Merge 32e6d58356 into aa67a5b095

Grid/DisableWarnings.h

@@ -51,13 +51,11 @@ directory
 #pragma nv_diag_suppress cast_to_qualified_type
  //disables nvcc specific warning in many files
 #pragma nv_diag_suppress esa_on_defaulted_function_ignored
-#pragma nv_diag_suppress declared_but_not_referenced
 #pragma nv_diag_suppress extra_semicolon
 #else
  //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
-#pragma diag_suppress declared_but_not_referenced
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon

Grid/algorithms/FFT.h

@@ -168,7 +168,6 @@ public:
   template<class vobj>
   void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
 #ifndef HAVE_FFTW
-    std::cerr << "FFTW is not compiled but is called"<<std::endl;
     assert(0);
 #else
     conformable(result.Grid(),vgrid);
@@ -191,8 +190,7 @@ public:
       
     Lattice<sobj> pgbuf(&pencil_g);
     autoView(pgbuf_v , pgbuf, CpuWrite);
-    //std::cout << "CPU view" << std::endl;
-    
+
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
       
@@ -215,7 +213,6 @@ public:
     else if ( sign == forward ) div = 1.0;
     else assert(0);
       
-    //std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
     FFTW_plan p;
     {
       FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
@@ -229,7 +226,6 @@ public:
     }
       
     // Barrel shift and collect global pencil
-    //std::cout << GridLogPerformance<<"Making pencil" << std::endl;
     Coordinate lcoor(Nd), gcoor(Nd);
     result = source;
     int pc = processor_coor[dim];
@@ -251,7 +247,6 @@ public:
       }
     }
       
-    //std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
     // Loop over orthog coords
     int NN=pencil_g.lSites();
     GridStopWatch timer;
@@ -274,7 +269,6 @@ public:
     usec += timer.useconds();
     flops+= flops_call*NN;
       
-    //std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
     // writing out result
     {
       autoView(pgbuf_v,pgbuf,CpuRead);
@@ -291,7 +285,6 @@ public:
     }
     result = result*div;
       
-    //std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
     // destroying plan
     FFTW<scalar>::fftw_destroy_plan(p);
 #endif

Grid/algorithms/LinearOperator.h

@@ -277,38 +277,6 @@ 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

@@ -269,9 +269,7 @@ 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

@@ -55,10 +55,10 @@ NAMESPACE_BEGIN(Grid);
   typedef cublasHandle_t gridblasHandle_t;
 #endif
 #ifdef GRID_SYCL
-  typedef sycl::queue *gridblasHandle_t;
+  typedef cl::sycl::queue *gridblasHandle_t;
 #endif
 #ifdef GRID_ONE_MKL
-  typedef sycl::queue *gridblasHandle_t;
+  typedef cl::sycl::queue *gridblasHandle_t;
 #endif
 #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
   typedef int32_t gridblasHandle_t;
@@ -89,9 +89,9 @@ public:
       gridblasHandle = theGridAccelerator;
 #endif
 #ifdef GRID_ONE_MKL
-      sycl::gpu_selector selector;
-      sycl::device selectedDevice { selector };
-      sycl::property_list q_prop{sycl::property::queue::in_order()};
+      cl::sycl::gpu_selector selector;
+      cl::sycl::device selectedDevice { selector };
+      cl::sycl::property_list q_prop{cl::sycl::property::queue::in_order()};
       gridblasHandle =new sycl::queue (selectedDevice,q_prop);
 #endif
       gridblasInit=1;
@@ -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
@@ -367,67 +367,28 @@ 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
-	  else
-	    eCmn = alpha * eAmk * eBkn ;
+	  eCmn = beta * 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);
-	  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 ;
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * 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);
-	  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() ;
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
 	  });
       } 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);
-	  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() ;
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
 	  } );
       } else { 
 	assert(0);
@@ -453,8 +414,8 @@ 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
@@ -553,70 +514,28 @@ 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
-	  else
-	    eCmn = alpha * eAmk * eBkn ;
+	  eCmn = beta * 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);
-	  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 ;
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * 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);
-	  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() ;
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
 	  });
       } 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);
-	  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() ;
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
 	  } );
       } else { 
 	assert(0);
@@ -742,41 +661,29 @@ 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
-	  else
-	    eCmn = alpha * eAmk * eBkn ;
+	  eCmn = beta * 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
-	  else
-	    eCmn = alpha * eAmk.transpose() * eBkn ;
+	  eCmn = beta * 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
-	  else
-	    eCmn = alpha * eAmk * eBkn.transpose() ;	  
+	  eCmn = beta * 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
-	  else
-	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
-	  });
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  } );
       } else { 
 	assert(0);
       }
@@ -902,40 +809,28 @@ 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
-	  else
-	    eCmn = alpha * eAmk * eBkn ;
+	  eCmn = beta * 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
-	  else
-	    eCmn = alpha * eAmk.transpose() * eBkn ;
+	  eCmn = beta * 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
-	  else
-	    eCmn = alpha * eAmk * eBkn.transpose() ;
+	  eCmn = beta * 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);
-	  if (std::abs(beta) != 0.0)
-	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
-	  else
-	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
+	  eCmn = beta * 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 <<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;
+    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;
   }
   
   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 <<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;
+    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;
 #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 <<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;
+    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;
 #endif
   }
 };

Grid/algorithms/iterative/AdefMrhs.h

@@ -63,12 +63,7 @@ class TwoLevelCGmrhs
   GridStopWatch SmoothTimer;
   GridStopWatch InsertTimer;
 
-  /*
-    Field rrr;
-  Field sss;
-  Field qqq;
-  Field zzz;
-  */  
+  
   // more most opertor functions
   TwoLevelCGmrhs(RealD tol,
 		 Integer maxit,
@@ -79,12 +74,6 @@ class TwoLevelCGmrhs
     MaxIterations(maxit),
     _FineLinop(FineLinop),
     _Smoother(Smoother)
-    /*
-    rrr(fine),
-    sss(fine),
-    qqq(fine),
-    zzz(fine)
-*/
   {
     grid       = fine;
   };
@@ -92,8 +81,8 @@ 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);
   }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -668,8 +657,6 @@ 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();
@@ -682,7 +669,6 @@ public:
       this->SmoothTimer.Stop();
     }
 #endif
-    //    this->sss=Min[0];
     
     for(int rhs=0;rhs<nrhs;rhs++) {
       
@@ -719,11 +705,9 @@ 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/ConjugateGradientMixedPrec.h

@@ -116,14 +116,14 @@ NAMESPACE_BEGIN(Grid);
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
       RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
-      std::cout<<GridLogMessage<<" rsd norm "<<norm<<std::endl;
+      
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
 
       if(norm < OuterLoopNormMult * stop){
 	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
 	break;
       }
-      while(norm * inner_tol * inner_tol < stop*1.01) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+      while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
       precisionChange(src_f, src_d, pc_wk_dp_to_sp);

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -102,11 +102,11 @@ public:
     assert(mass.size()==nshift);
     assert(mresidual.size()==nshift);
   
-    // remove dynamic sized arrays on stack; 2d is a pain with vector
-    std::vector<RealD>  bs(nshift);
-    std::vector<RealD>  rsq(nshift);
-    std::vector<std::array<RealD,2> >  z(nshift);
-    std::vector<int>     converged(nshift);
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
   
     const int       primary =0;
   

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -123,11 +123,11 @@ public:
     assert(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    std::vector<RealD>  bs(nshift);
-    std::vector<RealD>  rsq(nshift);
-    std::vector<RealD>  rsqf(nshift);
-    std::vector<std::array<RealD,2> >  z(nshift);
-    std::vector<int>     converged(nshift);
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  rsqf[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
   
     const int       primary =0;
   

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -156,11 +156,11 @@ public:
     assert(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    std::vector<RealD>  bs(nshift);
-    std::vector<RealD>  rsq(nshift);
-    std::vector<RealD>  rsqf(nshift);
-    std::vector<std::array<RealD,2> >  z(nshift);
-    std::vector<int>     converged(nshift);
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  rsqf[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
   
     const int       primary =0;
   

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -245,10 +245,9 @@ 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(tmp,tmp)); // HermOp^2.
+	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
 	RealD vden = norm2(src_n);
-	RealD na = std::sqrt(vnum/vden);
+	RealD na = vnum/vden;
 	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
 	  i=_MAX_ITER_IRL_MEVAPP_;
 	evalMaxApprox = na;
@@ -256,7 +255,6 @@ 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

@@ -30,8 +30,6 @@ 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)
@@ -97,7 +95,7 @@ public:
 
     RealD scale;
 
-    ConjugateGradient<FineField> CG(1.0e-3,400,false);
+    ConjugateGradient<FineField> CG(1.0e-2,100,false);
     FineField noise(FineGrid);
     FineField Mn(FineGrid);
 
@@ -110,7 +108,7 @@ public:
       
       hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
 
-      for(int i=0;i<4;i++){
+      for(int i=0;i<1;i++){
 
 	CG(hermop,noise,subspace[b]);
 
@@ -126,53 +124,6 @@ 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
@@ -209,21 +160,14 @@ 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);
-      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;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
       b++;
     }
 
@@ -269,18 +213,8 @@ public:
 	  Mn=*Tnp;
 	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
 	  subspace[b] = Mn;
-
-
-	  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;
-	  
+	  hermop.Op(Mn,tmp); 
+	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
 	  b++;
 	}
 
@@ -294,70 +228,6 @@ 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/CoarsenedMatrix.h

@@ -99,7 +99,7 @@ public:
   CoarseMatrix AselfInvEven;
   CoarseMatrix AselfInvOdd;
 
-  deviceVector<RealD> dag_factor;
+  Vector<RealD> dag_factor;
 
   ///////////////////////
   // Interface
@@ -124,13 +124,9 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
-    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
-    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+    Vector<Aview> AcceleratorViewContainer;
   
-    for(int p=0;p<geom.npoint;p++) {
-      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
-      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
-    }
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -165,7 +161,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   };
 
   void Mdag (const CoarseVector &in, CoarseVector &out)
@@ -194,14 +190,9 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
+    Vector<Aview> AcceleratorViewContainer;
 
-    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
-    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
-  
-    for(int p=0;p<geom.npoint;p++) {
-      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
-      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
-    }
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -210,10 +201,10 @@ public:
 
     int osites=Grid()->oSites();
 
-    deviceVector<int> points(geom.npoint);
-    for(int p=0; p<geom.npoint; p++) { 
-      acceleratorPut(points[p],geom.points_dagger[p]);
-    }
+    Vector<int> points(geom.npoint, 0);
+    for(int p=0; p<geom.npoint; p++)
+      points[p] = geom.points_dagger[p];
+
     auto points_p = &points[0];
 
     RealD* dag_factor_p = &dag_factor[0];
@@ -245,7 +236,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   }
 
   void MdirComms(const CoarseVector &in)
@@ -260,14 +251,8 @@ public:
     out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
-
-    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
-    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
-  
-    for(int p=0;p<geom.npoint;p++) {
-      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
-      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
-    }
+    Vector<Aview> AcceleratorViewContainer;
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     autoView( out_v , out, AcceleratorWrite);
@@ -300,7 +285,7 @@ public:
       }
       coalescedWrite(out_v[ss](b),res);
     });
-    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
   }
   void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
   {
@@ -484,20 +469,14 @@ public:
 
     // determine in what order we need the points
     int npoint = geom.npoint-1;
-    deviceVector<int> points(npoint);
-    for(int p=0; p<npoint; p++) {
-      int val = (dag && !hermitian) ? geom.points_dagger[p] : p;
-      acceleratorPut(points[p], val);
-    }
+    Vector<int> points(npoint, 0);
+    for(int p=0; p<npoint; p++)
+      points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
+
     auto points_p = &points[0];
 
-    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
-    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
-  
-    for(int p=0;p<geom.npoint;p++) {
-      hAcceleratorViewContainer[p] = a[p].View(AcceleratorRead);
-      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
-    }
+    Vector<Aview> AcceleratorViewContainer;
+    for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead));
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -560,7 +539,7 @@ public:
       });
     }
 
-    for(int p=0;p<npoint;p++) hAcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose();
   }
   
   CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
@@ -611,13 +590,11 @@ public:
     }
 
     // GPU readable prefactor
-    std::vector<RealD> h_dag_factor(nbasis*nbasis);
     thread_for(i, nbasis*nbasis, {
       int j = i/nbasis;
       int k = i%nbasis;
-      h_dag_factor[i] = dag_factor_eigen(j, k);
+      dag_factor[i] = dag_factor_eigen(j, k);
     });
-    acceleratorCopyToDevice(&h_dag_factor[0],&dag_factor[0],dag_factor.size()*sizeof(RealD));
   }
 
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -441,20 +441,8 @@ 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();
@@ -470,9 +458,11 @@ public:
     // Orthogonalise the subblocks over the basis
     /////////////////////////////////////////////////////////////
     CoarseScalar InnerProd(CoarseGrid()); 
-    blockOrthogonalise(InnerProd,V.subspace);
-    blockOrthogonalise(InnerProd,U.subspace);
+    blockOrthogonalise(InnerProd,Subspace.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();
@@ -552,7 +542,7 @@ 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]*V.subspace[i];
+	phaV = phaF[p]*Subspace.subspace[i];
 	tphaseBZ+=usecond();
 
 	/////////////////////////////////////////////////////////////////////
@@ -565,7 +555,7 @@ public:
 	//	std::cout << i << " " <<p << " MphaV "<<norm2(MphaV)<<" "<<norm2(phaV)<<std::endl;
 
 	tproj-=usecond();
-	blockProject(coarseInner,MphaV,U.subspace);
+	blockProject(coarseInner,MphaV,Subspace.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) { };
+  void construct(pointer __p, const _Tp& __val) { assert(0);};
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
@@ -174,10 +174,19 @@ 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> >;               // 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
+#ifdef ACCELERATOR_CSHIFT
+// Cshift on device
+template<class T> using cshiftAllocator = devAllocator<T>;
+#else
+// Cshift on host
+template<class T> using cshiftAllocator = std::allocator<T>;
+#endif
+
+template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
+template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
+template<class T> using commVector    = std::vector<T,devAllocator<T> >;
+template<class T> using deviceVector  = std::vector<T,devAllocator<T> >;
+template<class T> using cshiftVector  = std::vector<T,cshiftAllocator<T> >;
 
 /*
 template<class T> class vecView
@@ -188,9 +197,8 @@ template<class T> class vecView
   ViewMode mode;
   void * cpu_ptr;
  public:
-  // Rvalue accessor
   accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
-  vecView(Vector<T> &refer_to_me,ViewMode _mode)
+  vecView(std::vector<T> &refer_to_me,ViewMode _mode)
   {
     cpu_ptr = &refer_to_me[0];
     size = refer_to_me.size();
@@ -206,12 +214,22 @@ template<class T> class vecView
   }
 };
 
-template<class T> vecView<T> VectorView(Vector<T> &vec,ViewMode _mode)
+template<class T> vecView<T> VectorView(std::vector<T> &vec,ViewMode _mode)
 {
   vecView<T> ret(vec,_mode); // does the open
   return ret;                // must be closed
 }
 
+// Little autoscope assister
+template<class View> 
+class VectorViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+ public:
+  VectorViewCloser(View &_v) : v(_v) {};
+  ~VectorViewCloser() { auto ptr = v.cpu_ptr; v.ViewClose();  MemoryManager::NotifyDeletion(ptr);}
+};
+
 #define autoVecView(v_v,v,mode)					\
   auto v_v = VectorView(v,mode);				\
   ViewCloser<decltype(v_v)> _autoView##v_v(v_v);

Grid/allocator/MemoryManagerCache.cc

@@ -1,15 +1,16 @@
 #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 << std::endl;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug  << print_buffer << std::endl;
+#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 dprintf(...) 
-//#define mprintf(...) 
+
 
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
@@ -110,7 +111,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Discard(%lx) %lx",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -120,7 +121,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
     AccCache.AccPtr=(uint64_t) NULL;
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -140,7 +141,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld",
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld\n",
 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
   if (AccCache.accLock!=0) return;
@@ -154,7 +155,7 @@ 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 ",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   //  uint64_t CpuPtr = AccCache.CpuPtr;
   DeviceEvictions++;
@@ -168,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -183,9 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx",
-	  (uint64_t)AccCache.bytes,
-	  (uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone AccPtr %lx <- CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -211,7 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewClose %lx",(uint64_t)Ptr);
+    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -223,7 +222,7 @@ 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",(uint64_t)CpuPtr);
+    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -234,9 +233,6 @@ 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){
@@ -269,7 +265,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld",
+    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
@@ -309,7 +305,7 @@ 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",AccCache.accLock);
+    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -322,21 +318,21 @@ 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",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d\n",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",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d\n",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",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
@@ -345,7 +341,7 @@ 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 ");
+    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -368,10 +364,10 @@ 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",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
   } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)

Grid/allocator/MemoryStats.cc

@@ -15,10 +15,10 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
   uint64_t virt_pfn = (uint64_t)Buf / page_size;
   off_t offset = sizeof(uint64_t) * virt_pfn;
   uint64_t npages = (BYTES + page_size-1) / page_size;
-  std::vector<uint64_t> pagedata(npages);
+  uint64_t pagedata[npages];
   uint64_t ret = lseek(fd, offset, SEEK_SET);
   assert(ret == offset);
-  ret = ::read(fd, &pagedata[0], sizeof(uint64_t)*npages);
+  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
   assert(ret == sizeof(uint64_t) * npages);
   int nhugepages = npages / 512;
   int n4ktotal, nnothuge;

Grid/communicator/Communicator_base.cc

@@ -57,29 +57,18 @@ int                      CartesianCommunicator::ProcessorCount(void)    { return
 // very VERY rarely (Log, serial RNG) we need world without a grid
 ////////////////////////////////////////////////////////////////////////////////
 
-#ifdef USE_GRID_REDUCTION
-void CartesianCommunicator::GlobalSum(ComplexF &c)
-{
-  GlobalSumP2P(c);
-}
-void CartesianCommunicator::GlobalSum(ComplexD &c)
-{
-  GlobalSumP2P(c);
-}
-#else
 void CartesianCommunicator::GlobalSum(ComplexF &c)
 {
   GlobalSumVector((float *)&c,2);
 }
-void CartesianCommunicator::GlobalSum(ComplexD &c)
-{
-  GlobalSumVector((double *)&c,2);
-}
-#endif
 void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
 {
   GlobalSumVector((float *)c,2*N);
 }
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumVector((double *)&c,2);
+}
 void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);

Grid/communicator/Communicator_base.h

@@ -33,8 +33,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
-#define NVLINK_GET
-
 NAMESPACE_BEGIN(Grid);
 
 extern bool Stencil_force_mpi ;
@@ -129,36 +127,7 @@ 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;
-    obj accum = o;
-    int source,dest;
-    for(int d=0;d<_ndimension;d++){
-      column.resize(_processors[d]);
-      column[0] = accum;
-      std::vector<MpiCommsRequest_t> list;
-      for(int p=1;p<_processors[d];p++){
-	ShiftedRanks(d,p,source,dest);
-	SendToRecvFromBegin(list,
-			    &column[0],
-			    dest,
-			    &column[p],
-			    source,
-			    sizeof(obj),d*100+p);
-
-      }
-      if (!list.empty()) // avoid triggering assert in comms == none
-	CommsComplete(list);
-      for(int p=1;p<_processors[d];p++){
-	accum = accum + column[p];
-      }
-    }
-    Broadcast(0,accum);
-    o=accum;
-  }
-
+  
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
@@ -169,8 +138,8 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
-  void CommsComplete(std::vector<MpiCommsRequest_t> &list);
-  void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
+  void CommsComplete(std::vector<CommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 			   void *xmit,
 			   int dest,
 			   void *recv,
@@ -183,28 +152,16 @@ public:
 		      int recv_from_rank,
 		      int bytes);
   
-  int IsOffNode(int rank);
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,int do_xmit,
 			       void *recv,
 			       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,void *xmit_comp,
+				    void *xmit,
 				    int xmit_to_rank,int do_xmit,
-				    void *recv,void *recv_comp,
+				    void *recv,
 				    int recv_from_rank,int do_recv,
 				    int xbytes,int rbytes,int dir);
   

Grid/communicator/Communicator_mpi3.cc

@@ -30,7 +30,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
 ////////////////////////////////////////////
@@ -258,41 +257,15 @@ 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 float");
-  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 double");
-  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 uint32_t");
   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 uint64_t");
   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);
 }
@@ -301,36 +274,40 @@ void CartesianCommunicator::GlobalXOR(uint32_t &u){
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalXOR(uint64_t &u){
-  FlightRecorder::StepLog("GlobalXOR");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalMax(float &f)
 {
-  FlightRecorder::StepLog("GlobalMax");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
   assert(ierr==0);
 }
 void CartesianCommunicator::GlobalMax(double &d)
 {
-  FlightRecorder::StepLog("GlobalMax");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalSum(float &f){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSumVector(float *f,int N)
 {
-  FlightRecorder::StepLog("GlobalSumVector(float *)");
   int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSumVector(double *d,int N)
 {
-  FlightRecorder::StepLog("GlobalSumVector(double *)");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
   assert(ierr==0);
 }
 
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
 						int dest,
 						void *recv,
@@ -355,7 +332,7 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &
   assert(ierr==0);
   list.push_back(xrq);
 }
-void CartesianCommunicator::CommsComplete(std::vector<MpiCommsRequest_t> &list)
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
 {
   int nreq=list.size();
 
@@ -374,7 +351,9 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int from,
 					   int bytes)
 {
-  std::vector<MpiCommsRequest_t> reqs(0);
+  std::vector<CommsRequest_t> reqs(0);
+  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
 
   int myrank = _processor;
   int ierr;
@@ -390,6 +369,9 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 		    communicator,MPI_STATUS_IGNORE);
   assert(ierr==0);
 
+  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
+  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
@@ -399,34 +381,16 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
-  offbytes       += StencilSendToRecvFromBegin(list,xmit,xmit,dest,dox,recv,recv,from,dor,bytes,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
-int CartesianCommunicator::IsOffNode(int rank)
-{
-  int grank = ShmRanks[rank];
-  if ( grank == MPI_UNDEFINED ) return true;
-  else return false;
-}
 
-#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
-}
+#undef NVLINK_GET // Define to use get instead of put DMA
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,void *xmit_comp,
+							 void *xmit,
 							 int dest,int dox,
-							 void *recv,void *recv_comp,
+							 void *recv,
 							 int from,int dor,
 							 int xbytes,int rbytes,int dir)
 {
@@ -446,30 +410,27 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
   int tag;
-  
+
   if ( dor ) {
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
-      //      std::cout << " StencilSendToRecvFrom "<<dir<<" MPI_Irecv "<<std::hex<<recv<<std::dec<<std::endl;
-      ierr=MPI_Irecv(recv_comp, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      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);
-      //      std::cout << " StencilSendToRecvFrom "<<dir<<" CopyDeviceToDevice recv "<<std::hex<<recv<<" remote "<<shm <<std::dec<<std::endl;
       acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
-    }
 #endif
   }
-  // This is a NVLINK PUT  
+  
   if (dox) {
+    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit_comp, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
       assert(ierr==0);
       list.push_back(xrq);
       off_node_bytes+=xbytes;
@@ -479,341 +440,27 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       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,void *xmit_comp,
-							 int dest,int dox,
-							 void *recv,void *recv_comp,
-							 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;
-
-  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);
-
-      CommsRequest_t srq;
-
-      srq.ev = acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
-
-      srq.PacketType = IntraNodeRecv;
-      srq.bytes      = xbytes;
-      //      srq.req        = xrq;
-      srq.host_buf   = NULL;
-      srq.device_buf = xmit;
-      srq.tag        = -1;
-      srq.dest       = dest;
-      srq.commdir    = dir;
-      list.push_back(srq);
-    }
-  }  
-#else
-  if (dox) {
-
-    if ( !( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) ) {
-      // Intranode
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
-
-      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)
-{
-  acceleratorCopySynchronise(); // Complete all pending copy transfers D2D
-
-  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
-  }
-
-  int nreq=MpiRequests.size();
-
-  if (nreq>0) {
-    status.resize(MpiRequests.size());
-    int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.
-    assert(ierr==0);
-  }
-  
-  //  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)
@@ -821,13 +468,11 @@ 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,
@@ -841,13 +486,11 @@ int CartesianCommunicator::RankWorld(void){
   return r;
 }
 void CartesianCommunicator::BarrierWorld(void){
-  FlightRecorder::StepLog("BarrierWorld");
   int ierr = MPI_Barrier(communicator_world);
   assert(ierr==0);
 }
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
-  FlightRecorder::StepLog("BroadcastWorld");
   int ierr= MPI_Bcast(data,
 		      bytes,
 		      MPI_BYTE,
@@ -870,7 +513,6 @@ 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(list.size()==0);}
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
 void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
 						int dest,
@@ -124,8 +124,6 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   dest=0;
 }
 
-int CartesianCommunicator::IsOffNode(int rank) { return false; }
-
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int xmit_to_rank,int dox,
 						     void *recv,
@@ -134,17 +132,6 @@ 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

@@ -46,40 +46,8 @@ 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
@@ -137,7 +105,7 @@ 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

@@ -42,11 +42,6 @@ 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
@@ -542,22 +537,19 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
 #ifndef ACCELERATOR_AWARE_MPI
-  // printf("Host buffer allocate for GPU non-aware MPI\n");
-  HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
+  HostCommBuf= malloc(bytes);
 #endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << "SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << Mheader " acceleratorAllocDevice "<< bytes 
+    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
-  if ( WorldRank == 0 ){
-    std::cout<< Mheader "Setting up IPC"<<std::endl;
-  }
+  std::cout<< "Setting up IPC"<<std::endl;
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -577,8 +569,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
@@ -588,6 +580,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       if ( err != ZE_RESULT_SUCCESS ) {
 	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
+      } else {
+	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
@@ -646,12 +640,12 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef SHM_SOCKETS
       myfd=UnixSockets::RecvFileDescriptor();
 #else
-      //      std::cout<<"mapping seeking remote pid/fd "
-      //	       <<handle.pid<<"/"
-      //	       <<handle.fd<<std::endl;
+      std::cout<<"mapping seeking remote pid/fd "
+	       <<handle.pid<<"/"
+	       <<handle.fd<<std::endl;
 
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
-      //      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
+      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
       myfd  = syscall(438,pidfd,handle.fd,0);
       int err_t = errno;
@@ -661,7 +655,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	assert(0);
       }
 #endif
-      //      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
       memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
@@ -670,6 +664,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
 	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
+      } else {
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
+	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl;
       }
       assert(thisBuf!=nullptr);
     }
@@ -750,7 +747,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     WorldShmCommBufs[r] =ptr;
     //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
-  std::cout<< Mheader " Intra-node IPC setup is complete "<<std::endl;
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
 };
@@ -884,14 +880,14 @@ 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
@@ -927,7 +923,6 @@ 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();
 
@@ -958,7 +953,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
 #endif
 
-  //  SharedMemoryTest();
+  //SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier
@@ -980,18 +975,19 @@ void SharedMemory::SharedMemoryTest(void)
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
        check[2]=magic;
-       acceleratorCopyToDevice(check,ShmCommBufs[r],3*sizeof(uint64_t));
+       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
   for(uint64_t r=0;r<ShmSize;r++){
-    acceleratorCopyFromDevice(ShmCommBufs[r],check,3*sizeof(uint64_t));
+    ShmBarrier();
+    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
+    ShmBarrier();
     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)
@@ -1007,13 +1003,11 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 {
   int gpeer = ShmRanks[rank];
   assert(gpeer!=ShmRank); // never send to self
-  //  std::cout << "ShmBufferTranslate for rank " << rank<<" peer "<<gpeer<<std::endl;
   if (gpeer == MPI_UNDEFINED){
     return NULL;
   } else { 
     uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
     uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
-    //    std::cout << "ShmBufferTranslate : local,offset,remote "<<std::hex<<local_p<<" "<<offset<<" "<<remote<<std::dec<<std::endl;
     return (void *) remote;
   }
 }

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.h

@@ -51,6 +51,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif 
 
 NAMESPACE_BEGIN(Grid);
+
 template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
 auto Cshift(const Expression &expr,int dim,int shift)  -> decltype(closure(expr)) 
 {

Grid/cshift/Cshift_common.h

@@ -30,11 +30,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 extern std::vector<std::pair<int,int> > Cshift_table; 
-extern deviceVector<std::pair<int,int> > Cshift_table_device; 
+extern commVector<std::pair<int,int> > Cshift_table_device; 
 
 inline std::pair<int,int> *MapCshiftTable(void)
 {
   // GPU version
+#ifdef ACCELERATOR_CSHIFT    
   uint64_t sz=Cshift_table.size();
   if (Cshift_table_device.size()!=sz )    {
     Cshift_table_device.resize(sz);
@@ -44,13 +45,16 @@ inline std::pair<int,int> *MapCshiftTable(void)
 			  sizeof(Cshift_table[0])*sz);
 
   return &Cshift_table_device[0];
+#else 
+  return &Cshift_table[0];
+#endif
   // CPU version use identify map
 }
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -90,10 +94,17 @@ Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dim
   {
     auto buffer_p = & buffer[0];
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for(i,ent,{
+      buffer_p[table[i].first]=rhs_v[table[i].second];
+    });
+#endif
   }
 }
 
@@ -118,6 +129,7 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -128,10 +140,21 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+	int o      =   n*n1;
+	int offset = b+n*e2;
+	
+	vobj temp =rhs_v[so+o+b];
+	extract<vobj>(temp,pointers,offset);
+      });
+#endif
   } else { 
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
+#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -152,13 +175,33 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+
+	Coordinate coor;
+
+	int o=n*n1;
+	int oindex = o+b;
+
+       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
+
+	int ocb=1<<cb;
+	int offset = b+n*e2;
+
+	if ( ocb & cbmask ) {
+	  vobj temp =rhs_v[so+o+b];
+	  extract<vobj>(temp,pointers,offset);
+	}
+      });
+#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -202,10 +245,17 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<
   {
     auto buffer_p = & buffer[0];
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
     });
+#else
+    autoView( rhs_v, rhs, CpuWrite);
+    thread_for(i,ent,{
+      rhs_v[table[i].first]=buffer_p[table[i].second];
+    });
+#endif
   }
 }
 
@@ -228,6 +278,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   if(cbmask ==0x3 ) {
     int _slice_stride = rhs.Grid()->_slice_stride[dimension];
     int _slice_block = rhs.Grid()->_slice_block[dimension];
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v , rhs, AcceleratorWrite);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -236,6 +287,14 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
+#else
+    autoView( rhs_v , rhs, CpuWrite);
+    thread_for2d(n,e1,b,e2,{
+	int o      = n*_slice_stride;
+	int offset = b+n*_slice_block;
+	merge(rhs_v[so+o+b],pointers,offset);
+    });
+#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
@@ -301,11 +360,19 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
 
   {
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
       coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    autoView(lhs_v , lhs, CpuWrite);
+    thread_for(i,ent,{
+      lhs_v[table[i].first]=rhs_v[table[i].second];
+    });
+#endif
   }
 }
 
@@ -345,11 +412,19 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
 
   {
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
+#else
+    autoView( rhs_v, rhs, CpuRead);
+    autoView( lhs_v, lhs, CpuWrite);
+    thread_for(i,ent,{
+      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+    });
+#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

@@ -31,7 +31,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 
 NAMESPACE_BEGIN(Grid); 
-const int Cshift_verbose=0;
+
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
   typedef typename vobj::vector_type vector_type;
@@ -55,17 +55,17 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   RealD t1,t0;
   t0=usecond();
   if ( !comm_dim ) {
-    //    std::cout << "CSHIFT: Cshift_local" <<std::endl;
+    //std::cout << "CSHIFT: Cshift_local" <<std::endl;
     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
   } else if ( splice_dim ) {
-    //    std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
+    //std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift);
   } else {
-    //    std::cout << "CSHIFT: Cshift_comms" <<std::endl;
+    //std::cout << "CSHIFT: Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift);
   }
   t1=usecond();
-  if(Cshift_verbose) std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
+  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
 
@@ -94,16 +94,18 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
   sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
   sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
-  //  std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
+  //std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
   if ( sshift[0] == sshift[1] ) {
-    //    std::cout << "Single pass Cshift_comms" <<std::endl;
+    //std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
   } else {
-    //    std::cout << "Two pass Cshift_comms" <<std::endl;
+    //std::cout << "Two pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
 }
+#define ACCELERATOR_CSHIFT_NO_COPY
+#ifdef ACCELERATOR_CSHIFT_NO_COPY
 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;
@@ -123,13 +125,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   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
-  
+  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
+  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
+    
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
   RealD tcopy=0.0;
@@ -143,11 +141,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     int comm_proc = ((x+sshift)/rd)%pd;
     
     if (comm_proc==0) {
-      FlightRecorder::StepLog("Cshift_Copy_plane");
       tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
       tcopy+=usecond();
-      FlightRecorder::StepLog("Cshift_Copy_plane_complete");
     } else {
 
       int words = buffer_size;
@@ -155,57 +151,39 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
-      FlightRecorder::StepLog("Cshift_Gather_plane");
       tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
       tgather+=usecond();
-      FlightRecorder::StepLog("Cshift_Gather_plane_complete");
 
       //      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();
+      //      grid->Barrier();
 
-      FlightRecorder::StepLog("Cshift_SendRecv");
-#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
-      FlightRecorder::StepLog("Cshift_SendRecv_complete");
-
       xbytes+=bytes;
-      grid->Barrier();
+      //      grid->Barrier();
       tcomms+=usecond();
-      FlightRecorder::StepLog("Cshift_barrier_complete");
 
       tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,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;
-  }
+  /*
+  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)
@@ -223,9 +201,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   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;
+  //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);
@@ -246,20 +224,16 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  static std::vector<deviceVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<deviceVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
+  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;
-
+ 
   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); // 
@@ -307,29 +281,18 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
 	tcomms-=usecond();
-	grid->Barrier();
+	//	grid->Barrier();
 
 	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();
+	//	grid->Barrier();
 	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
@@ -342,15 +305,242 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     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 "<<2*xbytes<< " Bytes "<<std::endl;
+  /*
+  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 "<<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();
+    }
+  }
+  /*
+  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();
+
+  }
+  /*
+  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); 
 
 #endif

Grid/cshift/Cshift_table.cc

@@ -1,5 +1,5 @@
 #include <Grid/GridCore.h>       
 NAMESPACE_BEGIN(Grid);
 std::vector<std::pair<int,int> > Cshift_table; 
-deviceVector<std::pair<int,int> > Cshift_table_device; 
+commVector<std::pair<int,int> > Cshift_table_device; 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_arith.h

@@ -257,30 +257,17 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   });
 }
 
-#define FAST_AXPY_NORM
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpy_norm");
-#ifdef FAST_AXPY_NORM
-  return axpy_norm_fast(ret,a,x,y);
-#else
-  ret = a*x+y;
-  RealD nn=norm2(ret);
-  return nn;
-#endif
+    return axpy_norm_fast(ret,a,x,y);
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpby_norm");
-#ifdef FAST_AXPY_NORM
-  return axpby_norm_fast(ret,a,b,x,y);
-#else
-  ret = a*x+b*y;
-  RealD nn=norm2(ret);
-  return nn;
-#endif
+    return axpby_norm_fast(ret,a,b,x,y);
 }
 
 /// Trace product

Grid/lattice/Lattice_base.h

@@ -236,13 +236,10 @@ public:
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
     vobj vtmp;
     vtmp = r;
-#if 1
-    deviceVector<vobj> vvtmp(1);
-    acceleratorPut(vvtmp[0],vtmp);
-    vobj *vvtmp_p = & vvtmp[0];
+#if defined(GRID_HIP) || defined(GRID_CUDA) || defined (GRID_SYCL)
     auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
-	auto stmp=coalescedRead(*vvtmp_p);
+	auto stmp=coalescedRead(vtmp);
 	coalescedWrite(me[ss],stmp);
     });
 #else    

Grid/lattice/Lattice_basis.h

@@ -53,19 +53,36 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   typedef decltype(basis[0]) Field;
   typedef decltype(basis[0].View(AcceleratorRead)) View;
 
-  hostVector<View>  h_basis_v(basis.size());
-  deviceVector<View> d_basis_v(basis.size());
-  typedef typename std::remove_reference<decltype(h_basis_v[0][0])>::type vobj;
+  Vector<View> basis_v; basis_v.reserve(basis.size());
+  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
   typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
-
   GridBase* grid = basis[0].Grid();
       
   for(int k=0;k<basis.size();k++){
-    h_basis_v[k] = basis[k].View(AcceleratorWrite);
-    acceleratorPut(d_basis_v[k],h_basis_v[k]);
+    basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
-  View *basis_vp = &d_basis_v[0];
+#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
+  int max_threads = thread_max();
+  Vector < vobj > Bt(Nm * max_threads);
+  thread_region
+    {
+      vobj* B = &Bt[Nm * thread_num()];
+      thread_for_in_region(ss, grid->oSites(),{
+	  for(int j=j0; j<j1; ++j) B[j]=0.;
+      
+	  for(int j=j0; j<j1; ++j){
+	    for(int k=k0; k<k1; ++k){
+	      B[j] +=Qt(j,k) * basis_v[k][ss];
+	    }
+	  }
+	  for(int j=j0; j<j1; ++j){
+	    basis_v[j][ss] = B[j];
+	  }
+	});
+    }
+#else
+  View *basis_vp = &basis_v[0];
 
   int nrot = j1-j0;
   if (!nrot) // edge case not handled gracefully by Cuda
@@ -74,19 +91,17 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   uint64_t oSites   =grid->oSites();
   uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
 
-  deviceVector <vobj> Bt(siteBlock * nrot); 
+  Vector <vobj> Bt(siteBlock * nrot); 
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  hostVector<Coeff_t> h_Qt_jv(Nm*Nm);
-  deviceVector<Coeff_t> Qt_jv(Nm*Nm);
+  Vector<Coeff_t> Qt_jv(Nm*Nm);
   Coeff_t *Qt_p = & Qt_jv[0];
   thread_for(i,Nm*Nm,{
       int j = i/Nm;
       int k = i%Nm;
-      h_Qt_jv[i]=Qt(j,k);
+      Qt_p[i]=Qt(j,k);
   });
-  acceleratorCopyToDevice(&h_Qt_jv[0],Qt_p,Nm*Nm*sizeof(Coeff_t));
 
   // Block the loop to keep storage footprint down
   for(uint64_t s=0;s<oSites;s+=siteBlock){
@@ -122,8 +137,9 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
       });
   }
+#endif
 
-  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
 }
 
 // Extract a single rotated vector
@@ -136,19 +152,16 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
 
   result.Checkerboard() = basis[0].Checkerboard();
 
-  hostVector<View>  h_basis_v(basis.size());
-  deviceVector<View> d_basis_v(basis.size());
+  Vector<View> basis_v; basis_v.reserve(basis.size());
   for(int k=0;k<basis.size();k++){
-    h_basis_v[k]=basis[k].View(AcceleratorRead);
-    acceleratorPut(d_basis_v[k],h_basis_v[k]);
+    basis_v.push_back(basis[k].View(AcceleratorRead));
   }
-
   vobj zz=Zero();
-  deviceVector<double> Qt_jv(Nm);
+  Vector<double> Qt_jv(Nm);
   double * Qt_j = & Qt_jv[0];
-  for(int k=0;k<Nm;++k) acceleratorPut(Qt_j[k],Qt(j,k));
+  for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
-  auto basis_vp=& d_basis_v[0];
+  auto basis_vp=& basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
     vobj zzz=Zero();
@@ -158,7 +171,7 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
     }
     coalescedWrite(result_v[ss], B);
   });
-  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
 }
 
 template<class Field>

Grid/lattice/Lattice_reduction.h

@@ -46,7 +46,7 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
   //  const int Nsimd = vobj::Nsimd();
   const int nthread = GridThread::GetThreads();
 
-  std::vector<sobj> sumarray(nthread);
+  Vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -75,7 +75,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
 
   const int nthread = GridThread::GetThreads();
 
-  std::vector<sobj> sumarray(nthread);
+  Vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -290,10 +290,8 @@ template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
 
-  bool ok;
 #ifdef GRID_SYCL
   uint64_t csum=0;
-  uint64_t csum2=0;
   if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
   {
     // Hack
@@ -302,33 +300,13 @@ inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &righ
     Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
     uint64_t *base= (uint64_t *)&l_v[0];
     csum=svm_xor(base,words);
-    ok = FlightRecorder::CsumLog(csum);
-    if ( !ok ) {
-      csum2=svm_xor(base,words);
-      std::cerr<< " Bad CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
-    } else {
-      //      csum2=svm_xor(base,words);
-      //      std::cerr<< " ok CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
-    }
-    assert(ok);
   }
+  FlightRecorder::CsumLog(csum);
 #endif
-  FlightRecorder::StepLog("rank inner product");
   ComplexD nrm = rankInnerProduct(left,right);
-  //  ComplexD nrmck=nrm;
   RealD local = real(nrm);
-  ok = FlightRecorder::NormLog(real(nrm));
-  if ( !ok ) {
-    ComplexD nrm2 = rankInnerProduct(left,right);
-    RealD local2 = real(nrm2);
-    std::cerr<< " Bad NORM " << local << " recomputed as "<<local2<<std::endl;
-    assert(ok);
-  }
-  FlightRecorder::StepLog("Start global sum");
-  grid->GlobalSumP2P(nrm);
-  //  grid->GlobalSum(nrm);
-  FlightRecorder::StepLog("Finished global sum");
-  //  std::cout << " norm "<< nrm << " p2p norm "<<nrmck<<std::endl;
+  FlightRecorder::NormLog(real(nrm)); 
+  grid->GlobalSum(nrm);
   FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
@@ -365,6 +343,18 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
+#if 0
+  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
+  });
+  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
+#else
   typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
   deviceVector<inner_t> inner_tmp;
   inner_tmp.resize(sites);
@@ -375,44 +365,9 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
       coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
       coalescedWrite(z_v[ss],tmp);
   });
-  bool ok;
-#ifdef GRID_SYCL
-  uint64_t csum=0;
-  uint64_t csum2=0;
-  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
-  {
-    // z_v
-    {
-      Integer words = sites*sizeof(vobj)/sizeof(uint64_t);
-      uint64_t *base= (uint64_t *)&z_v[0];
-      csum=svm_xor(base,words);
-      ok = FlightRecorder::CsumLog(csum);
-      if ( !ok ) {
-	csum2=svm_xor(base,words);
-	std::cerr<< " Bad z_v CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
-      }
-      assert(ok);
-    }
-    // inner_v
-    {
-      Integer words = sites*sizeof(inner_t)/sizeof(uint64_t);
-      uint64_t *base= (uint64_t *)&inner_tmp_v[0];
-      csum=svm_xor(base,words);
-      ok = FlightRecorder::CsumLog(csum);
-      if ( !ok ) {
-	csum2=svm_xor(base,words);
-	std::cerr<< " Bad inner_tmp_v CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
-      }
-      assert(ok);
-    }
-  }
-#endif
   nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-  ok = FlightRecorder::NormLog(real(nrm));
-  assert(ok);
-  RealD local = real(nrm);
+#endif
   grid->GlobalSum(nrm);
-  FlightRecorder::ReductionLog(local,real(nrm));
   return nrm; 
 }
  
@@ -422,7 +377,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   conformable(left,right);
 
   typedef typename vobj::vector_typeD vector_type;
-  std::vector<ComplexD> tmp(2);
+  Vector<ComplexD> tmp(2);
 
   GridBase *grid = left.Grid();
 
@@ -432,8 +387,8 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   // GPU
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   typedef decltype(innerProductD(vobj(),vobj())) norm_t;
-  deviceVector<inner_t> inner_tmp(sites);
-  deviceVector<norm_t>  norm_tmp(sites);
+  Vector<inner_t> inner_tmp(sites);
+  Vector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
   {
@@ -483,9 +438,7 @@ inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
 // sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
-					  std::vector<typename vobj::scalar_object> &result,
-					  int orthogdim)
+template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
 {
   ///////////////////////////////////////////////////////
   // FIXME precision promoted summation
@@ -507,8 +460,8 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vobj> lvSum(rd); // will locally sum vectors first
-  std::vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  Vector<vobj> lvSum(rd); // will locally sum vectors first
+  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
   ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
 
   result.resize(fd); // And then global sum to return the same vector to every node 
@@ -556,8 +509,6 @@ 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);
-  //  std::cout << GridLogMessage << " sliceSum local"<<t_sum<<" us, host+mpi "<<t_rest<<std::endl;
-  
 }
 template<class vobj> inline
 std::vector<typename vobj::scalar_object> 
@@ -601,8 +552,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vector_type> lvSum(rd); // will locally sum vectors first
-  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  Vector<vector_type> lvSum(rd); // will locally sum vectors first
+  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
   ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
 
   result.resize(fd); // And then global sum to return the same vector to every node for IO to file

Grid/lattice/Lattice_reduction_gpu.h

@@ -214,12 +214,22 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   // Move out of UVM
   // Turns out I had messed up the synchronise after move to compute stream
   // as running this on the default stream fools the synchronise
-  deviceVector<sobj> buffer(numBlocks);
+#undef UVM_BLOCK_BUFFER  
+#ifndef UVM_BLOCK_BUFFER  
+  commVector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
   sobj result;
   reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
   acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
+#else
+  Vector<sobj> buffer(numBlocks);
+  sobj *buffer_v = &buffer[0];
+  sobj result;
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
+  accelerator_barrier();
+  result = *buffer_v;
+#endif
   return result;
 }
 
@@ -234,7 +244,7 @@ inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osi
   
   const int words = sizeof(vobj)/sizeof(vector);
 
-  deviceVector<vector> buffer(osites);
+  Vector<vector> buffer(osites);
   vector *dat = (vector *)lat;
   vector *buf = &buffer[0];
   iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];

Grid/lattice/Lattice_reduction_sycl.h

@@ -4,28 +4,33 @@ NAMESPACE_BEGIN(Grid);
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-
 template <class vobj>
 inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_objectD sobjD;
-
+  static Vector<sobj> mysum;
+  mysum.resize(1);
+  sobj *mysum_p = & mysum[0];
   sobj identity; zeroit(identity);
-  sobj ret; zeroit(ret);
+  mysum[0] = identity;
+  sobj ret ; 
+
   Integer nsimd= vobj::Nsimd();
-  { 
-    sycl::buffer<sobj, 1> abuff(&ret, {1});
-    theGridAccelerator->submit([&](sycl::handler &cgh) {
-      auto Reduction = sycl::reduction(abuff,cgh,identity,std::plus<>());
-      cgh.parallel_for(sycl::range<1>{osites},
-                      Reduction,
-                      [=] (sycl::id<1> item, auto &sum) {
-                        auto osite   = item[0];
-                        sum +=Reduce(lat[osite]);
-                      });
-    });
-  }
+
+  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+    auto Reduction = cl::sycl::reduction(mysum_p,identity,std::plus<>(),PropList);
+     cgh.parallel_for(cl::sycl::range<1>{osites},
+		      Reduction,
+		      [=] (cl::sycl::id<1> item, auto &sum) {
+      auto osite   = item[0];
+      sum +=Reduce(lat[osite]);
+     });
+   });
+  theGridAccelerator->wait();
+  ret = mysum[0];
+  //  free(mysum,*theGridAccelerator);
   sobjD dret; convertType(dret,ret);
   return dret;
 }
@@ -71,22 +76,59 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
 
 template<class Word> Word svm_xor(Word *vec,uint64_t L)
 {
+  Word xorResult; xorResult = 0;
+  static Vector<Word> d_sum;
+  d_sum.resize(1);
+  Word *d_sum_p=&d_sum[0];
   Word identity;  identity=0;
-  Word ret = 0;
-  { 
-    sycl::buffer<Word, 1> abuff(&ret, {1});
-    theGridAccelerator->submit([&](sycl::handler &cgh) {
-      auto Reduction = sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
-      cgh.parallel_for(sycl::range<1>{L},
-                      Reduction,
-                      [=] (sycl::id<1> index, auto &sum) {
-                        sum ^=vec[index];
-                      });
-    });
-  }
+  d_sum[0] = identity;
+  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+    auto Reduction = cl::sycl::reduction(d_sum_p,identity,std::bit_xor<>(),PropList);
+     cgh.parallel_for(cl::sycl::range<1>{L},
+		      Reduction,
+		      [=] (cl::sycl::id<1> index, auto &sum) {
+	 sum^=vec[index];
+     });
+   });
   theGridAccelerator->wait();
+  Word ret = d_sum[0];
+  //  free(d_sum,*theGridAccelerator);
   return ret;
 }
 
 NAMESPACE_END(Grid);
 
+/*
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_type  scalar;
+
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobjD;
+
+  sobjD ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int nsimd = vobj::Nsimd();
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<scalar> buffer(osites*nsimd);
+  scalar *buf = &buffer[0];
+  vector *dat = (vector *)lat;
+
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,nsimd,{
+	int lane = acceleratorSIMTlane(nsimd);
+	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
+    });
+    //Precision change at this point is to late to gain precision
+    ret_p[w] = svm_reduce(buf,nsimd*osites);
+  }
+  return ret;
+}
+*/

Grid/lattice/Lattice_slicesum_core.h

@@ -21,18 +21,9 @@ NAMESPACE_BEGIN(Grid);
 
 
 #if defined(GRID_CUDA) || defined(GRID_HIP)
-template<class vobj>
-inline void sliceSumReduction_cub_small(const vobj *Data,
-					std::vector<vobj> &lvSum,
-					const int rd,
-					const int e1,
-					const int e2,
-					const int stride,
-					const int ostride,
-					const int Nsimd)
-{
+template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
   size_t subvol_size = e1*e2;
-  deviceVector<vobj> reduction_buffer(rd*subvol_size);
+  commVector<vobj> reduction_buffer(rd*subvol_size);
   auto rb_p = &reduction_buffer[0];
   vobj zero_init;
   zeroit(zero_init);
@@ -55,7 +46,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
   d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
   
   //copy offsets to device
-  acceleratorCopyToDeviceAsynch(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  acceleratorCopyToDeviceAsync(&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 +79,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
     exit(EXIT_FAILURE);
   }
   
-  acceleratorCopyFromDeviceAsynch(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
   
   //sync after copy
   accelerator_barrier();
@@ -103,15 +94,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
 
 
 #if defined(GRID_SYCL)
-template<class vobj>
-inline void sliceSumReduction_sycl_small(const vobj *Data,
-					 std::vector <vobj> &lvSum,
-					 const int  &rd,
-					 const int &e1,
-					 const int &e2,
-					 const int &stride,
-					 const int &ostride,
-					 const int &Nsimd)
+template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
 {
   size_t subvol_size = e1*e2;
 
@@ -122,7 +105,7 @@ inline void sliceSumReduction_sycl_small(const vobj *Data,
     mysum[r] = vobj_zero; 
   }
 
-  deviceVector<vobj> reduction_buffer(rd*subvol_size);    
+  commVector<vobj> reduction_buffer(rd*subvol_size);    
 
   auto rb_p = &reduction_buffer[0];
 
@@ -141,11 +124,11 @@ inline void sliceSumReduction_sycl_small(const vobj *Data,
   });
 
   for (int r = 0; r < rd; r++) {
-      theGridAccelerator->submit([&](sycl::handler &cgh) {
-          auto Reduction = sycl::reduction(&mysum[r],std::plus<>());
-          cgh.parallel_for(sycl::range<1>{subvol_size},
+      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+          auto Reduction = cl::sycl::reduction(&mysum[r],std::plus<>());
+          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
           Reduction,
-          [=](sycl::id<1> item, auto &sum) {
+          [=](cl::sycl::id<1> item, auto &sum) {
               auto s = item[0];
               sum += rb_p[r*subvol_size+s];
           });
@@ -161,23 +144,14 @@ inline void sliceSumReduction_sycl_small(const vobj *Data,
 }
 #endif
 
-template<class vobj>
-inline void sliceSumReduction_large(const vobj *Data,
-				    std::vector<vobj> &lvSum,
-				    const int rd,
-				    const int e1,
-				    const int e2,
-				    const int stride,
-				    const int ostride,
-				    const int Nsimd)
-{
+template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
   typedef typename vobj::vector_type vector;
   const int words = sizeof(vobj)/sizeof(vector);
   const int osites = rd*e1*e2;
-  deviceVector<vector>buffer(osites);
+  commVector<vector>buffer(osites);
   vector *dat = (vector *)Data;
   vector *buf = &buffer[0];
-  std::vector<vector> lvSum_small(rd);
+  Vector<vector> lvSum_small(rd);
   vector *lvSum_ptr = (vector *)&lvSum[0];
 
   for (int w = 0; w < words; w++) {
@@ -194,18 +168,13 @@ inline void sliceSumReduction_large(const vobj *Data,
     for (int r = 0; r < rd; r++) {
       lvSum_ptr[w+words*r]=lvSum_small[r];
     }
+
   }
+
+  
 }
 
-template<class vobj>
-inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
-				  std::vector<vobj> &lvSum,
-				  const int rd,
-				  const int e1,
-				  const int e2,
-				  const int stride,
-				  const int ostride,
-				  const int Nsimd)
+template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
 {
   autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
     if constexpr (sizeof(vobj) <= 256) { 
@@ -223,15 +192,7 @@ inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
 }
 
 
-template<class vobj>
-inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
-				  std::vector<vobj> &lvSum,
-				  const int &rd,
-				  const int &e1,
-				  const int &e2,
-				  const int &stride,
-				  const int &ostride,
-				  const int &Nsimd)
+template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
 {
   // sum over reduced dimension planes, breaking out orthog dir
   // Parallel over orthog direction
@@ -247,20 +208,16 @@ inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
   });
 }
 
-template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data,
-						   std::vector<vobj> &lvSum,
-						   const int &rd,
-						   const int &e1,
-						   const int &e2,
-						   const int &stride,
-						   const int &ostride,
-						   const int &Nsimd) 
+template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) 
 {
-#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  
   sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-#else
+  
+  #else
   sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-#endif
+
+  #endif
 }
 
 

Grid/lattice/PaddedCell.h

@@ -54,7 +54,7 @@ struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::ve
  *
  */
 
-template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
+template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
 					      Lattice<vobj> &lat,
 					      int x,
 					      int dim,
@@ -140,7 +140,7 @@ template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
   });
 }
 
-template<class vobj> inline void GatherSlice(deviceVector<vobj> &buf,
+template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
 					     const Lattice<vobj> &lat,
 					     int x,
 					     int dim,
@@ -462,19 +462,13 @@ public:
     int rNsimd = Nsimd / simd[dimension];
     assert( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
 
-    static deviceVector<vobj> send_buf; 
-    static deviceVector<vobj> recv_buf;
+    static cshiftVector<vobj> send_buf; 
+    static cshiftVector<vobj> recv_buf;
     send_buf.resize(buffer_size*2*depth);    
     recv_buf.resize(buffer_size*2*depth);
-#ifndef ACCELERATOR_AWARE_MPI
-    static hostVector<vobj> hsend_buf; 
-    static hostVector<vobj> hrecv_buf;
-    hsend_buf.resize(buffer_size*2*depth);    
-    hrecv_buf.resize(buffer_size*2*depth);
-#endif    
 
-    std::vector<MpiCommsRequest_t> fwd_req;   
-    std::vector<MpiCommsRequest_t> bwd_req;   
+    std::vector<CommsRequest_t> fwd_req;   
+    std::vector<CommsRequest_t> bwd_req;   
 
     int words = buffer_size;
     int bytes = words * sizeof(vobj);
@@ -501,16 +495,9 @@ 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 ++ ) {
@@ -521,16 +508,9 @@ 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;
     }
 
@@ -553,13 +533,8 @@ 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++;
@@ -568,11 +543,6 @@ 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
   /////////////////////////////////////////////////////////////
@@ -132,10 +132,6 @@ 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

@@ -55,11 +55,6 @@ 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
@@ -97,11 +92,6 @@ 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

@@ -90,16 +90,16 @@ public:
   void M5D(const FermionField &psi,
 	   const FermionField &phi,
 	   FermionField &chi,
-	   std::vector<Coeff_t> &lower,
-	   std::vector<Coeff_t> &diag,
-	   std::vector<Coeff_t> &upper);
+	   Vector<Coeff_t> &lower,
+	   Vector<Coeff_t> &diag,
+	   Vector<Coeff_t> &upper);
 
   void M5Ddag(const FermionField &psi,
 	      const FermionField &phi,
 	      FermionField &chi,
-	      std::vector<Coeff_t> &lower,
-	      std::vector<Coeff_t> &diag,
-	      std::vector<Coeff_t> &upper);
+	      Vector<Coeff_t> &lower,
+	      Vector<Coeff_t> &diag,
+	      Vector<Coeff_t> &upper);
 
   virtual void   Instantiatable(void)=0;
 
@@ -119,51 +119,35 @@ public:
   RealD mass_plus, mass_minus;
 
   // Save arguments to SetCoefficientsInternal
-  std::vector<Coeff_t> _gamma;
+  Vector<Coeff_t> _gamma;
   RealD                _zolo_hi;
   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
-  std::vector<Coeff_t> cs;
-  std::vector<Coeff_t> as;
+  Vector<Coeff_t> omega;
+  Vector<Coeff_t> bs;    // S dependent coeffs
+  Vector<Coeff_t> cs;
+  Vector<Coeff_t> as;
   // For preconditioning Cayley form
-  std::vector<Coeff_t> bee;
-  std::vector<Coeff_t> cee;
-  std::vector<Coeff_t> aee;
-  std::vector<Coeff_t> beo;
-  std::vector<Coeff_t> ceo;
-  std::vector<Coeff_t> aeo;
+  Vector<Coeff_t> bee;
+  Vector<Coeff_t> cee;
+  Vector<Coeff_t> aee;
+  Vector<Coeff_t> beo;
+  Vector<Coeff_t> ceo;
+  Vector<Coeff_t> aeo;
   // LDU factorisation of the eeoo matrix
-  std::vector<Coeff_t> lee;
-  std::vector<Coeff_t> leem;
-  std::vector<Coeff_t> uee;
-  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;
+  Vector<Coeff_t> lee;
+  Vector<Coeff_t> leem;
+  Vector<Coeff_t> uee;
+  Vector<Coeff_t> ueem;
+  Vector<Coeff_t> dee;
 
   // Matrices of 5d ee inverse params
-  //  std::vector<iSinglet<Simd> >  MatpInv;
-  //  std::vector<iSinglet<Simd> >  MatmInv;
-  //  std::vector<iSinglet<Simd> >  MatpInvDag;
-  //  std::vector<iSinglet<Simd> >  MatmInvDag;
+  Vector<iSinglet<Simd> >  MatpInv;
+  Vector<iSinglet<Simd> >  MatmInv;
+  Vector<iSinglet<Simd> >  MatpInvDag;
+  Vector<iSinglet<Simd> >  MatmInvDag;
 
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
@@ -203,7 +187,7 @@ public:
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);
   virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c);
-  virtual void SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c);
+  virtual void SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t> & gamma,RealD b,RealD c);
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion5D.h

@@ -1,196 +0,0 @@
-/*************************************************************************************
-
-    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

@@ -60,50 +60,6 @@ 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);
@@ -134,12 +90,12 @@ protected:
   RealD mass;
   RealD R;
   RealD ZoloHiInv;
-  std::vector<double> Beta;
-  std::vector<double> cc;;
-  std::vector<double> cc_d;;
-  std::vector<double> sqrt_cc;
-  std::vector<double> See;
-  std::vector<double> Aee;
+  Vector<double> Beta;
+  Vector<double> cc;;
+  Vector<double> cc_d;;
+  Vector<double> sqrt_cc;
+  Vector<double> See;
+  Vector<double> Aee;
 
 };
 

Grid/qcd/action/fermion/DomainWallEOFAFermion.h

@@ -69,10 +69,10 @@ public:
   // Instantiate different versions depending on Impl
   /////////////////////////////////////////////////////
   void M5D(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	   std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
+	   Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
 
   void M5Ddag(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	      std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
+	      Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
 
   virtual void RefreshShiftCoefficients(RealD new_shift);
 
@@ -83,7 +83,7 @@ public:
 			RealD _M5, const ImplParams& p=ImplParams());
 
 protected:
-  void SetCoefficientsInternal(RealD zolo_hi, std::vector<Coeff_t>& gamma, RealD b, RealD c);
+  void SetCoefficientsInternal(RealD zolo_hi, Vector<Coeff_t>& gamma, RealD b, RealD c);
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -55,7 +55,6 @@ 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);
@@ -165,17 +164,12 @@ 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/ImprovedStaggeredFermion.h

@@ -102,11 +102,11 @@ public:
 		     GaugeField &mat, 
 		     const FermionField &A, const FermionField &B, int dag);
 
-  void DhopInternal(StencilImpl &st, DoubledGaugeField &U,DoubledGaugeField &UUU,
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
-  void DhopInternalSerialComms(StencilImpl &st, DoubledGaugeField &U,DoubledGaugeField &UUU,
+  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
-  void DhopInternalOverlappedComms(StencilImpl &st, DoubledGaugeField &U,DoubledGaugeField &UUU,
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
 
   //////////////////////////////////////////////////////////////////////////
@@ -154,12 +154,6 @@ public:
   StencilImpl Stencil;
   StencilImpl StencilEven;
   StencilImpl StencilOdd;
-  void SloppyComms(int sloppy)
-  {
-    Stencil.SetSloppyComms(sloppy);
-    StencilEven.SetSloppyComms(sloppy);
-    StencilOdd.SetSloppyComms(sloppy);
-  }
 
   // Copy of the gauge field , with even and odd subsets
   DoubledGaugeField Umu;
@@ -170,6 +164,8 @@ public:
   DoubledGaugeField UUUmuEven;
   DoubledGaugeField UUUmuOdd;
 
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
   
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -100,6 +100,7 @@ public:
 		     int dag);
     
   void DhopInternal(StencilImpl & st,
+		    LebesgueOrder &lo,
 		    DoubledGaugeField &U,
 		    DoubledGaugeField &UUU,
 		    const FermionField &in, 
@@ -107,6 +108,7 @@ public:
 		    int dag);
     
     void DhopInternalOverlappedComms(StencilImpl & st,
+		      LebesgueOrder &lo,
 		      DoubledGaugeField &U,
 		      DoubledGaugeField &UUU,
 		      const FermionField &in, 
@@ -114,6 +116,7 @@ public:
 		      int dag);
 
     void DhopInternalSerialComms(StencilImpl & st,
+		      LebesgueOrder &lo,
 		      DoubledGaugeField &U,
 		      DoubledGaugeField &UUU,
 		      const FermionField &in, 
@@ -179,12 +182,6 @@ public:
   StencilImpl Stencil; 
   StencilImpl StencilEven; 
   StencilImpl StencilOdd; 
-  void SloppyComms(int sloppy)
-  {
-    Stencil.SetSloppyComms(sloppy);
-    StencilEven.SetSloppyComms(sloppy);
-    StencilOdd.SetSloppyComms(sloppy);
-  }
     
   // Copy of the gauge field , with even and odd subsets
   DoubledGaugeField Umu;
@@ -195,6 +192,8 @@ public:
   DoubledGaugeField UUUmuEven;
   DoubledGaugeField UUUmuOdd;
     
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
   //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;

Grid/qcd/action/fermion/MobiusEOFAFermion.h

@@ -42,11 +42,11 @@ public:
 
 public:
   // Shift operator coefficients for red-black preconditioned Mobius EOFA
-  std::vector<Coeff_t> Mooee_shift;
-  std::vector<Coeff_t> MooeeInv_shift_lc;
-  std::vector<Coeff_t> MooeeInv_shift_norm;
-  std::vector<Coeff_t> MooeeInvDag_shift_lc;
-  std::vector<Coeff_t> MooeeInvDag_shift_norm;
+  Vector<Coeff_t> Mooee_shift;
+  Vector<Coeff_t> MooeeInv_shift_lc;
+  Vector<Coeff_t> MooeeInv_shift_norm;
+  Vector<Coeff_t> MooeeInvDag_shift_lc;
+  Vector<Coeff_t> MooeeInvDag_shift_norm;
 
   virtual void Instantiatable(void) {};
 
@@ -74,18 +74,18 @@ public:
   // Instantiate different versions depending on Impl
   /////////////////////////////////////////////////////
   void M5D(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	   std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
+	   Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
 
   void M5D_shift(const FermionField& psi, const FermionField& phi, FermionField& chi,
-		 std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper,
-		 std::vector<Coeff_t>& shift_coeffs);
+		 Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
+		 Vector<Coeff_t>& shift_coeffs);
 
   void M5Ddag(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	      std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
+	      Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
 
   void M5Ddag_shift(const FermionField& psi, const FermionField& phi, FermionField& chi,
-		    std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper,
-		    std::vector<Coeff_t>& shift_coeffs);
+		    Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
+		    Vector<Coeff_t>& shift_coeffs);
 
   virtual void RefreshShiftCoefficients(RealD new_shift);
 

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -102,11 +102,11 @@ public:
 		     GaugeField &mat, 
 		     const FermionField &A, const FermionField &B, int dag);
 
-  void DhopInternal(StencilImpl &st, DoubledGaugeField &U,
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
                     const FermionField &in, FermionField &out, int dag);
-  void DhopInternalSerialComms(StencilImpl &st, DoubledGaugeField &U,
+  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
 			       const FermionField &in, FermionField &out, int dag);
-  void DhopInternalOverlappedComms(StencilImpl &st, DoubledGaugeField &U,
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
 				   const FermionField &in, FermionField &out, int dag);
 
   //////////////////////////////////////////////////////////////////////////
@@ -146,18 +146,15 @@ public:
   StencilImpl Stencil;
   StencilImpl StencilEven;
   StencilImpl StencilOdd;
-  void SloppyComms(int sloppy)
-  {
-    Stencil.SetSloppyComms(sloppy);
-    StencilEven.SetSloppyComms(sloppy);
-    StencilOdd.SetSloppyComms(sloppy);
-  }
 
   // Copy of the gauge field , with even and odd subsets
   DoubledGaugeField Umu;
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
 
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+  
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
   ///////////////////////////////////////////////////////////////

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

@@ -41,10 +41,6 @@ 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

@@ -41,9 +41,6 @@ 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

@@ -40,9 +40,6 @@ 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

@@ -41,9 +41,6 @@ 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

@@ -40,11 +40,6 @@ 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=0;
+  const int part_frac_chroma_convention=1;
 
   void   Meooe_internal(const FermionField &in, FermionField &out,int dag);
   void   Mooee_internal(const FermionField &in, FermionField &out,int dag);
@@ -83,78 +83,19 @@ 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;
   RealD R;
   RealD amax;
   RealD scale;
-  std::vector<double> p; 
-  std::vector<double> q;
+  Vector<double> p; 
+  Vector<double> q;
 
 };
 

Grid/qcd/action/fermion/SchurDiagTwoKappa.h

@@ -35,7 +35,7 @@ template<class Matrix, class Field>
 class KappaSimilarityTransform {
 public:
   INHERIT_IMPL_TYPES(Matrix);
-  std::vector<Coeff_t> kappa, kappaDag, kappaInv, kappaInvDag;
+  Vector<Coeff_t> kappa, kappaDag, kappaInv, kappaInvDag;
 
   KappaSimilarityTransform (Matrix &zmob) {
     for (int i=0;i<(int)zmob.bs.size();i++) {

Grid/qcd/action/fermion/StaggeredKernels.h

@@ -49,10 +49,10 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    
  public:
 
-  void DhopImproved(StencilImpl &st,
+  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
 		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
 		    const FermionField &in, FermionField &out, int dag, int interior,int exterior);
-  void DhopNaive(StencilImpl &st,
+  void DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
 		 DoubledGaugeField &U,
 		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
   

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -32,6 +32,209 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+///////////////////////////////////////////////////////////////
+// Wilson compressor will need FaceGather policies for:
+// Periodic, Dirichlet, and partial Dirichlet for DWF
+///////////////////////////////////////////////////////////////
+const int dwf_compressor_depth=2;
+#define DWF_COMPRESS
+class FaceGatherPartialDWF
+{
+public:
+#ifdef DWF_COMPRESS
+  static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
+#else
+  static int PartialCompressionFactor(GridBase *grid) { return 1;}
+#endif
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
+				   const Lattice<vobj> &rhs,
+				   cobj *buffer,
+				   compressor &compress,
+				   int off,int so,int partial)
+  {
+    //DWF only hack: If a direction that is OFF node we use Partial Dirichlet
+    //  Shrinks local and remote comms buffers
+    GridBase *Grid = rhs.Grid();
+    int Ls = Grid->_rdimensions[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else 
+    int depth=Ls/2;
+#endif
+    std::pair<int,int> *table_v = & table[0];
+    auto rhs_v = rhs.View(AcceleratorRead);
+    int vol=table.size()/Ls;
+    accelerator_forNB( idx,table.size(), vobj::Nsimd(), {
+	Integer i=idx/Ls;
+	Integer s=idx%Ls;
+	Integer sc=depth+s-(Ls-depth);
+	if(s<depth)     compress.Compress(buffer[off+i+s*vol],rhs_v[so+table_v[idx].second]);
+	if(s>=Ls-depth) compress.Compress(buffer[off+i+sc*vol],rhs_v[so+table_v[idx].second]);
+    });
+    rhs_v.ViewClose();
+  }
+  template<class decompressor,class Decompression>
+  static void DecompressFace(decompressor decompress,Decompression &dd)
+  {
+    auto Ls = dd.dims[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else
+    int depth=Ls/2;
+#endif    
+    // Just pass in the Grid
+    auto kp = dd.kernel_p;
+    auto mp = dd.mpi_p;
+    int size= dd.buffer_size;
+    int vol= size/Ls;
+    accelerator_forNB(o,size,1,{
+	int idx=o/Ls;
+	int   s=o%Ls;
+	if ( s < depth ) {
+	  int oo=s*vol+idx;
+	  kp[o]=mp[oo];
+	} else if ( s >= Ls-depth ) {
+	  int sc = depth + s - (Ls-depth);
+	  int oo=sc*vol+idx; 
+	  kp[o]=mp[oo];
+	} else {
+	  kp[o] = Zero();//fill rest with zero if partial dirichlet
+	}
+    });
+  }
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Need to gather *interior portions* for ALL s-slices in simd directions
+  // Do the gather as need to treat SIMD lanes differently, and insert zeroes on receive side
+  // Reorder the fifth dim to be s=Ls-1 , s=0, s=1,...,Ls-2.
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    compressor &compress,int type,int partial)
+  {
+    GridBase *Grid = rhs.Grid();
+    int Ls = Grid->_rdimensions[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else
+    int depth = Ls/2;
+#endif
+    
+    // insertion of zeroes...
+    assert( (table.size()&0x1)==0);
+    int num=table.size()/2;
+    int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
+    
+    auto rhs_v = rhs.View(AcceleratorRead);
+    auto p0=&pointers[0][0];
+    auto p1=&pointers[1][0];
+    auto tp=&table[0];
+    int nnum=num/Ls;
+    accelerator_forNB(j, num, vobj::Nsimd(), {
+	//  Reorders both local and remote comms buffers
+	//  
+	int s  = j % Ls;
+	int sp1 = (s+depth)%Ls;  // peri incremented s slice
+	
+	int hxyz= j/Ls;
+
+	int xyz0= hxyz*2; // xyzt part of coor
+	int xyz1= hxyz*2+1;
+	
+	int jj= hxyz + sp1*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
+	
+	int kk0= xyz0*Ls + s ; // s=0 goes to s=1
+	int kk1= xyz1*Ls + s ; // s=Ls-1 -> s=0
+	compress.CompressExchange(p0[jj],p1[jj],
+				  rhs_v[so+tp[kk0 ].second], // Same s, consecutive xyz sites
+				  rhs_v[so+tp[kk1 ].second], 
+				  type);
+    });
+    rhs_v.ViewClose();
+  }
+  // Merge routine is for SIMD faces
+  template<class decompressor,class Merger>
+  static void MergeFace(decompressor decompress,Merger &mm)
+  {
+    auto Ls = mm.dims[0];
+#ifdef DWF_COMPRESS
+    int depth=dwf_compressor_depth;
+#else
+    int depth = Ls/2;
+#endif
+    int  num= mm.buffer_size/2; // relate vol and Ls to buffer size
+    auto mp = &mm.mpointer[0];
+    auto vp0= &mm.vpointers[0][0]; // First arg is exchange first
+    auto vp1= &mm.vpointers[1][0];
+    auto type= mm.type;
+    int nnum = num/Ls;
+    accelerator_forNB(o,num,Merger::Nsimd,{
+
+	int  s=o%Ls;
+	int hxyz=o/Ls; // xyzt related component
+	int xyz0=hxyz*2;
+	int xyz1=hxyz*2+1;
+
+	int sp = (s+depth)%Ls; 
+	int jj= hxyz + sp*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
+
+	int oo0= s+xyz0*Ls;
+	int oo1= s+xyz1*Ls;
+
+	// same ss0, ss1 pair goes to new layout
+	decompress.Exchange(mp[oo0],mp[oo1],vp0[jj],vp1[jj],type);
+      });
+  }
+};
+class FaceGatherDWFMixedBCs
+{
+public:
+#ifdef DWF_COMPRESS
+  static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
+#else 
+  static int PartialCompressionFactor(GridBase *grid) {return 1;}
+#endif
+  
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
+					 const Lattice<vobj> &rhs,
+					 cobj *buffer,
+					 compressor &compress,
+					 int off,int so,int partial)
+  {
+    //    std::cout << " face gather simple DWF partial "<<partial <<std::endl;
+    if(partial) FaceGatherPartialDWF::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
+    else        FaceGatherSimple::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
+  }
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    compressor &compress,int type,int partial)
+  {
+    //    std::cout << " face gather exch DWF partial "<<partial <<std::endl;
+    if(partial) FaceGatherPartialDWF::Gather_plane_exchange(table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
+    else        FaceGatherSimple::Gather_plane_exchange    (table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
+  }
+  template<class decompressor,class Merger>
+  static void MergeFace(decompressor decompress,Merger &mm)
+  {
+    int partial = mm.partial;
+    //    std::cout << " merge DWF partial "<<partial <<std::endl;
+    if ( partial ) FaceGatherPartialDWF::MergeFace(decompress,mm);
+    else           FaceGatherSimple::MergeFace(decompress,mm);
+  }
+
+  template<class decompressor,class Decompression>
+  static void DecompressFace(decompressor decompress,Decompression &dd)
+  {
+    int partial = dd.partial;
+    //    std::cout << " decompress DWF partial "<<partial <<std::endl;
+    if ( partial ) FaceGatherPartialDWF::DecompressFace(decompress,dd);
+    else           FaceGatherSimple::DecompressFace(decompress,dd);
+  }
+};
+
 /////////////////////////////////////////////////////////////////////////////////////////////
 // optimised versions supporting half precision too??? Deprecate
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -39,7 +242,8 @@ NAMESPACE_BEGIN(Grid);
 
 //Could make FaceGather a template param, but then behaviour is runtime not compile time
 template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
-class WilsonCompressorTemplate : public FaceGatherSimple
+class WilsonCompressorTemplate  : public FaceGatherDWFMixedBCs
+//  : public FaceGatherSimple
 {
 public:
   
@@ -198,6 +402,7 @@ public:
 
   typedef CartesianStencil<vobj,cobj,Parameters> Base;
   typedef typename Base::View_type View_type;
+  typedef typename Base::StencilVector StencilVector;
 
   //  Vector<int> surface_list;
   WilsonStencil(GridBase *grid,
@@ -210,6 +415,29 @@ 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) 
@@ -280,11 +508,6 @@ 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/WilsonFermion.h

@@ -126,17 +126,14 @@ public:
   void DerivInternal(StencilImpl &st, DoubledGaugeField &U, GaugeField &mat,
                      const FermionField &A, const FermionField &B, int dag);
 
-  void DhopInternal(StencilImpl &st,
-		    DoubledGaugeField &U,
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
                     const FermionField &in, FermionField &out, int dag);
 
-  void DhopInternalSerial(StencilImpl &st,
-			  DoubledGaugeField &U,
-			  const FermionField &in, FermionField &out, int dag);
+  void DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
 
-  void DhopInternalOverlappedComms(StencilImpl &st,
-				   DoubledGaugeField &U,
-				   const FermionField &in, FermionField &out, int dag);
+  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+                    const FermionField &in, FermionField &out, int dag);
 
   // Constructor
   WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
@@ -165,18 +162,15 @@ public:
   StencilImpl Stencil;
   StencilImpl StencilEven;
   StencilImpl StencilOdd;
-  void SloppyComms(int sloppy)
-  {
-    Stencil.SetSloppyComms(sloppy);
-    StencilEven.SetSloppyComms(sloppy);
-    StencilOdd.SetSloppyComms(sloppy);
-  }
 
   // Copy of the gauge field , with even and odd subsets
   DoubledGaugeField Umu;
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
 
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+
   WilsonAnisotropyCoefficients anisotropyCoeff;
 
   ///////////////////////////////////////////////////////////////

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);
-  virtual void   Mooee       (const FermionField &in, FermionField &out);
-  virtual void   MooeeInv    (const FermionField &in, FermionField &out);
+  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   MeooeDag    (const FermionField &in, FermionField &out);
-  virtual void   MooeeDag    (const FermionField &in, FermionField &out);
-  virtual void   MooeeInvDag (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   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
 
@@ -109,8 +109,6 @@ 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
@@ -119,9 +117,6 @@ 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);
@@ -140,18 +135,21 @@ public:
 		     int dag);
     
   void DhopInternal(StencilImpl & st,
+		    LebesgueOrder &lo,
 		    DoubledGaugeField &U,
 		    const FermionField &in, 
 		    FermionField &out,
 		    int dag);
 
   void DhopInternalOverlappedComms(StencilImpl & st,
+				   LebesgueOrder &lo,
 				   DoubledGaugeField &U,
 				   const FermionField &in, 
 				   FermionField &out,
 				   int dag);
 
   void DhopInternalSerialComms(StencilImpl & st,
+			       LebesgueOrder &lo,
 			       DoubledGaugeField &U,
 			       const FermionField &in, 
 			       FermionField &out,
@@ -204,14 +202,10 @@ public:
   DoubledGaugeField Umu;
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
-
-
-  void SloppyComms(int sloppy)
-  {
-    Stencil.SetSloppyComms(sloppy);
-    StencilEven.SetSloppyComms(sloppy);
-    StencilOdd.SetSloppyComms(sloppy);
-  }
+    
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+    
   // Comms buffer
   //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
 

Grid/qcd/action/fermion/WilsonKernels.h

@@ -57,10 +57,6 @@ 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/ZMobiusFermion.h

@@ -58,7 +58,7 @@ public:
   {
     //    RealD eps = 1.0;
     std::cout<<GridLogMessage << "ZMobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" gamma passed in"<<std::endl;
-    std::vector<Coeff_t> zgamma(this->Ls);
+    Vector<Coeff_t> zgamma(this->Ls);
     for(int s=0;s<this->Ls;s++){
       zgamma[s] = gamma[s];
     }

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

@@ -48,8 +48,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
   mass_plus(_mass), mass_minus(_mass)
-{
-  // qmu defaults to zero size;
+{ 
 }
 
 ///////////////////////////////////////////////////////////////
@@ -157,18 +156,18 @@ template<class Impl>
 void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag (Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
-  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
+  Vector<Coeff_t> diag (Ls,1.0);
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
 void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &Din)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = bs;
-  std::vector<Coeff_t> upper= cs;
-  std::vector<Coeff_t> lower= cs; 
+  Vector<Coeff_t> diag = bs;
+  Vector<Coeff_t> upper= cs;
+  Vector<Coeff_t> lower= cs; 
   upper[Ls-1]=-mass_minus*upper[Ls-1];
   lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
@@ -177,9 +176,9 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
 template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = beo;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = beo;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int i=0;i<Ls;i++) {
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
@@ -192,9 +191,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = bee;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = bee;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int i=0;i<Ls;i++) {
     upper[i]=-cee[i];
     lower[i]=-cee[i];
@@ -207,9 +206,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = bee;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = bee;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
 
   for (int s=0;s<Ls;s++){
     // Assemble the 5d matrix
@@ -237,9 +236,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0);
-  std::vector<Coeff_t> lower(Ls,-1.0);
+  Vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> upper(Ls,-1.0);
+  Vector<Coeff_t> lower(Ls,-1.0);
   upper[Ls-1]=-mass_plus*upper[Ls-1];
   lower[0]   =-mass_minus*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
@@ -249,9 +248,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField &Din)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag =bs;
-  std::vector<Coeff_t> upper=cs;
-  std::vector<Coeff_t> lower=cs; 
+  Vector<Coeff_t> diag =bs;
+  Vector<Coeff_t> upper=cs;
+  Vector<Coeff_t> lower=cs; 
 
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
@@ -271,34 +270,6 @@ 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)
 {
@@ -306,12 +277,8 @@ 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);
   
+  this->DW(Din,chi,DaggerNo);
   // ((b D_W + D_w hop terms +1) on s-diag
   axpby(chi,1.0,1.0,chi,psi); 
   
@@ -328,9 +295,6 @@ 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);
   
@@ -430,7 +394,7 @@ void CayleyFermion5D<Impl>::MeoDeriv(GaugeField &mat,const FermionField &U,const
 template<class Impl>
 void CayleyFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c)
 {
-  std::vector<Coeff_t> gamma(this->Ls);
+  Vector<Coeff_t> gamma(this->Ls);
   for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
   SetCoefficientsInternal(1.0,gamma,b,c);
 }
@@ -438,13 +402,13 @@ void CayleyFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,Re
 template<class Impl>
 void CayleyFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata,RealD b,RealD c)
 {
-  std::vector<Coeff_t> gamma(this->Ls);
+  Vector<Coeff_t> gamma(this->Ls);
   for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
   SetCoefficientsInternal(zolo_hi,gamma,b,c);
 }
 //Zolo
 template<class Impl>
-void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c)
+void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t> & gamma,RealD b,RealD c)
 {
   int Ls=this->Ls;
 
@@ -524,7 +488,7 @@ 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];
@@ -565,18 +529,6 @@ 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

@@ -43,9 +43,9 @@ void
 CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
 			       const FermionField &phi_i, 
 			       FermionField &chi_i,
-			       std::vector<Coeff_t> &lower,
-			       std::vector<Coeff_t> &diag,
-			       std::vector<Coeff_t> &upper)
+			       Vector<Coeff_t> &lower,
+			       Vector<Coeff_t> &diag,
+			       Vector<Coeff_t> &upper)
 {
   
   chi_i.Checkerboard()=psi_i.Checkerboard();
@@ -55,15 +55,11 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  int Ls =this->Ls;
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
 
-  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];
-  auto plower = &d_lower[0];
+  int Ls =this->Ls;
 
   // 10 = 3 complex mult + 2 complex add
   // Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
@@ -86,9 +82,9 @@ void
 CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
 			      const FermionField &phi_i, 
 			      FermionField &chi_i,
-			      std::vector<Coeff_t> &lower,
-			      std::vector<Coeff_t> &diag,
-			      std::vector<Coeff_t> &upper)
+			      Vector<Coeff_t> &lower,
+			      Vector<Coeff_t> &diag,
+			      Vector<Coeff_t> &upper)
 {
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
@@ -97,15 +93,11 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  int Ls=this->Ls;
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
 
-  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];
-  auto plower = &d_lower[0];
+  int Ls=this->Ls;
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   uint64_t nloop = grid->oSites();
@@ -134,17 +126,11 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
 
   int Ls=this->Ls;
 
-  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];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  auto plee  = & lee [0];
+  auto pdee  = & dee [0];
+  auto puee  = & uee [0];
+  auto pleem = & leem[0];
+  auto pueem = & ueem[0];
 
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -196,17 +182,11 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
   autoView(psi , psi_i,AcceleratorRead);
   autoView(chi , chi_i,AcceleratorWrite);
 
-  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];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
+  auto plee  = & lee [0];
+  auto pdee  = & dee [0];
+  auto puee  = & uee [0];
+  auto pleem = & leem[0];
+  auto pueem = & ueem[0];
 
   assert(psi.Checkerboard() == psi.Checkerboard());
 

Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h

@@ -1,5 +1,3 @@
-#if 0
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -820,5 +818,3 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
 }
 
 NAMESPACE_END(Grid);
-
-#endif

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

@@ -1,376 +0,0 @@
-/*************************************************************************************
-
-    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 << 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 << 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;
   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, 0);
+      ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
     }
     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, 0);
+      InsertSlice(input4d, tmp, Ls-1, Ls-1);
       tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
       this->Dminus(tmp,imported5d);
     }

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

@@ -41,7 +41,7 @@ NAMESPACE_BEGIN(Grid);
 // Pplus  backwards..
 template<class Impl>
 void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i, 
-				      std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper)
+				      Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   int Ls = this->Ls;
@@ -50,15 +50,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   autoView( psi , psi_i, AcceleratorRead);
   autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
-
-  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));
-
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   
   auto nloop=grid->oSites()/Ls;
@@ -79,7 +73,7 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
 
 template<class Impl>
 void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i, 
-					 std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper)
+					 Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase* grid = psi_i.Grid();
@@ -89,14 +83,9 @@ 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());
-  
-  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));
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
 
@@ -125,17 +114,12 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
   autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
-  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));
+  auto plee  = & this->lee[0];
+  auto pdee  = & this->dee[0];
+  auto puee  = & this->uee[0];
+
+  auto pleem = & this->leem[0];
+  auto pueem = & this->ueem[0];
 
   uint64_t nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{

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

@@ -131,9 +131,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi
     else{ shiftm = -shift*(mq3-mq2); }
   }
 
-  std::vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
-  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftp;
+  Vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftp;
 
 #if(0)
   std::cout << GridLogMessage << "DomainWallEOFAFermion::M5D(FF&,FF&):" << std::endl;
@@ -168,9 +168,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField&
     else{ shiftm = -shift*(mq3-mq2); }
   }
 
-  std::vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
-  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftm;
+  Vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftm;
 
   this->M5Ddag(psi, chi, chi, lower, diag, upper);
 }
@@ -181,9 +181,9 @@ void DomainWallEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& c
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
 
   for(int s=0; s<Ls; s++){
     upper[s] = -this->cee[s];
@@ -200,9 +200,9 @@ void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
 
   for(int s=0; s<Ls; s++){
     upper[s] = -this->cee[s];
@@ -218,7 +218,7 @@ void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField
 
 //Zolo
 template<class Impl>
-void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, std::vector<Coeff_t>& gamma, RealD b, RealD c)
+void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, Vector<Coeff_t>& gamma, RealD b, RealD c)
 {
   int   Ls    = this->Ls;
   int   pm    = this->pm;

Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h

@@ -61,6 +61,8 @@ ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GridCartesian
   UUUmu(&FourDimGrid),
   UUUmuEven(&FourDimRedBlackGrid),
   UUUmuOdd(&FourDimRedBlackGrid),
+  Lebesgue(&FourDimGrid),
+  LebesgueEvenOdd(&FourDimRedBlackGrid),
   _tmp(&FiveDimRedBlackGrid)
 {
 
@@ -275,18 +277,18 @@ void ImprovedStaggeredFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
 
 /*CHANGE */
 template<class Impl>
-void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, 
+void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
 						    const FermionField &in, FermionField &out,int dag)
 {
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,UUU,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-    DhopInternalSerialComms(st,U,UUU,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 
 template<class Impl>
-void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, 
+void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
 								   DoubledGaugeField & U,DoubledGaugeField & UUU,
 								   const FermionField &in, FermionField &out,int dag)
 {
@@ -311,7 +313,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   {
     int interior=1;
     int exterior=0;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 
   st.CommsMerge(compressor);
@@ -321,12 +323,12 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   {
     int interior=0;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 }
 
 template<class Impl>
-void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, 
+void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
 						    const FermionField &in, FermionField &out,int dag)
 {
@@ -339,7 +341,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   {
     int interior=1;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 }
 /*CHANGE END*/
@@ -355,7 +357,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
   assert(in.Checkerboard()==Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven,UmuOdd,UUUmuOdd,in,out,dag);
+  DhopInternal(StencilEven,LebesgueEvenOdd,UmuOdd,UUUmuOdd,in,out,dag);
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
@@ -366,7 +368,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
   assert(in.Checkerboard()==Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd,UmuEven,UUUmuEven,in,out,dag);
+  DhopInternal(StencilOdd,LebesgueEvenOdd,UmuEven,UUUmuEven,in,out,dag);
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
@@ -376,7 +378,7 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil,Umu,UUUmu,in,out,dag);
+  DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
 }
 
 /////////////////////////////////////////////////////////////////////////

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

@@ -48,6 +48,8 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GridCartesian &Fgrid, G
     StencilEven(&Hgrid, npoint, Even, directions, displacements,p),  // source is Even
     StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p),  // source is Odd
     mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
     Umu(&Fgrid),
     UmuEven(&Hgrid),
     UmuOdd(&Hgrid),
@@ -337,7 +339,7 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil, Umu, UUUmu, in, out, dag);
+  DhopInternal(Stencil, Lebesgue, Umu, UUUmu, in, out, dag);
 }
 
 template <class Impl>
@@ -349,7 +351,7 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
   assert(in.Checkerboard() == Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven, UmuOdd, UUUmuOdd, in, out, dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, UUUmuOdd, in, out, dag);
 }
 
 template <class Impl>
@@ -361,7 +363,7 @@ void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField
   assert(in.Checkerboard() == Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd, UmuEven, UUUmuEven, in, out, dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, UUUmuEven, in, out, dag);
 }
 
 template <class Impl>
@@ -392,19 +394,19 @@ void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionFiel
 
 
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, 
+void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
 						  DoubledGaugeField &U,
 						  DoubledGaugeField &UUU,
 						  const FermionField &in,
 						  FermionField &out, int dag) 
 {
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,UUU,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-    DhopInternalSerialComms(st,U,UUU,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, 
+void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 								 DoubledGaugeField &U,
 								 DoubledGaugeField &UUU,
 								 const FermionField &in,
@@ -427,7 +429,7 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   {
     int interior=1;
     int exterior=0;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 
   st.CommunicateComplete(requests);
@@ -438,13 +440,13 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   {
     int interior=0;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 }
 
 
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, 
+void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
 							     DoubledGaugeField &U,
 							     DoubledGaugeField &UUU,
 							     const FermionField &in,
@@ -458,7 +460,7 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st,
   {
     int interior=1;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 };
 

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

@@ -39,7 +39,7 @@ NAMESPACE_BEGIN(Grid);
  
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-				  std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper)
+				  Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -50,14 +50,10 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  auto pdiag  = &this->d_diag[0];
-  auto pupper = &this->d_upper[0];
-  auto plower = &this->d_lower[0];
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &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(),{
@@ -78,8 +74,8 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
 
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-					std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper,
-					std::vector<Coeff_t> &shift_coeffs)
+					Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper,
+					Vector<Coeff_t> &shift_coeffs)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -90,18 +86,13 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
 
   auto pm  = this->pm;
   int shift_s = (pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
-  
+
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  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));
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
+  auto pshift_coeffs = &shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -128,7 +119,7 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
 
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-				     std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper)
+				     Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -138,14 +129,10 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
   autoView(chi , chi_i, AcceleratorWrite);
 
   assert(phi.Checkerboard() == psi.Checkerboard());
-  
-  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));
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -167,8 +154,8 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
 
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-					   std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper,
-					   std::vector<Coeff_t> &shift_coeffs)
+					   Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper,
+					   Vector<Coeff_t> &shift_coeffs)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -180,16 +167,11 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  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];
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
+  auto pshift_coeffs = &shift_coeffs[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 +212,11 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  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));
+  auto plee = & this->lee [0];
+  auto pdee = & this->dee [0];
+  auto puee = & this->uee [0];
+  auto pleem= & this->leem[0];
+  auto pueem= & this->ueem[0];
 
   if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
 
@@ -292,23 +268,14 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  // Move into object and constructor
   auto pm = this->pm;
-  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];
-
-  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));
+  auto plee = & this->lee [0];
+  auto pdee = & this->dee [0];
+  auto puee = & this->uee [0];
+  auto pleem= & this->leem[0];
+  auto pueem= & this->ueem[0];
+  auto pMooeeInv_shift_lc   = &MooeeInv_shift_lc[0];
+  auto pMooeeInv_shift_norm = &MooeeInv_shift_norm[0];
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -366,17 +333,11 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  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));
+  auto plee = & this->lee [0];
+  auto pdee = & this->dee [0];
+  auto puee = & this->uee [0];
+  auto pleem= & this->leem[0];
+  auto pueem= & this->ueem[0];
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -426,25 +387,13 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
   int Ls = this->Ls;
 
   auto pm = this->pm;
-  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 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];
+  auto plee = & this->lee [0];
+  auto pdee = & this->dee [0];
+  auto puee = & this->uee [0];
+  auto pleem= & this->leem[0];
+  auto pueem= & this->ueem[0];
+  auto pMooeeInvDag_shift_lc   = &MooeeInvDag_shift_lc[0];
+  auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{

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

@@ -196,9 +196,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
-  std::vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
+  Vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
+  Vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
 
   // no shift term
   if(this->shift == 0.0){ this->M5D(psi, chi, chi, lower, diag, upper); }
@@ -212,9 +212,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
-  std::vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
+  Vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
+  Vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
 
   // no shift term
   if(this->shift == 0.0){ this->M5Ddag(psi, chi, chi, lower, diag, upper); }
@@ -230,9 +230,9 @@ void MobiusEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
   int Ls = this->Ls;
 
   // coefficients of Mooee
-  std::vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int s=0; s<Ls; s++){
     upper[s] = -this->cee[s];
     lower[s] = -this->cee[s];
@@ -253,9 +253,9 @@ void MobiusEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& ch
   int Ls = this->Ls;
 
   // coefficients of MooeeDag
-  std::vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int s=0; s<Ls; s++){
     if(s==0) {
       upper[s] = -this->cee[s+1];
@@ -314,10 +314,10 @@ void MobiusEOFAFermion<Impl>::SetCoefficientsPrecondShiftOps()
   // Tridiagonal solve for MooeeInvDag_shift_lc
   {
     Coeff_t m(0.0);
-    std::vector<Coeff_t> d = Mooee_shift;
-    std::vector<Coeff_t> u(Ls,0.0);
-    std::vector<Coeff_t> y(Ls,0.0);
-    std::vector<Coeff_t> q(Ls,0.0);
+    Vector<Coeff_t> d = Mooee_shift;
+    Vector<Coeff_t> u(Ls,0.0);
+    Vector<Coeff_t> y(Ls,0.0);
+    Vector<Coeff_t> q(Ls,0.0);
     if(pm == 1){ u[0] = 1.0; }
     else{ u[Ls-1] = 1.0; }
 

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

@@ -48,6 +48,8 @@ NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GridCartesian &Fgrid, GridRed
     StencilEven(&Hgrid, npoint, Even, directions, displacements,p),  // source is Even
     StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p),  // source is Odd
     mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
     Umu(&Fgrid),
     UmuEven(&Hgrid),
     UmuOdd(&Hgrid),
@@ -266,7 +268,7 @@ void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil, Umu, in, out, dag);
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
 }
 
 template <class Impl>
@@ -278,7 +280,7 @@ void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &o
   assert(in.Checkerboard() == Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven, UmuOdd, in, out, dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
 }
 
 template <class Impl>
@@ -290,7 +292,7 @@ void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &o
   assert(in.Checkerboard() == Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd, UmuEven, in, out, dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
 }
 
 template <class Impl>
@@ -321,18 +323,18 @@ void NaiveStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &
 
 
 template <class Impl>
-void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st,
+void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
 					       DoubledGaugeField &U,
 					       const FermionField &in,
 					       FermionField &out, int dag) 
 {
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
-    DhopInternalSerialComms(st,U,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,in,out,dag);
 }
 template <class Impl>
-void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
+void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 							      DoubledGaugeField &U,
 							      const FermionField &in,
 							      FermionField &out, int dag) 
@@ -354,7 +356,7 @@ void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
   {
     int interior=1;
     int exterior=0;
-    Kernels::DhopNaive(st,U,in,out,dag,interior,exterior);
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
 
   st.CommunicateComplete(requests);
@@ -365,12 +367,12 @@ void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
   {
     int interior=0;
     int exterior=1;
-    Kernels::DhopNaive(st,U,in,out,dag,interior,exterior);
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
 }
 
 template <class Impl>
-void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st,
+void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
 							  DoubledGaugeField &U,
 							  const FermionField &in,
 							  FermionField &out, int dag) 
@@ -383,7 +385,7 @@ void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st,
   {
     int interior=1;
     int exterior=1;
-    Kernels::DhopNaive(st,U,in,out,dag,interior,exterior);
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
 };
 

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

@@ -237,32 +237,7 @@ 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);
-
-  // 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;
-  }
+  this->DW(psi,D,DaggerNo); 
 
   int nblock=(Ls-1)/2;
   for(int b=0;b<nblock;b++){
@@ -280,55 +255,15 @@ 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] Hw
+    //R g5 psi[Ls] + p[0] H
     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);
     }
-   
   }
 
 }
@@ -476,18 +411,17 @@ 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, 0);
+      ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
     }
     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, 0);
+      InsertSlice(input4d, tmp, Ls-1, Ls-1);
       tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
       this->Dminus(tmp,imported5d);
     }
@@ -508,7 +442,7 @@ PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
 
 {
   int Ls = this->Ls;
-  qmu.resize(0);
+
   assert((Ls&0x1)==1); // Odd Ls required
   int nrational=Ls-1;
 
@@ -526,22 +460,6 @@ 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/StaggeredKernelsHand.h

@@ -375,6 +375,23 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
   }
 }
 
+/*
+#define DHOP_SITE_HAND_INSTANTIATE(IMPL)				\
+  template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionFieldView &in, FermionFieldView &out, int dag); \
+									\
+  template void StaggeredKernels<IMPL>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionFieldView &in, FermionFieldView &out, int dag); \
+									\
+  template void StaggeredKernels<IMPL>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionFieldView &in, FermionFieldView &out, int dag); \
+*/
 #undef LOAD_CHI
 #undef HAND_DECLARATIONS
 

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

@@ -256,7 +256,7 @@ void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldVie
   });
 
 template <class Impl> 
-void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, 
+void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
 					  DoubledGaugeField &U, DoubledGaugeField &UUU, 
 					  const FermionField &in, FermionField &out, int dag, int interior,int exterior)
 {
@@ -294,7 +294,7 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st,
   assert(0 && " Kernel optimisation case not covered ");
 }
 template <class Impl> 
-void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, 
+void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
 				       DoubledGaugeField &U,
 				       const FermionField &in, FermionField &out, int dag, int interior,int exterior)
 {

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

@@ -14,7 +14,6 @@ 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
@@ -59,9 +58,15 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   Umu(_FourDimGrid),
   UmuEven(_FourDimRedBlackGrid),
   UmuOdd (_FourDimRedBlackGrid),
+  Lebesgue(_FourDimGrid),
+  LebesgueEvenOdd(_FourDimRedBlackGrid),
   _tmp(&FiveDimRedBlackGrid),
   Dirichlet(0)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
+  
   // some assertions
   assert(FiveDimGrid._ndimension==5);
   assert(FourDimGrid._ndimension==4);
@@ -300,19 +305,19 @@ void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
 }
 
 template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st,
+void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
                                          DoubledGaugeField & U,
                                          const FermionField &in, FermionField &out,int dag)
 {
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else 
-    DhopInternalSerialComms(st,U,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,in,out,dag);
 }
 
 
 template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
+void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
 							DoubledGaugeField & U,
 							const FermionField &in, FermionField &out,int dag)
 {
@@ -326,22 +331,22 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   // Start comms  // Gather intranode and extra node differentiated??
   /////////////////////////////
   {
-    //    std::cout << " WilsonFermion5D gather " <<std::endl;
     GRID_TRACE("Gather");
     st.HaloExchangeOptGather(in,compressor); // Put the barrier in the routine
   }
   
-  //  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
   /////////////////////////////
-  st.CommunicateBegin(requests);
-  st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms 
-#endif
-
+  {
+    GRID_TRACE("MergeSHM");
+    st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+  }
+      
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
@@ -353,35 +358,22 @@ 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
   /////////////////////////////
   {
-    //    std::cout << " WilsonFermion5D Comms Merge " <<std::endl;
     GRID_TRACE("Merge");
     st.CommsMerge(compressor);
   }
   
 
-  //  std::cout << " WilsonFermion5D Exterior " <<std::endl;
   if (dag == DaggerYes) {
     GRID_TRACE("DhopDagExterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
@@ -389,12 +381,11 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
     GRID_TRACE("DhopExterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   }
-  //  std::cout << " WilsonFermion5D Done " <<std::endl;
 }
 
 
 template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, 
+void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,
 						    const FermionField &in, 
 						    FermionField &out,int dag)
@@ -404,13 +395,11 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
 
   int LLs = in.Grid()->_rdimensions[0];
 
-  //  std::cout << " WilsonFermion5D Halo exch " <<std::endl;
   {
     GRID_TRACE("HaloExchange");
     st.HaloExchangeOpt(in,compressor);
   }
   
-  //  std::cout << " WilsonFermion5D Dhop " <<std::endl;
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
     GRID_TRACE("DhopDag");
@@ -419,7 +408,6 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
     GRID_TRACE("Dhop");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   }
-  //  std::cout << " WilsonFermion5D Done " <<std::endl;
 }
 
 
@@ -432,7 +420,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
   assert(in.Checkerboard()==Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven,UmuOdd,in,out,dag);
+  DhopInternal(StencilEven,LebesgueEvenOdd,UmuOdd,in,out,dag);
 }
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
@@ -443,31 +431,8 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
   assert(in.Checkerboard()==Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd,UmuEven,in,out,dag);
+  DhopInternal(StencilOdd,LebesgueEvenOdd,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)
 {
@@ -476,7 +441,7 @@ void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int d
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil,Umu,in,out,dag);
+  DhopInternal(Stencil,Lebesgue,Umu,in,out,dag);
 }
 template<class Impl>
 void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag)
@@ -485,54 +450,6 @@ 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)
@@ -818,15 +735,6 @@ 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,
@@ -842,7 +750,6 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHwQ(FermionField &out,const F
     typedef typename FermionField::scalar_type ScalComplex;
 
     typedef Lattice<iSinglet<vector_type> > LatComplex;
-    typedef iSpinMatrix<ScalComplex> SpinMat;
 
 
     Coordinate latt_size   = _grid->_fdimensions;
@@ -860,10 +767,8 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHwQ(FermionField &out,const F
     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];
@@ -872,18 +777,9 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHwQ(FermionField &out,const F
       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); 
 
-      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 - sin(kmu)*ci*(Gamma(Gmu[mu])*in);
 
     }
     num = num + mass * in ;

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

@@ -52,12 +52,17 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
     StencilEven(&Hgrid, npoint, Even, directions,displacements,p),  // source is Even
     StencilOdd(&Hgrid, npoint, Odd, directions,displacements,p),  // source is Odd
     mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
     Umu(&Fgrid),
     UmuEven(&Hgrid),
     UmuOdd(&Hgrid),
       _tmp(&Hgrid),
       anisotropyCoeff(anis)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
   // Allocate the required comms buffer
   ImportGauge(_Umu);
   if  (anisotropyCoeff.isAnisotropic){
@@ -309,7 +314,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil, Umu, in, out, dag);
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
 }
 
 template <class Impl>
@@ -321,7 +326,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
   assert(in.Checkerboard() == Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven, UmuOdd, in, out, dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
 }
 
 template <class Impl>
@@ -333,7 +338,7 @@ void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int d
   assert(in.Checkerboard() == Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd, UmuEven, in, out, dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
 }
 
 template <class Impl>
@@ -386,21 +391,21 @@ void WilsonFermion<Impl>::DhopDirCalc(const FermionField &in, FermionField &out,
 };
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, 
+void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        DoubledGaugeField &U,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
 #ifdef GRID_OMP
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
 #endif
-    DhopInternalSerial(st,U,in,out,dag);
+    DhopInternalSerial(st,lo,U,in,out,dag);
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, 
+void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 						      DoubledGaugeField &U,
 						      const FermionField &in,
 						      FermionField &out, int dag)
@@ -469,10 +474,10 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
 
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, 
-					     DoubledGaugeField &U,
-					     const FermionField &in,
-					     FermionField &out, int dag)
+void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
+                                       DoubledGaugeField &U,
+                                       const FermionField &in,
+                                       FermionField &out, int dag)
 {
   GRID_TRACE("DhopSerial");
   assert((dag == DaggerNo) || (dag == DaggerYes));

Grid/qcd/action/fermion/implementation/WilsonKernelsAsmAvx512.h

@@ -40,11 +40,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /// Switch off the 5d vectorised code optimisations
 #undef DWFVEC5D
 
-static std::vector<vComplexF> signsF;
+static Vector<vComplexF> signsF;
 
   template<typename vtype>    
-  int setupSigns(std::vector<vtype>& signs ){
-    std::vector<vtype> bother(2);
+  int setupSigns(Vector<vtype>& signs ){
+    Vector<vtype> bother(2);
     signs = bother;
     vrsign(signs[0]);
     visign(signs[1]);
@@ -364,7 +364,7 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, Doubled
 
 #include <simd/Intel512double.h>
     
-static std::vector<vComplexD> signsD;
+static Vector<vComplexD> signsD;
 static int signInitD = setupSigns(signsD);
     
 #define MAYBEPERM(A,perm) if (perm) { A ; }

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);
 
@@ -411,46 +411,6 @@ 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;					\
@@ -458,7 +418,7 @@ extern "C" {
       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();
 
@@ -474,7 +434,7 @@ extern "C" {
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
-    int ss = sss; /*st.lo->Reorder(sss);*/			\
+    int ss = st.lo->Reorder(sss);					\
     int sU = ss;							\
     int sF = ss*Ls;							\
     WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v);		\
@@ -491,8 +451,6 @@ extern "C" {
     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,
@@ -517,7 +475,7 @@ 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;}
@@ -527,18 +485,6 @@ 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

@@ -1,45 +0,0 @@
-/*************************************************************************************
-
-    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

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

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

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

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

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

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

@@ -40,11 +40,6 @@ 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;
@@ -76,27 +71,27 @@ public:
     return action;
   };
 
-  virtual void deriv(const GaugeField &U, GaugeField &dSdU) {
+  virtual void deriv(const GaugeField &Umu,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 = U.Grid();
+    GridBase *grid = Umu.Grid();
 
-    std::vector<GaugeLinkField> Umu (Nd,grid);
+    std::vector<GaugeLinkField> U (Nd,grid);
     for(int mu=0;mu<Nd;mu++){
-      Umu[mu] = PeekIndex<LorentzIndex>(U,mu);
+      U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
     }
     std::vector<GaugeLinkField> RectStaple(Nd,grid), Staple(Nd,grid);
-    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, Umu, workspace);
+    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, U, workspace);
 
     GaugeLinkField dSdU_mu(grid);
     GaugeLinkField staple(grid);
 
     for (int mu=0; mu < Nd; mu++){
-      dSdU_mu = Ta(Umu[mu]*Staple[mu])*factor_p;
-      dSdU_mu = dSdU_mu + Ta(Umu[mu]*RectStaple[mu])*factor_r;
-
+      dSdU_mu = Ta(U[mu]*Staple[mu])*factor_p;
+      dSdU_mu = dSdU_mu + Ta(U[mu]*RectStaple[mu])*factor_r;
+	  
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);
     }
 

Grid/qcd/action/gauge/WilsonGaugeAction.h

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

Grid/qcd/action/pseudofermion/EvenOddSchurDifferentiable.h

@@ -86,8 +86,13 @@ public:
     assert(ForceE.Checkerboard()==Even);
     assert(ForceO.Checkerboard()==Odd);
 
+#if defined(GRID_CUDA) || defined(GRID_HIP)  || defined(GRID_SYCL)
+    acceleratorSetCheckerboard(Force,ForceE);
+    acceleratorSetCheckerboard(Force,ForceO);
+#else
     setCheckerboard(Force,ForceE); 
     setCheckerboard(Force,ForceO);
+#endif
     Force=-Force;
 
     delete forcecb;
@@ -130,8 +135,13 @@ public:
     assert(ForceE.Checkerboard()==Even);
     assert(ForceO.Checkerboard()==Odd);
 
+#if defined(GRID_CUDA) || defined(GRID_HIP)  || defined(GRID_SYCL)
+    acceleratorSetCheckerboard(Force,ForceE);
+    acceleratorSetCheckerboard(Force,ForceO);
+#else
     setCheckerboard(Force,ForceE); 
     setCheckerboard(Force,ForceO);
+#endif
     Force=-Force;
 
     delete forcecb;

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, smr;
-    this->build_filenames(traj, Params, config, smr, rng);
+    std::string config, rng;
+    this->build_filenames(traj, Params, config, 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, smr, rng);
+      this->build_filenames(traj, Params, config, 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, smr, smr);
+	_IldgWriter.writeConfiguration<GaugeStats>(SmartConfig.get_U(true), traj, config, config);
 	_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, smr;
-    this->build_filenames(traj, Params, config, smr, rng);
+    std::string config, rng;
+    this->build_filenames(traj, Params, config, 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, smr;
-    this->build_filenames(traj, Params, config, smr, rng);
+    std::string config, rng;
+    this->build_filenames(traj, Params, config, rng);
     this->check_filename(rng);
     this->check_filename(config);
 

Grid/qcd/representations/adjoint.h

@@ -40,7 +40,7 @@ public:
     U = Zero();
     LatticeColourMatrix tmp(Uin.Grid());
 
-    std::vector<typename SU<ncolour>::Matrix> ta(Dimension);
+    Vector<typename SU<ncolour>::Matrix> ta(Dimension);
 
     // Debug lines
     // LatticeMatrix uno(Uin.Grid());

Grid/qcd/representations/two_index.h

@@ -43,7 +43,7 @@ public:
     U = Zero();
     LatticeColourMatrix tmp(Uin.Grid());
 
-    std::vector<typename GaugeGroup<ncolour,group_name>::Matrix> eij(Dimension);
+    Vector<typename GaugeGroup<ncolour,group_name>::Matrix> eij(Dimension);
 
     for (int a = 0; a < Dimension; a++)
       GaugeGroupTwoIndex<ncolour, S, group_name>::base(a, eij[a]);

Grid/qcd/smearing/HISQSmearing.h

@@ -62,15 +62,15 @@ accelerator_inline int stencilIndex(int mu, int nu) {
 
 
 /*!  @brief structure holding the link treatment */
-struct HISQSmearingParameters{
-    HISQSmearingParameters(){}
+struct SmearingParameters{
+    SmearingParameters(){}
     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
-    HISQSmearingParameters(Real c1, Real cnaik, Real c3, Real c5, Real c7, Real clp) 
+    SmearingParameters(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;
-    HISQSmearingParameters _linkTreatment;
+    SmearingParameters _linkTreatment;
 
 public:
 
@@ -117,7 +117,7 @@ public:
     //          IN--u_thin
     void smear(GF& u_smr, GF& u_naik, GF& u_thin) const {
 
-        HISQSmearingParameters lt = this->_linkTreatment;
+        SmearingParameters 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

@@ -207,14 +207,11 @@ std::vector<RealD> WilsonFlowBase<Gimpl>::flowMeasureEnergyDensityCloverleaf(con
 }
 
 template <class Gimpl>
-void WilsonFlowBase<Gimpl>::setDefaultMeasurements(int meas_interval){
-  addMeasurement(meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
+void WilsonFlowBase<Gimpl>::setDefaultMeasurements(int topq_meas_interval){
+  addMeasurement(1, [](int step, RealD t, const typename Gimpl::GaugeField &U){
       std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "  << step << "  " << t << "  " << energyDensityPlaquette(t,U) << std::endl;
     });
-  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){
+  addMeasurement(topq_meas_interval, [](int step, RealD t, const typename Gimpl::GaugeField &U){
       std::cout << GridLogMessage << "[WilsonFlow] Top. charge           : "  << step << "  " << WilsonLoops<Gimpl>::TopologicalCharge(U) << std::endl;
     });
 }
@@ -252,11 +249,6 @@ 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);
@@ -341,11 +333,6 @@ 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/BaryonUtils.h

@@ -971,9 +971,7 @@ void BaryonUtils<FImpl>::BaryonGamma3pt(
   autoView( vq_ti , q_ti     , AcceleratorRead);
   autoView( vq_tf , q_tf     , AcceleratorRead);
 
-  deviceVector<mobj> my_Dq_spec(2);
-  acceleratorPut(my_Dq_spec[0],Dq_spec1);
-  acceleratorPut(my_Dq_spec[1],Dq_spec2);
+  Vector<mobj> my_Dq_spec{Dq_spec1,Dq_spec2};
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if (group == 1) {
@@ -1302,8 +1300,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonEye(const PropagatorField &qq_loop,
   autoView( vd_tf   , qd_tf    , AcceleratorRead);
   autoView( vs_ti   , qs_ti    , AcceleratorRead);
 
-  deviceVector<mobj> my_Dq_spec(1);
-  acceleratorPut(my_Dq_spec[0],Du_spec);
+  Vector<mobj> my_Dq_spec{Du_spec};
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if(op == "Q1"){
@@ -1356,8 +1353,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonNonEye(const PropagatorField &qq_ti,
   autoView( vd_tf , qd_tf    , AcceleratorRead  );
   autoView( vs_ti , qs_ti    , AcceleratorRead  );
   
-  deviceVector<mobj> my_Dq_spec(1);
-  acceleratorPut(my_Dq_spec[0],Du_spec);
+  Vector<mobj> my_Dq_spec{Du_spec};
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if(op == "Q1"){
@@ -1548,9 +1544,7 @@ void BaryonUtils<FImpl>::XiToSigmaEye(const PropagatorField &qq_loop,
   autoView( vd_tf   , qd_tf    , AcceleratorRead);
   autoView( vs_ti   , qs_ti    , AcceleratorRead);
 
-  deviceVector<mobj> my_Dq_spec(2);
-  acceleratorPut(my_Dq_spec[0],Dd_spec);
-  acceleratorPut(my_Dq_spec[0],Ds_spec);
+  Vector<mobj> my_Dq_spec{Dd_spec,Ds_spec};
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if(op == "Q1"){

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 accelerator_inline void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::SU) {
+static 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/SUnAdjoint.h

@@ -62,7 +62,7 @@ public:
     // returns i(T_Adj)^index necessary for the projectors
     // see definitions above
     iAdjTa = Zero();
-    iSUnMatrix<cplx> ta[ncolour * ncolour - 1];
+    Vector<iSUnMatrix<cplx> > ta(ncolour * ncolour - 1);
     iSUnMatrix<cplx> tmp;
 
     // FIXME not very efficient to get all the generators everytime