j < Nconv; j++){
+      finalevec[i] = eigenevec[j][i]*evec[j] + finalevec[i];
+    }
+  }
+  //normalize again;
+  for(int i = 0; i < Nconv; i++){
+    RealD tmp_RealD = norm2(finalevec[i]);
+    tmp_RealD = 1./pow(tmp_RealD, 0.5);
+    finalevec[i] = finalevec[i]*tmp_RealD;
+  }
+
+  //check
+  for(int i = 0; i < Nconv; i++){
+    G5R5Herm.HermOpAndNorm(finalevec[i], G5R5Mevec[i], eMe[i], eMMe[i]);
+  }
+
+  //**********************************************************************
+  //sort the eigenvectors
+  vector<LatticeFermion> finalevec_copy(Nconv, FGrid);
+  for(int i = 0; i < Nconv; i++){
+    finalevec_copy[i] = finalevec[i];
+  }
+  vector<RealD> eMe_copy(eMe);
+  for(int i = 0; i < Nconv; i++){
+    eMe[i] = fabs(eMe[i]);
+    eMe_copy[i] = eMe[i];
+  }
+  sort(eMe_copy.begin(), eMe_copy.end());
+  for(int i = 0; i < Nconv; i++){
+    for(int j = 0; j < Nconv; j++){
+      if(eMe[j] == eMe_copy[i]){
+        finalevec[i] = finalevec_copy[j];
+      }
+    }
+  }
+  for(int i = 0; i < Nconv; i++){
+    G5R5Herm.HermOpAndNorm(finalevec[i], G5R5Mevec[i], eMe[i], eMMe[i]);
+  }
+  cout << "Re<evec, G5R5M(evec)>: " << endl;
+  cout << eMe << endl;
+  cout << "<G5R5M(evec), G5R5M(evec)>" << endl;
+  cout << eMMe << endl;
+
+  
+
+//  vector<LatticeFermion> finalevec(Nconv, FGrid);
+// temporary, until doing rotation
+//  for(int i = 0; i < Nconv; i++)
+//	  finalevec[i]=evec[i];
+  //**********************************************************************
+  //calculate chirality matrix
+  vector<LatticeFermion> G5evec(Nconv, FGrid);
+  vector<vector<ComplexD>> chiral_matrix(Nconv);
+  vector<vector<RealD>> chiral_matrix_real(Nconv);
+  for(int i = 0; i < Nconv; i++){
+//    G5evec[i] = G5evec[i] - G5evec[i];
+    G5evec[i] = Zero();
+    for(int j = 0; j < Ls/2; j++){
+      axpby_ssp(G5evec[i], 1., finalevec[i], 0., G5evec[i], j, j);
+    }
+    for(int j = Ls/2; j < Ls; j++){
+      axpby_ssp(G5evec[i], -1., finalevec[i], 0., G5evec[i], j, j);
+    }
+  }
+  
+  for(int i = 0; i < Nconv; i++){
+    Ddwf.M(finalevec[i], G5R5Mevec[i]);
+    for(int j = 0; j < Nconv; j++){
+      std::cout << "<"<<j<<"|Ddwf|"<<i<<"> = "<<innerProduct(finalevec[j],G5R5Mevec[i])<<std::endl;
+    }
+  }
+  for(int i = 0; i < Nconv; i++){
+    RealD t1,t2;
+    G5R5Herm.HermOpAndNorm(finalevec[i], G5R5Mevec[i], t1, t2);
+    for(int j = 0; j < Nconv; j++){
+      std::cout << "<"<<j<<"|G5R5 M|"<<i<<"> = "<<innerProduct(finalevec[j],G5R5Mevec[i])<<std::endl;
+    }
+  }
+  
+  for(int i = 0; i < Nconv; i++){
+    chiral_matrix_real[i].resize(Nconv);
+    chiral_matrix[i].resize(Nconv);
+
+    std::string evfile("./evec_density");
+    evfile = evfile+"_"+std::to_string(i);
+    auto evdensity = localInnerProduct(finalevec[i],finalevec[i] );
+    writeFile(evdensity,evfile);
+
+    for(int j = 0; j < Nconv; j++){
+      chiral_matrix[i][j] = innerProduct(finalevec[i], G5evec[j]);
+      std::cout <<" chiral_matrix_real signed "<<i<<" "<<j<<" "<< chiral_matrix_real[i][j] << std::endl;
+      chiral_matrix_real[i][j] = abs(chiral_matrix[i][j]);
+      std::cout <<" chiral_matrix_real "<<i<<" "<<j<<" "<< chiral_matrix_real[i][j] << std::endl;
+      if ( chiral_matrix_real[i][j] > 0.8 ) {
+	auto g5density = localInnerProduct(finalevec[i], G5evec[j]);
+	std::string chfile("./chiral_density_");
+	chfile = chfile +std::to_string(i)+"_"+std::to_string(j);
+	writeFile(g5density,chfile);
+      }
+    }
+  }
+  for(int i = 0; i < Nconv; i++){
+    if(chiral_matrix[i][i].real() < 0.){
+      chiral_matrix_real[i][i] = -1. * chiral_matrix_real[i][i];
+    }
+  }
+
+  FILE *fp = fopen("lego-plot.py","w"); assert(fp!=NULL);
+#define PYTHON_LINE(A)  fprintf(fp,A"\n");
+  PYTHON_LINE("import matplotlib.pyplot as plt");
+  PYTHON_LINE("import numpy as np");
+  PYTHON_LINE("");
+  PYTHON_LINE("fig = plt.figure()");
+  PYTHON_LINE("ax = fig.add_subplot(projection='3d')");
+  PYTHON_LINE("");
+  PYTHON_LINE("x, y = np.random.rand(2, 100) * 4");
+  fprintf(fp,"hist, xedges, yedges = np.histogram2d(x, y, bins=%d, range=[[0, %d], [0, %d]])\n",Nconv,Nconv-1,Nconv-1);
+  PYTHON_LINE("");
+  PYTHON_LINE("# Construct arrays for the anchor positions of the 16 bars");
+  PYTHON_LINE("xpos, ypos = np.meshgrid(xedges[:-1] + 0.25, yedges[:-1] + 0.25, indexing=\"ij\")");
+  PYTHON_LINE("xpos = xpos.ravel()");
+  PYTHON_LINE("ypos = ypos.ravel()");
+  PYTHON_LINE("zpos = 0");
+  PYTHON_LINE("");
+  PYTHON_LINE("# Construct arrays with the dimensions for the 16 bars.");
+  PYTHON_LINE("dx = dy = 0.5 * np.ones_like(zpos)");
+  PYTHON_LINE("dz = np.array([");
+  for(int i = 0; i < Nconv; i++){
+    fprintf(fp,"\t[ ");
+    for(int j = 0; j < Nconv; j++){
+      fprintf(fp,"%lf ",chiral_matrix_real[i][j]);
+      if(j<Nconv-1) fprintf(fp,",");
+      else          fprintf(fp," ");
+    }
+    fprintf(fp,"]");
+    if(i<Nconv-1) fprintf(fp,",\n");
+    else          fprintf(fp,"\n");
+  }
+	      
+  PYTHON_LINE("\t])");
+  PYTHON_LINE("dz = dz.ravel()");
+  PYTHON_LINE("ax.bar3d(xpos, ypos, zpos, dx, dy, dz, zsort='average')");
+  PYTHON_LINE("plt.show()");
+  fclose(fp);
+  
+  Grid_finalize();
+}

tests/lanczos/Test_dwf_lanczos.cc

@@ -29,11 +29,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 using namespace std;
 using namespace Grid;
- ;
 
 template<typename Action>
 struct Setup{};
 
+#ifdef ENABLE_GPARITY
 template<>
 struct Setup<GparityMobiusFermionF>{
   static GparityMobiusFermionF* getAction(LatticeGaugeFieldF &Umu,
@@ -47,16 +47,24 @@ struct Setup<GparityMobiusFermionF>{
     return new GparityMobiusFermionF(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,params);
   }
 };
+#endif
 
 template<>
 struct Setup<DomainWallFermionF>{
   static DomainWallFermionF* getAction(LatticeGaugeFieldF &Umu,
+					  GridCartesian* FGrid, GridRedBlackCartesian* FrbGrid, GridCartesian* UGrid, GridRedBlackCartesian* UrbGrid){
+    RealD mass=0.00054;
+    RealD M5=1.8;
+    return new DomainWallFermionF(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  }
+};
+template<>
 struct Setup<DomainWallFermionD>{
   static DomainWallFermionD* getAction(LatticeGaugeField &Umu,
 					  GridCartesian* FGrid, GridRedBlackCartesian* FrbGrid, GridCartesian* UGrid, GridRedBlackCartesian* UrbGrid){
     RealD mass=0.00054;
     RealD M5=1.8;
-    return new DomainWallFermionF(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    return new DomainWallFermionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
   }
 };
 
@@ -168,7 +176,9 @@ int main (int argc, char ** argv)
   }
 
   if(action == "GparityMobius"){
+#ifdef ENABLE_GPARITY    
     run<GparityMobiusFermionF>();
+#endif    
   }else if(action == "DWF"){
     run<DomainWallFermionF>();
   }else if(action == "Mobius"){

tests/lanczos/Test_evec_compression.cc

@@ -555,6 +555,7 @@ int main (int argc, char ** argv) {
   double c = (args.mobius_scale - bmc)/2.;  // c = 1/2 [ (b+c) - (b-c) ]
     
   if(is_gparity){
+#ifdef ENABLE_GPARITY    
     GparityWilsonImplD::ImplParams Params = setupGparityParams(args.GparityDirs);
     readConfiguration<ConjugateGimplD>(Umu, config, args.is_cps_cfg);   //Read the gauge field
     
@@ -564,7 +565,10 @@ int main (int argc, char ** argv) {
     }else if(action_s == "Mobius"){
       GparityMobiusFermionD action(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, args.mass, args.M5, b, c, Params);
       run(action, config, args);	    
-    }      
+    }
+#else
+    assert(0);
+#endif
   }else{
     WilsonImplD::ImplParams Params = setupParams();
     readConfiguration<PeriodicGimplD>(Umu, config, args.is_cps_cfg);   //Read the gauge field

tests/lanczos/Test_wilson_DWFKernel.cc

@@ -0,0 +1,278 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_dwf_lanczos.cc
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+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;
+ ;
+
+typedef WilsonFermionD FermionOp;
+typedef typename WilsonFermionD::FermionField FermionField;
+
+
+RealD AllZero(RealD x) { return 0.; }
+
+namespace Grid {
+
+#if 0
+template<typename Field>
+class RationalHermOp : public LinearFunction<Field> {
+public:
+  using LinearFunction<Field>::operator();  
+//  OperatorFunction<Field>   & _poly;
+  LinearOperatorBase<Field> &_Linop;
+  RealD _massDen, _massNum;
+   
+  FunctionHermOp(LinearOperatorBase<Field>& linop, RealD massDen,RealD massNum)
+    :  _Linop(linop) ,_massDen(massDen),_massNum(massNum) {};
+      
+  void operator()(const Field& in, Field& out) {
+//    _poly(_Linop,in,out);
+  } 
+};
+#endif
+
+template<class Matrix,class Field>
+class InvG5LinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD _num;
+  RealD _Tol;
+  Integer _MaxIt;
+  Gamma g5;
+
+public:
+  InvG5LinearOperator(Matrix &Mat,RealD num): _Mat(Mat),_num(num), _Tol(1e-12),_MaxIt(10000), g5(Gamma::Algebra::Gamma5) {};
+
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    assert(0);
+    _Mat.Mdiag(in,out);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    assert(0);
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    assert(0);
+    _Mat.M(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    assert(0);
+    _Mat.Mdag(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+     Field tmp(in.Grid());
+     MdagMLinearOperator<Matrix,Field> denom(_Mat);
+     ConjugateGradient<Field> CG(_Tol,_MaxIt); 
+     _Mat.M(in,tmp);
+     tmp += _num*in;
+     _Mat.Mdag(tmp,out);
+     CG(denom,out,tmp);
+     out = g5*tmp;
+  }
+};
+
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass , 
+				RealD, resid,
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+//                                  Integer, StartTrajectory,
+//                                  Integer, Trajectories, /* @brief Number of sweeps in this run */
+//                                  bool, MetropolisTest,
+//                                  Integer, NoMetropolisUntil,
+//                                  std::string, StartingType,
+//                                  Integer, SW,
+//				  RealD, Kappa,
+//                                  IntegratorParameters, MD)
+
+  LanczosParameters() {
+    ////////////////////////////// Default values
+      mass = 0;
+//    MetropolisTest    = true;
+//    NoMetropolisUntil = 10;
+//    StartTrajectory   = 0;
+//    SW                = 2;
+//    Trajectories      = 10;
+//    StartingType      = "HotStart";
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass > 
+  void initialize(Reader<ReaderClass> &TheReader){
+//    std::cout << GridLogMessage << "Reading HMC\n";
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+  
+};
+
+}
+
+int main(int argc, char** argv) {
+  Grid_init(&argc, &argv);
+
+  GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(
+      GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()),
+      GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid =
+      SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian* FGrid = UGrid;
+  GridRedBlackCartesian* FrbGrid = UrbGrid;
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n", UGrid, UrbGrid, FGrid, FrbGrid);
+
+  std::vector<int> seeds4({1, 2, 3, 4});
+  std::vector<int> seeds5({5, 6, 7, 8});
+  GridParallelRNG RNG5(FGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG RNG5rb(FrbGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid);
+//  SU<Nc>::HotConfiguration(RNG4, Umu);
+
+  FieldMetaData header;
+  std::string file("./config");
+
+  int precision32 = 0;
+  int tworow      = 0;
+//  NerscIO::writeConfiguration(Umu,file,tworow,precision32);
+  NerscIO::readConfiguration(Umu,header,file);
+
+/*
+  std::vector<LatticeColourMatrix> U(4, UGrid);
+  for (int mu = 0; mu < Nd; mu++) {
+    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+  }
+*/
+
+  int Nstop = 5;
+  int Nk = 10;
+  int Np = 90;
+  int Nm = Nk + Np;
+  int MaxIt = 10000;
+  RealD resid = 1.0e-5;
+
+  RealD mass = -1.0;
+
+  LanczosParameters LanParams;
+#if 1
+  {
+    XmlReader  HMCrd("LanParams.xml");
+    read(HMCrd,"LanczosParameters",LanParams);
+  }
+#else
+  {
+    LanParams.mass = mass;
+  }
+#endif
+  std::cout << GridLogMessage<< LanParams <<std::endl;
+  { 
+    XmlWriter HMCwr("LanParams.xml.out");
+    write(HMCwr,"LanczosParameters",LanParams);
+  }
+
+  mass=LanParams.mass;
+  resid=LanParams.resid;
+
+
+while ( mass > - 5.0){
+  FermionOp WilsonOperator(Umu,*FGrid,*FrbGrid,2.+mass);
+  InvG5LinearOperator<FermionOp,LatticeFermion> HermOp(WilsonOperator,-2.); /// <-----
+  //SchurDiagTwoOperator<FermionOp,FermionField> HermOp(WilsonOperator);
+//  Gamma5HermitianLinearOperator <FermionOp,LatticeFermion> HermOp2(WilsonOperator); /// <-----
+
+  std::vector<double> Coeffs{0, 0, 1.};
+  Polynomial<FermionField> PolyX(Coeffs);
+  Chebyshev<FermionField> Cheby(LanParams.ChebyLow,LanParams.ChebyHigh,LanParams.ChebyOrder);
+
+       FunctionHermOp<FermionField> OpCheby(Cheby,HermOp);
+//     InvHermOp<FermionField> Op(WilsonOperator,HermOp);
+     PlainHermOp<FermionField> Op     (HermOp);
+//     PlainHermOp<FermionField> Op2     (HermOp2);
+
+  ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby, Op, Nstop, Nk, Nm, resid, MaxIt);
+
+  std::vector<RealD> eval(Nm);
+  FermionField src(FGrid);
+  gaussian(RNG5, src);
+  std::vector<FermionField> evec(Nm, FGrid);
+  for (int i = 0; i < 1; i++) {
+    std::cout << i << " / " << Nm << " grid pointer " << evec[i].Grid()
+              << std::endl;
+  };
+
+  int Nconv;
+  IRL.calc(eval, evec, src, Nconv);
+
+  std::cout << mass <<" : " << eval << std::endl;
+
+  Gamma g5(Gamma::Algebra::Gamma5) ;
+  ComplexD dot;
+  FermionField tmp(FGrid);
+  for (int i = 0; i < Nstop ; i++) {
+    tmp = g5*evec[i];
+    dot = innerProduct(tmp,evec[i]);
+    std::cout << mass << " : " << eval[i]  << " " << real(dot) << " " << imag(dot)  << std::endl ;
+  }
+  src  = evec[0]+evec[1]+evec[2];
+  mass += -0.1;
+}
+
+  Grid_finalize();
+}

tests/lanczos/Test_wilson_specflow.cc

@@ -0,0 +1,211 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_dwf_lanczos.cc
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+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;
+ ;
+
+typedef WilsonFermionD FermionOp;
+typedef typename WilsonFermionD::FermionField FermionField;
+
+
+RealD AllZero(RealD x) { return 0.; }
+
+namespace Grid {
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass , 
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+//                                  Integer, StartTrajectory,
+//                                  Integer, Trajectories, /* @brief Number of sweeps in this run */
+//                                  bool, MetropolisTest,
+//                                  Integer, NoMetropolisUntil,
+//                                  std::string, StartingType,
+//                                  Integer, SW,
+//				  RealD, Kappa,
+//                                  IntegratorParameters, MD)
+
+  LanczosParameters() {
+    ////////////////////////////// Default values
+      mass = 0;
+//    MetropolisTest    = true;
+//    NoMetropolisUntil = 10;
+//    StartTrajectory   = 0;
+//    SW                = 2;
+//    Trajectories      = 10;
+//    StartingType      = "HotStart";
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass > 
+  void initialize(Reader<ReaderClass> &TheReader){
+//    std::cout << GridLogMessage << "Reading HMC\n";
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+  
+};
+
+}
+
+int main(int argc, char** argv) {
+  Grid_init(&argc, &argv);
+
+  GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(
+      GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()),
+      GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid =
+      SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian* FGrid = UGrid;
+  GridRedBlackCartesian* FrbGrid = UrbGrid;
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n", UGrid, UrbGrid, FGrid, FrbGrid);
+
+  std::vector<int> seeds4({1, 2, 3, 4});
+  std::vector<int> seeds5({5, 6, 7, 8});
+  GridParallelRNG RNG5(FGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG RNG5rb(FrbGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid);
+//  SU<Nc>::HotConfiguration(RNG4, Umu);
+
+  FieldMetaData header;
+  std::string file("./config");
+
+  int precision32 = 0;
+  int tworow      = 0;
+//  NerscIO::writeConfiguration(Umu,file,tworow,precision32);
+  NerscIO::readConfiguration(Umu,header,file);
+
+/*
+  std::vector<LatticeColourMatrix> U(4, UGrid);
+  for (int mu = 0; mu < Nd; mu++) {
+    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+  }
+*/
+
+  int Nstop = 10;
+  int Nk = 20;
+  int Np = 80;
+  int Nm = Nk + Np;
+  int MaxIt = 10000;
+  RealD resid = 1.0e-5;
+
+  RealD mass = -1.0;
+
+  LanczosParameters LanParams;
+#if 1
+  {
+    XmlReader  HMCrd("LanParams.xml");
+    read(HMCrd,"LanczosParameters",LanParams);
+  }
+#else
+  {
+    LanParams.mass = mass;
+  }
+#endif
+  std::cout << GridLogMessage<< LanParams <<std::endl;
+  { 
+    XmlWriter HMCwr("LanParams.xml.out");
+    write(HMCwr,"LanczosParameters",LanParams);
+  }
+
+  mass=LanParams.mass;
+
+
+while ( mass > - 5.0){
+  FermionOp WilsonOperator(Umu,*FGrid,*FrbGrid,mass);
+  MdagMLinearOperator<FermionOp,FermionField> HermOp(WilsonOperator); /// <-----
+  //SchurDiagTwoOperator<FermionOp,FermionField> HermOp(WilsonOperator);
+  Gamma5HermitianLinearOperator <FermionOp,LatticeFermion> HermOp2(WilsonOperator); /// <-----
+
+  std::vector<double> Coeffs{0, 1.};
+  Polynomial<FermionField> PolyX(Coeffs);
+//  Chebyshev<FermionField> Cheby(0.5, 60., 31);
+//                                  RealD, ChebyLow,
+//                                RealD, ChebyHigh,
+//                                Integer, ChebyOrder)
+
+  Chebyshev<FermionField> Cheby(LanParams.ChebyLow,LanParams.ChebyHigh,LanParams.ChebyOrder);
+
+  FunctionHermOp<FermionField> OpCheby(Cheby,HermOp);
+     PlainHermOp<FermionField> Op     (HermOp);
+     PlainHermOp<FermionField> Op2     (HermOp2);
+
+  ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby, Op2, Nstop, Nk, Nm, resid, MaxIt);
+
+  std::vector<RealD> eval(Nm);
+  FermionField src(FGrid);
+  gaussian(RNG5, src);
+  std::vector<FermionField> evec(Nm, FGrid);
+  for (int i = 0; i < 1; i++) {
+    std::cout << i << " / " << Nm << " grid pointer " << evec[i].Grid()
+              << std::endl;
+  };
+
+  int Nconv;
+  IRL.calc(eval, evec, src, Nconv);
+
+  std::cout << mass <<" : " << eval << std::endl;
+
+  Gamma g5(Gamma::Algebra::Gamma5) ;
+  ComplexD dot;
+  FermionField tmp(FGrid);
+  for (int i = 0; i < Nstop ; i++) {
+    tmp = g5*evec[i];
+    dot = innerProduct(tmp,evec[i]);
+    std::cout << mass << " : " << eval[i]  << " " << real(dot) << " " << imag(dot)  << std::endl ;
+  }
+  src  = evec[0]+evec[1]+evec[2];
+  mass += -0.1;
+}
+
+  Grid_finalize();
+}