Scripts

Feature/wilson clover 5d

Grid/algorithms/FFT.h

@@ -191,7 +191,7 @@ public:
       
     Lattice<sobj> pgbuf(&pencil_g);
     autoView(pgbuf_v , pgbuf, CpuWrite);
-    std::cout << "CPU view" << std::endl;
+    //std::cout << "CPU view" << std::endl;
     
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -215,7 +215,7 @@ public:
     else if ( sign == forward ) div = 1.0;
     else assert(0);
       
-    std::cout << "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 << "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 << "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 << "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 << "Destroying plan " << std::endl;
+    //std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
     // destroying plan
     FFTW<scalar>::fftw_destroy_plan(p);
 #endif

Grid/algorithms/LinearOperator.h

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

Grid/algorithms/approx/Chebyshev.h

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

Grid/algorithms/blas/BatchedBlas.h

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

Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h

@@ -144,11 +144,11 @@ public:
       acceleratorCopyDeviceToDevice(&BLAS_Y[offset],&y_v[0],sizeof(scalar_object)*vol);
     }    
     RealD t4 = usecond();
-    std::cout << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
-    std::cout << "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
-    std::cout << "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
-    std::cout << "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
-    std::cout << "MulMatrix total "<< t4-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix total "<< t4-t0<<" us"<<std::endl;
   }
   
   void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y)
@@ -242,16 +242,16 @@ public:
     RealD flops = 8.0*M*N*K;
     flops = flops/(t4-t3)/1.e3;
     bytes = bytes/(t4-t3)/1.e3;
-    std::cout << "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
-    std::cout << "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
 #else
     int nrhs;
     GridBase *grid;
@@ -358,17 +358,17 @@ public:
     flops = flops/(t4-t3)/1.e3;
     bytes = bytes/(t4-t3)/1.e3;
     xybytes = 4*xybytes/(t2-t1)/1.e3;
-    std::cout << "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
-    std::cout << "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
 #endif
   }
 };

Grid/algorithms/iterative/AdefMrhs.h

@@ -63,7 +63,12 @@ class TwoLevelCGmrhs
   GridStopWatch SmoothTimer;
   GridStopWatch InsertTimer;
 
-  
+  /*
+    Field rrr;
+  Field sss;
+  Field qqq;
+  Field zzz;
+  */  
   // more most opertor functions
   TwoLevelCGmrhs(RealD tol,
 		 Integer maxit,
@@ -74,6 +79,12 @@ class TwoLevelCGmrhs
     MaxIterations(maxit),
     _FineLinop(FineLinop),
     _Smoother(Smoother)
+    /*
+    rrr(fine),
+    sss(fine),
+    qqq(fine),
+    zzz(fine)
+*/
   {
     grid       = fine;
   };
@@ -81,8 +92,8 @@ class TwoLevelCGmrhs
   // Vector case
   virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
   {
-    SolveSingleSystem(src,x);
-    //    SolvePrecBlockCG(src,x);
+    //    SolveSingleSystem(src,x);
+    SolvePrecBlockCG(src,x);
   }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -657,6 +668,8 @@ public:
     CoarseField PleftProjMrhs(this->coarsegridmrhs);
     CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
 
+    //    this->rrr=in[0];
+
 #undef SMOOTHER_BLOCK_SOLVE
 #if SMOOTHER_BLOCK_SOLVE
     this->SmoothTimer.Start();
@@ -669,6 +682,7 @@ public:
       this->SmoothTimer.Stop();
     }
 #endif
+    //    this->sss=Min[0];
     
     for(int rhs=0;rhs<nrhs;rhs++) {
       
@@ -705,9 +719,11 @@ public:
     this->_Projector.blockPromote(tmp,PleftMss_proj);// tmp= Q[in - A Min]  
     this->PromoteTimer.Stop();
     this->FineTimer.Start();
+    //    this->qqq=tmp[0];
     for(int rhs=0;rhs<nrhs;rhs++) {
       axpy(out[rhs],1.0,Min[rhs],tmp[rhs]); // Min+tmp
     }
+    //    this->zzz=out[0];
     this->FineTimer.Stop();
   }
 };

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/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -74,7 +74,7 @@ public:
 
   void operator() (const Field &src, Field &psi){
 
-    psi=Zero();
+    //    psi=Zero();
     RealD cp, ssq,rsq;
     ssq=norm2(src);
     rsq=Tolerance*Tolerance*ssq;

Grid/algorithms/multigrid/Aggregates.h

@@ -30,6 +30,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once
 
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+
 NAMESPACE_BEGIN(Grid);
 
 inline RealD AggregatePowerLaw(RealD x)
@@ -95,7 +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);
 
@@ -108,7 +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]);
 
@@ -124,6 +126,53 @@ public:
     }
   }
 
+  virtual void CreateSubspaceGCR(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &DiracOp,int nn=nbasis)
+  {
+    RealD scale;
+
+    TrivialPrecon<FineField> simple_fine;
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    FineField noise(FineGrid);
+    FineField src(FineGrid);
+    FineField guess(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
+
+      for(int i=0;i<2;i++){
+	//  void operator() (const Field &src, Field &psi){
+#if 1
+	std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	src = noise;
+	guess=Zero();
+	GCR(src,guess);
+	subspace[b] = guess;
+#else
+	std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	src=Zero();
+	guess = noise;
+	GCR(src,guess);
+	subspace[b] = guess;
+#endif
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
   ////////////////////////////////////////////////////////////////////////////////////////////////
   // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
   // and this is the best I found
@@ -160,14 +209,21 @@ public:
 
     int b =0;
     {
+      ComplexD ip;
       // Filter
       Chebyshev<FineField> Cheb(lo,hi,orderfilter);
       Cheb(hermop,noise,Mn);
       // normalise
       scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
       subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+      hermop.Op(Mn,tmp);
+      ip= innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+      hermop.AdjOp(Mn,tmp); 
+      ip = innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
       b++;
     }
 
@@ -213,8 +269,18 @@ public:
 	  Mn=*Tnp;
 	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
 	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+
+	  ComplexD ip;
+
+	  hermop.Op(Mn,tmp);
+	  ip= innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+	  hermop.AdjOp(Mn,tmp); 
+	  ip = innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+	  
 	  b++;
 	}
 
@@ -228,6 +294,70 @@ public:
     }
     assert(b==nn);
   }
+
+
+  virtual void CreateSubspacePolyCheby(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo1,
+				       int orderfilter,
+				       double lo2,
+				       int orderstep)
+  {
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" CreateSubspacePolyCheby "<<std::endl;
+    // Initial matrix element
+    hermop.Op(noise,Mn);
+    std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      // Filter
+      std::cout << GridLogMessage << "Cheby "<<lo1<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo1,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|n> "<<norm2(Mn)<<std::endl;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    for(int n=1;n<nn;n++){
+      std::cout << GridLogMessage << "Cheby "<<lo2<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo2,hi,orderstep);
+      Cheb(hermop,subspace[n-1],Mn);
+
+      for(int m=0;m<n;m++){
+	ComplexD c = innerProduct(subspace[m],Mn);
+	Mn = Mn - c*subspace[m];
+      }
+      
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5);
+      Mn=Mn*scale;
+      
+      subspace[n]=Mn;
+      
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|n> "<<norm2(Mn)<<std::endl;
+
+    }
+  }
+
   virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
 				       int nn,
 				       double hi,

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

@@ -1,16 +1,15 @@
 #include <Grid/GridCore.h>
 #ifndef GRID_UVM
 
-#warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
 
 #define MAXLINE 512
 static char print_buffer [ MAXLINE ];
 
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug << print_buffer;
+#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
@@ -111,7 +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);
@@ -121,7 +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);
@@ -141,7 +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;
@@ -155,7 +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++;
@@ -169,7 +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 AccPtr %lx -> CpuPtr %lx\n",(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;
@@ -184,7 +183,9 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: acceleratorCopyToDevice   Clone AccPtr %lx <- CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  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);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -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 : %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 ;
@@ -127,7 +129,7 @@ public:
   void GlobalSumVector(ComplexD *c,int N);
   void GlobalXOR(uint32_t &);
   void GlobalXOR(uint64_t &);
-  
+
   template<class obj> void GlobalSumP2P(obj &o)
   {
     std::vector<obj> column;
@@ -136,7 +138,7 @@ public:
     for(int d=0;d<_ndimension;d++){
       column.resize(_processors[d]);
       column[0] = accum;
-      std::vector<CommsRequest_t> list;
+      std::vector<MpiCommsRequest_t> list;
       for(int p=1;p<_processors[d];p++){
 	ShiftedRanks(d,p,source,dest);
 	SendToRecvFromBegin(list,
@@ -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];
       }
@@ -166,8 +169,8 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
-  void CommsComplete(std::vector<CommsRequest_t> &list);
-  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+  void CommsComplete(std::vector<MpiCommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
 			   void *xmit,
 			   int dest,
 			   void *recv,
@@ -186,6 +189,17 @@ public:
 			       int recv_from_rank,int do_recv,
 			       int bytes,int dir);
 
+  double StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+				      void *xmit,
+				      int xmit_to_rank,int do_xmit,
+				      void *recv,
+				      int recv_from_rank,int do_recv,
+				      int xbytes,int rbytes,int dir);
+
+  // Could do a PollHtoD and have a CommsMerge dependence
+  void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
+  void StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list);
+
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
 				    int xmit_to_rank,int do_xmit,

Grid/communicator/Communicator_mpi3.cc

@@ -30,6 +30,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
 ////////////////////////////////////////////
@@ -259,32 +260,39 @@ CartesianCommunicator::~CartesianCommunicator()
 }
 #ifdef USE_GRID_REDUCTION
 void CartesianCommunicator::GlobalSum(float &f){
+  FlightRecorder::StepLog("GlobalSumP2P");
   CartesianCommunicator::GlobalSumP2P(f);
 }
 void CartesianCommunicator::GlobalSum(double &d)
 {
+  FlightRecorder::StepLog("GlobalSumP2P");
   CartesianCommunicator::GlobalSumP2P(d);
 }
 #else
 void CartesianCommunicator::GlobalSum(float &f){
+  FlightRecorder::StepLog("AllReduce");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSum(double &d)
 {
+  FlightRecorder::StepLog("AllReduce");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
   assert(ierr==0);
 }
 #endif
 void CartesianCommunicator::GlobalSum(uint32_t &u){
+  FlightRecorder::StepLog("AllReduce");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSum(uint64_t &u){
+  FlightRecorder::StepLog("AllReduce");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
+  FlightRecorder::StepLog("AllReduceVector");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
@@ -317,7 +325,7 @@ void CartesianCommunicator::GlobalSumVector(double *d,int N)
   assert(ierr==0);
 }
 
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
 						void *xmit,
 						int dest,
 						void *recv,
@@ -342,7 +350,7 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
   assert(ierr==0);
   list.push_back(xrq);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
+void CartesianCommunicator::CommsComplete(std::vector<MpiCommsRequest_t> &list)
 {
   int nreq=list.size();
 
@@ -361,9 +369,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int from,
 					   int bytes)
 {
-  std::vector<CommsRequest_t> reqs(0);
-  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  std::vector<MpiCommsRequest_t> reqs(0);
 
   int myrank = _processor;
   int ierr;
@@ -379,9 +385,6 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 		    communicator,MPI_STATUS_IGNORE);
   assert(ierr==0);
 
-  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
@@ -391,12 +394,287 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  offbytes       += StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
-#undef NVLINK_GET // Define to use get instead of put DMA
+
+#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,
+							   void *recv,
+							   int from,int dor,
+							   int xbytes,int rbytes,int dir)
+{
+  return 0.0; // Do nothing -- no preparation required
+}
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,
+							 int dest,int dox,
+							 void *recv,
+							 int from,int dor,
+							 int xbytes,int rbytes,int dir)
+{
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+  
+  if ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      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;
+      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      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;
+}
+
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
+{
+  int nreq=list.size();
+  /*finishes Get/Put*/
+  acceleratorCopySynchronise();
+
+  if (nreq==0) return;
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+  list.resize(0);
+  this->StencilBarrier(); 
+}
+
+#else /* NOT     ... ACCELERATOR_AWARE_MPI */
+///////////////////////////////////////////
+// Pipeline mode through host memory
+///////////////////////////////////////////
+  /*
+   * In prepare (phase 1):
+   * PHASE 1: (prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   * - post device - device transfers
+   * PHASE 3: (Complete)
+   * - MPI_waitall
+   * - host-device transfers
+   *
+   *********************************
+   * NB could split this further:
+   *--------------------------------
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (BeginInterNode)
+   * - complete all copies 
+   * - post MPI send asynch
+   * PHASE 3: (BeginIntraNode)
+   * - post device - device transfers
+   * PHASE 4: (Complete)
+   * - MPI_waitall
+   * - host-device transfers asynch
+   * - (complete all copies) 
+   */
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int dest,int dox,
+							   void *recv,
+							   int from,int dor,
+							   int xbytes,int rbytes,int dir)
+{
+/*
+ * Bring sequence from Stencil.h down to lower level.
+ * Assume using XeLink is ok
+ */  
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_recv = NULL;
+  void * host_xmit = NULL;
+
+  /*
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   */
+  
+  if ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      host_recv = this->HostBufferMalloc(rbytes);
+      ierr=MPI_Irecv(host_recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeRecv;
+      srq.bytes      = rbytes;
+      srq.req        = rrq;
+      srq.host_buf   = host_recv;
+      srq.device_buf = recv;
+      list.push_back(srq);
+      off_node_bytes+=rbytes;
+    }
+  }
+  
+  if (dox) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+
+      tag= dir+_processor*32;
+
+      host_xmit = this->HostBufferMalloc(xbytes);
+      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;
+
+      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);
+    }
+  }
+
+  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,
 							 int dest,int dox,
@@ -421,56 +699,109 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( dor ) {
-    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+from*32;
-      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-      assert(ierr==0);
-      list.push_back(rrq);
-      off_node_bytes+=rbytes;
-    }
+  void * host_xmit = NULL;
+
+  ////////////////////////////////
+  // Receives already posted
+  // Copies already started
+  ////////////////////////////////
+  /*  
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   */
 #ifdef NVLINK_GET
+  if ( dor ) {
+
+    if ( ! ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
       void *shm = (void *) this->ShmBufferTranslate(from,xmit);
       assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
-#endif
-  }
-  
+
+      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) {
-    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-      tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
-      list.push_back(xrq);
-      off_node_bytes+=xbytes;
-    } else {
-#ifndef NVLINK_GET
+
+    if ( !( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
       assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
-#endif
+
+      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
 
-  acceleratorCopySynchronise();
+  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
+  }
 
-  if (nreq==0) return;
+  int nreq=MpiRequests.size();
 
-  std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
-  list.resize(0);
+  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);
+  }
+  
+  //  for(int r=0;r<nreq;r++){
+  //    if ( list[r].PacketType==InterNodeRecv ) {
+  //      acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
+  //    }
+  //  }
+  
+  
+  list.resize(0);               // Delete the list
+  this->HostBufferFreeAll();    // Clean up the buffer allocs
+#ifndef NVLINK_GET
+  this->StencilBarrier(); // if PUT must check our nbrs have filled our receive buffers.
+#endif   
 }
+#endif
+////////////////////////////////////////////
+// END PIPELINE MODE / NO CUDA AWARE MPI
+////////////////////////////////////////////
+
 void CartesianCommunicator::StencilBarrier(void)
 {
+  FlightRecorder::StepLog("NodeBarrier");
   MPI_Barrier  (ShmComm);
 }
 //void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
@@ -478,11 +809,13 @@ void CartesianCommunicator::StencilBarrier(void)
 //}
 void CartesianCommunicator::Barrier(void)
 {
+  FlightRecorder::StepLog("GridBarrier");
   int ierr = MPI_Barrier(communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 {
+  FlightRecorder::StepLog("Broadcast");
   int ierr=MPI_Bcast(data,
 		     bytes,
 		     MPI_BYTE,
@@ -501,6 +834,7 @@ void CartesianCommunicator::BarrierWorld(void){
 }
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
+  FlightRecorder::StepLog("BroadcastWorld");
   int ierr= MPI_Bcast(data,
 		      bytes,
 		      MPI_BYTE,
@@ -523,6 +857,7 @@ void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,
 }
 void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
 {
+  FlightRecorder::StepLog("AllToAll");
   // MPI is a pain and uses "int" arguments
   // 64*64*64*128*16 == 500Million elements of data.
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.

Grid/communicator/Communicator_none.cc

@@ -91,7 +91,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(list.size()==0);}
 void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
 						int dest,
@@ -132,6 +132,17 @@ 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,
+							   void *recv,
+							   int recv_from_rank,int dor,
+							   int xbytes,int rbytes, int dir)
+{
+  return 0.0;
+}
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int xmit_to_rank,int dox,

Grid/communicator/SharedMemory.h

@@ -46,8 +46,40 @@ NAMESPACE_BEGIN(Grid);
 
 #if defined (GRID_COMMS_MPI3) 
 typedef MPI_Comm    Grid_MPI_Comm;
+typedef MPI_Request MpiCommsRequest_t;
+#ifdef ACCELERATOR_AWARE_MPI
 typedef MPI_Request CommsRequest_t;
+#else
+/*
+ * 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
+
 #else 
+typedef int MpiCommsRequest_t;
 typedef int CommsRequest_t;
 typedef int Grid_MPI_Comm;
 #endif
@@ -105,7 +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

@@ -42,6 +42,11 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef ACCELERATOR_AWARE_MPI
 #define GRID_SYCL_LEVEL_ZERO_IPC
 #define SHM_SOCKETS
+#else
+#ifdef HAVE_NUMAIF_H
+  #warning " Using NUMAIF "
+#include <numaif.h>
+#endif 
 #endif 
 #include <syscall.h>
 #endif
@@ -537,7 +542,38 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
 #ifndef ACCELERATOR_AWARE_MPI
-  HostCommBuf= malloc(bytes);
+  // printf("Host buffer allocate for GPU non-aware MPI\n");
+#if 0
+  HostCommBuf= acceleratorAllocHost(bytes);
+#else 
+  HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
+#if 0
+  #warning "Moving host buffers to specific NUMA domain"
+  int numa;
+  char *numa_name=(char *)getenv("MPI_BUF_NUMA");
+  if(numa_name) {
+    unsigned long page_size = sysconf(_SC_PAGESIZE);
+    numa = atoi(numa_name);
+    unsigned long page_count = bytes/page_size;
+    std::vector<void *> pages(page_count);
+    std::vector<int>    nodes(page_count,numa);
+    std::vector<int>    status(page_count,-1);
+    for(unsigned long p=0;p<page_count;p++){
+      pages[p] =(void *) ((uint64_t) HostCommBuf + p*page_size);
+    }
+    int ret = move_pages(0,
+			 page_count,
+			 &pages[0],
+			 &nodes[0],
+			 &status[0],
+			 MPOL_MF_MOVE);
+    printf("Host buffer move to numa domain %d : move_pages returned %d\n",numa,ret);
+    if (ret) perror(" move_pages failed for reason:");
+  }
+#endif  
+  acceleratorPin(HostCommBuf,bytes);
+#endif  
+
 #endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
@@ -880,14 +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
@@ -923,6 +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();
 
@@ -953,7 +990,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
 #endif
 
-  //SharedMemoryTest();
+  //  SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier
@@ -975,19 +1012,18 @@ void SharedMemory::SharedMemoryTest(void)
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
        check[2]=magic;
-       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
+       acceleratorCopyToDevice(check,ShmCommBufs[r],3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
   for(uint64_t r=0;r<ShmSize;r++){
-    ShmBarrier();
-    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
-    ShmBarrier();
+    acceleratorCopyFromDevice(ShmCommBufs[r],check,3*sizeof(uint64_t));
     assert(check[0]==GlobalSharedMemory::WorldNode);
     assert(check[1]==r);
     assert(check[2]==magic);
-    ShmBarrier();
   }
+  ShmBarrier();
+  std::cout << GridLogDebug << " SharedMemoryTest has passed "<<std::endl;
 }
 
 void *SharedMemory::ShmBuffer(int rank)

Grid/communicator/SharedMemoryNone.cc

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

Grid/cshift/Cshift_mpi.h

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

Grid/lattice/Lattice_base.h

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

Grid/lattice/Lattice_reduction.h

@@ -376,9 +376,9 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
       coalescedWrite(z_v[ss],tmp);
   });
   bool ok;
+#ifdef GRID_SYCL
   uint64_t csum=0;
   uint64_t csum2=0;
-#ifdef GRID_SYCL
   if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
   {
     // z_v
@@ -522,14 +522,11 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
   int ostride=grid->_ostride[orthogdim];
   
   //Reduce Data down to lvSum
-  RealD t_sum =-usecond();
   sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
-  t_sum +=usecond();
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
   Coordinate icoor(Nd);
 
-  RealD t_rest =-usecond();
   for(int rt=0;rt<rd;rt++){
 
     extract(lvSum[rt],extracted);
@@ -559,8 +556,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
   scalar_type * ptr = (scalar_type *) &result[0];
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
-  t_rest +=usecond();
-  std::cout << GridLogMessage << " sliceSum local"<<t_sum<<" us, host+mpi "<<t_rest<<std::endl;
+  //  std::cout << GridLogMessage << " sliceSum local"<<t_sum<<" us, host+mpi "<<t_rest<<std::endl;
   
 }
 template<class vobj> inline

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

@@ -466,9 +466,15 @@ 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<CommsRequest_t> fwd_req;   
-    std::vector<CommsRequest_t> bwd_req;   
+    std::vector<MpiCommsRequest_t> fwd_req;   
+    std::vector<MpiCommsRequest_t> bwd_req;   
 
     int words = buffer_size;
     int bytes = words * sizeof(vobj);
@@ -495,9 +501,16 @@ public:
       t_gather+=usecond()-t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(fwd_req,
 				(void *)&send_buf[d*buffer_size], xmit_to_rank,
 				(void *)&recv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[d*buffer_size],&hsend_buf[d*buffer_size],bytes);
+      grid->SendToRecvFromBegin(fwd_req,
+				(void *)&hsend_buf[d*buffer_size], xmit_to_rank,
+				(void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#endif
       t_comms+=usecond()-t;
      }
     for ( int d=0;d < depth ; d ++ ) {
@@ -508,9 +521,16 @@ public:
       t_gather+= usecond() - t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(bwd_req,
 				(void *)&send_buf[(d+depth)*buffer_size], recv_from_rank,
 				(void *)&recv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[(d+depth)*buffer_size],&hsend_buf[(d+depth)*buffer_size],bytes);
+      grid->SendToRecvFromBegin(bwd_req,
+				(void *)&hsend_buf[(d+depth)*buffer_size], recv_from_rank,
+				(void *)&hrecv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#endif      
       t_comms+=usecond()-t;
     }
 
@@ -533,8 +553,13 @@ public:
 
     t=usecond();
     grid->CommsComplete(fwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
-
+    
     t=usecond();
     for ( int d=0;d < depth ; d ++ ) {
       ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++;
@@ -543,6 +568,11 @@ public:
 
     t=usecond();
     grid->CommsComplete(bwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
     
     t=usecond();

Grid/qcd/action/ActionBase.h

@@ -98,7 +98,7 @@ public:
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
   virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
-
+ 
   /////////////////////////////////////////////////////////////
   // virtual smeared interface through configuration container
   /////////////////////////////////////////////////////////////
@@ -132,6 +132,10 @@ public:
 template <class GaugeField >
 class EmptyAction : public Action <GaugeField>
 {
+  using Action<GaugeField>::refresh;
+  using Action<GaugeField>::Sinitial;
+  using Action<GaugeField>::deriv;
+
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) { assert(0);}; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) { return 0.0;};                             // evaluate the action
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) { assert(0); };        // evaluate the action derivative

Grid/qcd/action/fermion/AbstractEOFAFermion.h

@@ -55,6 +55,11 @@ public:
   RealD alpha; // Mobius scale
   RealD k;     // EOFA normalization constant
 
+  // Device resident
+  deviceVector<Coeff_t> d_shift_coefficients;
+  deviceVector<Coeff_t> d_MooeeInv_shift_lc;
+  deviceVector<Coeff_t> d_MooeeInv_shift_norm;
+  
   virtual void Instantiatable(void) = 0;
 
   // EOFA-specific operations
@@ -92,6 +97,11 @@ public:
     this->k = this->alpha * (_mq3-_mq2) * std::pow(this->alpha+1.0,2*Ls) /
       ( std::pow(this->alpha+1.0,Ls) + _mq2*std::pow(this->alpha-1.0,Ls) ) /
       ( std::pow(this->alpha+1.0,Ls) + _mq3*std::pow(this->alpha-1.0,Ls) );
+    
+    d_shift_coefficients.resize(Ls);
+    d_MooeeInv_shift_lc.resize(Ls);
+    d_MooeeInv_shift_norm.resize(Ls);
+
   };
 };
 

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -124,6 +124,11 @@ public:
   RealD                _b;
   RealD                _c;
 
+  // possible boost
+  std::vector<ComplexD> qmu;
+  void set_qmu(std::vector<ComplexD> _qmu) { qmu=_qmu; assert(qmu.size()==Nd);};
+  void addQmu(const FermionField &in, FermionField &out, int dag);
+  
   // Cayley form Moebius (tanh and zolotarev)
   std::vector<Coeff_t> omega;
   std::vector<Coeff_t> bs;    // S dependent coeffs
@@ -143,6 +148,17 @@ public:
   std::vector<Coeff_t> ueem;
   std::vector<Coeff_t> dee;
 
+  // Device memory
+  deviceVector<Coeff_t> d_diag;
+  deviceVector<Coeff_t> d_upper;
+  deviceVector<Coeff_t> d_lower;
+
+  deviceVector<Coeff_t> d_lee;
+  deviceVector<Coeff_t> d_dee;
+  deviceVector<Coeff_t> d_uee;
+  deviceVector<Coeff_t> d_leem;
+  deviceVector<Coeff_t> d_ueem;
+
   // Matrices of 5d ee inverse params
   //  std::vector<iSinglet<Simd> >  MatpInv;
   //  std::vector<iSinglet<Simd> >  MatmInv;

Grid/qcd/action/fermion/CompactWilsonCloverFermion5D.h

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

Grid/qcd/action/fermion/ContinuedFractionFermion5D.h

@@ -60,6 +60,50 @@ public:
   //      virtual void   Instantiatable(void)=0;
   virtual void   Instantiatable(void) =0;
 
+  void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary, std::vector<double> twist)
+  {
+    std::cout << "Free Propagator for PartialFraction"<<std::endl;
+    FermionField in_k(in.Grid());
+    FermionField prop_k(in.Grid());
+    
+    FFT theFFT((GridCartesian *) in.Grid());
+
+    //phase for boundary condition
+    ComplexField coor(in.Grid());
+    ComplexField ph(in.Grid());  ph = Zero();
+    FermionField in_buf(in.Grid()); in_buf = Zero();
+    typedef typename Simd::scalar_type Scalar;
+    Scalar ci(0.0,1.0);
+    assert(twist.size() == Nd);//check that twist is Nd
+    assert(boundary.size() == Nd);//check that boundary conditions is Nd
+    int shift = 0;
+    for(unsigned int nu = 0; nu < Nd; nu++)
+      {
+	// Shift coordinate lattice index by 1 to account for 5th dimension.
+	LatticeCoordinate(coor, nu + shift);
+	double boundary_phase = ::acos(real(boundary[nu]));
+	ph = ph + boundary_phase*coor*((1./(in.Grid()->_fdimensions[nu+shift])));
+	//momenta for propagator shifted by twist+boundary
+	twist[nu] = twist[nu] + boundary_phase/((2.0*M_PI));
+      }
+    in_buf = exp(ci*ph*(-1.0))*in;
+
+    theFFT.FFT_all_dim(in_k,in,FFT::forward);
+    this->MomentumSpacePropagatorHw(prop_k,in_k,mass,twist);
+    theFFT.FFT_all_dim(out,prop_k,FFT::backward);
+    
+    //phase for boundary condition
+    out = out * exp(ci*ph);
+  };
+
+  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass) {
+    std::vector<double> twist(Nd,0.0); //default: periodic boundarys in all directions
+    std::vector<Complex> boundary;
+    for(int i=0;i<Nd;i++) boundary.push_back(1);//default: periodic boundary conditions
+    FreePropagator(in,out,mass,boundary,twist);
+  };
+
+  
   // Efficient support for multigrid coarsening
   virtual void  Mdir (const FermionField &in, FermionField &out,int dir,int disp);
   virtual void  MdirAll(const FermionField &in, std::vector<FermionField> &out);

Grid/qcd/action/fermion/Fermion.h

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

Grid/qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h

@@ -42,7 +42,7 @@ public:
 
      void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
        this->MomentumSpacePropagatorHw(out,in,_m,twist);
-  };
+     };
 
   // Constructors
   OverlapWilsonCayleyTanhFermion(GaugeField &_Umu,

Grid/qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h

@@ -41,6 +41,10 @@ public:
 public:
 
   // Constructors
+  virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
 
   OverlapWilsonCayleyZolotarevFermion(GaugeField &_Umu,
 				      GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonContfracTanhFermion.h

@@ -41,6 +41,9 @@ public:
 public:
 
   virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
   // Constructors
   OverlapWilsonContFracTanhFermion(GaugeField &_Umu,
 				   GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h

@@ -40,6 +40,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
 
   virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
   // Constructors
   OverlapWilsonContFracZolotarevFermion(GaugeField &_Umu,
 					GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h

@@ -41,6 +41,9 @@ public:
 public:
 
   virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
   // Constructors
   OverlapWilsonPartialFractionTanhFermion(GaugeField &_Umu,
 					  GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h

@@ -40,6 +40,11 @@ public:
   INHERIT_IMPL_TYPES(Impl);
 
   virtual void   Instantiatable(void){};
+
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
+
   // Constructors
   OverlapWilsonPartialFractionZolotarevFermion(GaugeField &_Umu,
 					       GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/PartialFractionFermion5D.h

@@ -39,7 +39,7 @@ class PartialFractionFermion5D : public WilsonFermion5D<Impl>
 public:
   INHERIT_IMPL_TYPES(Impl);
 
-  const int part_frac_chroma_convention=1;
+  const int part_frac_chroma_convention=0;
 
   void   Meooe_internal(const FermionField &in, FermionField &out,int dag);
   void   Mooee_internal(const FermionField &in, FermionField &out,int dag);
@@ -83,11 +83,70 @@ public:
 			   GridRedBlackCartesian &FourDimRedBlackGrid,
 			   RealD _mass,RealD M5,const ImplParams &p= ImplParams());
 
+  PartialFractionFermion5D(GaugeField &_Umu,
+			   GridCartesian         &FiveDimGrid,
+			   GridRedBlackCartesian &FiveDimRedBlackGrid,
+			   GridCartesian         &FourDimGrid,
+			   GridRedBlackCartesian &FourDimRedBlackGrid,
+			   RealD _mass,RealD M5,std::vector<RealD> &_qmu,const ImplParams &p= ImplParams());
+
+  void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary, std::vector<double> twist)
+  {
+    std::cout << "Free Propagator for PartialFraction"<<std::endl;
+    FermionField in_k(in.Grid());
+    FermionField prop_k(in.Grid());
+    
+    FFT theFFT((GridCartesian *) in.Grid());
+
+    //phase for boundary condition
+    ComplexField coor(in.Grid());
+    ComplexField ph(in.Grid());  ph = Zero();
+    FermionField in_buf(in.Grid()); in_buf = Zero();
+    typedef typename Simd::scalar_type Scalar;
+    Scalar ci(0.0,1.0);
+    assert(twist.size() == Nd);//check that twist is Nd
+    assert(boundary.size() == Nd);//check that boundary conditions is Nd
+    int shift = 0;
+    for(unsigned int nu = 0; nu < Nd; nu++)
+      {
+	// Shift coordinate lattice index by 1 to account for 5th dimension.
+	LatticeCoordinate(coor, nu + shift);
+	double boundary_phase = ::acos(real(boundary[nu]));
+	ph = ph + boundary_phase*coor*((1./(in.Grid()->_fdimensions[nu+shift])));
+	//momenta for propagator shifted by twist+boundary
+	twist[nu] = twist[nu] + boundary_phase/((2.0*M_PI));
+      }
+    in_buf = exp(ci*ph*(-1.0))*in;
+
+    theFFT.FFT_all_dim(in_k,in,FFT::forward);
+    if ( this->qmu.size() ){
+      this->MomentumSpacePropagatorHwQ(prop_k,in_k,mass,twist,this->qmu);
+    } else {
+      this->MomentumSpacePropagatorHw(prop_k,in_k,mass,twist);
+    }
+    theFFT.FFT_all_dim(out,prop_k,FFT::backward);
+    
+    //phase for boundary condition
+    out = out * exp(ci*ph);
+  };
+
+  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass) {
+    std::vector<double> twist(Nd,0.0); //default: periodic boundarys in all directions
+    std::vector<Complex> boundary;
+    for(int i=0;i<Nd;i++) boundary.push_back(1);//default: periodic boundary conditions
+    FreePropagator(in,out,mass,boundary,twist);
+  };
+
+  void set_qmu(std::vector<RealD> _qmu) { qmu=_qmu; assert(qmu.size()==Nd);};
+  void addQmu(const FermionField &in, FermionField &out, int dag);
+
 protected:
 
   virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD scale);
   virtual void SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata);
 
+  std::vector<RealD> qmu;
+
   // Part frac
   RealD mass;
   RealD dw_diag;

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -414,29 +414,6 @@ public:
     //    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
-
-  /*
-  void BuildSurfaceList(int Ls,int vol4){
-
-    // find same node for SHM
-    // Here we know the distance is 1 for WilsonStencil
-    for(int point=0;point<this->_npoints;point++){
-      this->same_node[point] = this->SameNode(point);
-    }
-    
-    for(int site = 0 ;site< vol4;site++){
-      int local = 1;
-      for(int point=0;point<this->_npoints;point++){
-	if( (!this->GetNodeLocal(site*Ls,point)) && (!this->same_node[point]) ){ 
-	  local = 0;
-	}
-      }
-      if(local == 0) { 
-	surface_list.push_back(site);
-      }
-    }
-  }
-  */
   
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
@@ -507,6 +484,11 @@ public:
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
     accelerator_barrier();
+#ifdef NVLINK_GET
+    this->_grid->StencilBarrier(); // He can now get mu local gather, I can get his
+    // Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
+    // Or issue barrier AFTER the DMA is running
+#endif    
   }
 
 };

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -91,13 +91,13 @@ public:
   virtual void   Mdag (const FermionField &in, FermionField &out){assert(0);};
 
   // half checkerboard operations; leave unimplemented as abstract for now
-  virtual void   Meooe       (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   Mooee       (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   MooeeInv    (const FermionField &in, FermionField &out){assert(0);};
+  virtual void   Meooe       (const FermionField &in, FermionField &out);
+  virtual void   Mooee       (const FermionField &in, FermionField &out);
+  virtual void   MooeeInv    (const FermionField &in, FermionField &out);
 
-  virtual void   MeooeDag    (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   MooeeDag    (const FermionField &in, FermionField &out){assert(0);};
-  virtual void   MooeeInvDag (const FermionField &in, FermionField &out){assert(0);};
+  virtual void   MeooeDag    (const FermionField &in, FermionField &out);
+  virtual void   MooeeDag    (const FermionField &in, FermionField &out);
+  virtual void   MooeeInvDag (const FermionField &in, FermionField &out);
   virtual void   Mdir   (const FermionField &in, FermionField &out,int dir,int disp){assert(0);};   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
   virtual void   MdirAll(const FermionField &in, std::vector<FermionField> &out){assert(0);};   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
 
@@ -109,6 +109,8 @@ public:
   void MomentumSpacePropagatorHt_5d(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist) ;
   void MomentumSpacePropagatorHt(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist) ;
   void MomentumSpacePropagatorHw(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist) ;
+  void MomentumSpacePropagatorHwQ(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist,
+				  std::vector<double> qmu) ;
 
   // Implement hopping term non-hermitian hopping term; half cb or both
   // Implement s-diagonal DW
@@ -117,6 +119,9 @@ public:
   void DhopOE(const FermionField &in, FermionField &out,int dag);
   void DhopEO(const FermionField &in, FermionField &out,int dag);
 
+  void DhopComms  (const FermionField &in, FermionField &out);
+  void DhopCalc   (const FermionField &in, FermionField &out,uint64_t *ids);
+  
   // add a DhopComm
   // -- suboptimal interface will presently trigger multiple comms.
   void DhopDir(const FermionField &in, FermionField &out,int dir,int disp);

Grid/qcd/action/fermion/WilsonKernels.h

@@ -57,6 +57,10 @@ public:
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,
 			 int interior=1,int exterior=1) ;
 
+  static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
+			 int Ls, int Nsite, const FermionField &in, FermionField &out,
+			 uint64_t *ids);
+  
   static void DhopDagKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			    int Ls, int Nsite, const FermionField &in, FermionField &out,
 			    int interior=1,int exterior=1) ;

Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h

@@ -48,7 +48,8 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
   mass_plus(_mass), mass_minus(_mass)
-{ 
+{
+  // qmu defaults to zero size;
 }
 
 ///////////////////////////////////////////////////////////////
@@ -270,6 +271,34 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   M5Ddag(psi,psi,Din,lower,diag,upper);
 }
 
+template<class Impl>
+void CayleyFermion5D<Impl>::addQmu(const FermionField &psi,FermionField &chi, int dag)
+{
+  if ( qmu.size() ) {
+
+    Gamma::Algebra Gmu [] = {
+      Gamma::Algebra::GammaX,
+      Gamma::Algebra::GammaY,
+      Gamma::Algebra::GammaZ,
+      Gamma::Algebra::GammaT
+    };
+    std::vector<ComplexD> coeff(Nd);
+    ComplexD ci(0,1);
+
+    assert(qmu.size()==Nd);
+
+    for(int mu=0;mu<Nd;mu++){
+       coeff[mu] = ci*qmu[mu];
+       if ( dag ) coeff[mu] = conjugate(coeff[mu]);
+    }
+
+    chi = chi + Gamma(Gmu[0])*psi*coeff[0];
+    for(int mu=1;mu<Nd;mu++){
+      chi = chi + Gamma(Gmu[mu])*psi*coeff[mu];
+    }
+  }
+}
+
 template<class Impl>
 void CayleyFermion5D<Impl>::M    (const FermionField &psi, FermionField &chi)
 {
@@ -277,8 +306,12 @@ void CayleyFermion5D<Impl>::M    (const FermionField &psi, FermionField &chi)
   
   // Assemble Din
   Meooe5D(psi,Din);
-  
+
   this->DW(Din,chi,DaggerNo);
+
+  // add i q_mu gamma_mu here
+  addQmu(Din,chi,DaggerNo);
+  
   // ((b D_W + D_w hop terms +1) on s-diag
   axpby(chi,1.0,1.0,chi,psi); 
   
@@ -295,6 +328,9 @@ void CayleyFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
   FermionField Din(psi.Grid());
   // Apply Dw
   this->DW(psi,Din,DaggerYes); 
+
+  // add -i conj(q_mu) gamma_mu here ... if qmu is real, gammm_5 hermitian, otherwise not.
+  addQmu(psi,Din,DaggerYes);
   
   MeooeDag5D(Din,chi);
   
@@ -488,7 +524,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
   leem.resize(Ls);
   uee.resize(Ls);
   ueem.resize(Ls);
-  
+
   for(int i=0;i<Ls;i++){
     
     dee[i] = bee[i];
@@ -529,6 +565,18 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
     dee[Ls-1] += delta_d;
   }  
 
+  //////////////////////////////////////////
+  // Device buffers
+  //////////////////////////////////////////
+  d_diag.resize(Ls);
+  d_upper.resize(Ls);
+  d_lower.resize(Ls);
+
+  d_dee.resize(Ls);
+  d_lee.resize(Ls);
+  d_uee.resize(Ls);
+  d_leem.resize(Ls);
+  d_ueem.resize(Ls);
   //  int inv=1;
   //  this->MooeeInternalCompute(0,inv,MatpInv,MatmInv);
   //  this->MooeeInternalCompute(1,inv,MatpInvDag,MatmInvDag);

Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h

@@ -57,9 +57,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
 
   int Ls =this->Ls;
 
-  static deviceVector<Coeff_t> d_diag(Ls) ; acceleratorCopyToDevice(&diag[0] ,&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls); acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls); acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&diag[0] ,&this->d_diag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&this->d_upper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&this->d_lower[0],Ls*sizeof(Coeff_t));
   
   auto pdiag = &d_diag[0];
   auto pupper = &d_upper[0];
@@ -99,9 +99,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
 
   int Ls=this->Ls;
 
-  static deviceVector<Coeff_t> d_diag(Ls) ; acceleratorCopyToDevice(&diag[0] ,&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls); acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls); acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&diag[0] ,&this->d_diag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&this->d_upper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&this->d_lower[0],Ls*sizeof(Coeff_t));
   
   auto pdiag = &d_diag[0];
   auto pupper = &d_upper[0];
@@ -134,11 +134,11 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
 
   int Ls=this->Ls;
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
 
   auto plee  = & d_lee [0];
   auto pdee  = & d_dee [0];
@@ -196,11 +196,11 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
   autoView(psi , psi_i,AcceleratorRead);
   autoView(chi , chi_i,AcceleratorWrite);
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
 
   auto plee  = & d_lee [0];
   auto pdee  = & d_dee [0];

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

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

Grid/qcd/action/fermion/implementation/ContinuedFractionFermion5DImplementation.h

@@ -42,13 +42,13 @@ template<class Impl>
 void ContinuedFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata)
 {
   // How to check Ls matches??
-  //      std::cout<<GridLogMessage << Ls << " Ls"<<std::endl;
-  //      std::cout<<GridLogMessage << zdata->n  << " - n"<<std::endl;
-  //      std::cout<<GridLogMessage << zdata->da << " -da "<<std::endl;
-  //      std::cout<<GridLogMessage << zdata->db << " -db"<<std::endl;
-  //      std::cout<<GridLogMessage << zdata->dn << " -dn"<<std::endl;
-  //      std::cout<<GridLogMessage << zdata->dd << " -dd"<<std::endl;
+  std::cout<<GridLogMessage << zdata->n  << " - n"<<std::endl;
+  std::cout<<GridLogMessage << zdata->da << " -da "<<std::endl;
+  std::cout<<GridLogMessage << zdata->db << " -db"<<std::endl;
+  std::cout<<GridLogMessage << zdata->dn << " -dn"<<std::endl;
+  std::cout<<GridLogMessage << zdata->dd << " -dd"<<std::endl;
   int Ls = this->Ls;
+  std::cout<<GridLogMessage << Ls << " Ls"<<std::endl;
   assert(zdata->db==Ls);// Beta has Ls coeffs
 
   R=(1+this->mass)/(1-this->mass);
@@ -320,7 +320,7 @@ ContinuedFractionFermion5D<Impl>::ContinuedFractionFermion5D(
       int Ls = this->Ls;
       conformable(solution5d.Grid(),this->FermionGrid());
       conformable(exported4d.Grid(),this->GaugeGrid());
-      ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
+      ExtractSlice(exported4d, solution5d, Ls-1, 0);
     }
     template<class Impl>
     void ContinuedFractionFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
@@ -330,7 +330,7 @@ ContinuedFractionFermion5D<Impl>::ContinuedFractionFermion5D(
       conformable(input4d.Grid()   ,this->GaugeGrid());
       FermionField tmp(this->FermionGrid());
       tmp=Zero();
-      InsertSlice(input4d, tmp, Ls-1, Ls-1);
+      InsertSlice(input4d, tmp, Ls-1, 0);
       tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
       this->Dminus(tmp,imported5d);
     }

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

@@ -51,13 +51,13 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
+  auto pdiag  = &this->d_diag[0];
+  auto pupper = &this->d_upper[0];
+  auto plower = &this->d_lower[0];
+
+  acceleratorCopyToDevice(&diag[0],&pdiag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&pupper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&plower[0],Ls*sizeof(Coeff_t));
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   
@@ -89,14 +89,14 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
   autoView( phi , phi_i, AcceleratorRead);
   autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
-
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
   
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
+  auto pdiag  = &this->d_diag[0];
+  auto pupper = &this->d_upper[0];
+  auto plower = &this->d_lower[0];
+
+  acceleratorCopyToDevice(&diag[0] ,&pdiag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&pupper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&plower[0],Ls*sizeof(Coeff_t));
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
 
@@ -125,18 +125,18 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
   autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
-
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  auto plee  = & this->d_lee [0];
+  auto pdee  = & this->d_dee [0];
+  auto puee  = & this->d_uee [0];
+  auto pleem = & this->d_leem[0];
+  auto pueem = & this->d_ueem[0];
   
+  acceleratorCopyToDevice(&this->lee[0],&plee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->dee[0],&pdee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->uee[0],&puee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->leem[0],&pleem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->ueem[0],&pueem[0],Ls*sizeof(Coeff_t));
+
   uint64_t nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;

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

@@ -50,14 +50,14 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
+  auto pdiag  = &this->d_diag[0];
+  auto pupper = &this->d_upper[0];
+  auto plower = &this->d_lower[0];
 
+  acceleratorCopyToDevice(&diag[0],&pdiag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&pupper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&plower[0],Ls*sizeof(Coeff_t));
+  
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -93,15 +93,15 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
   
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_shift_coeffs(Ls);acceleratorCopyToDevice(&shift_coeffs[0],&d_shift_coeffs[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-  auto pshift_coeffs = &d_shift_coeffs[0];
+  auto pdiag  = &this->d_diag[0];
+  auto pupper = &this->d_upper[0];
+  auto plower = &this->d_lower[0];
+  auto pshift_coeffs = &this->d_shift_coefficients[0];
+
+  acceleratorCopyToDevice(&diag[0],&pdiag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&pupper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&plower[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&shift_coeffs[0],&pshift_coeffs[0],Ls*sizeof(Coeff_t));
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -138,14 +138,14 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
   autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
-
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
   
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
+  auto pdiag  = &this->d_diag[0];
+  auto pupper = &this->d_upper[0];
+  auto plower = &this->d_lower[0];
+
+  acceleratorCopyToDevice(&diag[0],&pdiag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&pupper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&plower[0],Ls*sizeof(Coeff_t));
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -180,16 +180,16 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_shift_coeffs(Ls);acceleratorCopyToDevice(&shift_coeffs[0],&d_shift_coeffs[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-  auto pshift_coeffs = &d_shift_coeffs[0];
+  auto pdiag  = &this->d_diag[0];
+  auto pupper = &this->d_upper[0];
+  auto plower = &this->d_lower[0];
+  auto pshift_coeffs = &this->d_shift_coefficients[0];
 
+  acceleratorCopyToDevice(&diag[0],&pdiag[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&upper[0],&pupper[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&lower[0],&plower[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&shift_coeffs[0],&pshift_coeffs[0],Ls*sizeof(Coeff_t));
+  
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   auto pm = this->pm;
 
@@ -230,17 +230,17 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  auto plee  = & this->d_lee [0];
+  auto pdee  = & this->d_dee [0];
+  auto puee  = & this->d_uee [0];
+  auto pleem = & this->d_leem[0];
+  auto pueem = & this->d_ueem[0];
 
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  acceleratorCopyToDevice(&this->lee[0],&plee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->dee[0],&pdee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->uee[0],&puee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->leem[0],&pleem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->ueem[0],&pueem[0],Ls*sizeof(Coeff_t));
 
   if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
 
@@ -293,23 +293,22 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   autoView(chi , chi_i, AcceleratorWrite);
 
   // Move into object and constructor
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
-
   auto pm = this->pm;
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  auto plee  = & this->d_lee [0];
+  auto pdee  = & this->d_dee [0];
+  auto puee  = & this->d_uee [0];
+  auto pleem = & this->d_leem[0];
+  auto pueem = & this->d_ueem[0];
+  auto pMooeeInv_shift_lc   = &this->d_MooeeInv_shift_lc[0];
+  auto pMooeeInv_shift_norm = &this->d_MooeeInv_shift_norm[0];
 
-  static deviceVector<Coeff_t> d_MooeeInv_shift_lc(Ls); acceleratorCopyToDevice(&MooeeInv_shift_lc[0],&d_MooeeInv_shift_lc[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_MooeeInv_shift_norm(Ls); acceleratorCopyToDevice(&MooeeInv_shift_norm[0],&d_MooeeInv_shift_norm[0],Ls*sizeof(Coeff_t));
-  auto pMooeeInv_shift_lc   = &d_MooeeInv_shift_lc[0];
-  auto pMooeeInv_shift_norm = &d_MooeeInv_shift_norm[0];
+  acceleratorCopyToDevice(&this->lee[0],&plee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->dee[0],&pdee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->uee[0],&puee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->leem[0],&pleem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->ueem[0],&pueem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&MooeeInv_shift_lc[0],&pMooeeInv_shift_lc[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&MooeeInv_shift_norm[0],&pMooeeInv_shift_norm[0],Ls*sizeof(Coeff_t));
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -367,17 +366,17 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  auto plee  = &this->d_lee [0];
+  auto pdee  = &this->d_dee [0];
+  auto puee  = &this->d_uee [0];
+  auto pleem = &this->d_leem[0];
+  auto pueem = &this->d_ueem[0];
 
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  acceleratorCopyToDevice(&this->lee[0],&plee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->dee[0],&pdee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->uee[0],&puee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->leem[0],&pleem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->ueem[0],&pueem[0],Ls*sizeof(Coeff_t));
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -426,25 +425,23 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
   autoView(chi , chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
-
   auto pm = this->pm;
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  auto plee  = & this->d_lee [0];
+  auto pdee  = & this->d_dee [0];
+  auto puee  = & this->d_uee [0];
+  auto pleem = & this->d_leem[0];
+  auto pueem = & this->d_ueem[0];
 
-  static deviceVector<Coeff_t> d_MooeeInvDag_shift_lc(Ls);
-  static deviceVector<Coeff_t> d_MooeeInvDag_shift_norm(Ls);
-  acceleratorCopyToDevice(&MooeeInvDag_shift_lc[0],&d_MooeeInvDag_shift_lc[0],Ls*sizeof(Coeff_t));
-  acceleratorCopyToDevice(&MooeeInvDag_shift_norm[0],&d_MooeeInvDag_shift_norm[0],Ls*sizeof(Coeff_t));
-  auto pMooeeInvDag_shift_lc   = &d_MooeeInvDag_shift_lc[0];
-  auto pMooeeInvDag_shift_norm = &d_MooeeInvDag_shift_norm[0];
+  auto pMooeeInvDag_shift_lc   = &this->d_MooeeInv_shift_lc[0];
+  auto pMooeeInvDag_shift_norm = &this->d_MooeeInv_shift_norm[0];
+
+  acceleratorCopyToDevice(&this->lee[0],&plee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->dee[0],&pdee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->uee[0],&puee[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->leem[0],&pleem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&this->ueem[0],&pueem[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&MooeeInvDag_shift_lc[0],&pMooeeInvDag_shift_lc[0],Ls*sizeof(Coeff_t));
+  acceleratorCopyToDevice(&MooeeInvDag_shift_norm[0],&pMooeeInvDag_shift_norm[0],Ls*sizeof(Coeff_t));
 
   //  auto pMooeeInvDag_shift_lc   = &MooeeInvDag_shift_lc[0];
   //  auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];

Grid/qcd/action/fermion/implementation/PartialFractionFermion5DImplementation.h

@@ -237,7 +237,32 @@ void   PartialFractionFermion5D<Impl>::M_internal(const FermionField &psi, Fermi
   //           ( 0     -sqrt(p_i)*amax   |  2 R gamma_5 + p0/amax 2H
   //
 
-  this->DW(psi,D,DaggerNo); 
+  this->DW(psi,D,DaggerNo);
+
+  // DW - DW+iqslash
+  //  (g5 Dw)^dag = g5 Dw
+  //  (iqmu g5 gmu)^dag = (-i qmu gmu^dag g5^dag) = i qmu g5 gmu
+  if ( qmu.size() ) {
+
+    std::cout<< "Mat" << "qmu ("<<qmu[0]<<","<<qmu[1]<<","<<qmu[2]<<","<<qmu[3]<<")"<<std::endl;
+    assert(qmu.size()==Nd);
+
+    FermionField qslash_psi(psi.Grid());
+
+    Gamma::Algebra Gmu [] = {
+			     Gamma::Algebra::GammaX,
+			     Gamma::Algebra::GammaY,
+			     Gamma::Algebra::GammaZ,
+			     Gamma::Algebra::GammaT
+    };
+    qslash_psi = qmu[0]*(Gamma(Gmu[0])*psi);
+    for(int mu=1;mu<Nd;mu++){
+      qslash_psi = qslash_psi + qmu[mu]*(Gamma(Gmu[mu])*psi);
+    }
+    ComplexD ci(0.0,1.0);
+    qslash_psi = ci*qslash_psi ; // i qslash
+    D = D + qslash_psi;
+  }
 
   int nblock=(Ls-1)/2;
   for(int b=0;b<nblock;b++){
@@ -255,15 +280,55 @@ void   PartialFractionFermion5D<Impl>::M_internal(const FermionField &psi, Fermi
   }
 	
   {
+    // The 'conventional' Cayley overlap operator is
+    //
+    // Dov = (1+m)/2 + (1-m)/2 g5 sgn Hw
+    //
+    //
+    // With massless limit 1/2(1+g5 sgnHw)
+    //
+    // Luscher shows quite neatly that 1+g5 sgn Hw has tree level propagator i qslash +O(a^2)
+    //
+    // However, the conventional normalisation has both a leading order factor of 2 in Zq
+    // at tree level AND a mass dependent (1-m) that are convenient to absorb.
+    //
+    // In WilsonFermion5DImplementation.h, the tree level propagator for Hw is
+    //
+    // num = -i sin kmu gmu
+    //
+    // denom ( sqrt(sk^2 + (2shk^2 - 1)^2
+    //    b_k = sk2 - M5;
+    //     
+    //    w_k = sqrt(sk + b_k*b_k);
+    //
+    //    denom= ( w_k + b_k + mass*mass) ;
+    //
+    //    denom= one/denom;
+    //    out = num*denom;
+    //
+    // Chroma, and Grid define partial fraction via 4d operator
+    //
+    //   Dpf = 2/(1-m) x Dov = (1+m)/(1-m) + g5 sgn Hw
+    //
+    // Now since:
+    //
+    //      (1+m)/(1-m) = (1-m)/(1-m) + 2m/(1-m) = 1 + 2m/(1-m)
+    //
+    // This corresponds to a modified mass parameter
+    //
+    // It has an annoying 
+    //
+    // 
     double R=(1+this->mass)/(1-this->mass);
-    //R g5 psi[Ls] + p[0] H
+    //R g5 psi[Ls] + p[0] Hw
     ag5xpbg5y_ssp(chi,R*scale,psi,p[nblock]*scale/amax,D,Ls-1,Ls-1);
-	
+    
     for(int b=0;b<nblock;b++){
       int s = 2*b+1;
       double pp = p[nblock-1-b];
       axpby_ssp(chi,1.0,chi,-sqrt(amax*pp)*scale*sign,psi,Ls-1,s);
     }
+   
   }
 
 }
@@ -411,17 +476,18 @@ void  PartialFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,App
       int Ls = this->Ls;
       conformable(solution5d.Grid(),this->FermionGrid());
       conformable(exported4d.Grid(),this->GaugeGrid());
-      ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
+      ExtractSlice(exported4d, solution5d, Ls-1, 0);
     }
     template<class Impl>
     void PartialFractionFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
     {
+      //void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int orthog)
       int Ls = this->Ls;
       conformable(imported5d.Grid(),this->FermionGrid());
       conformable(input4d.Grid()   ,this->GaugeGrid());
       FermionField tmp(this->FermionGrid());
       tmp=Zero();
-      InsertSlice(input4d, tmp, Ls-1, Ls-1);
+      InsertSlice(input4d, tmp, Ls-1, 0);
       tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
       this->Dminus(tmp,imported5d);
     }
@@ -442,7 +508,7 @@ PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
 
 {
   int Ls = this->Ls;
-
+  qmu.resize(0);
   assert((Ls&0x1)==1); // Odd Ls required
   int nrational=Ls-1;
 
@@ -460,6 +526,22 @@ PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
   Approx::zolotarev_free(zdata);
 
 }
+template<class Impl>
+PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
+							 GridCartesian         &FiveDimGrid,
+							 GridRedBlackCartesian &FiveDimRedBlackGrid,
+							 GridCartesian         &FourDimGrid,
+							 GridRedBlackCartesian &FourDimRedBlackGrid,
+							 RealD _mass,RealD M5,
+							 std::vector<RealD> &_qmu,
+							 const ImplParams &p)
+  : PartialFractionFermion5D<Impl>(_Umu,
+			     FiveDimGrid,FiveDimRedBlackGrid,
+			     FourDimGrid,FourDimRedBlackGrid,
+			     _mass,M5,p)
+{
+  qmu=_qmu;
+}
 
 NAMESPACE_END(Grid);
 

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

@@ -14,6 +14,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 Author: Guido Cossu <guido.cossu@ed.ac.uk>
 Author: Andrew Lawson <andrew.lawson1991@gmail.com>
 Author: Vera Guelpers <V.M.Guelpers@soton.ac.uk>
+Author: Christoph Lehner <christoph@lhnr.de>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -332,22 +333,18 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   
   //  std::cout << " WilsonFermion5D Communicate Begin " <<std::endl;
   std::vector<std::vector<CommsRequest_t> > requests;
-  auto id=traceStart("Communicate overlapped");
-  st.CommunicateBegin(requests);
 
+#if 1
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
-  {
-    //  std::cout << " WilsonFermion5D Comms merge " <<std::endl;
-    GRID_TRACE("MergeSHM");
-    st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
-  }
-      
+  st.CommunicateBegin(requests);
+  st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms 
+#endif
+
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
-  //  std::cout << " WilsonFermion5D Interior " <<std::endl;
   int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
   if (dag == DaggerYes) {
     GRID_TRACE("DhopDagInterior");
@@ -356,13 +353,23 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
     GRID_TRACE("DhopInterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   }
-
+  
+  //ifdef GRID_ACCELERATED
+#if 0
+  /////////////////////////////
+  // Overlap with comms -- on GPU the interior kernel call is nonblocking
+  /////////////////////////////
+  st.CommunicateBegin(requests);
+  st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+#endif
+  
+  
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   //  std::cout << " WilsonFermion5D Comms Complete " <<std::endl;
   st.CommunicateComplete(requests);
-  traceStop(id);
+  //  traceStop(id);
 
   /////////////////////////////
   // do the compute exterior
@@ -438,6 +445,29 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
 
   DhopInternal(StencilOdd,UmuEven,in,out,dag);
 }
+template<class Impl>
+void WilsonFermion5D<Impl>::DhopComms(const FermionField &in, FermionField &out)
+{
+  int dag =0 ;
+  conformable(in.Grid(),FermionGrid()); // verifies full grid
+  conformable(in.Grid(),out.Grid());
+  out.Checkerboard() = in.Checkerboard();
+  Compressor compressor(dag);
+  Stencil.HaloExchangeOpt(in,compressor);
+}
+template<class Impl>
+void WilsonFermion5D<Impl>::DhopCalc(const FermionField &in, FermionField &out,uint64_t *ids)
+{
+  conformable(in.Grid(),FermionGrid()); // verifies full grid
+  conformable(in.Grid(),out.Grid());
+
+  out.Checkerboard() = in.Checkerboard();
+
+  int LLs = in.Grid()->_rdimensions[0];
+  int Opt = WilsonKernelsStatic::Opt;
+  Kernels::DhopKernel(Opt,Stencil,Umu,Stencil.CommBuf(),LLs,Umu.oSites(),in,out,ids);
+}
+
 template<class Impl>
 void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
 {
@@ -455,6 +485,54 @@ void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag
   Dhop(in,out,dag); // -0.5 is included
   axpy(out,4.0-M5,in,out);
 }
+template <class Impl>
+void WilsonFermion5D<Impl>::Meooe(const FermionField &in, FermionField &out)
+{
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerNo);
+  } else {
+    DhopOE(in, out, DaggerNo);
+  }
+}
+
+template <class Impl>
+void WilsonFermion5D<Impl>::MeooeDag(const FermionField &in, FermionField &out)
+{
+  if (in.Checkerboard() == Odd) {
+    DhopEO(in, out, DaggerYes);
+  } else {
+    DhopOE(in, out, DaggerYes);
+  }
+}
+
+template <class Impl>
+void WilsonFermion5D<Impl>::Mooee(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  typename FermionField::scalar_type scal(4.0 + M5);
+  out = scal * in;
+}
+
+template <class Impl>
+void WilsonFermion5D<Impl>::MooeeDag(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  Mooee(in, out);
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::MooeeInv(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  out = (1.0/(4.0 + M5))*in;
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
+{
+  out.Checkerboard() = in.Checkerboard();
+  MooeeInv(in,out);
+}
 
 template<class Impl>
 void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const FermionField &in, RealD mass,std::vector<double> twist)
@@ -740,6 +818,15 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
 
 template<class Impl>
 void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist)
+{
+  std::vector<double> empty_q(Nd,0.0);
+  MomentumSpacePropagatorHwQ(out,in,mass,twist,empty_q);
+}
+template<class Impl>
+void WilsonFermion5D<Impl>::MomentumSpacePropagatorHwQ(FermionField &out,const FermionField &in,
+						       RealD mass,
+						       std::vector<double> twist,
+						       std::vector<double> qmu)
 {
     Gamma::Algebra Gmu [] = {
       Gamma::Algebra::GammaX,
@@ -755,6 +842,7 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const Fe
     typedef typename FermionField::scalar_type ScalComplex;
 
     typedef Lattice<iSinglet<vector_type> > LatComplex;
+    typedef iSpinMatrix<ScalComplex> SpinMat;
 
 
     Coordinate latt_size   = _grid->_fdimensions;
@@ -772,8 +860,10 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const Fe
     LatComplex kmu(_grid); 
     ScalComplex ci(0.0,1.0);
 
+    std::cout<< "Feynman Rule" << "qmu ("<<qmu[0]<<","<<qmu[1]<<","<<qmu[2]<<","<<qmu[3]<<")"<<std::endl;
+    
     for(int mu=0;mu<Nd;mu++) {
-
+      
       LatticeCoordinate(kmu,mu);
 
       RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
@@ -782,9 +872,18 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const Fe
       kmu = kmu + TwoPiL * one * twist[mu];//momentum for twisted boundary conditions
 
       sk2 = sk2 + 2.0*sin(kmu*0.5)*sin(kmu*0.5);
-      sk  = sk  + sin(kmu)*sin(kmu); 
 
-      num = num - sin(kmu)*ci*(Gamma(Gmu[mu])*in);
+      sk = sk + (sin(kmu)+qmu[mu])*(sin(kmu)+qmu[mu]); 
+
+      // Terms for boosted Fermion
+      // 1/2 [ -i gamma.(sin p + q )     ]
+      //     [ --------------------- + 1 ]
+      //     [         wq + b            ]
+      //
+      // wq = sqrt( (sinp+q)^2 + b^2 )
+      //
+      
+      num = num - (sin(kmu)+qmu[mu])*ci*(Gamma(Gmu[mu])*in);
 
     }
     num = num + mass * in ;

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);
 
@@ -411,6 +411,46 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
+#ifdef GRID_SYCL
+extern "C" {
+    ulong SYCL_EXTERNAL __attribute__((overloadable)) intel_get_cycle_counter( void );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_active_channel_mask( void );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_grf_register( uint reg );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_flag_register( uint flag );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_control_register( uint reg );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_hw_thread_id( void );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_slice_id( void );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_subslice_id( void );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_eu_id( void );
+    uint  SYCL_EXTERNAL __attribute__((overloadable)) intel_get_eu_thread_id( void );
+    void  SYCL_EXTERNAL __attribute__((overloadable)) intel_eu_thread_pause( uint value );
+}
+#ifdef GRID_SIMT
+#define MAKE_ID(A) (intel_get_eu_id()<<16)|(intel_get_slice_id()<<8)|(intel_get_subslice_id())
+#else
+#define MAKE_ID(A) (0)
+#endif
+
+#else
+
+#define MAKE_ID(A) (0)
+
+#endif
+
+
+#define KERNEL_CALL_ID(A)						\
+  const uint64_t    NN = Nsite*Ls;					\
+  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
+      int sF = ss;							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
+      const int Nsimd = SiteHalfSpinor::Nsimd();			\
+      const int lane=acceleratorSIMTlane(Nsimd);                        \
+      int idx=sF*Nsimd+lane;						\
+      uint64_t id = MAKE_ID();						\
+      ids[idx]=id;							\
+    });									\
+  accelerator_barrier();
 
 #define KERNEL_CALLNB(A)						\
   const uint64_t    NN = Nsite*Ls;					\
@@ -418,7 +458,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
       int sF = ss;							\
       int sU = ss/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-  });
+    });
 
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
@@ -451,6 +491,8 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
     WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v);		\
     });}
 
+
+
 template <class Impl>
 void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 				     int Ls, int Nsite, const FermionField &in, FermionField &out,
@@ -475,7 +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;}
@@ -485,6 +527,18 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    }
    assert(0 && " Kernel optimisation case not covered ");
   }
+
+template <class Impl>
+void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
+				     int Ls, int Nsite, const FermionField &in, FermionField &out,
+				     uint64_t *ids)
+{
+    autoView(U_v  ,  U,AcceleratorRead);
+    autoView(in_v , in,AcceleratorRead);
+    autoView(out_v,out,AcceleratorWrite);
+    autoView(st_v , st,AcceleratorRead);
+    KERNEL_CALL_ID(GenericDhopSite);
+}
   template <class Impl>
   void WilsonKernels<Impl>::DhopDagKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 					  int Ls, int Nsite, const FermionField &in, FermionField &out,

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

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

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

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

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

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

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

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

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

@@ -40,6 +40,11 @@ public:
 
   INHERIT_GIMPL_TYPES(Gimpl);
 
+  using Action<GaugeField>::S;
+  using Action<GaugeField>::Sinitial;
+  using Action<GaugeField>::deriv;
+  using Action<GaugeField>::refresh;
+
 private:
   RealD c_plaq;
   RealD c_rect;
@@ -71,27 +76,27 @@ public:
     return action;
   };
 
-  virtual void deriv(const GaugeField &Umu,GaugeField & dSdU) {
+  virtual void deriv(const GaugeField &U, GaugeField &dSdU) {
     //extend Ta to include Lorentz indexes
     RealD factor_p = c_plaq/RealD(Nc)*0.5;
     RealD factor_r = c_rect/RealD(Nc)*0.5;
 
-    GridBase *grid = Umu.Grid();
+    GridBase *grid = U.Grid();
 
-    std::vector<GaugeLinkField> U (Nd,grid);
+    std::vector<GaugeLinkField> Umu (Nd,grid);
     for(int mu=0;mu<Nd;mu++){
-      U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+      Umu[mu] = PeekIndex<LorentzIndex>(U,mu);
     }
     std::vector<GaugeLinkField> RectStaple(Nd,grid), Staple(Nd,grid);
-    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, U, workspace);
+    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, Umu, workspace);
 
     GaugeLinkField dSdU_mu(grid);
     GaugeLinkField staple(grid);
 
     for (int mu=0; mu < Nd; mu++){
-      dSdU_mu = Ta(U[mu]*Staple[mu])*factor_p;
-      dSdU_mu = dSdU_mu + Ta(U[mu]*RectStaple[mu])*factor_r;
-	  
+      dSdU_mu = Ta(Umu[mu]*Staple[mu])*factor_p;
+      dSdU_mu = dSdU_mu + Ta(Umu[mu]*RectStaple[mu])*factor_r;
+
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);
     }
 

Grid/qcd/action/gauge/WilsonGaugeAction.h

@@ -43,6 +43,11 @@ class WilsonGaugeAction : public Action<typename Gimpl::GaugeField> {
 public:  
   INHERIT_GIMPL_TYPES(Gimpl);
 
+  using Action<GaugeField>::S;
+  using Action<GaugeField>::Sinitial;
+  using Action<GaugeField>::deriv;
+  using Action<GaugeField>::refresh;
+  
   /////////////////////////// constructors
   explicit WilsonGaugeAction(RealD beta_):beta(beta_){};
 
@@ -68,20 +73,23 @@ public:
     // extend Ta to include Lorentz indexes
 
     RealD factor = 0.5 * beta / RealD(Nc);
+    GridBase *grid = U.Grid();
 
-    GaugeLinkField Umu(U.Grid());
-    GaugeLinkField dSdU_mu(U.Grid());
+    GaugeLinkField dSdU_mu(grid);
+    std::vector<GaugeLinkField> Umu(Nd, grid);
     for (int mu = 0; mu < Nd; mu++) {
+      Umu[mu] = PeekIndex<LorentzIndex>(U, mu);
+    }
 
-      Umu = PeekIndex<LorentzIndex>(U, mu);
-      
+    for (int mu = 0; mu < Nd; mu++) {
       // Staple in direction mu
-      WilsonLoops<Gimpl>::Staple(dSdU_mu, U, mu);
-      dSdU_mu = Ta(Umu * dSdU_mu) * factor;
-      
+      WilsonLoops<Gimpl>::Staple(dSdU_mu, Umu, mu);
+      dSdU_mu = Ta(Umu[mu] * dSdU_mu) * factor;
+
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);
     }
   }
+
 private:
   RealD beta;  
  };

Grid/qcd/hmc/checkpointers/BinaryCheckpointer.h

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

Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h

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

Grid/qcd/hmc/checkpointers/ScidacCheckpointer.h

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

Grid/qcd/smearing/HISQSmearing.h

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

Grid/qcd/smearing/WilsonFlow.h

@@ -207,11 +207,14 @@ std::vector<RealD> WilsonFlowBase<Gimpl>::flowMeasureEnergyDensityCloverleaf(con
 }
 
 template <class Gimpl>
-void WilsonFlowBase<Gimpl>::setDefaultMeasurements(int topq_meas_interval){
-  addMeasurement(1, [](int step, RealD t, const typename Gimpl::GaugeField &U){
+void WilsonFlowBase<Gimpl>::setDefaultMeasurements(int meas_interval){
+  addMeasurement(meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
       std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "  << step << "  " << t << "  " << energyDensityPlaquette(t,U) << std::endl;
     });
-  addMeasurement(topq_meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
+  addMeasurement(meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
+      std::cout << GridLogMessage << "[WilsonFlow] Energy density (cloverleaf) : "  << step << "  " << t << "  " << energyDensityCloverleaf(t,U) << std::endl;
+    });
+  addMeasurement(meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
       std::cout << GridLogMessage << "[WilsonFlow] Top. charge           : "  << step << "  " << WilsonLoops<Gimpl>::TopologicalCharge(U) << std::endl;
     });
 }
@@ -249,6 +252,11 @@ void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const{
 
   out = in;
   RealD taus = 0.;
+
+  // Perform initial t=0 measurements
+  for(auto const &meas : this->functions)
+    meas.second(0,taus,out);
+  
   for (unsigned int step = 1; step <= Nstep; step++) { //step indicates the number of smearing steps applied at the time of measurement
     auto start = std::chrono::high_resolution_clock::now();
     evolve_step(out, taus);
@@ -333,6 +341,11 @@ void WilsonFlowAdaptive<Gimpl>::smear(GaugeField& out, const GaugeField& in) con
   RealD taus = 0.;
   RealD eps = init_epsilon;
   unsigned int step = 0;
+
+  // Perform initial t=0 measurements
+  for(auto const &meas : this->functions)
+    meas.second(step,taus,out);
+  
   do{
     int step_success = evolve_step_adaptive(out, taus, eps); 
     step += step_success; //step will not be incremented if the integration step fails

Grid/qcd/utils/SUn.impl.h

@@ -118,7 +118,7 @@ static void generatorDiagonal(int diagIndex, iGroupMatrix<cplx> &ta) {
 ////////////////////////////////////////////////////////////////////////
 // Map a su2 subgroup number to the pair of rows that are non zero
 ////////////////////////////////////////////////////////////////////////
-static void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::SU) {
+static accelerator_inline void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::SU) {
   assert((su2_index >= 0) && (su2_index < (ncolour * (ncolour - 1)) / 2));
 
   int spare = su2_index;

Grid/qcd/utils/Sp2n.impl.h

@@ -207,7 +207,7 @@ static void generatorZtype(int zIndex, iGroupMatrix<cplx> &ta) {
 // Map a su2 subgroup number to the pair of rows that are non zero
 ////////////////////////////////////////////////////////////////////////
 template <ONLY_IF_Sp>
-static void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::Sp) {
+static accelerator_inline void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::Sp) {
   const int nsp=ncolour/2;
   assert((su2_index >= 0) && (su2_index < (nsp * (nsp - 1)) / 2));
 

Grid/qcd/utils/WilsonLoops.h

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

Grid/stencil/Stencil.h

@@ -121,17 +121,22 @@ class CartesianStencilAccelerator {
   StencilVector same_node;
   Coordinate    _simd_layout;
   Parameters    parameters;
+  ViewMode mode;
   StencilEntry*  _entries_p;
+  StencilEntry*  _entries_host_p;
   cobj* u_recv_buf_p;
   cobj* u_send_buf_p;
 
   accelerator_inline cobj *CommBuf(void) const { return u_recv_buf_p; }
 
-  accelerator_inline int GetNodeLocal(int osite,int point) const {
-    return this->_entries_p[point+this->_npoints*osite]._is_local;
+  // Not a device function
+  inline int GetNodeLocal(int osite,int point) const {
+    StencilEntry SE=this->_entries_host_p[point+this->_npoints*osite];
+    return SE._is_local;
   }
   accelerator_inline StencilEntry * GetEntry(int &ptype,int point,int osite) const {
-    ptype = this->_permute_type[point]; return & this->_entries_p[point+this->_npoints*osite];
+    ptype = this->_permute_type[point];
+    return & this->_entries_p[point+this->_npoints*osite];
   }
 
   accelerator_inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) const {
@@ -164,28 +169,22 @@ class CartesianStencilView : public CartesianStencilAccelerator<vobj,cobj,Parame
 {
 public:
   int *closed;
-  StencilEntry *cpu_ptr;
-  ViewMode      mode;
+  //  StencilEntry *cpu_ptr;
  public:
   // default copy constructor
   CartesianStencilView (const CartesianStencilView &refer_to_me) = default;
 
   CartesianStencilView (const CartesianStencilAccelerator<vobj,cobj,Parameters> &refer_to_me,ViewMode _mode)
-    : CartesianStencilAccelerator<vobj,cobj,Parameters>(refer_to_me),
-    cpu_ptr(this->_entries_p),
-    mode(_mode)
+    : CartesianStencilAccelerator<vobj,cobj,Parameters>(refer_to_me)
   {
-    this->_entries_p =(StencilEntry *)
-      MemoryManager::ViewOpen(this->_entries_p,
-			      this->_npoints*this->_osites*sizeof(StencilEntry),
-			      mode,
-			      AdviseDefault);
+    this->ViewOpen(_mode);
+  }
+  void ViewOpen(ViewMode _mode)
+  {
+    this->mode = _mode;
   }
 
-  void ViewClose(void)
-  {
-    MemoryManager::ViewClose(this->cpu_ptr,this->mode);
-  }
+  void ViewClose(void)  {  }
 
 };
 
@@ -274,8 +273,8 @@ public:
   std::vector<deviceVector<std::pair<int,int> > > face_table ;
   deviceVector<int> surface_list;
 
-  std::vector<StencilEntry>  _entries; // Resident in host memory
-  deviceVector<StencilEntry>     _entries_device; // Resident in device memory
+  std::vector<StencilEntry>   _entries; // Resident in host memory
+  deviceVector<StencilEntry>  _entries_device; // Resident in device memory
   std::vector<Packet> Packets;
   std::vector<Merge> Mergers;
   std::vector<Merge> MergersSHM;
@@ -364,12 +363,32 @@ 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,
+					  Packets[i].recv_buf,
+					  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,
@@ -377,6 +396,7 @@ public:
 					Packets[i].from_rank,Packets[i].do_recv,
 					Packets[i].xbytes,Packets[i].rbytes,i);
     }
+    FlightRecorder::StepLog("Communicate begin has finished");
     // Get comms started then run checksums
     // Having this PRIOR to the dslash seems to make Sunspot work... (!)
     for(int i=0;i<Packets.size();i++){
@@ -387,15 +407,20 @@ 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();
     else DslashLogFull();
     //    acceleratorCopySynchronise();// is in the StencilSendToRecvFromComplete
     //    accelerator_barrier(); 
-    _grid->StencilBarrier(); 
-    // run any checksums
     for(int i=0;i<Packets.size();i++){
       if ( Packets[i].do_recv )
 	FlightRecorder::recvLog(Packets[i].recv_buf,Packets[i].rbytes,Packets[i].from_rank);
@@ -422,6 +447,7 @@ public:
     Communicate();
     CommsMergeSHM(compress);
     CommsMerge(compress);
+    accelerator_barrier();
   }
 
   template<class compressor> int HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point,int & face_idx)
@@ -477,6 +503,9 @@ public:
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
     //    accelerator_barrier();
+    //////////////////////////////////
+    // I will overwrite my send buffers
+    //////////////////////////////////
     _grid->StencilBarrier();// Synch shared memory on a single nodes
 
     assert(source.Grid()==_grid);
@@ -490,7 +519,11 @@ public:
       HaloGatherDir(source,compress,point,face_idx);
     }
     accelerator_barrier(); // All my local gathers are complete
-    //    _grid->StencilBarrier();// Synch shared memory on a single nodes
+#ifdef NVLINK_GET
+    _grid->StencilBarrier(); // He can now get mu local gather, I can get his
+    // Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
+    // Or issue barrier AFTER the DMA is running
+#endif    
     face_table_computed=1;
     assert(u_comm_offset==_unified_buffer_size);
   }
@@ -529,6 +562,7 @@ public:
 	  coalescedWrite(to[j] ,coalescedRead(from [j]));
       });
       acceleratorFenceComputeStream();
+      // Also fenced in WilsonKernels
     }
   }
   
@@ -626,10 +660,10 @@ public:
   ////////////////////////////////////////
   void PrecomputeByteOffsets(void){
     for(int i=0;i<_entries.size();i++){
-      if( _entries[i]._is_local ) {
-	_entries[i]._byte_offset = _entries[i]._offset*sizeof(vobj);
+      if( this->_entries[i]._is_local ) {
+	this->_entries[i]._byte_offset = this->_entries[i]._offset*sizeof(vobj);
       } else {
-	_entries[i]._byte_offset = _entries[i]._offset*sizeof(cobj);
+	this->_entries[i]._byte_offset = this->_entries[i]._offset*sizeof(cobj);
       }
     }
   };
@@ -657,7 +691,7 @@ public:
 	}
       }
     }
-    std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
+    //    std::cout << "BuildSurfaceList size is "<<surface_list_size<<std::endl;
     surface_list.resize(surface_list_size);
     std::vector<int> surface_list_host(surface_list_size);
     int32_t ss=0;
@@ -677,6 +711,7 @@ public:
       }
     }
     acceleratorCopyToDevice(&surface_list_host[0],&surface_list[0],surface_list_size*sizeof(int));
+    //    std::cout << GridLogMessage<<"BuildSurfaceList size is "<<surface_list_size<<std::endl;
   }
   /// Introduce a block structure and switch off comms on boundaries
   void DirichletBlock(const Coordinate &dirichlet_block)
@@ -764,7 +799,13 @@ public:
     this->_osites  = _grid->oSites();
 
     _entries.resize(this->_npoints* this->_osites);
-    this->_entries_p = &_entries[0];
+    _entries_device.resize(this->_npoints* this->_osites);
+    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;
+    
     for(int ii=0;ii<npoints;ii++){
 
       int i = ii; // reverse direction to get SIMD comms done first
@@ -841,6 +882,7 @@ public:
       u_simd_send_buf[l] = (cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
     }
     PrecomputeByteOffsets();
+    acceleratorCopyToDevice(&this->_entries[0],&this->_entries_device[0],this->_entries.size()*sizeof(StencilEntry));
   }
 
   void Local     (int point, int dimension,int shiftpm,int cbmask)
@@ -996,10 +1038,10 @@ public:
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
 	  int idx=point+(lo+o+b)*this->_npoints;
-	  _entries[idx]._offset  =ro+o+b;
-	  _entries[idx]._permute=permute;
-	  _entries[idx]._is_local=1;
-	  _entries[idx]._around_the_world=wrap;
+	  this->_entries[idx]._offset  =ro+o+b;
+	  this->_entries[idx]._permute=permute;
+	  this->_entries[idx]._is_local=1;
+	  this->_entries[idx]._around_the_world=wrap;
 	}
 	o +=_grid->_slice_stride[dimension];
       }
@@ -1017,10 +1059,10 @@ public:
 
 	  if ( ocb&cbmask ) {
 	    int idx = point+(lo+o+b)*this->_npoints;
-	    _entries[idx]._offset =ro+o+b;
-	    _entries[idx]._is_local=1;
-	    _entries[idx]._permute=permute;
-	    _entries[idx]._around_the_world=wrap;
+	    this->_entries[idx]._offset =ro+o+b;
+	    this->_entries[idx]._is_local=1;
+	    this->_entries[idx]._permute=permute;
+	    this->_entries[idx]._around_the_world=wrap;
 	  }
 
 	}
@@ -1044,10 +1086,10 @@ public:
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
 	  int idx=point+(so+o+b)*this->_npoints;
-	  _entries[idx]._offset  =offset+(bo++);
-	  _entries[idx]._is_local=0;
-	  _entries[idx]._permute=0;
-	  _entries[idx]._around_the_world=wrap;
+	  this->_entries[idx]._offset  =offset+(bo++);
+	  this->_entries[idx]._is_local=0;
+	  this->_entries[idx]._permute=0;
+	  this->_entries[idx]._around_the_world=wrap;
 	}
 	o +=_grid->_slice_stride[dimension];
       }
@@ -1064,10 +1106,10 @@ public:
 	  int ocb=1<<_grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	  if ( ocb & cbmask ) {
 	    int idx = point+(so+o+b)*this->_npoints;
-	    _entries[idx]._offset  =offset+(bo++);
-	    _entries[idx]._is_local=0;
-	    _entries[idx]._permute =0;
-	    _entries[idx]._around_the_world=wrap;
+	    this->_entries[idx]._offset  =offset+(bo++);
+	    this->_entries[idx]._is_local=0;
+	    this->_entries[idx]._permute =0;
+	    this->_entries[idx]._around_the_world=wrap;
 	  }
 	}
 	o +=_grid->_slice_stride[dimension];

Grid/threads/Accelerator.h

@@ -209,6 +209,17 @@ void Lambda6Apply(uint64_t num1, uint64_t num2, uint64_t num3,
     }									\
   }
 
+inline void *acceleratorAllocHost(size_t bytes)
+{
+  void *ptr=NULL;
+  auto err = cudaMallocHost((void **)&ptr,bytes);
+  if( err != cudaSuccess ) {
+    ptr = (void *) NULL;
+    printf(" cudaMallocHost failed for %d %s \n",bytes,cudaGetErrorString(err));
+    assert(0);
+  }
+  return ptr;
+}
 inline void *acceleratorAllocShared(size_t bytes)
 {
   void *ptr=NULL;
@@ -230,18 +241,34 @@ 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 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 acceleratorFreeHost(void *ptr){ cudaFree(ptr);};
+inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
+inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
+inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) {
+  acceleratorCopyToDevice(from,to,bytes);
+  return 0;
+}
+inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) {
+  acceleratorCopyFromDevice(from,to,bytes);
+  return 0;
+}
+inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
   cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
+  return 0;
 }
 inline void acceleratorCopySynchronise(void) { cudaStreamSynchronize(copyStream); };
+inline void acceleratorEventWait(acceleratorEvent_t ev)
+{
+  //auto discard=cudaStreamSynchronize(ev);
+}
+inline int acceleratorEventIsComplete(acceleratorEvent_t ev){ acceleratorEventWait(ev) ; return 1;}
 
 
 inline int  acceleratorIsCommunicable(void *ptr)
@@ -310,7 +337,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
     cgh.parallel_for(							\
 		     sycl::nd_range<3>(global,local),			\
 		     [=] (sycl::nd_item<3> item) /*mutable*/		\
-		     [[intel::reqd_sub_group_size(16)]]			\
+		     [[sycl::reqd_sub_group_size(16)]]			\
 		     {							\
 		       auto iter1    = item.get_global_id(0);		\
 		       auto iter2    = item.get_global_id(1);		\
@@ -322,14 +349,36 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
+inline void *acceleratorAllocHost(size_t bytes)  { return malloc_host(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
+inline void acceleratorFreeHost(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 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 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)
@@ -340,8 +389,10 @@ inline int  acceleratorIsCommunicable(void *ptr)
   else return 0;
 #endif
   return 1;
+
 }
 
+
 #endif
 
 //////////////////////////////////////////////
@@ -438,6 +489,16 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
     }								\
   }
 
+inline void *acceleratorAllocHost(size_t bytes)
+{
+  void *ptr=NULL;
+  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);
+  }
+  return ptr;
+};
 inline void *acceleratorAllocShared(size_t bytes)
 {
   void *ptr=NULL;
@@ -461,28 +522,47 @@ inline void *acceleratorAllocDevice(size_t bytes)
   return ptr;
 };
 
+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 acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
-//inline void acceleratorCopySynchronise(void) {  }
+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)
+{
+#ifdef GRID_SYCL
+  sycl::ext::oneapi::experimental::prepare_for_device_copy(ptr,bytes,theCopyAccelerator->get_context());
+#endif
+}
+
 //////////////////////////////////////////////
 // Common on all GPU targets
 //////////////////////////////////////////////
@@ -510,6 +590,8 @@ inline void acceleratorCopySynchronise(void) { auto discard=hipStreamSynchronize
 
 #undef GRID_SIMT
 
+typedef int acceleratorEvent_t;
+
 inline void acceleratorMem(void)
 {
   /*
@@ -530,15 +612,22 @@ inline void acceleratorMem(void)
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
 
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes);}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
+inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from,void *to,size_t bytes)        { acceleratorCopyToDevice(from,to,bytes); return 0; }
+inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from,void *to,size_t bytes)      { acceleratorCopyFromDevice(from,to,bytes); return 0; }
+inline void acceleratorEventWait(acceleratorEvent_t ev){}
+inline int acceleratorEventIsComplete(acceleratorEvent_t ev){ acceleratorEventWait(ev); return 1;}
+inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); return 0;}
+
 inline void acceleratorCopySynchronise(void) {};
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { memset(base,value,bytes);}
 #ifdef HAVE_MM_MALLOC_H
+inline void *acceleratorAllocHost(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
 inline void *acceleratorAllocShared(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
 inline void *acceleratorAllocDevice(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
+inline void acceleratorFreeHost(void *ptr){_mm_free(ptr);};
 inline void acceleratorFreeShared(void *ptr){_mm_free(ptr);};
 inline void acceleratorFreeDevice(void *ptr){_mm_free(ptr);};
 #else
@@ -618,9 +707,9 @@ inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)
   acceleratorCopySynchronise();
 }
 
-template<class T> void acceleratorPut(T& dev,T&host)
+template<class T> void acceleratorPut(T& dev,const T&host)
 {
-  acceleratorCopyToDevice(&host,&dev,sizeof(T));
+  acceleratorCopyToDevice((void *)&host,&dev,sizeof(T));
 }
 template<class T> T acceleratorGet(T& dev)
 {

Grid/threads/Threads.h

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

Grid/util/FlightRecorder.cc

@@ -280,10 +280,11 @@ void FlightRecorder::xmitLog(void *buf,uint64_t bytes)
   if(LoggingMode == LoggingModeNone) return;
 
   if ( ChecksumCommsSend ){
-  uint64_t *ubuf = (uint64_t *)buf;
-  if(LoggingMode == LoggingModeNone) return;
+
+    if(LoggingMode == LoggingModeNone) return;
   
 #ifdef GRID_SYCL
+  uint64_t *ubuf = (uint64_t *)buf;
   uint64_t _xor = svm_xor(ubuf,bytes/sizeof(uint64_t));
   if(LoggingMode == LoggingModePrint) {
     std::cerr<<"FlightRecorder::xmitLog : "<< XmitLoggingCounter <<" "<< std::hex << _xor <<std::dec <<std::endl;
@@ -327,9 +328,9 @@ void FlightRecorder::xmitLog(void *buf,uint64_t bytes)
 void FlightRecorder::recvLog(void *buf,uint64_t bytes,int rank)
 {
   if ( ChecksumComms ){
-  uint64_t *ubuf = (uint64_t *)buf;
   if(LoggingMode == LoggingModeNone) return;
 #ifdef GRID_SYCL
+  uint64_t *ubuf = (uint64_t *)buf;
   uint64_t _xor = svm_xor(ubuf,bytes/sizeof(uint64_t));
   if(LoggingMode == LoggingModePrint) {
     std::cerr<<"FlightRecorder::recvLog : "<< RecvLoggingCounter <<" "<< std::hex << _xor <<std::dec <<std::endl;

Grid/util/Init.cc

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

Grid/util/Lexicographic.h

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

HMC/ComputeWilsonFlow.cc

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

HMC/FTHMC2p1f.cc

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

HMC/FTHMC2p1f_3GeV.cc

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

HMC/HMC2p1f_3GeV.cc

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

Makefile.am

@@ -1,5 +1,5 @@
 # additional include paths necessary to compile the C++ library
-SUBDIRS = Grid HMC benchmarks tests examples
+SUBDIRS = Grid  benchmarks tests examples HMC
 
 include $(top_srcdir)/doxygen.inc
 

benchmarks/Benchmark_usqcd.cc

@@ -118,7 +118,7 @@ public:
     fprintf(FP,"Packet bytes, direction, GB/s per node\n");
     for(int lat=16;lat<=maxlat;lat+=8){
       //      for(int Ls=8;Ls<=8;Ls*=2){
-      { int Ls=12;
+      { int Ls=8;
 
 	Coordinate latt_size  ({lat*mpi_layout[0],
 	      lat*mpi_layout[1],
@@ -175,8 +175,8 @@ public:
 	    timestat.statistics(t_time);
 	  
 	    dbytes=dbytes*ppn;
-	    double xbytes    = dbytes*0.5;
-	    double bidibytes = dbytes;
+	    double xbytes    = dbytes;
+	    double bidibytes = dbytes*2.0;
 	  
 	    std::cout<<GridLogMessage << lat<<"\t"<<Ls<<"\t "
 		     << bytes << " \t "
@@ -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;
       
       }
 

configure.ac

@@ -72,6 +72,7 @@ AC_CHECK_HEADERS(malloc/malloc.h)
 AC_CHECK_HEADERS(malloc.h)
 AC_CHECK_HEADERS(endian.h)
 AC_CHECK_HEADERS(execinfo.h)
+AC_CHECK_HEADERS(numaif.h)
 AC_CHECK_DECLS([ntohll],[], [], [[#include <arpa/inet.h>]])
 AC_CHECK_DECLS([be64toh],[], [], [[#include <arpa/inet.h>]])
 
@@ -150,7 +151,7 @@ AC_ARG_ENABLE([tracing],
 case ${ac_TRACING} in
     nvtx)
         AC_DEFINE([GRID_TRACING_NVTX],[1],[use NVTX])
-	LIBS="${LIBS} -lnvToolsExt64_1"
+	LIBS="${LIBS} -lnvToolsExt"
 	;;
     roctx)
         AC_DEFINE([GRID_TRACING_ROCTX],[1],[use ROCTX])
@@ -240,6 +241,20 @@ case ${ac_SFW_FP16} in
 esac
 
 
+############### MPI BOUNCE TO HOST
+AC_ARG_ENABLE([accelerator-aware-mpi],
+    [AS_HELP_STRING([--enable-accelerator-aware-mpi=yes|no],[run mpi transfers from device])],
+    [ac_ACCELERATOR_AWARE_MPI=${enable_accelerator_aware_mpi}], [ac_ACCELERATOR_AWARE_MPI=yes])
+
+# Force accelerator CSHIFT now
+AC_DEFINE([ACCELERATOR_CSHIFT],[1],[ Cshift runs on device])
+
+case ${ac_ACCELERATOR_AWARE_MPI} in
+    yes)
+      AC_DEFINE([ACCELERATOR_AWARE_MPI],[1],[ Stencil can use device pointers]);;
+    *);;
+esac
+
 ############### SYCL/CUDA/HIP/none
 AC_ARG_ENABLE([accelerator],
     [AS_HELP_STRING([--enable-accelerator=cuda|sycl|hip|none],[enable none,cuda,sycl,hip acceleration])],

examples/Example_Laplacian_smearing.cc

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

examples/Example_taku.cc

@@ -1,383 +0,0 @@
-/*
- * Warning: This code illustrative only: not well tested, and not meant for production use
- * without regression / tests being applied
- */
-
-#include <Grid/Grid.h>
-
-using namespace std;
-using namespace Grid;
-
-RealD LLscale =1.0;
-RealD LCscale =1.0;
-
-template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
-{
-public:
-  INHERIT_GIMPL_TYPES(Gimpl);
-
-  GridBase *grid;
-  GaugeField U;
-  
-  CovariantLaplacianCshift(GaugeField &_U)    :
-    grid(_U.Grid()),
-    U(_U) {  };
-
-  virtual GridBase *Grid(void) { return grid; };
-
-  virtual void  M    (const Field &in, Field &out)
-  {
-    out=Zero();
-    for(int mu=0;mu<Nd-1;mu++) {
-      GaugeLinkField Umu = PeekIndex<LorentzIndex>(U, mu); // NB: Inefficent
-      out = out - Gimpl::CovShiftForward(Umu,mu,in);    
-      out = out - Gimpl::CovShiftBackward(Umu,mu,in);    
-      out = out + 2.0*in;
-    }
-  };
-  virtual void  Mdag (const Field &in, Field &out) { M(in,out);}; // Laplacian is hermitian
-  virtual  void Mdiag    (const Field &in, Field &out)                  {assert(0);}; // Unimplemented need only for multigrid
-  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);}; // Unimplemented need only for multigrid
-  virtual  void MdirAll  (const Field &in, std::vector<Field> &out)     {assert(0);}; // Unimplemented need only for multigrid
-};
-
-void MakePhase(Coordinate mom,LatticeComplex &phase)
-{
-  GridBase *grid = phase.Grid();
-  auto latt_size = grid->GlobalDimensions();
-  ComplexD ci(0.0,1.0);
-  phase=Zero();
-
-  LatticeComplex coor(phase.Grid());
-  for(int mu=0;mu<Nd;mu++){
-    RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
-    LatticeCoordinate(coor,mu);
-    phase = phase + (TwoPiL * mom[mu]) * coor;
-  }
-  phase = exp(phase*ci);
-}
-void PointSource(Coordinate &coor,LatticePropagator &source)
-{
-  //  Coordinate coor({0,0,0,0});
-  source=Zero();
-  SpinColourMatrix kronecker; kronecker=1.0;
-  pokeSite(kronecker,source,coor);
-}
-void Z2WallSource(GridParallelRNG &RNG,int tslice,LatticePropagator &source)
-{
-  GridBase *grid = source.Grid();
-  LatticeComplex noise(grid);
-  LatticeComplex zz(grid); zz=Zero();
-  LatticeInteger t(grid);
-
-  RealD nrm=1.0/sqrt(2);
-  bernoulli(RNG, noise); // 0,1 50:50
-
-  noise = (2.*noise - Complex(1,1))*nrm;
-
-  LatticeCoordinate(t,Tdir);
-  noise = where(t==Integer(tslice), noise, zz);
-
-  source = 1.0;
-  source = source*noise;
-  std::cout << " Z2 wall " << norm2(source) << std::endl;
-}
-template<class Field>
-void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
-{
-  typedef CovariantLaplacianCshift <PeriodicGimplR,Field> Laplacian_t;
-  Laplacian_t Laplacian(U);
-
-  Integer Iterations = 40;
-  Real width = 2.0;
-  Real coeff = (width*width) / Real(4*Iterations);
-
-  Field tmp(U.Grid());
-  smeared=unsmeared;
-  //  chi = (1-p^2/2N)^N kronecker
-  for(int n = 0; n < Iterations; ++n) {
-    Laplacian.M(smeared,tmp);
-    smeared = smeared - coeff*tmp;
-    std::cout << " smear iter " << n<<" " <<norm2(smeared)<<std::endl;
-  }
-}
-void GaussianSource(Coordinate &site,LatticeGaugeField &U,LatticePropagator &source)
-{
-  LatticePropagator tmp(source.Grid());
-  PointSource(site,source);
-  std::cout << " GaussianSource Kronecker "<< norm2(source)<<std::endl;
-  tmp = source;
-  GaussianSmear(U,tmp,source);
-  std::cout << " GaussianSource Smeared "<< norm2(source)<<std::endl;
-}
-void GaussianWallSource(GridParallelRNG &RNG,int tslice,LatticeGaugeField &U,LatticePropagator &source)
-{
-  Z2WallSource(RNG,tslice,source);
-  auto tmp = source;
-  GaussianSmear(U,tmp,source);
-}
-void SequentialSource(int tslice,Coordinate &mom,LatticePropagator &spectator,LatticePropagator &source)
-{
-  assert(mom.size()==Nd);
-  assert(mom[Tdir] == 0);
-
-  GridBase * grid = spectator.Grid();
-
-
-  LatticeInteger ts(grid);
-  LatticeCoordinate(ts,Tdir);
-  source = Zero();
-  source = where(ts==Integer(tslice),spectator,source); // Stick in a slice of the spectator, zero everywhere else
-
-  LatticeComplex phase(grid);
-  MakePhase(mom,phase);
-
-  source = source *phase;
-}
-template<class Action>
-void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
-{
-  GridBase *UGrid = D.GaugeGrid();
-  GridBase *FGrid = D.FermionGrid();
-
-  LatticeFermion src4  (UGrid); 
-  LatticeFermion src5  (FGrid); 
-  LatticeFermion result5(FGrid);
-  LatticeFermion result4(UGrid);
-  LatticePropagator prop5(FGrid);
-  
-  ConjugateGradient<LatticeFermion> CG(1.0e-8,100000);
-  SchurRedBlackDiagMooeeSolve<LatticeFermion> schur(CG);
-  ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
-   for(int s=0;s<Nd;s++){
-    for(int c=0;c<Nc;c++){
-      PropToFerm<Action>(src4,source,s,c);
-
-      D.ImportPhysicalFermionSource(src4,src5);
-
-      result5=Zero();
-      schur(D,src5,result5,ZG);
-      std::cout<<GridLogMessage
-	       <<"spin "<<s<<" color "<<c
-	       <<" norm2(src5d) "   <<norm2(src5)
-               <<" norm2(result5d) "<<norm2(result5)<<std::endl;
-
-      D.ExportPhysicalFermionSolution(result5,result4);
-
-      FermToProp<Action>(prop5,result5,s,c);
-      FermToProp<Action>(propagator,result4,s,c);
-    }
-  }
-  LatticePropagator Axial_mu(UGrid); 
-  LatticePropagator Vector_mu(UGrid); 
-
-  LatticeComplex    PA (UGrid); 
-  LatticeComplex    VV (UGrid); 
-  LatticeComplex    PJ5q(UGrid);
-  LatticeComplex    PP (UGrid);
-
-  std::vector<TComplex> sumPA;
-  std::vector<TComplex> sumVV;
-  std::vector<TComplex> sumPP;
-  std::vector<TComplex> sumPJ5q;
-
-  Gamma g5(Gamma::Algebra::Gamma5);
-  D.ContractConservedCurrent(prop5,prop5,Axial_mu,source,Current::Axial,Tdir);
-  PA       = trace(g5*Axial_mu);      // Pseudoscalar-Axial conserved current
-  sliceSum(PA,sumPA,Tdir);
-
-  int Nt{static_cast<int>(sumPA.size())};
-
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PAc["<<t<<"] "<<real(TensorRemove(sumPA[t]))*LCscale<<std::endl;
-
-  PP       = trace(adj(propagator)*propagator); // Pseudoscalar density
-  sliceSum(PP,sumPP,Tdir);
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PP["<<t<<"] "<<real(TensorRemove(sumPP[t]))*LCscale<<std::endl;
-  
-  D.ContractJ5q(prop5,PJ5q);
-  sliceSum(PJ5q,sumPJ5q,Tdir);
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PJ5q["<<t<<"] "<<real(TensorRemove(sumPJ5q[t]))<<std::endl;
-
-  Gamma::Algebra GammaV[3] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ
-  };
-  for( int mu=0;mu<3;mu++ ) {
-    Gamma gV(GammaV[mu]);
-    D.ContractConservedCurrent(prop5,prop5,Vector_mu,source,Current::Vector,mu);
-    //    auto ss=sliceSum(Vector_mu,Tdir);
-    //    for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"ss["<<mu<<"]["<<t<<"] "<<ss[t]<<std::endl;
-    VV       = trace(gV*Vector_mu);     // (local) Vector-Vector conserved current
-    sliceSum(VV,sumVV,Tdir);
-    for(int t=0;t<Nt;t++){
-      RealD Ct = real(TensorRemove(sumVV[t]))*LCscale;
-      std::cout<<GridLogMessage <<"VVc["<<mu<<"]["<<t<<"] "<< Ct
-	       << " 2 pi^2 t^3 C(t) "<< 2 * M_PI *M_PI * t*t*t *Ct<<std::endl;
-    }
-  }
-
-}
-
-class MesonFile: Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(MesonFile, std::vector<std::vector<Complex> >, data);
-};
-
-void MesonTrace(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
-{
-  const int nchannel=3;
-  Gamma::Algebra Gammas[nchannel][2] = {
-    {Gamma::Algebra::GammaX,Gamma::Algebra::GammaX},
-    {Gamma::Algebra::GammaY,Gamma::Algebra::GammaY},
-    {Gamma::Algebra::GammaZ,Gamma::Algebra::GammaZ}
-  };
-
-  Gamma G5(Gamma::Algebra::Gamma5);
-
-  LatticeComplex meson_CF(q1.Grid());
-  MesonFile MF;
-
-  for(int ch=0;ch<nchannel;ch++){
-
-    Gamma Gsrc(Gammas[ch][0]);
-    Gamma Gsnk(Gammas[ch][1]);
-
-    meson_CF = trace(G5*adj(q1)*G5*Gsnk*q2*adj(Gsrc));
-
-    std::vector<TComplex> meson_T;
-    sliceSum(meson_CF,meson_T, Tdir);
-
-    int nt=meson_T.size();
-
-    std::vector<Complex> corr(nt);
-    for(int t=0;t<nt;t++){
-      corr[t] = TensorRemove(meson_T[t])*LLscale; // Yes this is ugly, not figured a work around
-      std::cout << " channel "<<ch<<" t "<<t<<" " <<real(corr[t])<< " 2 pi^2 t^3 C(t) "<< 2 * M_PI *M_PI * t*t*t *real(corr[t])<<std::endl;
-    }
-    MF.data.push_back(corr);
-  }
-
-  {
-    XmlWriter WR(file);
-    write(WR,"MesonFile",MF);
-  }
-}
-
-int main (int argc, char ** argv)
-{
-  const int Ls=32;
-
-  Grid_init(&argc,&argv);
-
-  // Double precision grids
-  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), 
-								   GridDefaultSimd(Nd,vComplex::Nsimd()),
-								   GridDefaultMpi());
-  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
-  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
-
-  //////////////////////////////////////////////////////////////////////
-  // You can manage seeds however you like.
-  // Recommend SeedUniqueString.
-  //////////////////////////////////////////////////////////////////////
-  std::vector<int> seeds4({1,2,3,4}); 
-  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
-
-  LatticeGaugeField Umu(UGrid);
-  std::string config;
-  RealD M5=1.8;
-  if( argc > 1 && argv[1][0] != '-' )
-  {
-    std::cout<<GridLogMessage <<"Loading configuration from "<<argv[1]<<std::endl;
-    FieldMetaData header;
-    NerscIO::readConfiguration(Umu, header, argv[1]);
-    config=argv[1];
-    M5=1.8;
-  }
-  else
-  {
-    SU<Nc>::ColdConfiguration(Umu);
-    config="ColdConfig";
-    //    RealD P=1.0; // Don't scale
-    RealD P=0.5871119; // 48I
-    //    RealD P=0.6153342; // 64I
-    //    RealD P=0.6388238 // 32Ifine
-    RealD u0 = sqrt(sqrt(P));
-    RealD M5mf = M5 - 4.0*(1.0-u0);
-    RealD w0   = 1.0 - M5mf;
-#if 0
-    // M5=1.8 with U=u0
-    Umu = Umu * u0;
-    LLscale = 1.0;
-    LCscale = 1.0;
-    std::cout<<GridLogMessage <<"Gauge links are u=u0= "<<u0<<std::endl;
-    std::cout<<GridLogMessage <<"M5 =  "<<M5<<std::endl;
-#else
-    M5 = M5mf;
-    std::cout<<GridLogMessage <<"Gauge links are u=1  "<<std::endl;
-    std::cout<<GridLogMessage <<"u0="<<u0<<std::endl;
-    std::cout<<GridLogMessage <<"M5=M5mf =  "<<M5<<std::endl;
-    LLscale = 1.0/(1-w0*w0)/(1-w0*w0);
-    LCscale = 1.0/(1-w0*w0)/(1-w0*w0);
-#endif
-    std::cout<<GridLogMessage <<"LLscale =  "<<LLscale<<std::endl;
-    std::cout<<GridLogMessage <<"LCscale =  "<<LCscale<<std::endl;
-  }
-
-  std::vector<RealD> masses({ 0.00} ); // u/d, s, c ??
-
-  int nmass = masses.size();
-
-  std::vector<MobiusFermionD *> FermActs;
-  
-  std::cout<<GridLogMessage <<"======================"<<std::endl;
-  std::cout<<GridLogMessage <<"MobiusFermion action as Scaled Shamir kernel"<<std::endl;
-  std::cout<<GridLogMessage <<"======================"<<std::endl;
-
-  for(auto mass: masses) {
-
-    RealD b=1.5;// Scale factor b+c=2, b-c=1
-    RealD c=0.5;
-    
-    FermActs.push_back(new MobiusFermionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c));
-   
-  }
-
-  LatticePropagator point_source(UGrid);
-  //  LatticePropagator wall_source(UGrid);
-
-  Coordinate Origin({0,0,0,0});
-  PointSource   (Origin,point_source);
-  //  Z2WallSource  (RNG4,0,wall_source);
-  
-  std::vector<LatticePropagator> PointProps(nmass,UGrid);
-  //  std::vector<LatticePropagator> GaussProps(nmass,UGrid);
-  //  std::vector<LatticePropagator> Z2Props   (nmass,UGrid);
-
-  for(int m=0;m<nmass;m++) {
-    
-    Solve(*FermActs[m],point_source   ,PointProps[m]);
-  }
-
-  LatticeComplex phase(UGrid);
-  Coordinate mom({0,0,0,0});
-  MakePhase(mom,phase);
-  
-  for(int m1=0 ;m1<nmass;m1++) {
-  for(int m2=m1;m2<nmass;m2++) {
-    std::stringstream ssp,ssg,ssz;
-
-    ssp<<config<< "_m" << m1 << "_m"<< m2 << "_point_meson.xml";
-    ssz<<config<< "_m" << m1 << "_m"<< m2 << "_wall_meson.xml";
-
-    MesonTrace(ssp.str(),PointProps[m1],PointProps[m2],phase);
-    //    MesonTrace(ssz.str(),Z2Props[m1],Z2Props[m2],phase);
-  }}
-
-  Grid_finalize();
-}
-
-
-

examples/Example_taku1.cc

@@ -1,479 +0,0 @@
-/*
- * Warning: This code illustrative only: not well tested, and not meant for production use
- * without regression / tests being applied
- */
-
-#include <Grid/Grid.h>
-
-using namespace std;
-using namespace Grid;
-
-RealD LLscale =1.0;
-RealD LCscale =1.0;
-
-template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
-{
-public:
-  INHERIT_GIMPL_TYPES(Gimpl);
-
-  GridBase *grid;
-  GaugeField U;
-  
-  CovariantLaplacianCshift(GaugeField &_U)    :
-    grid(_U.Grid()),
-    U(_U) {  };
-
-  virtual GridBase *Grid(void) { return grid; };
-
-  virtual void  M    (const Field &in, Field &out)
-  {
-    out=Zero();
-    for(int mu=0;mu<Nd-1;mu++) {
-      GaugeLinkField Umu = PeekIndex<LorentzIndex>(U, mu); // NB: Inefficent
-      out = out - Gimpl::CovShiftForward(Umu,mu,in);    
-      out = out - Gimpl::CovShiftBackward(Umu,mu,in);    
-      out = out + 2.0*in;
-    }
-  };
-  virtual void  Mdag (const Field &in, Field &out) { M(in,out);}; // Laplacian is hermitian
-  virtual  void Mdiag    (const Field &in, Field &out)                  {assert(0);}; // Unimplemented need only for multigrid
-  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);}; // Unimplemented need only for multigrid
-  virtual  void MdirAll  (const Field &in, std::vector<Field> &out)     {assert(0);}; // Unimplemented need only for multigrid
-};
-
-void MakePhase(Coordinate mom,LatticeComplex &phase)
-{
-  GridBase *grid = phase.Grid();
-  auto latt_size = grid->GlobalDimensions();
-  ComplexD ci(0.0,1.0);
-  phase=Zero();
-
-  LatticeComplex coor(phase.Grid());
-  for(int mu=0;mu<Nd;mu++){
-    RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
-    LatticeCoordinate(coor,mu);
-    phase = phase + (TwoPiL * mom[mu]) * coor;
-  }
-  phase = exp(phase*ci);
-}
-
-void PointSource(Coordinate &coor,LatticePropagator &source)
-{
-  //  Coordinate coor({0,0,0,0});
-  source=Zero();
-  SpinColourMatrix kronecker; kronecker=1.0;
-  pokeSite(kronecker,source,coor);
-}
-void Z2WallSource(GridParallelRNG &RNG,int tslice,LatticePropagator &source)
-{
-  GridBase *grid = source.Grid();
-  LatticeComplex noise(grid);
-  LatticeComplex zz(grid); zz=Zero();
-  LatticeInteger t(grid);
-
-  RealD nrm=1.0/sqrt(2);
-  bernoulli(RNG, noise); // 0,1 50:50
-
-  noise = (2.*noise - Complex(1,1))*nrm;
-
-  LatticeCoordinate(t,Tdir);
-  noise = where(t==Integer(tslice), noise, zz);
-
-  source = 1.0;
-  source = source*noise;
-  std::cout << " Z2 wall " << norm2(source) << std::endl;
-}
-template<class Field>
-void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
-{
-  typedef CovariantLaplacianCshift <PeriodicGimplR,Field> Laplacian_t;
-  Laplacian_t Laplacian(U);
-
-  Integer Iterations = 40;
-  Real width = 2.0;
-  Real coeff = (width*width) / Real(4*Iterations);
-
-  Field tmp(U.Grid());
-  smeared=unsmeared;
-  //  chi = (1-p^2/2N)^N kronecker
-  for(int n = 0; n < Iterations; ++n) {
-    Laplacian.M(smeared,tmp);
-    smeared = smeared - coeff*tmp;
-    std::cout << " smear iter " << n<<" " <<norm2(smeared)<<std::endl;
-  }
-}
-void GaussianSource(Coordinate &site,LatticeGaugeField &U,LatticePropagator &source)
-{
-  LatticePropagator tmp(source.Grid());
-  PointSource(site,source);
-  std::cout << " GaussianSource Kronecker "<< norm2(source)<<std::endl;
-  tmp = source;
-  GaussianSmear(U,tmp,source);
-  std::cout << " GaussianSource Smeared "<< norm2(source)<<std::endl;
-}
-void GaussianWallSource(GridParallelRNG &RNG,int tslice,LatticeGaugeField &U,LatticePropagator &source)
-{
-  Z2WallSource(RNG,tslice,source);
-  auto tmp = source;
-  GaussianSmear(U,tmp,source);
-}
-void SequentialSource(int tslice,Coordinate &mom,LatticePropagator &spectator,LatticePropagator &source)
-{
-  assert(mom.size()==Nd);
-  assert(mom[Tdir] == 0);
-
-  GridBase * grid = spectator.Grid();
-
-
-  LatticeInteger ts(grid);
-  LatticeCoordinate(ts,Tdir);
-  source = Zero();
-  source = where(ts==Integer(tslice),spectator,source); // Stick in a slice of the spectator, zero everywhere else
-
-  LatticeComplex phase(grid);
-  MakePhase(mom,phase);
-
-  source = source *phase;
-}
-
-template<class Action>
-void MasslessFreePropagator(Action &D,LatticePropagator &source,LatticePropagator &propagator)
-{			   
- GridBase *UGrid = source.Grid();
-  GridBase *FGrid = D.FermionGrid();
-  bool fiveD = true; //calculate 5d free propagator
-  RealD mass = D.Mass();
-  LatticeFermion src4  (UGrid);
-  LatticeFermion result4  (UGrid);
-  LatticeFermion result5(FGrid);
-  LatticeFermion src5(FGrid);
-  LatticePropagator prop5(FGrid);
-  for(int s=0;s<Nd;s++){
-    for(int c=0;c<Nc;c++){
- 
-      PropToFerm<Action>(src4,source,s,c);
-
-      D.ImportPhysicalFermionSource(src4,src5);
-      D.FreePropagator(src5,result5,mass,true);
-      std::cout<<GridLogMessage
-               <<"Free 5D prop spin "<<s<<" color "<<c
-               <<" norm2(src5d) "   <<norm2(src5)
-               <<" norm2(result5d) "<<norm2(result5)<<std::endl;
-
-      D.ExportPhysicalFermionSolution(result5,result4);
-
-      FermToProp<Action>(prop5,result5,s,c);
-      FermToProp<Action>(propagator,result4,s,c);
-    }
-  }
-
-  LatticePropagator Vector_mu(UGrid);
-  LatticeComplex    VV (UGrid);
-  std::vector<TComplex> sumVV;
-  Gamma::Algebra GammaV[3] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ
-  };
-  for( int mu=0;mu<3;mu++ ) {
-    Gamma gV(GammaV[mu]);
-    D.ContractConservedCurrent(prop5,prop5,Vector_mu,source,Current::Vector,mu);
-    VV       = trace(gV*Vector_mu);     // (local) Vector-Vector conserved current
-    sliceSum(VV,sumVV,Tdir);
-    int Nt = sumVV.size();
-    for(int t=0;t<Nt;t++){
-      RealD Ct = real(TensorRemove(sumVV[t]))*LCscale;
-      RealD Cont=0;
-      if(t) Cont=1.0/(2 * M_PI *M_PI * t*t*t);
-      std::cout<<GridLogMessage <<"VVc["<<mu<<"]["<<t<<"] "<< Ct
-               << " 2 pi^2 t^3 C(t) "<< Ct/Cont << " delta Ct "<< Ct-Cont <<std::endl;
-    }
-  }
-}
-template<class Action>
-void MasslessFreePropagator1(Action &D,LatticePropagator &source,LatticePropagator &propagator)
-{			   
-  bool fiveD = false; //calculate 4d free propagator
-  RealD mass = D.Mass();
-  GridBase *UGrid = source.Grid();
-  LatticeFermion src4  (UGrid); 
-  LatticeFermion result4  (UGrid); 
-  for(int s=0;s<Nd;s++){
-    for(int c=0;c<Nc;c++){
-      PropToFerm<Action>(src4,source,s,c);
-      D.FreePropagator(src4,result4,mass,false);
-      FermToProp<Action>(propagator,result4,s,c);
-    }
-  }
-}
-
-template<class Action>
-void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
-{
-  GridBase *UGrid = D.GaugeGrid();
-  GridBase *FGrid = D.FermionGrid();
-
-  LatticeFermion src4  (UGrid); 
-  LatticeFermion src5  (FGrid); 
-  LatticeFermion result5(FGrid);
-  LatticeFermion result4(UGrid);
-  LatticePropagator prop5(FGrid);
-  
-  ConjugateGradient<LatticeFermion> CG(1.0e-10,100000);
-  SchurRedBlackDiagMooeeSolve<LatticeFermion> schur(CG);
-  ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
-   for(int s=0;s<Nd;s++){
-    for(int c=0;c<Nc;c++){
-      PropToFerm<Action>(src4,source,s,c);
-
-      D.ImportPhysicalFermionSource(src4,src5);
-
-      result5=Zero();
-      schur(D,src5,result5,ZG);
-      std::cout<<GridLogMessage
-	       <<"spin "<<s<<" color "<<c
-	       <<" norm2(src5d) "   <<norm2(src5)
-               <<" norm2(result5d) "<<norm2(result5)<<std::endl;
-
-      D.ExportPhysicalFermionSolution(result5,result4);
-
-      FermToProp<Action>(prop5,result5,s,c);
-      FermToProp<Action>(propagator,result4,s,c);
-    }
-  }
-  LatticePropagator Axial_mu(UGrid); 
-  LatticePropagator Vector_mu(UGrid); 
-
-  LatticeComplex    PA (UGrid); 
-  LatticeComplex    VV (UGrid); 
-  LatticeComplex    PJ5q(UGrid);
-  LatticeComplex    PP (UGrid);
-
-  std::vector<TComplex> sumPA;
-  std::vector<TComplex> sumVV;
-  std::vector<TComplex> sumPP;
-  std::vector<TComplex> sumPJ5q;
-
-  Gamma g5(Gamma::Algebra::Gamma5);
-  D.ContractConservedCurrent(prop5,prop5,Axial_mu,source,Current::Axial,Tdir);
-  PA       = trace(g5*Axial_mu);      // Pseudoscalar-Axial conserved current
-  sliceSum(PA,sumPA,Tdir);
-
-  int Nt{static_cast<int>(sumPA.size())};
-
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PAc["<<t<<"] "<<real(TensorRemove(sumPA[t]))*LCscale<<std::endl;
-
-  PP       = trace(adj(propagator)*propagator); // Pseudoscalar density
-  sliceSum(PP,sumPP,Tdir);
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PP["<<t<<"] "<<real(TensorRemove(sumPP[t]))*LCscale<<std::endl;
-  
-  D.ContractJ5q(prop5,PJ5q);
-  sliceSum(PJ5q,sumPJ5q,Tdir);
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PJ5q["<<t<<"] "<<real(TensorRemove(sumPJ5q[t]))<<std::endl;
-
-  Gamma::Algebra GammaV[3] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ
-  };
-  for( int mu=0;mu<3;mu++ ) {
-    Gamma gV(GammaV[mu]);
-    D.ContractConservedCurrent(prop5,prop5,Vector_mu,source,Current::Vector,mu);
-    //    auto ss=sliceSum(Vector_mu,Tdir);
-    //    for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"ss["<<mu<<"]["<<t<<"] "<<ss[t]<<std::endl;
-    VV       = trace(gV*Vector_mu);     // (local) Vector-Vector conserved current
-    sliceSum(VV,sumVV,Tdir);
-    for(int t=0;t<Nt;t++){
-      RealD Ct = real(TensorRemove(sumVV[t]))*LCscale;
-      RealD Cont=0;
-      if(t) Cont=1.0/(2 * M_PI *M_PI * t*t*t);
-      std::cout<<GridLogMessage <<"VVc["<<mu<<"]["<<t<<"] "<< Ct
-               << " 2 pi^2 t^3 C(t) "<< Ct/Cont << " delta Ct "<< Ct-Cont <<std::endl;
-    }
-  }
-
-}
-
-class MesonFile: Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(MesonFile, std::vector<std::vector<Complex> >, data);
-};
-
-void MesonTrace(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
-{
-  const int nchannel=4;
-  Gamma::Algebra Gammas[nchannel][2] = {
-    {Gamma::Algebra::GammaXGamma5,Gamma::Algebra::GammaXGamma5},
-    {Gamma::Algebra::GammaYGamma5,Gamma::Algebra::GammaYGamma5},
-    {Gamma::Algebra::GammaZGamma5,Gamma::Algebra::GammaZGamma5},
-    {Gamma::Algebra::Identity,Gamma::Algebra::Identity}
-  };
-
-  LatticeComplex meson_CF(q1.Grid());
-  MesonFile MF;
-
-  for(int ch=0;ch<nchannel;ch++){
-
-    Gamma Gsrc(Gammas[ch][0]);
-    Gamma Gsnk(Gammas[ch][1]);
-
-    meson_CF = trace(adj(q1)*Gsnk*q2*adj(Gsrc));
-
-    std::vector<TComplex> meson_T;
-    sliceSum(meson_CF,meson_T, Tdir);
-
-    int nt=meson_T.size();
-
-    std::vector<Complex> corr(nt);
-    for(int t=0;t<nt;t++){
-      corr[t] = TensorRemove(meson_T[t])*LLscale; // Yes this is ugly, not figured a work around
-      RealD Ct = real(corr[t]);
-      RealD Cont=0;
-      if(t) Cont=1.0/(2 * M_PI *M_PI * t*t*t);
-      std::cout << " channel "<<ch<<" t "<<t<<" " <<real(corr[t])<< " 2 pi^2 t^3 C(t) "<< 2 * M_PI *M_PI * t*t*t * Ct
-		<< " deltaC " <<Ct-Cont<<std::endl;
-    }
-    MF.data.push_back(corr);
-  }
-
-  {
-    XmlWriter WR(file);
-    write(WR,"MesonFile",MF);
-  }
-}
-
-int main (int argc, char ** argv)
-{
-  const int Ls=10;
-
-  Grid_init(&argc,&argv);
-
-  // Double precision grids
-  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), 
-								   GridDefaultSimd(Nd,vComplex::Nsimd()),
-								   GridDefaultMpi());
-  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
-  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
-
-  //////////////////////////////////////////////////////////////////////
-  // You can manage seeds however you like.
-  // Recommend SeedUniqueString.
-  //////////////////////////////////////////////////////////////////////
-  //  std::vector<int> seeds4({1,2,3,4}); 
-  //  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
-
-  LatticeGaugeField Umu(UGrid);
-  std::string config;
-  RealD M5=atof(getenv("M5"));
-  RealD mq = atof(getenv("mass"));
-  int   tadpole = atof(getenv("tadpole"));
-  std::vector<RealD> masses({ mq} ); // u/d, s, c ??
-  if( argc > 1 && argv[1][0] != '-' )
-  {
-    std::cout<<GridLogMessage <<"Loading configuration from "<<argv[1]<<std::endl;
-    FieldMetaData header;
-    NerscIO::readConfiguration(Umu, header, argv[1]);
-    config=argv[1];
-    LLscale = 1.0;
-    LCscale = 1.0;
-  }
-  else
-  {
-    SU<Nc>::ColdConfiguration(Umu);
-    config="ColdConfig";
-    //    RealD P=1.0; // Don't scale
-    //    RealD P=0.6388238 // 32Ifine
-    //    RealD P=0.6153342; // 64I
-    RealD P=0.5871119; // 48I
-    RealD u0 = sqrt(sqrt(P));
-    RealD w0 = 1 - M5;
-    std::cout<<GridLogMessage <<"For plaquette P="<<P<<" u0= "<<u0<<std::endl;
-    if ( tadpole == 1 ) {
-      Umu = Umu * u0;
-      //      LLscale = 1.0/(1-w0*w0)/(1-w0*w0)/u0/u0;
-      //      LCscale = 1.0/(1-w0*w0)/(1-w0*w0)/u0/u0;
-      LLscale = 1.0;
-      LCscale = 1.0;
-      std::cout<<GridLogMessage <<"Gauge links are u= u0 "<<std::endl;
-      std::cout<<GridLogMessage <<"M5 =  "<<M5<<std::endl;
-    } else if ( tadpole == 2) {
-      std::cout<<GridLogMessage <<"Gauge links are u=1 "<<std::endl;
-      LLscale = 1.0;
-      LCscale = 1.0;
-      std::cout<<GridLogMessage <<"M5 =  "<<M5<<std::endl;
-    } else {
-      LLscale = 1.0/u0/u0;
-      LCscale = 1.0/u0/u0;
-      M5 = M5 - 4.0 * (1-u0);
-      std::cout<<GridLogMessage <<"Gauge links are u=1 "<<std::endl;
-      std::cout<<GridLogMessage <<"M5mf =  "<<M5<<std::endl;
-    }
-    std::cout<<GridLogMessage <<"mq =  "<<mq<<std::endl;
-    std::cout<<GridLogMessage <<"LLscale =  "<<LLscale<<std::endl;
-    std::cout<<GridLogMessage <<"LCscale =  "<<LCscale<<std::endl;
-  }
-
-  int nmass = masses.size();
-
-  typedef DomainWallFermionD FermionActionD;
-  //  typedef MobiusFermionD FermionActionD;
-  std::vector<FermionActionD *> FermActs;
-  std::vector<DomainWallFermionD *> DWFActs;
-  
-  std::cout<<GridLogMessage <<"======================"<<std::endl;
-  std::cout<<GridLogMessage <<"DomainWallFermion action"<<std::endl;
-  std::cout<<GridLogMessage <<"======================"<<std::endl;
-
-  for(auto mass: masses) {
-    std::vector<Complex> boundary = {1,1,1,-1};
-    FermionActionD::ImplParams Params(boundary);
-    RealD b=1.5;
-    RealD c=0.5;
-    std::cout<<GridLogMessage <<"Making DomainWallFermion action"<<std::endl;
-    //    DWFActs.push_back(new DomainWallFermionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5));
-    FermActs.push_back(new FermionActionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,Params));
-    //    FermActs.push_back(new FermionActionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass+0.001,M5,b,c));
-    std::cout<<GridLogMessage <<"Made DomainWallFermion action"<<std::endl;
-  }
-
-  LatticePropagator point_source(UGrid);
-
-  Coordinate Origin({0,0,0,0});
-  PointSource   (Origin,point_source);
-  
-  std::vector<LatticePropagator> PointProps(nmass,UGrid);
-  //  std::vector<LatticePropagator> FreeProps(nmass,UGrid);
-  //  LatticePropagator delta(UGrid);
-
-  for(int m=0;m<nmass;m++) {
-    Solve(*FermActs[m],point_source   ,PointProps[m]);
-    //    MasslessFreePropagator(*FermActs[m],point_source   ,FreeProps[m]);
-
-    //    delta = PointProps[m] - FreeProps[m];
-    //    std::cout << " delta "<<norm2(delta) << " FFT "<<norm2(FreeProps[m])<< " CG " <<norm2(PointProps[m])<<std::endl;
-  }
-
-  LatticeComplex phase(UGrid);
-  Coordinate mom({0,0,0,0});
-  MakePhase(mom,phase);
-  
-  for(int m1=0 ;m1<nmass;m1++) {
-  for(int m2=m1;m2<nmass;m2++) {
-    std::stringstream ssp,ssg,ssz;
-
-    ssp<<config<< "_m" << m1 << "_m"<< m2 << "_point_meson.xml";
-    ssz<<config<< "_m" << m1 << "_m"<< m2 << "_free_meson.xml";
-
-    std::cout << "CG determined VV correlation function"<<std::endl;
-    MesonTrace(ssp.str(),PointProps[m1],PointProps[m2],phase);
-    
-    //    std::cout << "FFT derived VV correlation function"<<std::endl;
-    //    MesonTrace(ssz.str(),FreeProps[m1],FreeProps[m2],phase);
-  }}
-
-  Grid_finalize();
-}
-
-
-

examples/Example_taku2.cc

@@ -1,433 +0,0 @@
-/*
- * Warning: This code illustrative only: not well tested, and not meant for production use
- * without regression / tests being applied
- */
-
-#include <Grid/Grid.h>
-
-using namespace std;
-using namespace Grid;
-
-RealD LLscale =1.0;
-RealD LCscale =1.0;
-
-template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
-{
-public:
-  INHERIT_GIMPL_TYPES(Gimpl);
-
-  GridBase *grid;
-  GaugeField U;
-  
-  CovariantLaplacianCshift(GaugeField &_U)    :
-    grid(_U.Grid()),
-    U(_U) {  };
-
-  virtual GridBase *Grid(void) { return grid; };
-
-  virtual void  M    (const Field &in, Field &out)
-  {
-    out=Zero();
-    for(int mu=0;mu<Nd-1;mu++) {
-      GaugeLinkField Umu = PeekIndex<LorentzIndex>(U, mu); // NB: Inefficent
-      out = out - Gimpl::CovShiftForward(Umu,mu,in);    
-      out = out - Gimpl::CovShiftBackward(Umu,mu,in);    
-      out = out + 2.0*in;
-    }
-  };
-  virtual void  Mdag (const Field &in, Field &out) { M(in,out);}; // Laplacian is hermitian
-  virtual  void Mdiag    (const Field &in, Field &out)                  {assert(0);}; // Unimplemented need only for multigrid
-  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);}; // Unimplemented need only for multigrid
-  virtual  void MdirAll  (const Field &in, std::vector<Field> &out)     {assert(0);}; // Unimplemented need only for multigrid
-};
-
-void MakePhase(Coordinate mom,LatticeComplex &phase)
-{
-  GridBase *grid = phase.Grid();
-  auto latt_size = grid->GlobalDimensions();
-  ComplexD ci(0.0,1.0);
-  phase=Zero();
-
-  LatticeComplex coor(phase.Grid());
-  for(int mu=0;mu<Nd;mu++){
-    RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
-    LatticeCoordinate(coor,mu);
-    phase = phase + (TwoPiL * mom[mu]) * coor;
-  }
-  phase = exp(phase*ci);
-}
-
-void PointSource(Coordinate &coor,LatticePropagator &source)
-{
-  //  Coordinate coor({0,0,0,0});
-  source=Zero();
-  SpinColourMatrix kronecker; kronecker=1.0;
-  pokeSite(kronecker,source,coor);
-}
-void Z2WallSource(GridParallelRNG &RNG,int tslice,LatticePropagator &source)
-{
-  GridBase *grid = source.Grid();
-  LatticeComplex noise(grid);
-  LatticeComplex zz(grid); zz=Zero();
-  LatticeInteger t(grid);
-
-  RealD nrm=1.0/sqrt(2);
-  bernoulli(RNG, noise); // 0,1 50:50
-
-  noise = (2.*noise - Complex(1,1))*nrm;
-
-  LatticeCoordinate(t,Tdir);
-  noise = where(t==Integer(tslice), noise, zz);
-
-  source = 1.0;
-  source = source*noise;
-  std::cout << " Z2 wall " << norm2(source) << std::endl;
-}
-template<class Field>
-void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
-{
-  typedef CovariantLaplacianCshift <PeriodicGimplR,Field> Laplacian_t;
-  Laplacian_t Laplacian(U);
-
-  Integer Iterations = 40;
-  Real width = 2.0;
-  Real coeff = (width*width) / Real(4*Iterations);
-
-  Field tmp(U.Grid());
-  smeared=unsmeared;
-  //  chi = (1-p^2/2N)^N kronecker
-  for(int n = 0; n < Iterations; ++n) {
-    Laplacian.M(smeared,tmp);
-    smeared = smeared - coeff*tmp;
-    std::cout << " smear iter " << n<<" " <<norm2(smeared)<<std::endl;
-  }
-}
-void GaussianSource(Coordinate &site,LatticeGaugeField &U,LatticePropagator &source)
-{
-  LatticePropagator tmp(source.Grid());
-  PointSource(site,source);
-  std::cout << " GaussianSource Kronecker "<< norm2(source)<<std::endl;
-  tmp = source;
-  GaussianSmear(U,tmp,source);
-  std::cout << " GaussianSource Smeared "<< norm2(source)<<std::endl;
-}
-void GaussianWallSource(GridParallelRNG &RNG,int tslice,LatticeGaugeField &U,LatticePropagator &source)
-{
-  Z2WallSource(RNG,tslice,source);
-  auto tmp = source;
-  GaussianSmear(U,tmp,source);
-}
-void SequentialSource(int tslice,Coordinate &mom,LatticePropagator &spectator,LatticePropagator &source)
-{
-  assert(mom.size()==Nd);
-  assert(mom[Tdir] == 0);
-
-  GridBase * grid = spectator.Grid();
-
-
-  LatticeInteger ts(grid);
-  LatticeCoordinate(ts,Tdir);
-  source = Zero();
-  source = where(ts==Integer(tslice),spectator,source); // Stick in a slice of the spectator, zero everywhere else
-
-  LatticeComplex phase(grid);
-  MakePhase(mom,phase);
-
-  source = source *phase;
-}
-
-template<class Action>
-void MasslessFreePropagator(Action &D,LatticePropagator &source,LatticePropagator &propagator)
-{			   
- GridBase *UGrid = source.Grid();
-  GridBase *FGrid = D.FermionGrid();
-  bool fiveD = true; //calculate 4d free propagator                                                                                                                 
-  RealD mass = D.Mass();
-  LatticeFermion src4  (UGrid);
-  LatticeFermion result4  (UGrid);
-  LatticeFermion result5(FGrid);
-  LatticeFermion src5(FGrid);
-  LatticePropagator prop5(FGrid);
-  for(int s=0;s<Nd;s++){
-    for(int c=0;c<Nc;c++){
- 
-      PropToFerm<Action>(src4,source,s,c);
-
-      D.ImportPhysicalFermionSource(src4,src5);
-      D.FreePropagator(src5,result5,mass,true);
-      std::cout<<GridLogMessage
-               <<"spin "<<s<<" color "<<c
-               <<" norm2(src5d) "   <<norm2(src5)
-               <<" norm2(result5d) "<<norm2(result5)<<std::endl;
-
-      D.ExportPhysicalFermionSolution(result5,result4);
-
-      FermToProp<Action>(prop5,result5,s,c);
-      FermToProp<Action>(propagator,result4,s,c);
-    }
-  }
-
-  LatticePropagator Vector_mu(UGrid);
-  LatticeComplex    VV (UGrid);
-  std::vector<TComplex> sumVV;
-  Gamma::Algebra GammaV[3] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ
-  };
-  for( int mu=0;mu<3;mu++ ) {
-    Gamma gV(GammaV[mu]);
-    D.ContractConservedCurrent(prop5,prop5,Vector_mu,source,Current::Vector,mu);
-    VV       = trace(gV*Vector_mu);     // (local) Vector-Vector conserved current
-    sliceSum(VV,sumVV,Tdir);
-    int Nt = sumVV.size();
-    for(int t=0;t<Nt;t++){
-      RealD Ct = real(TensorRemove(sumVV[t]))*LCscale;
-      std::cout<<GridLogMessage <<"VVc["<<mu<<"]["<<t<<"] "<< Ct
-               << " 2 pi^2 t^3 C(t) "<< 2 * M_PI *M_PI * t*t*t *Ct<<std::endl;
-    }
-  }
-}
-
-template<class Action>
-void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
-{
-  GridBase *UGrid = D.GaugeGrid();
-  GridBase *FGrid = D.FermionGrid();
-
-  LatticeFermion src4  (UGrid); 
-  LatticeFermion src5  (FGrid); 
-  LatticeFermion result5(FGrid);
-  LatticeFermion result4(UGrid);
-  LatticePropagator prop5(FGrid);
-  
-  ConjugateGradient<LatticeFermion> CG(1.0e-6,100000);
-  SchurRedBlackDiagMooeeSolve<LatticeFermion> schur(CG);
-  ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
-   for(int s=0;s<Nd;s++){
-    for(int c=0;c<Nc;c++){
-      PropToFerm<Action>(src4,source,s,c);
-
-      D.ImportPhysicalFermionSource(src4,src5);
-
-      result5=Zero();
-      schur(D,src5,result5,ZG);
-      std::cout<<GridLogMessage
-	       <<"spin "<<s<<" color "<<c
-	       <<" norm2(src5d) "   <<norm2(src5)
-               <<" norm2(result5d) "<<norm2(result5)<<std::endl;
-
-      D.ExportPhysicalFermionSolution(result5,result4);
-
-      FermToProp<Action>(prop5,result5,s,c);
-      FermToProp<Action>(propagator,result4,s,c);
-    }
-  }
-  LatticePropagator Axial_mu(UGrid); 
-  LatticePropagator Vector_mu(UGrid); 
-
-  LatticeComplex    PA (UGrid); 
-  LatticeComplex    VV (UGrid); 
-  LatticeComplex    PJ5q(UGrid);
-  LatticeComplex    PP (UGrid);
-
-  std::vector<TComplex> sumPA;
-  std::vector<TComplex> sumVV;
-  std::vector<TComplex> sumPP;
-  std::vector<TComplex> sumPJ5q;
-
-  Gamma g5(Gamma::Algebra::Gamma5);
-  D.ContractConservedCurrent(prop5,prop5,Axial_mu,source,Current::Axial,Tdir);
-  PA       = trace(g5*Axial_mu);      // Pseudoscalar-Axial conserved current
-  sliceSum(PA,sumPA,Tdir);
-
-  int Nt{static_cast<int>(sumPA.size())};
-
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PAc["<<t<<"] "<<real(TensorRemove(sumPA[t]))*LCscale<<std::endl;
-
-  PP       = trace(adj(propagator)*propagator); // Pseudoscalar density
-  sliceSum(PP,sumPP,Tdir);
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PP["<<t<<"] "<<real(TensorRemove(sumPP[t]))*LCscale<<std::endl;
-  
-  D.ContractJ5q(prop5,PJ5q);
-  sliceSum(PJ5q,sumPJ5q,Tdir);
-  for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"PJ5q["<<t<<"] "<<real(TensorRemove(sumPJ5q[t]))<<std::endl;
-
-  Gamma::Algebra GammaV[3] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ
-  };
-  for( int mu=0;mu<3;mu++ ) {
-    Gamma gV(GammaV[mu]);
-    D.ContractConservedCurrent(prop5,prop5,Vector_mu,source,Current::Vector,mu);
-    //    auto ss=sliceSum(Vector_mu,Tdir);
-    //    for(int t=0;t<Nt;t++) std::cout<<GridLogMessage <<"ss["<<mu<<"]["<<t<<"] "<<ss[t]<<std::endl;
-    VV       = trace(gV*Vector_mu);     // (local) Vector-Vector conserved current
-    sliceSum(VV,sumVV,Tdir);
-    for(int t=0;t<Nt;t++){
-      RealD Ct = real(TensorRemove(sumVV[t]))*LCscale;
-      std::cout<<GridLogMessage <<"VVc["<<mu<<"]["<<t<<"] "<< Ct
-	       << " 2 pi^2 t^3 C(t) "<< 2 * M_PI *M_PI * t*t*t *Ct<<std::endl;
-    }
-  }
-
-}
-
-class MesonFile: Serializable {
-public:
-  GRID_SERIALIZABLE_CLASS_MEMBERS(MesonFile, std::vector<std::vector<Complex> >, data);
-};
-
-void MesonTrace(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
-{
-  const int nchannel=3;
-  Gamma::Algebra Gammas[nchannel][2] = {
-    {Gamma::Algebra::GammaX,Gamma::Algebra::GammaX},
-    {Gamma::Algebra::GammaY,Gamma::Algebra::GammaY},
-    //    {Gamma::Algebra::GammaZ,Gamma::Algebra::GammaZ}
-    {Gamma::Algebra::Gamma5,Gamma::Algebra::Gamma5}
-  };
-
-  Gamma G5(Gamma::Algebra::Gamma5);
-
-  LatticeComplex meson_CF(q1.Grid());
-  MesonFile MF;
-
-  for(int ch=0;ch<nchannel;ch++){
-
-    Gamma Gsrc(Gammas[ch][0]);
-    Gamma Gsnk(Gammas[ch][1]);
-
-    meson_CF = trace(G5*adj(q1)*G5*Gsnk*q2*adj(Gsrc));
-
-    std::vector<TComplex> meson_T;
-    sliceSum(meson_CF,meson_T, Tdir);
-
-    int nt=meson_T.size();
-
-    std::vector<Complex> corr(nt);
-    for(int t=0;t<nt;t++){
-      corr[t] = TensorRemove(meson_T[t])*LLscale; // Yes this is ugly, not figured a work around
-      std::cout << " channel "<<ch<<" t "<<t<<" " <<real(corr[t])<< " 2 pi^2 t^3 C(t) "<< 2 * M_PI *M_PI * t*t*t *real(corr[t])<<std::endl;
-    }
-    MF.data.push_back(corr);
-  }
-
-  {
-    XmlWriter WR(file);
-    write(WR,"MesonFile",MF);
-  }
-}
-
-int main (int argc, char ** argv)
-{
-  const int Ls=8;
-
-  Grid_init(&argc,&argv);
-
-  // Double precision grids
-  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), 
-								   GridDefaultSimd(Nd,vComplex::Nsimd()),
-								   GridDefaultMpi());
-  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
-  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
-
-  //////////////////////////////////////////////////////////////////////
-  // You can manage seeds however you like.
-  // Recommend SeedUniqueString.
-  //////////////////////////////////////////////////////////////////////
-  //  std::vector<int> seeds4({1,2,3,4}); 
-  //  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
-
-  LatticeGaugeField Umu(UGrid);
-  std::string config;
-  RealD M5=atof(getenv("M5"));
-  RealD mq = atof(getenv("mass"));
-  std::vector<RealD> masses({ mq} ); // u/d, s, c ??
-  if( argc > 1 && argv[1][0] != '-' )
-  {
-    std::cout<<GridLogMessage <<"Loading configuration from "<<argv[1]<<std::endl;
-    FieldMetaData header;
-    NerscIO::readConfiguration(Umu, header, argv[1]);
-    config=argv[1];
-    LLscale = 1.0;
-    LCscale = 1.0;
-  }
-  else
-  {
-    SU<Nc>::ColdConfiguration(Umu);
-    config="ColdConfig";
-    //    RealD P=1.0; // Don't scale
-    //    RealD P=0.6153342; // 64I
-    //    RealD P=0.6388238 // 32Ifine
-    //    RealD P=0.5871119; // 48I
-    //    RealD u0 = sqrt(sqrt(P));
-    //    Umu = Umu * u0;
-    RealD w0 = 1 - M5;
-    LLscale = 1.0/(1-w0*w0)/(1-w0*w0);
-    LCscale = 1.0/(1-w0*w0)/(1-w0*w0);
-    std::cout<<GridLogMessage <<"Gauge links are u=1 "<<std::endl;
-    std::cout<<GridLogMessage <<"M5 =  "<<M5<<std::endl;
-    std::cout<<GridLogMessage <<"mq =  "<<mq<<std::endl;
-    std::cout<<GridLogMessage <<"LLscale =  "<<LLscale<<std::endl;
-    std::cout<<GridLogMessage <<"LCscale =  "<<LCscale<<std::endl;
-  }
-
-  int nmass = masses.size();
-
-  std::vector<DomainWallFermionD *> FermActs;
-  
-  std::cout<<GridLogMessage <<"======================"<<std::endl;
-  std::cout<<GridLogMessage <<"DomainWallFermion action"<<std::endl;
-  std::cout<<GridLogMessage <<"======================"<<std::endl;
-
-  for(auto mass: masses) {
-
-    std::cout<<GridLogMessage <<"Making DomainWallFermion action"<<std::endl;
-    FermActs.push_back(new DomainWallFermionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5));
-    std::cout<<GridLogMessage <<"Made DomainWallFermion action"<<std::endl;
-   
-  }
-
-  LatticePropagator point_source(UGrid);
-
-  Coordinate Origin({0,0,0,0});
-  PointSource   (Origin,point_source);
-  
-  //  std::vector<LatticePropagator> PointProps(nmass,UGrid);
-  std::vector<LatticePropagator> FreeProps(nmass,UGrid);
-  LatticePropagator delta(UGrid);
-
-  for(int m=0;m<nmass;m++) {
-    //    Solve(*FermActs[m],point_source   ,PointProps[m]);
-    MasslessFreePropagator(*FermActs[m],point_source   ,FreeProps[m]);
-
-    //    delta = PointProps[m] - FreeProps[m];
-    //    std::cout << " delta "<<norm2(delta) << " FFT "<<norm2(FreeProps[m])<< " CG " <<norm2(PointProps[m])<<std::endl;
-  }
-
-  LatticeComplex phase(UGrid);
-  Coordinate mom({0,0,0,0});
-  MakePhase(mom,phase);
-  
-  for(int m1=0 ;m1<nmass;m1++) {
-  for(int m2=m1;m2<nmass;m2++) {
-    std::stringstream ssp,ssg,ssz;
-
-    ssp<<config<< "_m" << m1 << "_m"<< m2 << "_point_meson.xml";
-    ssz<<config<< "_m" << m1 << "_m"<< m2 << "_free_meson.xml";
-
-    //    std::cout << "CG determined VV correlation function"<<std::endl;
-    //    MesonTrace(ssp.str(),PointProps[m1],PointProps[m2],phase);
-    
-    std::cout << "FFT derived VV correlation function"<<std::endl;
-    MesonTrace(ssz.str(),FreeProps[m1],FreeProps[m2],phase);
-  }}
-
-  Grid_finalize();
-}
-
-
-

systems/Aurora/benchmarks/bench1.pbs

@@ -1,6 +1,7 @@
 #!/bin/bash
 
-#PBS -q EarlyAppAccess
+##PBS -q EarlyAppAccess
+#PBS -q debug
 #PBS -l select=1
 #PBS -l walltime=00:20:00
 #PBS -A LatticeQCD_aesp_CNDA
@@ -12,27 +13,24 @@ source ../sourceme.sh
 cp $PBS_NODEFILE nodefile
 
 export OMP_NUM_THREADS=4
-export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
-unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
-unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
-unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+export MPICH_OFI_NIC_POLICY=GPU
+
+#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
 #export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
 #export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
 #export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
-export MPICH_OFI_NIC_POLICY=GPU
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
 
 CMD="mpiexec -np 12 -ppn 12  -envall \
-	     ./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.48 \
-		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --debug-signals"
+	     ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.96 \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 8 "
 
-#for f in 1 2 3 4 5 6 7 8
-for f in 1
-do
 echo $CMD
-$CMD | tee 1node.32.32.64.48.dwf.hbm.$f
-done
+$CMD
 

systems/Aurora/benchmarks/bench16.pbs

@@ -0,0 +1,74 @@
+#!/bin/bash
+
+##PBS -q LatticeQCD_aesp_CNDA
+#PBS -q debug-scaling
+##PBS -q prod
+#PBS -l select=16
+#PBS -l walltime=00:20:00
+#PBS -A LatticeQCD_aesp_CNDA
+
+cd $PBS_O_WORKDIR
+
+source ../sourceme.sh
+
+cp $PBS_NODEFILE nodefile
+
+export OMP_NUM_THREADS=4
+export MPICH_OFI_NIC_POLICY=GPU
+
+#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
+#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
+
+#
+# Local vol 16.16.16.32
+#
+
+LX=16
+LY=16
+LZ=16
+LT=32
+
+NX=2
+NY=2
+NZ=4
+NT=1
+
+GX=2
+GY=2
+GZ=1
+GT=3
+
+PX=$((NX * GX ))
+PY=$((NY * GY ))
+PZ=$((NZ * GZ ))
+PT=$((NT * GT ))
+
+VX=$((PX * LX ))
+VY=$((PY * LY ))
+VZ=$((PZ * LZ ))
+VT=$((PT * LT ))
+
+NP=$((PX*PY*PZ*PT))
+VOL=${VX}.${VY}.${VZ}.${VT}
+AT=8
+MPI=${PX}.${PY}.${PZ}.${PT}
+
+CMD="mpiexec -np $NP -ppn 12  -envall \
+	     ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi $MPI --grid $VOL \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads $AT --comms-overlap "
+
+echo VOL $VOL
+echo MPI $MPI
+echo NPROC $NP
+echo $CMD
+$CMD
+

systems/Aurora/benchmarks/bench2.pbs

@@ -1,58 +1,48 @@
 #!/bin/bash
 
-#PBS -q EarlyAppAccess
+##PBS -q EarlyAppAccess
+#PBS -q debug
 #PBS -l select=2
 #PBS -l walltime=00:20:00
 #PBS -A LatticeQCD_aesp_CNDA
 
-#export OMP_PROC_BIND=spread
-#unset OMP_PLACES
-
 cd $PBS_O_WORKDIR
 
 source ../sourceme.sh
-#module load pti-gpu
-
 
 cp $PBS_NODEFILE nodefile
 
 export OMP_NUM_THREADS=4
-export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+export MPICH_OFI_NIC_POLICY=GPU
+
+#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
 #unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
-export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
-export MPICH_OFI_NIC_POLICY=GPU
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
+#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
 
-# 12 ppn, 2 nodes, 24 ranks
 #
-CMD="mpiexec -np 24 -ppn 12  -envall \
-	     ./gpu_tile.sh \
-	     ./Benchmark_comms_host_device --mpi 2.2.2.3 --grid 24.32.32.24 \
-		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32" 
-#$CMD | tee 2node.comms.hbm
+# Local vol 16.16.16.32
+#
 
+#VOL=32.64.64.96
 
-CMD="mpiexec -np 24 -ppn 12  -envall \
-	     ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.32.64.48 \
-		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap --debug-signals"
-
-#for f in 1 2 3 4 5 6 7 8
-for f in 1
+for VOL in 32.32.32.96 32.64.64.96
 do
+for AT in 32
+do
+CMD="mpiexec -np 24 -ppn 12  -envall \
+	     ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid $VOL \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads $AT --comms-overlap "
+
 echo $CMD
-$CMD | tee 2node.32.32.64.48.dwf.hbm.$f
+$CMD
+done
 done
 
-CMD="mpiexec -np 24 -ppn 12  -envall \
-	     ./gpu_tile.sh \
-	     ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 64.64.64.96 \
-		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
-#$CMD | tee 2node.64.64.64.96.dwf.hbm
-

systems/Aurora/benchmarks/gpu_tile.sh

@@ -4,10 +4,12 @@
 #export NUMA_MAP=(0 0 1 1 0 0 1 1 0 0 1 1);
 #export  GPU_MAP=(0.0 0.1 3.0 3.1 1.0 1.1 4.0 4.1 2.0 2.1 5.0 5.1)
 
-export NUMA_MAP=(0 0 0 0 0 0 1 1 1 1 1 1 );
+export NUMA_PMAP=(0 0 0 1 1 1 0 0 0 1 1 1 );
+export NUMA_HMAP=(2 2 2 3 3 3 3 2 2 2 2 3 3 3 );
 export  GPU_MAP=(0.0 1.0 2.0 3.0 4.0 5.0 0.1 1.1 2.1 3.1 4.1 5.1 )
 
-export NUMA=${NUMA_MAP[$PALS_LOCAL_RANKID]}
+export NUMAP=${NUMA_PMAP[$PALS_LOCAL_RANKID]}
+export NUMAH=${NUMA_HMAP[$PALS_LOCAL_RANKID]}
 export gpu_id=${GPU_MAP[$PALS_LOCAL_RANKID]}
   
 unset EnableWalkerPartition
@@ -17,18 +19,19 @@ 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=0:5
-#export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:2
+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
 
+#export MPI_BUF_NUMA=$NUMAH
+
 echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "
 
 if [ $PALS_RANKID = "0" ]
 then
-#    numactl -m $NUMA -N $NUMA onetrace --chrome-device-timeline  "$@"
-#    numactl -m $NUMA -N $NUMA unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
-    numactl -m $NUMA -N $NUMA  "$@"
+#    numactl -p $NUMAP -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
+    numactl -p $NUMAP -N $NUMAP  "$@"
 else 
-    numactl -m $NUMA -N $NUMA  "$@"
+    numactl -p $NUMAP -N $NUMAP  "$@"
 fi

systems/Aurora/config-command

@@ -1,23 +1,25 @@
 #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${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl " 
-export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions "
+
+export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs -device -Xs pvc -fsycl-device-lib=all -lze_loader -L${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 \
 	--enable-shm=nvlink \
 	--enable-accelerator=sycl \
-	--enable-accelerator-aware-mpi=yes\
+	--enable-accelerator-aware-mpi=no\
 	--enable-unified=no \
 	MPICXX=mpicxx \
 	CXX=icpx 

systems/Aurora/sourceme.sh

@@ -2,6 +2,7 @@
 #module load mpich/icc-all-debug-pmix-gpu/52.2
 #module load mpich-config/mode/deterministic
 #module load intel_compute_runtime/release/821.35
+module load pti-gpu
 
 source ~/spack/share/spack/setup-env.sh 
 spack load c-lime

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

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

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

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

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

systems/Jupiter/benchmarks/dwf1.slurm

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

systems/Jupiter/benchmarks/dwf4.slurm

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

systems/Jupiter/config-command

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

systems/Jupiter/sourceme.sh

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

systems/Linux-cuda/config-command

@@ -4,6 +4,8 @@
     --enable-gen-simd-width=64 \
     --enable-shm=nvlink \
     --with-lime=$CLIME \
+    --with-hdf5=$HDF5 \
+    --with-fftw=$FFTW \
     --with-gmp=$GMP \
     --with-mpfr=$MPFR \
     --enable-accelerator=cuda \

systems/Linux-cuda/sourceme.sh

@@ -3,10 +3,14 @@ spack load cuda@12.0.0
 spack load c-lime
 spack load gmp
 spack load mpfr
+spack load hdf5
+spack load fftw
 spack load openmpi
+export FFTW=`spack find --paths fftw | grep fftw | cut -c 14-`
+export HDF5=`spack find --paths hdf5 | grep hdf5 | cut -c 14-`
 export CUDA=`spack find --paths cuda@11.8.0 | grep cuda | cut -c 14-`
 export CLIME=`spack find --paths c-lime | grep c-lime| cut -c 15-`
 export GMP=`spack find --paths gmp | grep gmp | cut -c 12-`
 export MPFR=`spack find --paths mpfr | grep mpfr | cut -c 13-`
 export NVIDIALIB=$CUDA/targets/x86_64-linux/lib/
-export LD_LIBRARY_PATH=$NVIDIALIB:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$NVIDIALIB:$LD_LIBRARY_PATH:$HDF5/lib:$FFTW/lib:$CLIME/lib/:$MPFR/lib

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
+--------------------------------------------------------------------