Adding DWF evec Chirality measurement

Const improvements in Accelerator.h

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/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;
 
 ////////////////////////////////////////////
@@ -362,8 +363,6 @@ 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 +378,6 @@ 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 +395,8 @@ 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 +438,15 @@ 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 +455,11 @@ 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 +468,7 @@ 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 +568,105 @@ 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 +703,89 @@ 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
 ////////////////////////////////////////////

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/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_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 +501,17 @@ 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);
+      acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
+#endif
       t_comms+=usecond()-t;
      }
     for ( int d=0;d < depth ; d ++ ) {
@@ -508,9 +522,17 @@ 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);
+      acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
+#endif      
       t_comms+=usecond()-t;
     }
 

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/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/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,
@@ -395,7 +406,14 @@ 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 +446,7 @@ 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 +502,9 @@ 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 +518,11 @@ 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 +561,7 @@ public:
 	  coalescedWrite(to[j] ,coalescedRead(from [j]));
       });
       acceleratorFenceComputeStream();
+      // Also fenced in WilsonKernels
     }
   }
   
@@ -663,7 +690,7 @@ 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 +710,7 @@ 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 +802,8 @@ 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(to,from,bytes, cudaMemcpyHostToDevice);
+  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)
@@ -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)
 {
   /*
@@ -564,9 +611,12 @@ 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 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 +705,9 @@ 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;
@@ -651,3 +658,4 @@ 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;

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;

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/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/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 +1,16 @@
+../../configure \
+--enable-comms=mpi-auto \
+--enable-unified=yes \
+--enable-shm=shmopen \
+--enable-shm-fast-path=shmopen \
+--enable-accelerator=none \
+--enable-simd=AVX512 \
+--disable-accelerator-cshift \
+--disable-fermion-reps \
+--disable-gparity \
+CXX=clang++ \
+MPICXX=mpicxx \
+CXXFLAGS="-std=c++17"
+
+
+

systems/sdcc-genoa/sourceme.sh

@@ -0,0 +1,4 @@
+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

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/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/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_dwf_G5R5.cc

@@ -0,0 +1,346 @@
+/*************************************************************************************
+
+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;
+
+
+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 = -1.0;
+
+  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 Nstop = 10;
+  int Nk = 20;
+  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;
+
+#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++){
+    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++){
+    chiral_matrix_real[i].resize(Nconv);
+    chiral_matrix[i].resize(Nconv);
+    for(int j = 0; j < Nconv; j++){
+      chiral_matrix[i][j] = innerProduct(finalevec[i], G5evec[j]);
+      chiral_matrix_real[i][j] = abs(chiral_matrix[i][j]);
+      std::cout <<" chiral_matrix_real "<<i<<" "<<j<<" "<< chiral_matrix_real[i][j] << std::endl;
+    }
+  }
+  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];
+    }
+  }
+
+
+  Grid_finalize();
+}

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();
+}