Merge branch 'feature/dirichlet' into feature/block_lanczos22

Grid/DisableWarnings.h

@@ -45,7 +45,7 @@ directory
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
 
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
  //disables nvcc specific warning in json.hpp
 #pragma nv_diag_suppress unsigned_compare_with_zero
 #pragma nv_diag_suppress cast_to_qualified_type

Grid/GridCore.h

@@ -44,10 +44,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridStd.h>
 #include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
-#include <Grid/perfmon/Tracing.h>
 //#include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
+#include <Grid/perfmon/Tracing.h>
 #include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
 #include <Grid/threads/ThreadReduction.h>

Grid/Grid_Eigen_Dense.h

@@ -14,7 +14,7 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
 #pragma nv_diag_suppress code_is_unreachable
 #else
 #pragma diag_suppress code_is_unreachable

Grid/algorithms/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
         });
@@ -324,9 +324,9 @@ public:
   GridBase*        _cbgrid;
   int hermitian;
 
-  CartesianStencil<siteVector,siteVector,int> Stencil; 
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> Stencil; 
-  CartesianStencil<siteVector,siteVector,int> StencilEven;
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilEven;
-  CartesianStencil<siteVector,siteVector,int> StencilOdd;
+  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilOdd;
 
   std::vector<CoarseMatrix> A;
   std::vector<CoarseMatrix> Aeven;
@@ -631,7 +631,7 @@ public:
     assert(Aself != nullptr);
   }
 
-  void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
+  void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
                        const CoarseVector &in, CoarseVector &out, int dag) {
     int point = geom.npoint-1;
     autoView( out_v, out, AcceleratorWrite);
@@ -694,7 +694,7 @@ public:
     }
   }
 
-  void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
+  void DhopInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, std::vector<CoarseMatrix> &a,
                     const CoarseVector &in, CoarseVector &out, int dag) {
     SimpleCompressor<siteVector> compressor;
 
@@ -784,9 +784,9 @@ public:
     _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
-    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements),
-    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements),
     A(geom.npoint,&CoarseGrid),
     Aeven(geom.npoint,_cbgrid),
     Aodd(geom.npoint,_cbgrid),
@@ -804,9 +804,9 @@ public:
     _cbgrid(&CoarseRBGrid),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
-    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
+    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements),
-    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
+    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements),
     A(geom.npoint,&CoarseGrid),
     Aeven(geom.npoint,&CoarseRBGrid),
     Aodd(geom.npoint,&CoarseRBGrid),

Grid/algorithms/LinearOperator.h

@@ -526,6 +526,7 @@ public:
       (*this)(Linop,in[k],out[k]);
     }
   };
+  virtual ~OperatorFunction(){};
 };
 
 template<class Field> class LinearFunction {

Grid/algorithms/approx/Chebyshev.h

@@ -258,26 +258,12 @@ public:
     for(int n=2;n<order;n++){
 
       Linop.HermOp(*Tn,y);
-#if 0
-      auto y_v = y.View();
-      auto Tn_v = Tn->View();
-      auto Tnp_v = Tnp->View();
-      auto Tnm_v = Tnm->View();
-      constexpr int Nsimd = vector_type::Nsimd();
-      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
-	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
-	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
-      });
-      if ( Coeffs[n] != 0.0) {
-	axpy(out,Coeffs[n],*Tnp,out);
-      }
-#else
       axpby(y,xscale,mscale,y,(*Tn));
       axpby(*Tnp,2.0,-1.0,y,(*Tnm));
       if ( Coeffs[n] != 0.0) {
 	axpy(out,Coeffs[n],*Tnp,out);
       }
-#endif
+
       // Cycle pointers to avoid copies
       Field *swizzle = Tnm;
       Tnm    =Tn;

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -109,6 +109,9 @@ NAMESPACE_BEGIN(Grid);
     
     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
 
+    precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
+    
     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
@@ -123,7 +126,7 @@ NAMESPACE_BEGIN(Grid);
       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d);
+      precisionChange(src_f, src_d, pc_wk_dp_to_sp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -142,7 +145,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f);
+      precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -0,0 +1,373 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Christopher Kelly <ckelly@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
+//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
+//Every update_freq iterations the residual is corrected in double precision. 
+//For safety the a final regular CG is applied to clean up if necessary
+
+//PB Pure single, then double fixup
+
+template<class FieldD, class FieldF,
+	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class ConjugateGradientMultiShiftMixedPrecCleanup : public OperatorMultiFunction<FieldD>,
+					     public OperatorFunction<FieldD>
+{
+public:                                                
+
+  using OperatorFunction<FieldD>::operator();
+
+  RealD   Tolerance;
+  Integer MaxIterationsMshift;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
+  int verbose;
+  MultiShiftFunction shifts;
+  std::vector<RealD> TrueResidualShift;
+
+  int ReliableUpdateFreq; //number of iterations between reliable updates
+
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  LinearOperatorBase<FieldF> &Linop_f; //single precision
+
+  ConjugateGradientMultiShiftMixedPrecCleanup(Integer maxit, const MultiShiftFunction &_shifts,
+				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
+				       int _ReliableUpdateFreq) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+    MaxIterations(20000)
+  { 
+    verbose=1;
+    IterationsToCompleteShift.resize(_shifts.order);
+    TrueResidualShift.resize(_shifts.order);
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
+  {
+    GridBase *grid = src.Grid();
+    int nshift = shifts.order;
+    std::vector<FieldD> results(nshift,grid);
+    (*this)(Linop,src,results,psi);
+  }
+  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
+  {
+    int nshift = shifts.order;
+
+    (*this)(Linop,src,results);
+  
+    psi = shifts.norm*src;
+    for(int i=0;i<nshift;i++){
+      psi = psi + shifts.residues[i]*results[i];
+    }
+
+    return;
+  }
+
+  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
+  { 
+    GRID_TRACE("ConjugateGradientMultiShiftMixedPrecCleanup");
+    GridBase *DoublePrecGrid = src_d.Grid();
+
+    ////////////////////////////////////////////////////////////////////////
+    // Convenience references to the info stored in "MultiShiftFunction"
+    ////////////////////////////////////////////////////////////////////////
+    int nshift = shifts.order;
+
+    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
+    std::vector<RealD> &mresidual(shifts.tolerances);
+    std::vector<RealD> alpha(nshift,1.0);
+
+    //Double precision search directions
+    FieldD p_d(DoublePrecGrid);
+    std::vector<FieldF> ps_f (nshift, SinglePrecGrid);// Search directions (single precision)
+    std::vector<FieldF> psi_f(nshift, SinglePrecGrid);// solutions (single precision)
+
+    FieldD tmp_d(DoublePrecGrid);
+    FieldD r_d(DoublePrecGrid);
+    FieldF r_f(SinglePrecGrid);
+    FieldD mmp_d(DoublePrecGrid);
+
+    assert(psi_d.size()==nshift);
+    assert(mass.size()==nshift);
+    assert(mresidual.size()==nshift);
+  
+    // dynamic sized arrays on stack; 2d is a pain with vector
+    RealD  bs[nshift];
+    RealD  rsq[nshift];
+    RealD  rsqf[nshift];
+    RealD  z[nshift][2];
+    int     converged[nshift];
+  
+    const int       primary =0;
+  
+    //Primary shift fields CG iteration
+    RealD a,b,c,d;
+    RealD cp,bp,qq; //prev
+  
+    // Matrix mult fields
+    FieldF p_f(SinglePrecGrid);
+    FieldF mmp_f(SinglePrecGrid);
+
+    // Check lightest mass
+    for(int s=0;s<nshift;s++){
+      assert( mass[s]>= mass[primary] );
+      converged[s]=0;
+    }
+  
+    // Wire guess to zero
+    // Residuals "r" are src
+    // First search direction "p" is also src
+    cp = norm2(src_d);
+
+    // Handle trivial case of zero src.
+    if( cp == 0. ){
+      for(int s=0;s<nshift;s++){
+	psi_d[s] = Zero();
+	psi_f[s] = Zero();
+	IterationsToCompleteShift[s] = 1;
+	TrueResidualShift[s] = 0.;
+      }
+      return;
+    }
+
+    for(int s=0;s<nshift;s++){
+      rsq[s] = cp * mresidual[s] * mresidual[s];
+      rsqf[s] =rsq[s];
+      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
+      //      ps_d[s] = src_d;
+      precisionChangeFast(ps_f[s],src_d);
+    }
+    // r and p for primary
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
+    
+    //MdagM+m[0]
+    precisionChangeFast(p_f,p_d);
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    precisionChangeFast(tmp_d,mmp_f);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    //    assert(norm2(tmp_d)< 1.0e-4);
+
+    axpy(mmp_d,mass[0],p_d,mmp_d);
+    RealD rn = norm2(p_d);
+    d += rn*mass[0];
+
+    b = -cp /d;
+  
+    // Set up the various shift variables
+    int       iz=0;
+    z[0][1-iz] = 1.0;
+    z[0][iz]   = 1.0;
+    bs[0]      = b;
+    for(int s=1;s<nshift;s++){
+      z[s][1-iz] = 1.0;
+      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
+      bs[s]      = b*z[s][iz]; 
+    }
+  
+    // r += b[0] A.p[0]
+    // c= norm(r)
+    c=axpy_norm(r_d,b,mmp_d,r_d);
+  
+    for(int s=0;s<nshift;s++) {
+      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
+      precisionChangeFast(psi_f[s],psi_d[s]);
+    }
+  
+    ///////////////////////////////////////
+    // Timers
+    ///////////////////////////////////////
+    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
+
+    SolverTimer.Start();
+  
+    // Iteration loop
+    int k;
+  
+    for (k=1;k<=MaxIterationsMshift;k++){    
+
+      a = c /cp;
+      AXPYTimer.Start();
+      axpy(p_d,a,p_d,r_d); 
+      AXPYTimer.Stop();
+
+      PrecChangeTimer.Start();
+      precisionChangeFast(r_f, r_d);
+      PrecChangeTimer.Stop();
+
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	if ( ! converged[s] ) { 
+	  if (s==0){
+	    axpy(ps_f[s],a,ps_f[s],r_f);
+	  } else{
+	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
+	    axpby(ps_f[s],z[s][iz],as,r_f,ps_f[s]);
+	  }
+	}
+      }
+      AXPYTimer.Stop();
+
+      cp=c;
+      PrecChangeTimer.Start();
+      precisionChangeFast(p_f, p_d); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f);
+      MatrixTimer.Stop();  
+      PrecChangeTimer.Start();
+      precisionChangeFast(mmp_d, mmp_f); // From Float to Double
+      PrecChangeTimer.Stop();
+
+      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_d,mass[0],p_d,mmp_d);
+      RealD rn = norm2(p_d);
+      d += rn*mass[0];
+    
+      bp=b;
+      b=-cp/d;
+
+      // Toggle the recurrence history
+      bs[0] = b;
+      iz = 1-iz;
+      ShiftTimer.Start();
+      for(int s=1;s<nshift;s++){
+	if((!converged[s])){
+	  RealD z0 = z[s][1-iz];
+	  RealD z1 = z[s][iz];
+	  z[s][iz] = z0*z1*bp
+	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
+	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
+	}
+      }
+      ShiftTimer.Stop();
+
+      //Update single precision solutions
+      AXPYTimer.Start();
+      for(int s=0;s<nshift;s++){
+	int ss = s;
+	if( (!converged[s]) ) { 
+	  axpy(psi_f[ss],-bs[s]*alpha[s],ps_f[s],psi_f[ss]);
+	}
+      }
+      c = axpy_norm(r_d,b,mmp_d,r_d);
+      AXPYTimer.Stop();
+    
+      // Convergence checks
+      int all_converged = 1;
+      for(int s=0;s<nshift;s++){
+      
+	if ( (!converged[s]) ){
+	  IterationsToCompleteShift[s] = k;
+	
+	  RealD css  = c * z[s][iz]* z[s][iz];
+	
+	  if(css<rsqf[s]){
+	    if ( ! converged[s] )
+	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
+	    converged[s]=1;
+	  } else {
+	    all_converged=0;
+	  }
+
+	}
+      }
+
+      if ( all_converged || k == MaxIterationsMshift-1){
+
+	SolverTimer.Stop();
+
+	for(int s=0;s<nshift;s++){
+	  precisionChangeFast(psi_d[s],psi_f[s]);
+	}
+
+	
+	if ( all_converged ){
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: All shifts have converged iteration "<<k<<std::endl;
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Not all shifts have converged iteration "<<k<<std::endl;
+	}
+	
+	// Check answers 
+	for(int s=0; s < nshift; s++) { 
+	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
+	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
+	  axpy(r_d,-alpha[s],src_d,tmp_d);
+	  RealD rn = norm2(r_d);
+	  RealD cn = norm2(src_d);
+	  TrueResidualShift[s] = std::sqrt(rn/cn);
+	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
+
+	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
+	  if(rn >= rsq[s]){
+	    CleanupTimer.Start();
+	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: performing cleanup step for shift " << s << std::endl;
+
+	    //Setup linear operators for final cleanup
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
+	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
+					       
+	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
+	    cg(src_d, psi_d[s]);
+	    
+	    TrueResidualShift[s] = cg.TrueResidual;
+	    CleanupTimer.Stop();
+	  }
+	}
+
+	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrecCleanup: Time Breakdown for body"<<std::endl;
+	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
+	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
+
+	IterationsToComplete = k;	
+
+	return;
+      }
+   
+    }
+    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
+    assert(0);
+  }
+
+};
+NAMESPACE_END(Grid);
+

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -81,6 +81,7 @@ public:
   using OperatorFunction<FieldD>::operator();
 
   RealD   Tolerance;
+  Integer MaxIterationsMshift;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -95,9 +96,9 @@ public:
 
   ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
 				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
-				       int _ReliableUpdateFreq
+				       int _ReliableUpdateFreq) : 
-				       ) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
-    MaxIterations(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
+    MaxIterations(20000)
   { 
     verbose=1;
     IterationsToCompleteShift.resize(_shifts.order);
@@ -130,6 +131,9 @@ public:
     GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
     GridBase *DoublePrecGrid = src_d.Grid();
 
+    precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid);
+    
     ////////////////////////////////////////////////////////////////////////
     // Convenience references to the info stored in "MultiShiftFunction"
     ////////////////////////////////////////////////////////////////////////
@@ -154,6 +158,7 @@ public:
     // dynamic sized arrays on stack; 2d is a pain with vector
     RealD  bs[nshift];
     RealD  rsq[nshift];
+    RealD  rsqf[nshift];
     RealD  z[nshift][2];
     int     converged[nshift];
   
@@ -164,12 +169,8 @@ public:
     RealD cp,bp,qq; //prev
   
     // Matrix mult fields
-    FieldF r_f(SinglePrecGrid);
     FieldF p_f(SinglePrecGrid);
-    FieldF tmp_f(SinglePrecGrid);
     FieldF mmp_f(SinglePrecGrid);
-    FieldF src_f(SinglePrecGrid);
-    precisionChange(src_f, src_d);
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
@@ -194,18 +195,26 @@ public:
 
     for(int s=0;s<nshift;s++){
       rsq[s] = cp * mresidual[s] * mresidual[s];
+      rsqf[s] =rsq[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
       ps_d[s] = src_d;
     }
     // r and p for primary
-    r_f=src_f; //residual maintained in single
-    p_f=src_f;
     p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
     
     //MdagM+m[0]
+    precisionChange(p_f, p_d, pc_wk_d_to_s);
+
     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    axpy(mmp_f,mass[0],p_f,mmp_f);
+    precisionChange(tmp_d, mmp_f, pc_wk_s_to_d);
-    RealD rn = norm2(p_f);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    //    assert(norm2(tmp_d)< 1.0e-4);
+
+    axpy(mmp_d,mass[0],p_d,mmp_d);
+    RealD rn = norm2(p_d);
     d += rn*mass[0];
 
     b = -cp /d;
@@ -223,7 +232,7 @@ public:
   
     // r += b[0] A.p[0]
     // c= norm(r)
-    c=axpy_norm(r_f,b,mmp_f,r_f);
+    c=axpy_norm(r_d,b,mmp_d,r_d);
   
     for(int s=0;s<nshift;s++) {
       axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
@@ -239,14 +248,9 @@ public:
     // Iteration loop
     int k;
   
-    for (k=1;k<=MaxIterations;k++){    
+    for (k=1;k<=MaxIterationsMshift;k++){    
+
       a = c /cp;
-
-      //Update double precision search direction by residual
-      PrecChangeTimer.Start();
-      precisionChange(r_d, r_f);
-      PrecChangeTimer.Stop();
-
       AXPYTimer.Start();
       axpy(p_d,a,p_d,r_d); 
 
@@ -263,19 +267,23 @@ public:
       AXPYTimer.Stop();
 
       PrecChangeTimer.Start();
-      precisionChange(p_f, p_d); //get back single prec search direction for linop
+      precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop
       PrecChangeTimer.Stop();
 
       cp=c;
       MatrixTimer.Start();  
       Linop_f.HermOp(p_f,mmp_f);
-      d=real(innerProduct(p_f,mmp_f));    
       MatrixTimer.Stop();  
 
+      PrecChangeTimer.Start();
+      precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double
+      PrecChangeTimer.Stop();
+
       AXPYTimer.Start();
-      axpy(mmp_f,mass[0],p_f,mmp_f);
+      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_d,mass[0],p_d,mmp_d);
       AXPYTimer.Stop();
-      RealD rn = norm2(p_f);
+      RealD rn = norm2(p_d);
       d += rn*mass[0];
     
       bp=b;
@@ -306,12 +314,12 @@ public:
       }
 
       //Perform reliable update if necessary; otherwise update residual from single-prec mmp
-      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
+      c = axpy_norm(r_d,b,mmp_d,r_d);
+
       AXPYTimer.Stop();
 
-      c = c_f;
-
       if(k % ReliableUpdateFreq == 0){
+	RealD c_old = c;
 	//Replace r with true residual
 	MatrixTimer.Start();  
 	Linop_d.HermOp(psi_d[0],mmp_d); 
@@ -320,15 +328,10 @@ public:
 	AXPYTimer.Start();
 	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
 
-	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
+	c = axpy_norm(r_d, -1.0, mmp_d, src_d);
 	AXPYTimer.Stop();
 
-	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
+	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_old <<" with |r|^2 = "<<c<<std::endl;
-	
-	PrecChangeTimer.Start();
-	precisionChange(r_f, r_d);
-	PrecChangeTimer.Stop();
-	c = c_d;
       }
     
       // Convergence checks
@@ -340,7 +343,7 @@ public:
 	
 	  RealD css  = c * z[s][iz]* z[s][iz];
 	
-	  if(css<rsq[s]){
+	  if(css<rsqf[s]){
 	    if ( ! converged[s] )
 	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
 	    converged[s]=1;
@@ -351,11 +354,16 @@ public:
 	}
       }
 
-      if ( all_converged ){
+      if ( all_converged || k == MaxIterationsMshift-1){
 
 	SolverTimer.Stop();
+
+	if ( all_converged ){
 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
+	}
 	
 	// Check answers 
 	for(int s=0; s < nshift; s++) { 
@@ -398,11 +406,9 @@ public:
 	return;
       }
    
-   
     }
-    // ugly hack
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
+    assert(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -48,7 +48,7 @@ public:
   LinearOperatorBase<FieldF> &Linop_f;
   LinearOperatorBase<FieldD> &Linop_d;
   GridBase* SinglePrecGrid;
-  RealD Delta; //reliable update parameter
+  RealD Delta; //reliable update parameter. A reliable update is performed when the residual drops by a factor of Delta relative to its value at the last update
 
   //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
   LinearOperatorBase<FieldF> *Linop_fallback;
@@ -65,7 +65,9 @@ public:
       ErrorOnNoConverge(err_on_no_conv),
       DoFinalCleanup(true),
       Linop_fallback(NULL)
-  {};
+  {
+    assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
+  };
 
   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
     Linop_fallback = &_Linop_fallback;
@@ -116,9 +118,12 @@ public:
     }
 
     //Single prec initialization
+    precisionChangeWorkspace pc_wk_sp_to_dp(src.Grid(), SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, src.Grid());
+    
     FieldF r_f(SinglePrecGrid);
     r_f.Checkerboard() = r.Checkerboard();
-    precisionChange(r_f, r);
+    precisionChange(r_f, r, pc_wk_dp_to_sp);
 
     FieldF psi_f(r_f);
     psi_f = Zero();
@@ -134,6 +139,7 @@ public:
     GridStopWatch LinalgTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
+    GridStopWatch PrecChangeTimer;
     
     SolverTimer.Start();
     int k = 0;
@@ -173,7 +179,9 @@ public:
       // Stopping condition
       if (cp <= rsq) {
 	//Although not written in the paper, I assume that I have to add on the final solution
-	precisionChange(mmp, psi_f);
+	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
+	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 	
 	
@@ -194,6 +202,9 @@ public:
 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPrecChange " << PrecChangeTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPrecChange avg time " << PrecChangeTimer.Elapsed()/(2*l+1) <<std::endl;
+
 	
 	IterationsToComplete = k;	
 	ReliableUpdatesPerformed = l;
@@ -214,14 +225,21 @@ public:
       else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
 		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
-	precisionChange(mmp, psi_f);
+	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
+	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 
+	MatrixTimer.Start();
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	MatrixTimer.Stop();
+	
 	r = src - mmp;
 
 	psi_f = Zero();
-	precisionChange(r_f, r);
+	PrecChangeTimer.Start();
+	precisionChange(r_f, r, pc_wk_dp_to_sp);
+	PrecChangeTimer.Stop();
 	cp = norm2(r);
 	MaxResidSinceLastRelUp = cp;
 

Grid/allocator/MemoryManager.h

@@ -36,6 +36,11 @@ NAMESPACE_BEGIN(Grid);
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
+
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
 // Advise the LatticeAccelerator class
@@ -92,8 +97,9 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
+  static void Audit(std::string s);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -113,6 +119,8 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
  
  private:
 #ifndef GRID_UVM
@@ -170,6 +178,7 @@ private:
 
  public:
   static void Print(void);
+  static void PrintAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -8,9 +8,8 @@ NAMESPACE_BEGIN(Grid);
 static char print_buffer [ MAXLINE ];
 
 #define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-//#define dprintf(...) printf (__VA_ARGS__ ); fflush(stdout);
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) 
+//#define dprintf(...) 
-
 
 
 ////////////////////////////////////////////////////////////
@@ -29,6 +28,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -116,8 +117,10 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t) NULL;
     dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
@@ -127,26 +130,36 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove entry
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
-  // Cannot be locked. If allocated must be in LRU pool.
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
-  assert(AccCache.accLock==0);
+	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
-  assert(AccCache.cpuLock==0);
+	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  if (AccCache.accLock!=0) return;
+  if (AccCache.cpuLock!=0) return;
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
     dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  uint64_t CpuPtr = AccCache.CpuPtr;
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
-  EntryErase(CpuPtr);
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -197,6 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -208,6 +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\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -218,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back();
+      uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
+    } else {
+      return;
     }
   }
 }
@@ -247,11 +265,12 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld\n",
+    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-		    (uint64_t)bytes);
+	            (uint64_t)bytes,
+		    (uint64_t)AccCache.accLock);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -286,6 +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\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -298,28 +318,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",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\n",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 into device accLock %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  // If view is opened on device remove from LRU
+  assert(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
+    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -340,10 +362,12 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
-
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
+  } else {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -380,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  if (!AccCache.AccPtr) {
+  // CPU doesn't need to free space
-     EvictVictims(bytes);
+  //  if (!AccCache.AccPtr) {
-  }
+  //    EvictVictims(bytes);
+  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -436,20 +461,28 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
-  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -459,13 +492,13 @@ void  MemoryManager::Print(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -479,6 +512,64 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
+void MemoryManager::Audit(std::string s)
+{
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  uint64_t LockedBytes=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    assert(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    assert(AccCache.LRU_valid==1);
+    assert(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      assert(AccCache.LRU_valid==0);
+
+      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+		<< "\t cpuLock  " << AccCache.cpuLock
+		<< "\t accLock  " << AccCache.accLock
+		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
+    }
+
+    assert( AccCache.cpuLock== 0 ) ;
+    assert( AccCache.accLock== 0 ) ;
+  }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  assert(LruBytes1==LruBytes2);
+  assert(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  assert(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  assert(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+
+}
 
 void MemoryManager::PrintState(void* _CpuPtr)
 {
@@ -495,8 +586,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
     if ( AccCache.state==EvictNext) str = std::string("EvictNext");
 
     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
     << "\t" << AccCache.cpuLock
     << "\t" << AccCache.accLock
     << "\t" << AccCache.LRU_valid<<std::endl;

Grid/allocator/MemoryManagerShared.cc

@@ -12,7 +12,10 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
+void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
@@ -21,6 +24,7 @@ void  MemoryManager::PrintState(void* CpuPtr)
 std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
 };
 void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_base.h

@@ -131,7 +131,7 @@ public:
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
-    scalar_type * ptr = (scalar_type *)& o;
+    scalar_type * ptr = (scalar_type *)& o; // Safe alias 
     GlobalSumVector(ptr,words);
   }
   
@@ -155,7 +155,7 @@ public:
 				    int xmit_to_rank,int do_xmit,
 				    void *recv,
 				    int recv_from_rank,int do_recv,
-				    int bytes,int dir);
+				    int xbytes,int rbytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);

Grid/communicator/Communicator_mpi3.cc

@@ -343,7 +343,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
@@ -353,7 +353,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int dest,int dox,
 							 void *recv,
 							 int from,int dor,
-							 int bytes,int dir)
+							 int xbytes,int rbytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -375,32 +375,27 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   if ( dor ) {
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
-      ierr=MPI_Irecv(recv, bytes, 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+=bytes;
+      off_node_bytes+=rbytes;
     }
   }
   
   if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, bytes, 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+=bytes;
+      off_node_bytes+=xbytes;
     } else {
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
       assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
     }
   }
 
-  /*  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-   *    this->StencilSendToRecvFromComplete(list,dir);
-   *    list.resize(0);
-   *  }
-   */
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)

Grid/communicator/Communicator_none.cc

@@ -126,7 +126,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int xmit_to_rank,int dox,
 							 void *recv,
 							 int recv_from_rank,int dor,
-							 int bytes, int dir)
+							 int xbytes,int rbytes, int dir)
 {
   return 2.0*bytes;
 }

Grid/communicator/SharedMemoryMPI.cc

@@ -29,6 +29,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 
 #include <Grid/GridCore.h>
 #include <pwd.h>
+#include <syscall.h>
 
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>

Grid/lattice/Lattice_ET.h

@@ -63,7 +63,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
+  //  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   const int Nsimd = vobj::vector_type::Nsimd();

Grid/lattice/Lattice_base.h

@@ -291,8 +291,8 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
@@ -306,8 +306,8 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });

Grid/lattice/Lattice_matrix_reduction.h

@@ -32,7 +32,6 @@ template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -82,7 +81,6 @@ template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -130,7 +128,6 @@ template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs.Grid();

Grid/lattice/Lattice_peekpoke.h

@@ -96,9 +96,6 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   GridBase *grid=l.Grid();
 
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
   int Nsimd = grid->Nsimd();
 
   assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
@@ -136,9 +133,6 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         
   GridBase *grid=l.Grid();
 
-  typedef typename vobj::scalar_type scalar_type;
-  typedef typename vobj::vector_type vector_type;
-
   int Nsimd = grid->Nsimd();
 
   assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
@@ -179,11 +173,11 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  scalar_type * vp = (scalar_type *)&l[odx];
+  const vector_type *vp = (const vector_type *) &l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
-    pt[w] = vp[idx+w*Nsimd];
+    pt[w] = getlane(vp[w],idx);
   }
       
   return;
@@ -216,10 +210,10 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  scalar_type * vp = (scalar_type *)&l[odx];
+  vector_type * vp = (vector_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
   for(int w=0;w<words;w++){
-    vp[idx+w*Nsimd] = pt[w];
+    putlane(vp[w],pt[w],idx);
   }
   return;
 };

Grid/lattice/Lattice_reduction.h

@@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
+#if defined(GRID_SYCL)
+#include <Grid/lattice/Lattice_reduction_sycl.h>
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -124,7 +127,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -133,7 +136,7 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -142,7 +145,7 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu_large(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -152,13 +155,13 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu(&arg_v[0],osites);
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  typename vobj::scalar_object ssum;
+  autoView( arg_v, arg, AcceleratorRead);
+  ssum= sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
   auto ssum= sum_cpu(&arg_v[0],osites);
 #endif  
   arg.Grid()->GlobalSum(ssum);
@@ -168,7 +171,7 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 template<class vobj>
 inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
   auto ssum= sum_gpu_large(&arg_v[0],osites);
@@ -219,7 +222,6 @@ template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
 template<class vobj>
 inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
 {
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
   ComplexD  nrm;
   
@@ -233,11 +235,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-    
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
-
+    // This code could read coalesce
     // GPU - SIMT lane compliance...
     accelerator_for( ss, sites, nsimd,{
 	auto x_l = left_v(ss);
@@ -296,7 +297,6 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   conformable(z,x);
   conformable(x,y);
 
-  typedef typename vobj::scalar_type scalar_type;
   //  typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
@@ -341,7 +341,6 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
 {
   conformable(left,right);
 
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
   Vector<ComplexD> tmp(2);
 
@@ -598,6 +597,7 @@ template<class vobj>
 static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
 			    int orthogdim,RealD scale=1.0) 
 {
+  // perhaps easier to just promote A to a field and use regular madd
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
@@ -628,8 +628,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
     for(int l=0;l<Nsimd;l++){
       grid->iCoorFromIindex(icoor,l);
       int ldx =r+icoor[orthogdim]*rd;
-      scalar_type *as =(scalar_type *)&av;
+      av.putlane(scalar_type(a[ldx])*zscale,l);
-      as[l] = scalar_type(a[ldx])*zscale;
     }
 
     tensor_reduced at; at=av;
@@ -669,7 +668,6 @@ template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -723,7 +721,6 @@ template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -777,7 +774,6 @@ template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
   typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs.Grid();

Grid/lattice/Lattice_reduction_gpu.h

@@ -211,13 +211,28 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   assert(ok);
 
   Integer smemSize = numThreads * sizeof(sobj);
-
+  // UVM seems to be buggy under later CUDA drivers
-  Vector<sobj> buffer(numBlocks);
+  // This fails on A100 and driver 5.30.02 / CUDA 12.1
+  // Fails with multiple NVCC versions back to 11.4,
+  // which worked with earlier drivers.
+  // Not sure which driver had first fail and this bears checking
+  // Is awkward as must install multiple driver versions
+#undef UVM_BLOCK_BUFFER  
+#ifndef UVM_BLOCK_BUFFER  
+  commVector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
-  
+  sobj result;
   reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
   accelerator_barrier();
-  auto result = buffer_v[0];
+  acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
+#else
+  Vector<sobj> buffer(numBlocks);
+  sobj *buffer_v = &buffer[0];
+  sobj result;
+  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  accelerator_barrier();
+  result = *buffer_v;
+#endif
   return result;
 }
 
@@ -250,8 +265,6 @@ inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osi
 template <class vobj>
 inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
 {
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
   typedef typename vobj::scalar_objectD sobj;
   sobj ret;
   

Grid/lattice/Lattice_reduction_sycl.h

@@ -0,0 +1,125 @@
+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;
+  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
+  sobj identity; zeroit(identity);
+  sobj ret ; 
+
+  Integer nsimd= vobj::Nsimd();
+  
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
+     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;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_tensor(lat,osites);
+}
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Return as same precision as input performing reduction in double precision though
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu(lat,osites);
+  return result;
+}
+
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu_large(lat,osites);
+  return result;
+}
+
+NAMESPACE_END(Grid);
+
+/*
+template<class Double> Double svm_reduce(Double *vec,uint64_t L)
+{
+  Double sumResult; zeroit(sumResult);
+  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
+  Double identity;  zeroit(identity);
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{L},
+		      Reduction,
+		      [=] (cl::sycl::id<1> index, auto &sum) {
+	 sum +=vec[index];
+     });
+   });
+  theGridAccelerator->wait();
+  Double ret = d_sum[0];
+  free(d_sum,*theGridAccelerator);
+  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
+  return ret;
+}
+
+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_transfer.h

@@ -194,11 +194,11 @@ accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
 #endif
 
 accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) {
-  out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v);
+  precisionChange(out,in);
 }
 
 accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
-  Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
+  precisionChange(out,in);
 }
 
 template<typename T1,typename T2>
@@ -677,10 +677,10 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
       Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
       Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
-      scalar_type * fp = (scalar_type *)&f_v[odx_f];
+      vector_type * fp = (vector_type *)&f_v[odx_f];
-      scalar_type * tp = (scalar_type *)&t_v[odx_t];
+      vector_type * tp = (vector_type *)&t_v[odx_t];
       for(int w=0;w<words;w++){
-	tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd];  // FIXME IF RRII layout, type pun no worke
+	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
       }
     }
   });
@@ -1080,9 +1080,27 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//Convert a Lattice from one precision to another
+//Very fast precision change. Requires in/out objects to reside on same Grid (e.g. by using double2 for the double-precision field)
 template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+{
+  typedef typename VobjOut::vector_type Vout;
+  typedef typename VobjIn::vector_type Vin;
+  const int N = sizeof(VobjOut)/sizeof(Vout);
+  conformable(out.Grid(),in.Grid());
+  out.Checkerboard() = in.Checkerboard();
+  int nsimd = out.Grid()->Nsimd();
+  autoView( out_v  , out, AcceleratorWrite);
+  autoView(  in_v ,   in, AcceleratorRead);
+  accelerator_for(idx,out.Grid()->oSites(),1,{
+      Vout *vout = (Vout *)&out_v[idx];
+      Vin  *vin  = (Vin  *)&in_v[idx];
+      precisionChange(vout,vin,N);
+  });
+}
+//Convert a Lattice from one precision to another (original, slow implementation)
+template<class VobjOut, class VobjIn>
+void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 {
   assert(out.Grid()->Nd() == in.Grid()->Nd());
   for(int d=0;d<out.Grid()->Nd();d++){
@@ -1097,7 +1115,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 
   int ndim = out.Grid()->Nd();
   int out_nsimd = out_grid->Nsimd();
-    
+  int in_nsimd = in_grid->Nsimd();
   std::vector<Coordinate > out_icoor(out_nsimd);
       
   for(int lane=0; lane < out_nsimd; lane++){
@@ -1128,6 +1146,128 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   });
 }
 
+//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
+class precisionChangeWorkspace{
+  std::pair<Integer,Integer>* fmap_device; //device pointer
+  //maintain grids for checking
+  GridBase* _out_grid;
+  GridBase* _in_grid;
+public:
+  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
+    //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
+    assert(out_grid->Nd() == in_grid->Nd());
+    for(int d=0;d<out_grid->Nd();d++){
+      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
+    }
+    int Nsimd_out = out_grid->Nsimd();
+
+    std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
+    for(int lane=0; lane < out_grid->Nsimd(); lane++)
+      out_grid->iCoorFromIindex(out_icorrs[lane], lane);
+  
+    std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
+    thread_for(out_oidx,out_grid->oSites(),{
+	Coordinate out_ocorr; 
+	out_grid->oCoorFromOindex(out_ocorr, out_oidx);
+      
+	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
+	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
+	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
+	
+	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
+	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
+	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
+	  int in_oidx = 0, in_lane = 0;
+	  for(int d=0;d<in_grid->_ndimension;d++){
+	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
+	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
+	  }
+	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
+	}
+      });
+
+    //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
+    size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
+    fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
+    acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes); 
+  }
+
+  //Prevent moving or copying
+  precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
+  precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
+  
+  std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
+
+  void checkGrids(GridBase* out, GridBase* in) const{
+    conformable(out, _out_grid);
+    conformable(in, _in_grid);
+  }
+  
+  ~precisionChangeWorkspace(){
+    acceleratorFreeDevice(fmap_device);
+  }
+};
+
+
+//We would like to use precisionChangeFast when possible. However usage of this requires the Grids to be the same (runtime check)
+//*and* the precisionChange(VobjOut::vector_type, VobjIn, int) function to be defined for the types; this requires an extra compile-time check which we do using some SFINAE trickery
+template<class VobjOut, class VobjIn>
+auto _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, int dummy)->decltype( precisionChange( ((typename VobjOut::vector_type*)0), ((typename VobjIn::vector_type*)0), 1), int()){
+  if(out.Grid() == in.Grid()){
+    precisionChangeFast(out,in);
+    return 1;
+  }else{
+    return 0;
+  }
+}
+template<class VobjOut, class VobjIn>
+int _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, long dummy){ //note long here is intentional; it means the above is preferred if available
+  return 0;
+}
+
+
+//Convert a lattice of one precision to another. Much faster than original implementation but requires a pregenerated workspace
+//which contains the mapping data.
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
+  if(_precisionChangeFastWrap(out,in,0)) return;
+  
+  static_assert( std::is_same<typename VobjOut::scalar_typeD, typename VobjIn::scalar_typeD>::value == 1, "precisionChange: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+
+  out.Checkerboard() = in.Checkerboard();
+  constexpr int Nsimd_out = VobjOut::Nsimd();
+
+  workspace.checkGrids(out.Grid(),in.Grid());
+  std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
+
+  //Do the copy/precision change
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( in_v , in, AcceleratorRead);
+
+  accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
+      std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
+      for(int out_lane=0; out_lane < Nsimd_out; out_lane++){      
+	int in_oidx = fmap_osite[out_lane].first;
+	int in_lane = fmap_osite[out_lane].second;
+	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
+      }
+    });
+}
+
+//Convert a Lattice from one precision to another. Much faster than original implementation but slower than precisionChangeFast
+//or precisionChange called with pregenerated workspace, as it needs to internally generate the workspace on the host and copy to device
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
+  if(_precisionChangeFastWrap(out,in,0)) return;   
+  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
+  precisionChange(out, in, workspace);
+}
+
+
+
+
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////

Grid/log/Log.cc

@@ -66,6 +66,7 @@ GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
 GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
+GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
 GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
@@ -77,7 +78,8 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
-  GridLogMemory.Active(0); // at least the messages should be always on
+  GridLogMemory.Active(0); 
+  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
@@ -87,6 +89,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
+    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
     if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);

Grid/log/Log.h

@@ -186,6 +186,7 @@ extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
 extern GridLogger GridLogHMC;
 extern GridLogger GridLogMemory;
+extern GridLogger GridLogTracing;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/perfmon/PerfCount.h

@@ -30,6 +30,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
+
+#ifndef __SSC_START
+#define __SSC_START
+#define __SSC_STOP
+#endif
+
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>

Grid/perfmon/Tracing.h

@@ -1,4 +1,7 @@
 #pragma once
+
+NAMESPACE_BEGIN(Grid);
+
 #ifdef GRID_TRACING_NVTX
 #include <nvToolsExt.h>
 class GridTracer {
@@ -64,3 +67,4 @@ inline void traceStop(int ID) {  }
 #else
 #define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
 #endif
+NAMESPACE_END(Grid);

Grid/pugixml/pugixml.cc

@@ -16,7 +16,7 @@
 
 #ifdef __NVCC__
 #pragma push
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
 #pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
 #else
 #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"

Grid/qcd/QCD.h

@@ -126,6 +126,7 @@ typedef iSpinMatrix<ComplexD >          SpinMatrixD;
 typedef iSpinMatrix<vComplex >          vSpinMatrix;
 typedef iSpinMatrix<vComplexF>          vSpinMatrixF;
 typedef iSpinMatrix<vComplexD>          vSpinMatrixD;
+typedef iSpinMatrix<vComplexD2>         vSpinMatrixD2;
 
 // Colour Matrix
 typedef iColourMatrix<Complex  >        ColourMatrix;
@@ -135,6 +136,7 @@ typedef iColourMatrix<ComplexD >        ColourMatrixD;
 typedef iColourMatrix<vComplex >        vColourMatrix;
 typedef iColourMatrix<vComplexF>        vColourMatrixF;
 typedef iColourMatrix<vComplexD>        vColourMatrixD;
+typedef iColourMatrix<vComplexD2>       vColourMatrixD2;
 
 // SpinColour matrix
 typedef iSpinColourMatrix<Complex  >    SpinColourMatrix;
@@ -144,6 +146,7 @@ typedef iSpinColourMatrix<ComplexD >    SpinColourMatrixD;
 typedef iSpinColourMatrix<vComplex >    vSpinColourMatrix;
 typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF;
 typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD;
+typedef iSpinColourMatrix<vComplexD2>   vSpinColourMatrixD2;
 
 // SpinColourSpinColour matrix
 typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
@@ -153,6 +156,7 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2;
 
 // SpinColourSpinColour matrix
 typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
@@ -162,6 +166,7 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
+typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2;
 
 // LorentzColour
 typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix;
@@ -171,6 +176,7 @@ typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
 typedef iLorentzColourMatrix<vComplex >  vLorentzColourMatrix;
 typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF;
 typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD;
+typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2;
 
 // DoubleStored gauge field
 typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix;
@@ -180,6 +186,7 @@ typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD;
 typedef iDoubleStoredColourMatrix<vComplex >  vDoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<vComplexF>  vDoubleStoredColourMatrixF;
 typedef iDoubleStoredColourMatrix<vComplexD>  vDoubleStoredColourMatrixD;
+typedef iDoubleStoredColourMatrix<vComplexD2> vDoubleStoredColourMatrixD2;
 
 //G-parity flavour matrix
 typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
@@ -189,6 +196,7 @@ typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
 typedef iGparityFlavourMatrix<vComplex>   vGparityFlavourMatrix;
 typedef iGparityFlavourMatrix<vComplexF>  vGparityFlavourMatrixF;
 typedef iGparityFlavourMatrix<vComplexD>  vGparityFlavourMatrixD;
+typedef iGparityFlavourMatrix<vComplexD2> vGparityFlavourMatrixD2;
 
 
 // Spin vector
@@ -199,6 +207,7 @@ typedef iSpinVector<ComplexD>           SpinVectorD;
 typedef iSpinVector<vComplex >           vSpinVector;
 typedef iSpinVector<vComplexF>           vSpinVectorF;
 typedef iSpinVector<vComplexD>           vSpinVectorD;
+typedef iSpinVector<vComplexD2>          vSpinVectorD2;
 
 // Colour vector
 typedef iColourVector<Complex >         ColourVector;
@@ -208,6 +217,7 @@ typedef iColourVector<ComplexD>         ColourVectorD;
 typedef iColourVector<vComplex >         vColourVector;
 typedef iColourVector<vComplexF>         vColourVectorF;
 typedef iColourVector<vComplexD>         vColourVectorD;
+typedef iColourVector<vComplexD2>        vColourVectorD2;
 
 // SpinColourVector
 typedef iSpinColourVector<Complex >     SpinColourVector;
@@ -217,6 +227,7 @@ typedef iSpinColourVector<ComplexD>     SpinColourVectorD;
 typedef iSpinColourVector<vComplex >     vSpinColourVector;
 typedef iSpinColourVector<vComplexF>     vSpinColourVectorF;
 typedef iSpinColourVector<vComplexD>     vSpinColourVectorD;
+typedef iSpinColourVector<vComplexD2>    vSpinColourVectorD2;
 
 // HalfSpin vector
 typedef iHalfSpinVector<Complex >       HalfSpinVector;
@@ -226,6 +237,7 @@ typedef iHalfSpinVector<ComplexD>       HalfSpinVectorD;
 typedef iHalfSpinVector<vComplex >       vHalfSpinVector;
 typedef iHalfSpinVector<vComplexF>       vHalfSpinVectorF;
 typedef iHalfSpinVector<vComplexD>       vHalfSpinVectorD;
+typedef iHalfSpinVector<vComplexD2>      vHalfSpinVectorD2;
 
 // HalfSpinColour vector
 typedef iHalfSpinColourVector<Complex > HalfSpinColourVector;
@@ -235,6 +247,7 @@ typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
 typedef iHalfSpinColourVector<vComplex >  vHalfSpinColourVector;
 typedef iHalfSpinColourVector<vComplexF>  vHalfSpinColourVectorF;
 typedef iHalfSpinColourVector<vComplexD>  vHalfSpinColourVectorD;
+typedef iHalfSpinColourVector<vComplexD2> vHalfSpinColourVectorD2;
 
 //G-parity flavour vector
 typedef iGparityFlavourVector<Complex >         GparityFlavourVector;
@@ -244,7 +257,7 @@ typedef iGparityFlavourVector<ComplexD>         GparityFlavourVectorD;
 typedef iGparityFlavourVector<vComplex >         vGparityFlavourVector;
 typedef iGparityFlavourVector<vComplexF>         vGparityFlavourVectorF;
 typedef iGparityFlavourVector<vComplexD>         vGparityFlavourVectorD;
-
+typedef iGparityFlavourVector<vComplexD2>        vGparityFlavourVectorD2;
     
 // singlets
 typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type.
@@ -254,6 +267,7 @@ typedef iSinglet<ComplexD>         TComplexD;    // FIXME This is painful. Tenso
 typedef iSinglet<vComplex >        vTComplex ;   // what if we don't know the tensor structure
 typedef iSinglet<vComplexF>        vTComplexF;   // what if we don't know the tensor structure
 typedef iSinglet<vComplexD>        vTComplexD;   // what if we don't know the tensor structure
+typedef iSinglet<vComplexD2>       vTComplexD2;   // what if we don't know the tensor structure
 
 typedef iSinglet<Real >            TReal;        // Shouldn't need these; can I make it work without?
 typedef iSinglet<RealF>            TRealF;       // Shouldn't need these; can I make it work without?
@@ -271,47 +285,58 @@ typedef iSinglet<Integer >         TInteger;
 typedef Lattice<vColourMatrix>          LatticeColourMatrix;
 typedef Lattice<vColourMatrixF>         LatticeColourMatrixF;
 typedef Lattice<vColourMatrixD>         LatticeColourMatrixD;
+typedef Lattice<vColourMatrixD2>        LatticeColourMatrixD2;
 
 typedef Lattice<vSpinMatrix>            LatticeSpinMatrix;
 typedef Lattice<vSpinMatrixF>           LatticeSpinMatrixF;
 typedef Lattice<vSpinMatrixD>           LatticeSpinMatrixD;
+typedef Lattice<vSpinMatrixD2>          LatticeSpinMatrixD2;
 
 typedef Lattice<vSpinColourMatrix>      LatticeSpinColourMatrix;
 typedef Lattice<vSpinColourMatrixF>     LatticeSpinColourMatrixF;
 typedef Lattice<vSpinColourMatrixD>     LatticeSpinColourMatrixD;
+typedef Lattice<vSpinColourMatrixD2>    LatticeSpinColourMatrixD2;
 
 typedef Lattice<vSpinColourSpinColourMatrix>      LatticeSpinColourSpinColourMatrix;
 typedef Lattice<vSpinColourSpinColourMatrixF>     LatticeSpinColourSpinColourMatrixF;
 typedef Lattice<vSpinColourSpinColourMatrixD>     LatticeSpinColourSpinColourMatrixD;
+typedef Lattice<vSpinColourSpinColourMatrixD2>    LatticeSpinColourSpinColourMatrixD2;
 
 typedef Lattice<vLorentzColourMatrix>   LatticeLorentzColourMatrix;
 typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF;
 typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD;
+typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2;
 
 // DoubleStored gauge field
 typedef Lattice<vDoubleStoredColourMatrix>   LatticeDoubleStoredColourMatrix;
 typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF;
 typedef Lattice<vDoubleStoredColourMatrixD>  LatticeDoubleStoredColourMatrixD;
+typedef Lattice<vDoubleStoredColourMatrixD2> LatticeDoubleStoredColourMatrixD2;
 
 typedef Lattice<vSpinVector>            LatticeSpinVector;
 typedef Lattice<vSpinVectorF>           LatticeSpinVectorF;
 typedef Lattice<vSpinVectorD>           LatticeSpinVectorD;
+typedef Lattice<vSpinVectorD2>          LatticeSpinVectorD2;
 
 typedef Lattice<vColourVector>          LatticeColourVector;
 typedef Lattice<vColourVectorF>         LatticeColourVectorF;
 typedef Lattice<vColourVectorD>         LatticeColourVectorD;
+typedef Lattice<vColourVectorD2>        LatticeColourVectorD2;
 
 typedef Lattice<vSpinColourVector>      LatticeSpinColourVector;
 typedef Lattice<vSpinColourVectorF>     LatticeSpinColourVectorF;
 typedef Lattice<vSpinColourVectorD>     LatticeSpinColourVectorD;
+typedef Lattice<vSpinColourVectorD2>    LatticeSpinColourVectorD2;
 
 typedef Lattice<vHalfSpinVector>        LatticeHalfSpinVector;
 typedef Lattice<vHalfSpinVectorF>       LatticeHalfSpinVectorF;
 typedef Lattice<vHalfSpinVectorD>       LatticeHalfSpinVectorD;
+typedef Lattice<vHalfSpinVectorD2>      LatticeHalfSpinVectorD2;
 
 typedef Lattice<vHalfSpinColourVector>   LatticeHalfSpinColourVector;
 typedef Lattice<vHalfSpinColourVectorF>  LatticeHalfSpinColourVectorF;
 typedef Lattice<vHalfSpinColourVectorD>  LatticeHalfSpinColourVectorD;
+typedef Lattice<vHalfSpinColourVectorD2> LatticeHalfSpinColourVectorD2;
 
 typedef Lattice<vTReal>            LatticeReal;
 typedef Lattice<vTRealF>           LatticeRealF;
@@ -320,6 +345,7 @@ typedef Lattice<vTRealD>           LatticeRealD;
 typedef Lattice<vTComplex>         LatticeComplex;
 typedef Lattice<vTComplexF>        LatticeComplexF;
 typedef Lattice<vTComplexD>        LatticeComplexD;
+typedef Lattice<vTComplexD2>       LatticeComplexD2;
 
 typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where"
 
@@ -329,35 +355,40 @@ typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where"
 ///////////////////////////////////////////
 typedef LatticeHalfSpinColourVector   LatticeHalfFermion;
 typedef LatticeHalfSpinColourVectorF  LatticeHalfFermionF;
-typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD;
+typedef LatticeHalfSpinColourVectorD  LatticeHalfFermionD;
+typedef LatticeHalfSpinColourVectorD2 LatticeHalfFermionD2;
 
 typedef LatticeSpinColourVector      LatticeFermion;
 typedef LatticeSpinColourVectorF     LatticeFermionF;
 typedef LatticeSpinColourVectorD     LatticeFermionD;
+typedef LatticeSpinColourVectorD2    LatticeFermionD2;
 
 typedef LatticeSpinColourMatrix                LatticePropagator;
 typedef LatticeSpinColourMatrixF               LatticePropagatorF;
 typedef LatticeSpinColourMatrixD               LatticePropagatorD;
+typedef LatticeSpinColourMatrixD2              LatticePropagatorD2;
 
 typedef LatticeLorentzColourMatrix             LatticeGaugeField;
 typedef LatticeLorentzColourMatrixF            LatticeGaugeFieldF;
 typedef LatticeLorentzColourMatrixD            LatticeGaugeFieldD;
+typedef LatticeLorentzColourMatrixD2           LatticeGaugeFieldD2;
 
 typedef LatticeDoubleStoredColourMatrix        LatticeDoubledGaugeField;
 typedef LatticeDoubleStoredColourMatrixF       LatticeDoubledGaugeFieldF;
 typedef LatticeDoubleStoredColourMatrixD       LatticeDoubledGaugeFieldD;
+typedef LatticeDoubleStoredColourMatrixD2      LatticeDoubledGaugeFieldD2;
 
 template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >;
 
-// Uhgg... typing this hurt  ;)
-// (my keyboard got burning hot when I typed this, must be the anti-Fermion)
 typedef Lattice<vColourVector>          LatticeStaggeredFermion;    
 typedef Lattice<vColourVectorF>         LatticeStaggeredFermionF;    
 typedef Lattice<vColourVectorD>         LatticeStaggeredFermionD;    
+typedef Lattice<vColourVectorD2>        LatticeStaggeredFermionD2;    
 
 typedef Lattice<vColourMatrix>          LatticeStaggeredPropagator; 
 typedef Lattice<vColourMatrixF>         LatticeStaggeredPropagatorF; 
 typedef Lattice<vColourMatrixD>         LatticeStaggeredPropagatorD; 
+typedef Lattice<vColourMatrixD2>        LatticeStaggeredPropagatorD2; 
 
 //////////////////////////////////////////////////////////////////////////////
 // Peek and Poke named after physics attributes
@@ -476,9 +507,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
+#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuWrite);
+  autoView(f_v,f,CpuRead);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -490,12 +532,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
+#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuRead);
+  autoView(f_v,f,CpuWrite);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -507,6 +560,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/ActionBase.h

@@ -50,19 +50,23 @@ public:
   RealD refresh_us;
   void  reset_timer(void)        {
     deriv_us = S_us = refresh_us = 0.0;
-    deriv_num=0;
     deriv_norm_sum = deriv_max_sum=0.0;
     Fdt_max_sum =  Fdt_norm_sum = 0.0;
+    deriv_num=0;
   }
   void  deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) {
-    deriv_max_sum+=max;
+    if ( max > deriv_max_sum ) {
+      deriv_max_sum=max;
+    }
     deriv_norm_sum+=nrm;
-    Fdt_max_sum+=Fdt_max;
+    if ( Fdt_max > Fdt_max_sum ) {
+      Fdt_max_sum=Fdt_max;
+    }
     Fdt_norm_sum+=Fdt_nrm; deriv_num++;
   }
-  RealD deriv_max_average(void)       { return deriv_max_sum/deriv_num; };
+  RealD deriv_max_average(void)       { return deriv_max_sum; };
   RealD deriv_norm_average(void)      { return deriv_norm_sum/deriv_num; };
-  RealD Fdt_max_average(void)         { return Fdt_max_sum/deriv_num; };
+  RealD Fdt_max_average(void)         { return Fdt_max_sum; };
   RealD Fdt_norm_average(void)        { return Fdt_norm_sum/deriv_num; };
   RealD deriv_timer(void)        { return deriv_us; };
   RealD S_timer(void)            { return S_us; };

Grid/qcd/action/ActionParams.h

@@ -34,34 +34,43 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
-// These can move into a params header and be given MacroMagic serialisation
+
 struct GparityWilsonImplParams {
   Coordinate twists;
                      //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
   Coordinate dirichlet; // Blocksize of dirichlet BCs
-  GparityWilsonImplParams() : twists(Nd, 0) { dirichlet.resize(0); };
+  int  partialDirichlet;
+  GparityWilsonImplParams() : twists(Nd, 0) {
+    dirichlet.resize(0);
+    partialDirichlet=0;
+  };
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
   Coordinate dirichlet; // Blocksize of dirichlet BCs
+  int  partialDirichlet;
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
     dirichlet.resize(0);
+    partialDirichlet=0;
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
+    partialDirichlet=0;
     dirichlet.resize(0);
   }
 };
 
 struct StaggeredImplParams {
   Coordinate dirichlet; // Blocksize of dirichlet BCs
+  int  partialDirichlet;
   StaggeredImplParams()
   {
+    partialDirichlet=0;
     dirichlet.resize(0);
   };
 };

Grid/qcd/action/fermion/CloverHelpers.h

@@ -140,6 +140,7 @@ public:
     return NMAX;
   }
 
+  static int getNMAX(Lattice<iImplClover<vComplexD2>> &t, RealD R) {return getNMAX(1e-12,R);}
   static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
   static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
@@ -204,15 +205,18 @@ public:
   typedef WilsonCloverHelpers<Impl> Helpers;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
     Clover += diag_mass;
   }
 
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
+  static void InvertClover(CloverField& InvClover,
-                          CloverTriangleField& Triangle,
+                            const CloverDiagonalField& diagonal,
-                          RealD csw_t, RealD diag_mass) {
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
 
-    // Do nothing
+    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
   }
 
   // TODO: implement Cmunu for better performances with compact layout, but don't do it
@@ -237,9 +241,17 @@ public:
   template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
+  // Can this be avoided?
-    // do nothing!
+  static void IdentityTimesC(const CloverField& in, RealD c) {
-    // mass term is multiplied to exp(Clover) below
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
   }
 
   static int getNMAX(RealD prec, RealD R) {
@@ -254,175 +266,62 @@ public:
     return NMAX;
   }
 
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
-  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
 
-  	  typedef iMatrix<ComplexD,6> mat;
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
 
-  	  RealD qn[6];
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-  	  RealD qnold[6];
-  	  RealD p[5];
-  	  RealD trA2, trA3, trA4;
 
-  	  mat A2, A3, A4, A5;
+    Clover *= (1.0/diag_mass);
-  	  A2 = alpha * alpha * arg * arg;
-  	  A3 = alpha * arg * A2;
-  	  A4 = A2 * A2;
-  	  A5 = A2 * A3;
 
-  	  trA2 = toReal( trace(A2) );
+    // Taylor expansion, slow but generic
-  	  trA3 = toReal( trace(A3) );
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-  	  trA4 = toReal( trace(A4));
+    // qN = cN
-
+    // qn = cn + qn+1 X
-  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
-  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
-  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
-  	  p[3] = trA3 / 3.0;
-  	  p[4] = 0.5 * trA2;
-
-  	  qnold[0] = cN[Niter];
-  	  qnold[1] = 0.0;
-  	  qnold[2] = 0.0;
-  	  qnold[3] = 0.0;
-  	  qnold[4] = 0.0;
-  	  qnold[5] = 0.0;
-
-  	  for(int i = Niter-1; i >= 0; i--)
-  	  {
-  	   qn[0] = p[0] * qnold[5] + cN[i];
-  	   qn[1] = p[1] * qnold[5] + qnold[0];
-  	   qn[2] = p[2] * qnold[5] + qnold[1];
-  	   qn[3] = p[3] * qnold[5] + qnold[2];
-  	   qn[4] = p[4] * qnold[5] + qnold[3];
-  	   qn[5] = qnold[4];
-
-  	   qnold[0] = qn[0];
-  	   qnold[1] = qn[1];
-  	   qnold[2] = qn[2];
-  	   qnold[3] = qn[3];
-  	   qnold[4] = qn[4];
-  	   qnold[5] = qn[5];
-  	  }
-
-  	  mat unit(1.0);
-
-  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
-
-    }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Diagonal.Grid();
-    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
-
-    //
-    // Implementation completely in Daniel's layout
-    //
-
-    // Taylor expansion with Cayley-Hamilton recursion
-    // underlying Horner scheme as above
     std::vector<RealD> cn(NMAX+1);
     cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++){
+    for (int i=1; i<=NMAX; i++)
       cn[i] = cn[i-1] / RealD(i);
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
+
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
+
+    Clover = ExpClover * diag_mass;
+
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
+
+    CloverInv = ExpClover * (1.0/diag_mass);
+
   }
 
-      // Taken over from Daniel's implementation
+  static void InvertClover(CloverField& InvClover,
-      conformable(Diagonal, Triangle);
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
 
-      long lsites = grid->lSites();
+    if (fixedBoundaries)
     {
-      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
+      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
-      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-      typedef iMatrix<ComplexD,6> mat;
-
-      autoView(diagonal_v,  Diagonal,  CpuRead);
-      autoView(triangle_v,  Triangle,  CpuRead);
-      autoView(diagonalExp_v, Diagonal, CpuWrite);
-      autoView(triangleExp_v, Triangle, CpuWrite);
-
-      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
-
-    	  mat srcCloverOpUL(0.0); // upper left block
-    	  mat srcCloverOpLR(0.0); // lower right block
-    	  mat ExpCloverOp;
-
-        scalar_object_diagonal diagonal_tmp     = Zero();
-        scalar_object_diagonal diagonal_exp_tmp = Zero();
-        scalar_object_triangle triangle_tmp     = Zero();
-        scalar_object_triangle triangle_exp_tmp = Zero();
-
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-
-        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-        peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-        int block;
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-        		if (i == j){
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
     }
-        		else{
+    else
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
+    {
+      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
     }
   }
-        }
-        block = 1;
-        for(int i = 0; i < 6; i++){
-          	for(int j = 0; j < 6; j++){
-           		if (i == j){
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-           		}
-           		else{
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-           		}
-            }
-        }
-
-        // exp(Clover)
-
-        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-            	if (i == j){
-            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-            	}
-            	else if(i < j){
-            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-            	}
-           	}
-        }
-
-        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 1;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-              	if (i == j){
-              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-               	}
-               	else if(i < j){
-               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-               	}
-            }
-        }
-
-        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
-        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
-      });
-    }
-
-    Diagonal *= diag_mass;
-    Triangle *= diag_mass;
-  }
-
 
   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
     assert(0);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -225,7 +225,7 @@ public:
   RealD csw_t;
   RealD cF;
 
-  bool open_boundaries;
+  bool fixedBoundaries;
 
   CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
   CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;

Grid/qcd/action/fermion/DWFSlow.h

@@ -0,0 +1,291 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DWFSlow.h
+
+Copyright (C) 2022
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template <class Impl>
+class DWFSlowFermion : public FermionOperator<Impl>
+{
+public:
+  INHERIT_IMPL_TYPES(Impl);
+
+  ///////////////////////////////////////////////////////////////
+  // Implement the abstract base
+  ///////////////////////////////////////////////////////////////
+  GridBase *GaugeGrid(void) { return _grid4; }
+  GridBase *GaugeRedBlackGrid(void) { return _cbgrid4; }
+  GridBase *FermionGrid(void) { return _grid; }
+  GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
+
+  FermionField _tmp;
+  FermionField &tmp(void) { return _tmp; }
+
+  //////////////////////////////////////////////////////////////////
+  // override multiply; cut number routines if pass dagger argument
+  // and also make interface more uniformly consistent
+  //////////////////////////////////////////////////////////////////
+  virtual void  M(const FermionField &in, FermionField &out)
+  {
+    FermionField tmp(_grid);
+    out = (5.0 - M5) * in;
+    Dhop(in,tmp,DaggerNo);
+    out = out + tmp;
+  }
+  virtual void  Mdag(const FermionField &in, FermionField &out)
+  {
+    FermionField tmp(_grid);
+    out = (5.0 - M5) * in;
+    Dhop(in,tmp,DaggerYes);
+    out = out + tmp;
+  };
+
+  /////////////////////////////////////////////////////////
+  // half checkerboard operations 5D redblack so just site identiy
+  /////////////////////////////////////////////////////////
+  void Meooe(const FermionField &in, FermionField &out)
+  {
+    if ( in.Checkerboard() == Odd ) {
+      this->DhopEO(in,out,DaggerNo);
+    } else {
+      this->DhopOE(in,out,DaggerNo);
+    }
+  }
+  void MeooeDag(const FermionField &in, FermionField &out)
+  {
+    if ( in.Checkerboard() == Odd ) {
+      this->DhopEO(in,out,DaggerYes);
+    } else {
+      this->DhopOE(in,out,DaggerYes);
+    }
+  };
+
+  // allow override for twisted mass and clover
+  virtual void Mooee(const FermionField &in, FermionField &out)
+  {
+    out = (5.0 - M5) * in;
+  }
+  virtual void MooeeDag(const FermionField &in, FermionField &out)
+  {
+    out = (5.0 - M5) * in;
+  }
+  virtual void MooeeInv(const FermionField &in, FermionField &out)
+  {
+    out = (1.0/(5.0 - M5)) * in;
+  };
+  virtual void MooeeInvDag(const FermionField &in, FermionField &out)
+  {
+    out = (1.0/(5.0 - M5)) * in;
+  };
+
+  virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _mass,std::vector<double> twist) {} ;
+
+  ////////////////////////
+  // Derivative interface
+  ////////////////////////
+  // Interface calls an internal routine
+  void DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)  { assert(0);};
+  void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
+  void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
+
+  ///////////////////////////////////////////////////////////////
+  // non-hermitian hopping term; half cb or both
+  ///////////////////////////////////////////////////////////////
+  void Dhop(const FermionField &in, FermionField &out, int dag)
+  {
+    FermionField tmp(in.Grid());
+    Dhop5(in,out,MassField,MassField,dag );
+    for(int mu=0;mu<4;mu++){
+      DhopDirU(in,Umu[mu],Umu[mu],tmp,mu,dag );    out = out + tmp;
+    }
+  };
+  void DhopOE(const FermionField &in, FermionField &out, int dag)
+  {
+    FermionField tmp(in.Grid());
+    assert(in.Checkerboard()==Even);
+    Dhop5(in,out,MassFieldOdd,MassFieldEven,dag);
+    for(int mu=0;mu<4;mu++){
+      DhopDirU(in,UmuOdd[mu],UmuEven[mu],tmp,mu,dag );    out = out + tmp;
+    }
+  };
+  void DhopEO(const FermionField &in, FermionField &out, int dag)
+  {
+    FermionField tmp(in.Grid());
+    assert(in.Checkerboard()==Odd);
+    Dhop5(in,out, MassFieldEven,MassFieldOdd ,dag );  
+    for(int mu=0;mu<4;mu++){
+      DhopDirU(in,UmuEven[mu],UmuOdd[mu],tmp,mu,dag );    out = out + tmp;
+    }
+  };
+
+  ///////////////////////////////////////////////////////////////
+  // Multigrid assistance; force term uses too
+  ///////////////////////////////////////////////////////////////
+  void Mdir(const FermionField &in, FermionField &out, int dir, int disp){ assert(0);};
+  void MdirAll(const FermionField &in, std::vector<FermionField> &out)   { assert(0);};
+  void DhopDir(const FermionField &in, FermionField &out, int dir, int disp) { assert(0);};
+  void DhopDirAll(const FermionField &in, std::vector<FermionField> &out)    { assert(0);};
+  void DhopDirCalc(const FermionField &in, FermionField &out, int dirdisp,int gamma, int dag) { assert(0);};
+
+  void DhopDirU(const FermionField &in, const GaugeLinkField &U5e, const GaugeLinkField &U5o, FermionField &out, int mu, int dag)
+  {
+    RealD     sgn= 1.0;
+    if (dag ) sgn=-1.0;
+
+    Gamma::Algebra Gmu [] = {
+			 Gamma::Algebra::GammaX,
+			 Gamma::Algebra::GammaY,
+			 Gamma::Algebra::GammaZ,
+			 Gamma::Algebra::GammaT
+    };
+
+    //    mass is  1,1,1,1,-m has to multiply the round the world term
+    FermionField tmp (in.Grid());
+    tmp = U5e * Cshift(in,mu+1,1);
+    out = tmp - Gamma(Gmu[mu])*tmp*sgn;
+    
+    tmp = Cshift(adj(U5o)*in,mu+1,-1);
+    out = out + tmp + Gamma(Gmu[mu])*tmp*sgn;
+
+    out = -0.5*out;
+  };
+
+  void Dhop5(const FermionField &in, FermionField &out, ComplexField &massE, ComplexField &massO, int dag)
+  {
+    // Mass term.... must multiple the round world with mass = 1,1,1,1, -m
+    RealD     sgn= 1.0;
+    if (dag ) sgn=-1.0;
+
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    FermionField tmp (in.Grid());
+    tmp = massE*Cshift(in,0,1);
+    out = tmp - G5*tmp*sgn;
+    
+    tmp = Cshift(massO*in,0,-1);
+    out = out + tmp + G5*tmp*sgn;
+    out = -0.5*out;
+  };
+
+  // Constructor
+  DWFSlowFermion(GaugeField &_Umu, GridCartesian &Fgrid,
+		 GridRedBlackCartesian &Hgrid, RealD _mass, RealD _M5)
+    :
+    _grid(&Fgrid),
+    _cbgrid(&Hgrid),
+    _grid4(_Umu.Grid()),
+    Umu(Nd,&Fgrid),
+    UmuEven(Nd,&Hgrid),
+    UmuOdd(Nd,&Hgrid),
+    MassField(&Fgrid),
+    MassFieldEven(&Hgrid),
+    MassFieldOdd(&Hgrid),
+    M5(_M5),
+    mass(_mass),
+    _tmp(&Hgrid)
+    {
+      Ls=Fgrid._fdimensions[0];
+      ImportGauge(_Umu);
+
+      typedef typename FermionField::scalar_type scalar;
+
+      Lattice<iScalar<vInteger> > coor(&Fgrid);
+      LatticeCoordinate(coor, 0); // Scoor
+      ComplexField one(&Fgrid);
+      MassField =scalar(-mass);
+      one       =scalar(1.0);
+      MassField =where(coor==Integer(Ls-1),MassField,one);
+      for(int mu=0;mu<Nd;mu++){
+	pickCheckerboard(Even,UmuEven[mu],Umu[mu]);
+	pickCheckerboard(Odd ,UmuOdd[mu],Umu[mu]);
+      }
+      pickCheckerboard(Even,MassFieldEven,MassField);
+      pickCheckerboard(Odd ,MassFieldOdd,MassField);
+      
+    }
+  
+  // DoubleStore impl dependent
+  void ImportGauge(const GaugeField &_Umu4)
+  {
+    GaugeLinkField U4(_grid4);
+    for(int mu=0;mu<Nd;mu++){
+      U4 = PeekIndex<LorentzIndex>(_Umu4, mu);
+      for(int s=0;s<this->Ls;s++){
+	InsertSlice(U4,Umu[mu],s,0);
+      }
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////
+  // Data members require to support the functionality
+  ///////////////////////////////////////////////////////////////
+
+public:
+  virtual RealD Mass(void) { return mass; }
+  virtual int   isTrivialEE(void) { return 1; };
+  RealD mass;
+  RealD M5;
+  int Ls;
+
+  GridBase *_grid4;
+  GridBase *_grid;
+  GridBase *_cbgrid4;
+  GridBase *_cbgrid;
+
+  // Copy of the gauge field , with even and odd subsets
+  std::vector<GaugeLinkField> Umu;
+  std::vector<GaugeLinkField> UmuEven;
+  std::vector<GaugeLinkField> UmuOdd;
+  ComplexField MassField;
+  ComplexField MassFieldEven;
+  ComplexField MassFieldOdd;
+
+  ///////////////////////////////////////////////////////////////
+  // Conserved current utilities
+  ///////////////////////////////////////////////////////////////
+  void ContractConservedCurrent(PropagatorField &q_in_1,
+                                PropagatorField &q_in_2,
+                                PropagatorField &q_out,
+                                PropagatorField &phys_src,
+                                Current curr_type,
+                                unsigned int mu){}
+  void SeqConservedCurrent(PropagatorField &q_in,
+                           PropagatorField &q_out,
+                           PropagatorField &phys_src,
+                           Current curr_type,
+                           unsigned int mu,
+                           unsigned int tmin,
+			   unsigned int tmax,
+			   ComplexField &lattice_cmplx){}
+};
+
+typedef DWFSlowFermion<WilsonImplF> DWFSlowFermionF;
+typedef DWFSlowFermion<WilsonImplD> DWFSlowFermionD;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -47,6 +47,7 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 ////////////////////////////////////////////
 // Fermion operators / actions
 ////////////////////////////////////////////
+#include <Grid/qcd/action/fermion/DWFSlow.h>       // Slow DWF
 
 #include <Grid/qcd/action/fermion/WilsonFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(Wilson);
@@ -112,28 +113,21 @@ NAMESPACE_CHECK(DWFutils);
 // Cayley 5d
 NAMESPACE_BEGIN(Grid);
 
-typedef WilsonFermion<WilsonImplR> WilsonFermionR;
+typedef WilsonFermion<WilsonImplD2> WilsonFermionD2;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
-//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
-//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
-//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
-
-typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
 typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD;
 
-typedef WilsonFermion<WilsonTwoIndexSymmetricImplR> WilsonTwoIndexSymmetricFermionR;
 typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD;
 
-typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonTwoIndexAntiSymmetricFermionR;
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD;
 
 // Twisted mass fermion
-typedef WilsonTMFermion<WilsonImplR> WilsonTMFermionR;
+typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
@@ -141,23 +135,20 @@ typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
 template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
 
-typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
+typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2;
 typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
 typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
+typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2;
 typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
 typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
 
-typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
 typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
 typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
 
-typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
 typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
 typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
 
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
 typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
 typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
@@ -165,161 +156,108 @@ typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiS
 template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
 template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
 
-typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
+typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
 typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
 typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
 
-typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
+typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
 typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
 typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
 
-typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
 typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
 typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
 
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
 typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
 typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
 
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
 typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
 typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Domain Wall fermions
-typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
+typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2;
 
-//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
+typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2;
-//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
-//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
-
-typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
+typedef MobiusFermion<WilsonImplD2> MobiusFermionD2;
-//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
-//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
-
-typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
+typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2;
-//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
-//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
-
-typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
+typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2;
-//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
-//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
-
-typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
+typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2;
-//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
-//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
-
-// Ls vectorised
-typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
 typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
 typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD;
 
-typedef MobiusZolotarevFermion<WilsonImplR> MobiusZolotarevFermionR;
+typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2;
 typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF;
 typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD;
-typedef ShamirZolotarevFermion<WilsonImplR> ShamirZolotarevFermionR;
+typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2;
 typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF;
 typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD;
 
-typedef OverlapWilsonCayleyTanhFermion<WilsonImplR> OverlapWilsonCayleyTanhFermionR;
+typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD;
-typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplR> OverlapWilsonCayleyZolotarevFermionR;
+typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD;
 
 // Continued fraction
-typedef OverlapWilsonContFracTanhFermion<WilsonImplR> OverlapWilsonContFracTanhFermionR;
+typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD;
-typedef OverlapWilsonContFracZolotarevFermion<WilsonImplR> OverlapWilsonContFracZolotarevFermionR;
+typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD;
 
 // Partial fraction
-typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplR> OverlapWilsonPartialFractionTanhFermionR;
+typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD;
 
-typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplR> OverlapWilsonPartialFractionZolotarevFermionR;
+typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD;
 
 // Gparity cases; partial list until tested
-typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 
-//typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
-//typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
-//typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
-
-typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
+typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2;
-//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
-//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
-
-typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
+typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2;
-//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
-//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
-
-typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
+typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2;
-//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
-//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
-
-typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
+typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2;
-//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
-//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
-
-typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 
-//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
-//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
-//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
-
-typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
 typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
 
-typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR;
 typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
 typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
 
-typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
 typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
 typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
 

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -32,17 +32,218 @@ 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
+// optimised versions supporting half precision too??? Deprecate
 /////////////////////////////////////////////////////////////////////////////////////////////
 
-template<class _HCspinor,class _Hspinor,class _Spinor, class projector,typename SFINAE = void >
-class WilsonCompressorTemplate;
-
 
+//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< _HCspinor, _Hspinor, _Spinor, projector,
+class WilsonCompressorTemplate  : public FaceGatherDWFMixedBCs
-				typename std::enable_if<std::is_same<_HCspinor,_Hspinor>::value>::type >
+//  : public FaceGatherSimple
 {
 public:
   
@@ -79,18 +280,19 @@ public:
   /*****************************************************/
   /* Exchange includes precision change if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Exchange(SiteHalfSpinor *mp,
+  accelerator_inline void Exchange(SiteHalfSpinor &mp0,
-				   const SiteHalfSpinor * __restrict__ vp0,
+				   SiteHalfSpinor &mp1,
-				   const SiteHalfSpinor * __restrict__ vp1,
+				   const SiteHalfSpinor & vp0,
-				   Integer type,Integer o) const {
+				   const SiteHalfSpinor & vp1,
+				   Integer type) const {
 #ifdef GRID_SIMT
-    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+    exchangeSIMT(mp0,mp1,vp0,vp1,type);
 #else
     SiteHalfSpinor tmp1;
     SiteHalfSpinor tmp2;
-    exchange(tmp1,tmp2,vp0[o],vp1[o],type);
+    exchange(tmp1,tmp2,vp0,vp1,type);
-    vstream(mp[2*o  ],tmp1);
+    vstream(mp0,tmp1);
-    vstream(mp[2*o+1],tmp2);
+    vstream(mp1,tmp2);
 #endif
   }
   
@@ -98,153 +300,61 @@ public:
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
+  accelerator_inline void Decompress(SiteHalfSpinor &out,
-				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
+				     SiteHalfSpinor &in) const {    
-    assert(0);
+    out = in;
   }
 
   /*****************************************************/
   /* Compress Exchange                                 */
   /*****************************************************/
-  accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
+  accelerator_inline void CompressExchange(SiteHalfSpinor &out0,
-					   SiteHalfSpinor * __restrict__ out1,
+					   SiteHalfSpinor &out1,
-					   const SiteSpinor * __restrict__ in,
+					   const SiteSpinor &in0,
-					   Integer j,Integer k, Integer m,Integer type) const
+					   const SiteSpinor &in1,
+					   Integer type) const
   {
 #ifdef GRID_SIMT
     typedef SiteSpinor vobj;
     typedef SiteHalfSpinor hvobj;
-    typedef decltype(coalescedRead(*in))    sobj;
+    typedef decltype(coalescedRead(in0))    sobj;
-    typedef decltype(coalescedRead(*out0)) hsobj;
+    typedef decltype(coalescedRead(out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
+    const vobj *vp0 = &in0;
-    const vobj *vp1 = &in[m];
+    const vobj *vp1 = &in1;
     const vobj *vp = (lane&mask) ? vp1:vp0;
     auto sa = coalescedRead(*vp,j0);
     auto sb = coalescedRead(*vp,j1);
     hsobj psa, psb;
     projector::Proj(psa,sa,mu,dag);
     projector::Proj(psb,sb,mu,dag);
-    coalescedWrite(out0[j],psa);
+    coalescedWrite(out0,psa);
-    coalescedWrite(out1[j],psb);
+    coalescedWrite(out1,psb);
 #else
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
-    projector::Proj(temp1,in[k],mu,dag);
+    projector::Proj(temp1,in0,mu,dag);
-    projector::Proj(temp2,in[m],mu,dag);
+    projector::Proj(temp2,in1,mu,dag);
     exchange(temp3,temp4,temp1,temp2,type);
-    vstream(out0[j],temp3);
+    vstream(out0,temp3);
-    vstream(out1[j],temp4);
+    vstream(out1,temp4);
 #endif
   }
 
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return false; }
+  accelerator_inline bool DecompressionStep(void) const {
+    return false;
+  }
 
 };
 
-#if 0
-template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
-class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
-				typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
-{
-public:
-  
-  int mu,dag;  
-
-  void Point(int p) { mu=p; };
-
-  WilsonCompressorTemplate(int _dag=0){
-    dag = _dag;
-  }
-
-  typedef _Spinor         SiteSpinor;
-  typedef _Hspinor     SiteHalfSpinor;
-  typedef _HCspinor SiteHalfCommSpinor;
-  typedef typename SiteHalfCommSpinor::vector_type vComplexLow;
-  typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
-  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
-
-  accelerator_inline int CommDatumSize(void) const {
-    return sizeof(SiteHalfCommSpinor);
-  }
-
-  /*****************************************************/
-  /* Compress includes precision change if mpi data is not same */
-  /*****************************************************/
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-    SiteHalfSpinor hsp;
-    SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
-    projector::Proj(hsp,in,mu,dag);
-    precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
-  }
-  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
-#ifdef GRID_SIMT
-    typedef decltype(coalescedRead(buf)) sobj;
-    sobj sp;
-    auto sin = coalescedRead(in);
-    projector::Proj(sp,sin,mu,dag);
-    coalescedWrite(buf,sp);
-#else
-    projector::Proj(buf,in,mu,dag);
-#endif
-  }
-
-  /*****************************************************/
-  /* Exchange includes precision change if mpi data is not same */
-  /*****************************************************/
-  accelerator_inline void Exchange(SiteHalfSpinor *mp,
-                       SiteHalfSpinor *vp0,
-                       SiteHalfSpinor *vp1,
-		       Integer type,Integer o) const {
-    SiteHalfSpinor vt0,vt1;
-    SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
-    SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
-    precisionChange((vComplexHigh *)&vt0,(vComplexLow *)&vpp0[o],Nw);
-    precisionChange((vComplexHigh *)&vt1,(vComplexLow *)&vpp1[o],Nw);
-    exchange(mp[2*o],mp[2*o+1],vt0,vt1,type);
-  }
-
-  /*****************************************************/
-  /* Have a decompression step if mpi data is not same */
-  /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
-    SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
-    precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
-  }
-
-  /*****************************************************/
-  /* Compress Exchange                                 */
-  /*****************************************************/
-  accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
-			       SiteHalfSpinor *out1,
-			       const SiteSpinor *in,
-			       Integer j,Integer k, Integer m,Integer type) const {
-    SiteHalfSpinor temp1, temp2,temp3,temp4;
-    SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
-    SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
-    projector::Proj(temp1,in[k],mu,dag);
-    projector::Proj(temp2,in[m],mu,dag);
-    exchange(temp3,temp4,temp1,temp2,type);
-    precisionChange((vComplexLow *)&hout0[j],(vComplexHigh *)&temp3,Nw);
-    precisionChange((vComplexLow *)&hout1[j],(vComplexHigh *)&temp4,Nw);
-  }
-
-  /*****************************************************/
-  /* Pass the info to the stencil */
-  /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const { return true; }
-
-};
-#endif
-
 #define DECLARE_PROJ(Projector,Compressor,spProj)			\
   class Projector {							\
   public:								\
@@ -374,24 +484,26 @@ public:
 
     int dag = compress.dag;
     int face_idx=0;
+#define vet_same_node(a,b) \
+      { auto tmp = b;  }
     if ( dag ) { 
-      assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
+      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
-      assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
+      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
-      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
+      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
-      assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
+      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
-      assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
+      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
-      assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
+      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
-      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
+      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
-      assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
+      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
     } else {
-      assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
+      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
-      assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
+      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
-      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
+      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
-      assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
+      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
-      assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
+      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
-      assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
+      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
-      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
+      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
-      assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
+      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);

Grid/qcd/action/fermion/WilsonImpl.h

@@ -242,18 +242,12 @@ public:
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR;  // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
-
+typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffReal > WilsonImplD2;  // Double
-//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
-//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
-//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
 
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
-
+typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffComplex > ZWilsonImplD2; // Double
-//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
-//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
-//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
 
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float

Grid/qcd/action/fermion/WilsonKernels.h

@@ -52,13 +52,6 @@ public:
   typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;   
 public:
 
-#ifdef GRID_SYCL
-#define SYCL_HACK
-#endif  
-#ifdef SYCL_HACK
-  static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
-			       int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
-#endif
   
   static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,

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

@@ -905,88 +905,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 #undef TopRowWithSource
 
 
-
-#if 0
-template<class Impl>
-void CayleyFermion5D<Impl>::MooeeInternalCompute(int dag, int inv,
-						 Vector<iSinglet<Simd> > & Matp,
-						 Vector<iSinglet<Simd> > & Matm)
-{
-  int Ls=this->Ls;
-
-  GridBase *grid = this->FermionRedBlackGrid();
-  int LLs = grid->_rdimensions[0];
-
-  if ( LLs == Ls ) {
-    return; // Not vectorised in 5th direction
-  }
-
-  Eigen::MatrixXcd Pplus  = Eigen::MatrixXcd::Zero(Ls,Ls);
-  Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
-  
-  for(int s=0;s<Ls;s++){
-    Pplus(s,s) = bee[s];
-    Pminus(s,s)= bee[s];
-  }
-  
-  for(int s=0;s<Ls-1;s++){
-    Pminus(s,s+1) = -cee[s];
-  }
-  
-  for(int s=0;s<Ls-1;s++){
-    Pplus(s+1,s) = -cee[s+1];
-  }
-  Pplus (0,Ls-1) = mass*cee[0];
-  Pminus(Ls-1,0) = mass*cee[Ls-1];
-  
-  Eigen::MatrixXcd PplusMat ;
-  Eigen::MatrixXcd PminusMat;
-  
-  if ( inv ) {
-    PplusMat =Pplus.inverse();
-    PminusMat=Pminus.inverse();
-  } else { 
-    PplusMat =Pplus;
-    PminusMat=Pminus;
-  }
-  
-  if(dag){
-    PplusMat.adjointInPlace();
-    PminusMat.adjointInPlace();
-  }
-  
-  typedef typename SiteHalfSpinor::scalar_type scalar_type;
-  const int Nsimd=Simd::Nsimd();
-  Matp.resize(Ls*LLs);
-  Matm.resize(Ls*LLs);
-
-  for(int s2=0;s2<Ls;s2++){
-    for(int s1=0;s1<LLs;s1++){
-      int istride = LLs;
-      int ostride = 1;
-      Simd Vp;
-      Simd Vm;
-      scalar_type *sp = (scalar_type *)&Vp;
-      scalar_type *sm = (scalar_type *)&Vm;
-      for(int l=0;l<Nsimd;l++){
-	if ( switcheroo<Coeff_t>::iscomplex() ) {
-	  sp[l] = PplusMat (l*istride+s1*ostride,s2);
-	  sm[l] = PminusMat(l*istride+s1*ostride,s2);
-	} else { 
-	  // if real
-	  scalar_type tmp;
-	  tmp = PplusMat (l*istride+s1*ostride,s2);
-	  sp[l] = scalar_type(tmp.real(),tmp.real());
-	  tmp = PminusMat(l*istride+s1*ostride,s2);
-	  sm[l] = scalar_type(tmp.real(),tmp.real());
-	}
-      }
-      Matp[LLs*s2+s1] = Vp;
-      Matm[LLs*s2+s1] = Vm;
-    }}
-}
-#endif
-
 NAMESPACE_END(Grid);
 
 

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

@@ -48,7 +48,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , csw_r(_csw_r)
   , csw_t(_csw_t)
   , cF(_cF)
-  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
   , Diagonal(&Fgrid),        Triangle(&Fgrid)
   , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
   , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
@@ -67,7 +67,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries) {
+  if (fixedBoundaries) {
     this->BoundaryMaskEven.Checkerboard() = Even;
     this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
@@ -77,31 +77,31 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->open_boundaries) ApplyBoundaryMask(out);
+  if(this->fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
-  if(this->open_boundaries) {
+  if(this->fixedBoundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
@@ -112,7 +112,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in,
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -121,19 +121,19 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& i
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -147,7 +147,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -166,7 +166,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -186,7 +186,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!open_boundaries); // TODO check for changes required for open bc
+  assert(!fixedBoundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -305,6 +305,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   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();
@@ -324,24 +325,27 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  // Handle mass term based on clover policy
+
-  CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
+  // 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, this->diag_mass);
 
   // Convert the data layout of the clover term
-  double t4 = usecond();
+  double t5 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Exponentiate the clover (nothing happens in case of the standard clover)
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
-  double t5 = usecond();
-  CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
-
-  // Possible modify the boundary values
   double t6 = usecond();
-  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
+  // 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();
-  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
 
   // Fill the remaining clover fields
   double t8 = usecond();
@@ -362,10 +366,10 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   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 << "convert =                    " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "exponentiation =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "boundaries =                 " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "inversions =                 " << (t8 - t7) / 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;
 }

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

@@ -63,6 +63,10 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   _tmp(&FiveDimRedBlackGrid),
   Dirichlet(0)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
+  
   // some assertions
   assert(FiveDimGrid._ndimension==5);
   assert(FourDimGrid._ndimension==4);
@@ -96,6 +100,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
     Coordinate block = p.dirichlet;
     if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
       Dirichlet = 1;
+      std::cout << GridLogMessage << " WilsonFermion: non-trivial Dirichlet condition "<< block << std::endl;
+      std::cout << GridLogMessage << " WilsonFermion: partial Dirichlet "<< p.partialDirichlet << std::endl;
       Block = block;
     }
   } else {
@@ -137,9 +143,6 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   StencilEven.BuildSurfaceList(LLs,vol4);
    StencilOdd.BuildSurfaceList(LLs,vol4);
 
-   //  std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
-   //                       <<" " << StencilEven.surface_list.size()<<std::endl;
-
 }
 
 template<class Impl>
@@ -148,12 +151,29 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
   if ( Dirichlet ) {
-    std::cout << GridLogDslash << " Dirichlet BCs 5d " <<Block<<std::endl;
+
+    if ( this->Params.partialDirichlet ) {
+      std::cout << GridLogMessage << " partialDirichlet BCs " <<Block<<std::endl;
+    } else {
+      std::cout << GridLogMessage << " FULL Dirichlet BCs " <<Block<<std::endl;
+    }
+    
+    std:: cout << GridLogMessage << "Checking block size multiple of rank boundaries for Dirichlet"<<std::endl;
+    for(int d=0;d<Nd;d++) {
+      int GaugeBlock = Block[d+1];
+      int ldim=GaugeGrid()->LocalDimensions()[d];
+      if (GaugeBlock) assert( (GaugeBlock%ldim)==0);
+    }
+
+    if (!this->Params.partialDirichlet) {
+      std::cout << GridLogMessage << " Dirichlet filtering gauge field BCs block " <<Block<<std::endl;
       Coordinate GaugeBlock(Nd);
       for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
-    std::cout << GridLogDslash << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
       DirichletFilter<GaugeField> Filter(GaugeBlock);
       Filter.applyFilter(HUmu);
+    } else {
+      std::cout << GridLogMessage << " Dirichlet "<< Dirichlet << " NOT filtered gauge field" <<std::endl;
+    }
   }
   Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
   pickCheckerboard(Even,UmuEven,Umu);

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

@@ -60,6 +60,9 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
       _tmp(&Hgrid),
       anisotropyCoeff(anis)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
   // Allocate the required comms buffer
   ImportGauge(_Umu);
   if  (anisotropyCoeff.isAnisotropic){

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

@@ -433,11 +433,23 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
     });									
 
 #define ASM_CALL(A)							\
-  thread_for( ss, Nsite, {						\
+  thread_for( sss, Nsite, {						\
+    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);		\
   });
+#define ASM_CALL_SLICE(A)						\
+  auto grid = in.Grid() ;						\
+  int nt = grid->LocalDimensions()[4];					\
+  int nxyz = Nsite/nt ;							\
+  for(int t=0;t<nt;t++){						\
+  thread_for( sss, nxyz, {						\
+    int ss = t*nxyz+sss;						\
+    int sU = ss;							\
+    int sF = ss*Ls;							\
+    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,
@@ -490,6 +502,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
+     acceleratorFenceComputeStream();
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
@@ -497,11 +510,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
+     acceleratorFenceComputeStream();
    }
    assert(0 && " Kernel optimisation case not covered ");
   }

Grid/qcd/action/fermion/instantiation/WilsonImplD2/CayleyFermion5DInstantiationWilsonImplD2.cc

@@ -0,0 +1 @@
+../CayleyFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/ContinuedFractionFermion5DInstantiationWilsonImplD2.cc

@@ -0,0 +1 @@
+../ContinuedFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/DomainWallEOFAFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/MobiusEOFAFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/PartialFractionFermion5DInstantiationWilsonImplD2.cc

@@ -0,0 +1 @@
+../PartialFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonCloverFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonFermion5DInstantiationWilsonImplD2.cc

@@ -0,0 +1 @@
+../WilsonFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonKernelsInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonTMFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/impl.h

@@ -0,0 +1 @@
+#define IMPLEMENTATION WilsonImplD2

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/CayleyFermion5DInstantiationZWilsonImplD2.cc

@@ -0,0 +1 @@
+../CayleyFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/ContinuedFractionFermion5DInstantiationZWilsonImplD2.cc

@@ -0,0 +1 @@
+../ContinuedFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/DomainWallEOFAFermionInstantiationZWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/MobiusEOFAFermionInstantiationZWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/PartialFractionFermion5DInstantiationZWilsonImplD2.cc

@@ -0,0 +1 @@
+../PartialFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/WilsonFermion5DInstantiationZWilsonImplD2.cc

@@ -0,0 +1 @@
+../WilsonFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/WilsonKernelsInstantiationZWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/impl.h

@@ -0,0 +1 @@
+#define IMPLEMENTATION ZWilsonImplD2

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

@@ -9,6 +9,7 @@ STAG5_IMPL_LIST=""
 WILSON_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
+	   WilsonImplD2 \
 	   WilsonAdjImplF \
 	   WilsonAdjImplD \
 	   WilsonTwoIndexSymmetricImplF \
@@ -25,8 +26,9 @@ COMPACT_WILSON_IMPL_LIST=" \
 DWF_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
+	   WilsonImplD2 \
 	   ZWilsonImplF \
-	   ZWilsonImplD "
+	   ZWilsonImplD2 "
 
 GDWF_IMPL_LIST=" \
 	   GparityWilsonImplF \

Grid/qcd/action/gauge/Photon.h

@@ -49,7 +49,7 @@ NAMESPACE_BEGIN(Grid);
     
     typedef Lattice<SiteLink>  LinkField;
     typedef Lattice<SiteField> Field;
-    typedef Field              ComplexField;
+    typedef LinkField          ComplexField;
   };
   
   typedef QedGImpl<vComplex> QedGImplR;

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h

@@ -59,7 +59,7 @@ NAMESPACE_BEGIN(Grid);
 
       typedef RationalActionParams Params;
       Params param;
-
+      RealD  RefreshAction;
       //For action evaluation
       MultiShiftFunction ApproxPowerAction   ;  //rational approx for X^{1/inv_pow}
       MultiShiftFunction ApproxNegPowerAction;  //rational approx for X^{-1/inv_pow}
@@ -115,6 +115,56 @@ NAMESPACE_BEGIN(Grid);
       
     public:
 
+      // allow non-uniform tolerances 
+      void SetTolerances(std::vector<RealD> action_tolerance,std::vector<RealD> md_tolerance)
+      {
+	assert(action_tolerance.size()==ApproxPowerAction.tolerances.size());
+	assert(    md_tolerance.size()==ApproxPowerMD.tolerances.size());
+	
+	// Fix up the tolerances
+	for(int i=0;i<ApproxPowerAction.tolerances.size();i++){
+	  ApproxPowerAction.tolerances[i]       = action_tolerance[i];
+	  ApproxNegPowerAction.tolerances[i]    = action_tolerance[i];
+	  ApproxHalfPowerAction.tolerances[i]   = action_tolerance[i];
+	  ApproxNegHalfPowerAction.tolerances[i]= action_tolerance[i];
+	}
+	for(int i=0;i<ApproxPowerMD.tolerances.size();i++){
+	  ApproxPowerMD.tolerances[i]       = md_tolerance[i];
+	  ApproxNegPowerMD.tolerances[i]    = md_tolerance[i];
+	  ApproxHalfPowerMD.tolerances[i]   = md_tolerance[i];
+	  ApproxNegHalfPowerMD.tolerances[i]= md_tolerance[i];
+	}
+	
+	// Print out - could deprecate
+	for(int i=0;i<ApproxPowerMD.tolerances.size();i++) {
+	  std::cout<<GridLogMessage << " ApproxPowerMD shift["<<i<<"] "
+		   <<" pole    "<<ApproxPowerMD.poles[i]
+		   <<" residue "<<ApproxPowerMD.residues[i]
+		   <<" tol     "<<ApproxPowerMD.tolerances[i]<<std::endl;
+	}
+	/*
+	  for(int i=0;i<ApproxNegPowerMD.tolerances.size();i++) {
+	  std::cout<<GridLogMessage << " ApproxNegPowerMD shift["<<i<<"] "
+		   <<" pole    "<<ApproxNegPowerMD.poles[i]
+		   <<" residue "<<ApproxNegPowerMD.residues[i]
+		   <<" tol     "<<ApproxNegPowerMD.tolerances[i]<<std::endl;
+	}
+	for(int i=0;i<ApproxHalfPowerMD.tolerances.size();i++) {
+	  std::cout<<GridLogMessage << " ApproxHalfPowerMD shift["<<i<<"] "
+		   <<" pole    "<<ApproxHalfPowerMD.poles[i]
+		   <<" residue "<<ApproxHalfPowerMD.residues[i]
+		   <<" tol     "<<ApproxHalfPowerMD.tolerances[i]<<std::endl;
+	}
+	for(int i=0;i<ApproxNegHalfPowerMD.tolerances.size();i++) {
+	  std::cout<<GridLogMessage << " ApproxNegHalfPowerMD shift["<<i<<"] "
+		   <<" pole    "<<ApproxNegHalfPowerMD.poles[i]
+		   <<" residue "<<ApproxNegHalfPowerMD.residues[i]
+		   <<" tol     "<<ApproxNegHalfPowerMD.tolerances[i]<<std::endl;
+	}
+	*/
+	
+      }
+      
       GeneralEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
 						     FermionOperator<Impl>  &_DenOp, 
 						     const Params & p
@@ -149,6 +199,11 @@ NAMESPACE_BEGIN(Grid);
 	    ApproxNegHalfPowerMD.tolerances[i] = ApproxHalfPowerMD.tolerances[i] = param.md_tolerance;
 	}
 
+	std::vector<RealD> action_tolerance(ApproxHalfPowerAction.tolerances.size(),param.action_tolerance);
+	std::vector<RealD> md_tolerance    (ApproxHalfPowerMD.tolerances.size(),param.md_tolerance);
+
+	SetTolerances(action_tolerance, md_tolerance);
+	
 	std::cout<<GridLogMessage << action_name() << " initialize: complete" << std::endl;
       };
 
@@ -217,12 +272,19 @@ NAMESPACE_BEGIN(Grid);
 	assert(NumOp.ConstEE() == 1);
 	assert(DenOp.ConstEE() == 1);
 	PhiEven = Zero();
-	std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
+
+	RefreshAction = norm2( etaOdd );
+        std::cout<<GridLogMessage << action_name() << " refresh: action is " << RefreshAction << std::endl;
       };
 
       //////////////////////////////////////////////////////
       // S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi       
       //////////////////////////////////////////////////////
+      virtual RealD Sinitial(const GaugeField &U) {
+	std::cout << GridLogMessage << "Returning stored two flavour refresh action "<<RefreshAction<<std::endl;
+	return RefreshAction;
+      }
+
       virtual RealD S(const GaugeField &U) {
 	std::cout<<GridLogMessage << action_name() << " compute action: starting" << std::endl;
 	ImportGauge(U);

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h

@@ -29,6 +29,8 @@
 #ifndef QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
 #define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
 
+#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h>
+
 NAMESPACE_BEGIN(Grid);
 
     /////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -36,15 +38,19 @@ NAMESPACE_BEGIN(Grid);
     // cf. GeneralEvenOddRational.h for details
     /////////////////////////////////////////////////////////////////////////////////////////////////////////////
       
-    template<class ImplD, class ImplF>
+    template<class ImplD, class ImplF, class ImplD2>
     class GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<ImplD> {
     private:
+      typedef typename ImplD2::FermionField FermionFieldD2;
       typedef typename ImplD::FermionField FermionFieldD;
       typedef typename ImplF::FermionField FermionFieldF;
 
       FermionOperator<ImplD> & NumOpD;
       FermionOperator<ImplD> & DenOpD;
 
+      FermionOperator<ImplD2> & NumOpD2;
+      FermionOperator<ImplD2> & DenOpD2;
+     
       FermionOperator<ImplF> & NumOpF;
       FermionOperator<ImplF> & DenOpF;
 
@@ -53,37 +59,71 @@ NAMESPACE_BEGIN(Grid);
 
       //Allow derived classes to override the multishift CG
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
-	SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
+#if 0
+	SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOpD : DenOpD);
+	ConjugateGradientMultiShift<FermionFieldD> msCG(MaxIter, approx);
+	msCG(schurOp,in, out);
+#else
+	SchurDifferentiableOperator<ImplD2> schurOpD2(numerator ? NumOpD2 : DenOpD2);
 	SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
+	FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
+	FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
 
-	ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	// Action better with higher precision?
-	msCG(schurOpD, in, out);
+	ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	precisionChange(inD2,in);
+	std::cout << "msCG single solve "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
+	msCG(schurOpD2, inD2, outD2);
+	precisionChange(out,outD2);
+#endif
       }
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
-	SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
+	SchurDifferentiableOperator<ImplD2> schurOpD2(numerator ? NumOpD2 : DenOpD2);
 	SchurDifferentiableOperator<ImplF>  schurOpF (numerator ? NumOpF  : DenOpF);
 
-	ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
-	msCG(schurOpD, in, out_elems, out);
+	FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
+	std::vector<FermionFieldD2> out_elemsD2(out_elems.size(),NumOpD2.FermionRedBlackGrid());
+	ConjugateGradientMultiShiftMixedPrecCleanup<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	precisionChange(inD2,in);
+	std::cout << "msCG in "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
+	msCG(schurOpD2, inD2, out_elemsD2, outD2);
+	precisionChange(out,outD2);
+	for(int i=0;i<out_elems.size();i++){
+	  precisionChange(out_elems[i],out_elemsD2[i]);
+	}
       }
       //Allow derived classes to override the gauge import
       virtual void ImportGauge(const typename ImplD::GaugeField &Ud){
+
 	typename ImplF::GaugeField Uf(NumOpF.GaugeGrid());
+	typename ImplD2::GaugeField Ud2(NumOpD2.GaugeGrid());
 	precisionChange(Uf, Ud);
+	precisionChange(Ud2, Ud);
+
+	std::cout << "Importing "<<norm2(Ud)<<" "<< norm2(Uf)<<" " << norm2(Ud2)<<std::endl;
 	
 	NumOpD.ImportGauge(Ud);
 	DenOpD.ImportGauge(Ud);
 
 	NumOpF.ImportGauge(Uf);
 	DenOpF.ImportGauge(Uf);
+
+	NumOpD2.ImportGauge(Ud2);
+	DenOpD2.ImportGauge(Ud2);
       }
       
     public:
       GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction(FermionOperator<ImplD>  &_NumOpD, FermionOperator<ImplD>  &_DenOpD, 
 							      FermionOperator<ImplF>  &_NumOpF, FermionOperator<ImplF>  &_DenOpF, 
+							      FermionOperator<ImplD2>  &_NumOpD2, FermionOperator<ImplD2>  &_DenOpD2, 
 							      const RationalActionParams & p, Integer _ReliableUpdateFreq
 							      ) : GeneralEvenOddRatioRationalPseudoFermionAction<ImplD>(_NumOpD, _DenOpD, p),
-								  ReliableUpdateFreq(_ReliableUpdateFreq), NumOpD(_NumOpD), DenOpD(_DenOpD), NumOpF(_NumOpF), DenOpF(_DenOpF){}
+								  ReliableUpdateFreq(_ReliableUpdateFreq),
+								  NumOpD(_NumOpD), DenOpD(_DenOpD),
+								  NumOpF(_NumOpF), DenOpF(_DenOpF),
+								  NumOpD2(_NumOpD2), DenOpD2(_DenOpD2)
+      {}
       
       virtual std::string action_name(){return "GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction";}
     };

Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h

@@ -67,8 +67,9 @@ NAMESPACE_BEGIN(Grid);
       virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
     };
 
-    template<class Impl,class ImplF>
+    template<class Impl,class ImplF,class ImplD2>
-    class OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction : public GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF> {
+    class OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction
+      : public GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF,ImplD2> {
     public:
       typedef OneFlavourRationalParams Params;
     private:
@@ -90,9 +91,11 @@ NAMESPACE_BEGIN(Grid);
 								 FermionOperator<Impl>  &_DenOp, 
 								 FermionOperator<ImplF>  &_NumOpF, 
 								 FermionOperator<ImplF>  &_DenOpF, 
+								 FermionOperator<ImplD2>  &_NumOpD2, 
+								 FermionOperator<ImplD2>  &_DenOpD2, 
 								 const Params & p, Integer ReliableUpdateFreq
 							) : 
-	GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF>(_NumOp, _DenOp,_NumOpF, _DenOpF, transcribe(p),ReliableUpdateFreq){}
+	GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF,ImplD2>(_NumOp, _DenOp,_NumOpF, _DenOpF,_NumOpD2, _DenOpD2, transcribe(p),ReliableUpdateFreq){}
 
       virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
     };

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -50,6 +50,8 @@ NAMESPACE_BEGIN(Grid);
       FermionField PhiOdd;   // the pseudo fermion field for this trajectory
       FermionField PhiEven;  // the pseudo fermion field for this trajectory
 
+      RealD RefreshAction;
+      
     public:
       TwoFlavourEvenOddRatioPseudoFermionAction(FermionOperator<Impl>  &_NumOp, 
                                                 FermionOperator<Impl>  &_DenOp, 
@@ -110,33 +112,60 @@ NAMESPACE_BEGIN(Grid);
         // NumOp == V
         // DenOp == M
         //
+    AUDIT();
         FermionField etaOdd (NumOp.FermionRedBlackGrid());
         FermionField etaEven(NumOp.FermionRedBlackGrid());
         FermionField tmp    (NumOp.FermionRedBlackGrid());
 
+    AUDIT();
         pickCheckerboard(Even,etaEven,eta);
+    AUDIT();
         pickCheckerboard(Odd,etaOdd,eta);
 
+    AUDIT();
         NumOp.ImportGauge(U);
+    AUDIT();
         DenOp.ImportGauge(U);
+	std::cout << " TwoFlavourRefresh:  Imported gauge "<<std::endl;
+    AUDIT();
 
         SchurDifferentiableOperator<Impl> Mpc(DenOp);
+    AUDIT();
         SchurDifferentiableOperator<Impl> Vpc(NumOp);
+    AUDIT();
 
+	std::cout << " TwoFlavourRefresh: Diff ops "<<std::endl;
+    AUDIT();
         // Odd det factors
         Mpc.MpcDag(etaOdd,PhiOdd);
+    AUDIT();
+	std::cout << " TwoFlavourRefresh: MpcDag "<<std::endl;
         tmp=Zero();
+    AUDIT();
+	std::cout << " TwoFlavourRefresh: Zero() guess "<<std::endl;
+    AUDIT();
         HeatbathSolver(Vpc,PhiOdd,tmp);
+    AUDIT();
+	std::cout << " TwoFlavourRefresh: Heatbath solver "<<std::endl;
         Vpc.Mpc(tmp,PhiOdd);            
+	std::cout << " TwoFlavourRefresh: Mpc "<<std::endl;
 
         // Even det factors
         DenOp.MooeeDag(etaEven,tmp);
         NumOp.MooeeInvDag(tmp,PhiEven);
+	std::cout << " TwoFlavourRefresh: Mee "<<std::endl;
+
+	RefreshAction = norm2(etaEven)+norm2(etaOdd);
+	std::cout << " refresh " <<action_name()<< " action "<<RefreshAction<<std::endl;
       };
 
       //////////////////////////////////////////////////////
       // S = phi^dag V (Mdag M)^-1 Vdag phi
       //////////////////////////////////////////////////////
+      virtual RealD Sinitial(const GaugeField &U) {
+	std::cout << GridLogMessage << "Returning stored two flavour refresh action "<<RefreshAction<<std::endl;
+	return RefreshAction;
+      }
       virtual RealD S(const GaugeField &U) {
 
         NumOp.ImportGauge(U);

Grid/qcd/action/scalar/ScalarInteractionAction.h

@@ -47,7 +47,7 @@ private:
   const unsigned int N = Impl::Group::Dimension;
 
   typedef typename Field::vector_object vobj;
-  typedef CartesianStencil<vobj, vobj,int> Stencil;
+  typedef CartesianStencil<vobj, vobj,DefaultImplParams> Stencil;
 
   SimpleCompressor<vobj> compressor;
   int npoint = 2 * Ndim;
@@ -82,7 +82,7 @@ public:
   virtual RealD S(const Field &p)
   {
     assert(p.Grid()->Nd() == Ndim);
-    static Stencil phiStencil(p.Grid(), npoint, 0, directions, displacements,0);
+    static Stencil phiStencil(p.Grid(), npoint, 0, directions, displacements);
     phiStencil.HaloExchange(p, compressor);
     Field action(p.Grid()), pshift(p.Grid()), phisquared(p.Grid());
     phisquared = p * p;
@@ -133,7 +133,7 @@ public:
     double interm_t = usecond();
 
     // move this outside
-    static Stencil phiStencil(p.Grid(), npoint, 0, directions, displacements,0);
+    static Stencil phiStencil(p.Grid(), npoint, 0, directions, displacements);
 
     phiStencil.HaloExchange(p, compressor);
     double halo_t = usecond();

Grid/qcd/hmc/HMC.h

@@ -143,6 +143,7 @@ private:
     GridBase *Grid = U.Grid();
 
     if(Params.PerformRandomShift){
+#if 0
       //////////////////////////////////////////////////////////////////////////////////////////////////////
       // Mainly for DDHMC perform a random translation of U modulo volume
       //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -167,11 +168,11 @@ private:
 	//shift all fields together in a way that respects the gauge BCs
 	for(int mu=0; mu < Grid->Nd(); mu++)
 	  Umu[mu] = FieldImplementation::CshiftLink(Umu[mu],d,shift);
-      }
 
 	for(int mu=0;mu<Grid->Nd();mu++) PokeIndex<LorentzIndex>(U,Umu[mu],mu);
-      
+      }
       std::cout << GridLogMessage << "--------------------------------------------------\n";
+#endif	
     }
 
     TheIntegrator.reset_timer();
@@ -189,7 +190,7 @@ private:
     //////////////////////////////////////////////////////////////////////////////////////////////////////
     std::cout << GridLogMessage << "--------------------------------------------------\n";
     std::cout << GridLogMessage << "Compute initial action";
-    RealD H0 = TheIntegrator.S(U);  
+    RealD H0 = TheIntegrator.Sinitial(U);  
     std::cout << GridLogMessage << "--------------------------------------------------\n";
 
     std::streamsize current_precision = std::cout.precision();

Grid/qcd/hmc/integrators/Integrator.h

@@ -132,10 +132,17 @@ protected:
 
       Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
+
+      std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] before"<<std::endl;
+      AUDIT();
+      
       as[level].actions.at(a)->deriv_timer_start();
       as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
       as[level].actions.at(a)->deriv_timer_stop();
 
+      std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] after"<<std::endl;
+      AUDIT();
+
       std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
       auto name = as[level].actions.at(a)->action_name();
       if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
@@ -145,7 +152,7 @@ protected:
 
       //      DumpSliceNorm("force ",force,Nd-1);
       MomFilter->applyFilter(force);
-      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<< std::endl;
+      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<<" dt "<<ep<<  std::endl;
       DumpSliceNorm("force filtered ",force,Nd-1);
       
       Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
@@ -156,6 +163,7 @@ protected:
 
       as[level].actions.at(a)->deriv_log(force_abs,force_max,impulse_abs,impulse_max);
       
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] dt           : " << ep <<" "<<name<<std::endl;
       std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force average: " << force_abs <<" "<<name<<std::endl;
       std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force max    : " << force_max <<" "<<name<<std::endl;
       std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt average  : " << impulse_abs <<" "<<name<<std::endl;
@@ -276,6 +284,15 @@ public:
 		  << as[level].actions.at(actionID)->deriv_us*1.0e-6<<" s"<< std::endl;
       }
     }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Dslash counts "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    uint64_t full, partial, dirichlet;
+    DslashGetCounts(dirichlet,partial,full);
+    std::cout << GridLogMessage << " Full BCs               : "<<full<<std::endl;
+    std::cout << GridLogMessage << " Partial dirichlet BCs  : "<<partial<<std::endl;
+    std::cout << GridLogMessage << " Dirichlet BCs          : "<<dirichlet<<std::endl;
+
     std::cout << GridLogMessage << "--------------------------- "<<std::endl;
     std::cout << GridLogMessage << " Force average size "<<std::endl;
     std::cout << GridLogMessage << "------------------------- "<<std::endl;
@@ -283,7 +300,7 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
 	std::cout << GridLogMessage 
 		  << as[level].actions.at(actionID)->action_name()
-		  <<"["<<level<<"]["<< actionID<<"] : "
+		  <<"["<<level<<"]["<< actionID<<"] :\n\t\t "
 		  <<" force max " << as[level].actions.at(actionID)->deriv_max_average()
 		  <<" norm "      << as[level].actions.at(actionID)->deriv_norm_average()
 		  <<" Fdt max  "  << as[level].actions.at(actionID)->Fdt_max_average()
@@ -363,9 +380,14 @@ public:
         std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
 
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
+
+	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] before"<<std::endl;
+	AUDIT();
 	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
 	as[level].actions.at(actionID)->refresh_timer_stop();
+	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] after"<<std::endl;
+	AUDIT();
       }
 
       // Refresh the higher representation actions
@@ -402,6 +424,7 @@ public:
     // Actions
     for (int level = 0; level < as.size(); ++level) {
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	AUDIT();
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
@@ -411,6 +434,7 @@ public:
    	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
+	AUDIT();
       }
       as[level].apply(S_hireps, Representations, level, H);
     }
@@ -418,8 +442,55 @@ public:
     return H;
   }
 
+  struct _Sinitial {
+    template <class FieldType, class Repr>
+    void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, int level, RealD& H) {
+      
+      for (int a = 0; a < repr_set.size(); ++a) {
+	AUDIT();
+        RealD Hterm = repr_set.at(a)->Sinitial(Rep.U);
+	AUDIT();
+        std::cout << GridLogMessage << "Sinitial Level " << level << " term " << a << " H Hirep = " << Hterm << std::endl;
+        H += Hterm;
+
+      }
+    }
+  } Sinitial_hireps{};
+
+  RealD Sinitial(Field& U) 
+  {  // here also U not used
+
+    std::cout << GridLogIntegrator << "Integrator initial action\n";
+
+    RealD H = - FieldImplementation::FieldSquareNorm(P)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
+
+    RealD Hterm;
+
+    // Actions
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+        // get gauge field from the SmearingPolicy and
+        // based on the boolean is_smeared in actionID
+        Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
+        std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
+	        as[level].actions.at(actionID)->S_timer_start();
+	AUDIT();
+        Hterm = as[level].actions.at(actionID)->Sinitial(Us);
+   	        as[level].actions.at(actionID)->S_timer_stop();
+	AUDIT();
+        std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
+        H += Hterm;
+      }
+      as[level].apply(Sinitial_hireps, Representations, level, H);
+    }
+
+    return H;
+  }
+
+  
   void integrate(Field& U) 
   {
+    AUDIT();
     // reset the clocks
     t_U = 0;
     for (int level = 0; level < as.size(); ++level) {
@@ -437,8 +508,10 @@ public:
       assert(fabs(t_U - t_P[level]) < 1.0e-6);  // must be the same
       std::cout << GridLogIntegrator << " times[" << level << "]= " << t_P[level] << " " << t_U << std::endl;
     }
+    AUDIT();
 
     FieldImplementation::Project(U);
+    AUDIT();
 
     // and that we indeed got to the end of the trajectory
     assert(fabs(t_U - Params.trajL) < 1.0e-6);

Grid/qcd/hmc/integrators/Integrator_algorithm.h

@@ -230,7 +230,8 @@ public:
     // Presently 4 force evals, and should have 3, so 1.33x too expensive.
     // could reduce this with sloppy CG to perhaps 1.15x too expensive
     // even without prediction.
-    this->update_P(Pfg, Ufg, level, 1.0);
+    this->update_P(Pfg, Ufg, level, fg_dt);
+    Pfg = Pfg*(1.0/fg_dt);
     this->update_U(Pfg, Ufg, fg_dt);
     this->update_P(Ufg, level, ep);
   }

Grid/qcd/modules/Registration.h

@@ -78,13 +78,13 @@ static Registrar<OneFlavourRatioEOFModule<FermionImplementationPolicy>,
 // Now a specific registration with a fermion field
 // here must instantiate CG and CR for every new fermion field type (macro!!)
 
-static Registrar< ConjugateGradientModule<WilsonFermionR::FermionField>,   
+static Registrar< ConjugateGradientModule<WilsonFermionD::FermionField>,   
-                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CGWFmodXMLInit("ConjugateGradient"); 
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionD::FermionField, Serialiser> > __CGWFmodXMLInit("ConjugateGradient"); 
 
-static Registrar< BiCGSTABModule<WilsonFermionR::FermionField>,   
+static Registrar< BiCGSTABModule<WilsonFermionD::FermionField>,   
-                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __BiCGWFmodXMLInit("BiCGSTAB"); 
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionD::FermionField, Serialiser> > __BiCGWFmodXMLInit("BiCGSTAB"); 
-static Registrar< ConjugateResidualModule<WilsonFermionR::FermionField>,   
+static Registrar< ConjugateResidualModule<WilsonFermionD::FermionField>,   
-                  HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CRWFmodXMLInit("ConjugateResidual"); 
+                  HMC_SolverModuleFactory<solver_string, WilsonFermionD::FermionField, Serialiser> > __CRWFmodXMLInit("ConjugateResidual"); 
 
 // add the staggered, scalar versions here
 

Grid/qcd/utils/SUn.h

@@ -615,7 +615,6 @@ public:
     GridBase *grid = out.Grid();
 
     typedef typename LatticeMatrixType::vector_type vector_type;
-    typedef typename LatticeMatrixType::scalar_type scalar_type;
 
     typedef iSinglet<vector_type> vTComplexType;
 

Grid/serialisation/JSON_IO.cc

@@ -26,7 +26,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef GRID_HIP
 
 NAMESPACE_BEGIN(Grid);
 
@@ -82,7 +82,7 @@ void JSONWriter::writeDefault(const std::string &s,	const std::string &x)
   if (s.size())
     ss_ << "\""<< s << "\" : \"" << os.str() << "\" ," ;
   else
-    ss_ << os.str() << " ," ;
+    ss_ << "\""<< os.str() << "\" ," ;
 }
 
 // Reader implementation ///////////////////////////////////////////////////////

Grid/serialisation/JSON_IO.h

@@ -54,7 +54,7 @@ namespace Grid
     void pop(void);
     template <typename U>
     void writeDefault(const std::string &s, const U &x);
-#ifdef __NVCC__
+#if defined(GRID_CUDA) || defined(GRID_HIP)
     void writeDefault(const std::string &s, const Grid::ComplexD &x) 
     { 
       std::complex<double> z(real(x),imag(x));
@@ -101,7 +101,7 @@ namespace Grid
     void readDefault(const std::string &s, std::vector<U> &output);
     template <typename U, typename P>
     void readDefault(const std::string &s, std::pair<U,P> &output);
-#ifdef __NVCC__
+#if defined(GRID_CUDA) || defined(GRID_HIP)
     void readDefault(const std::string &s, ComplexD &output)
     { 
       std::complex<double> z;

Grid/serialisation/Serialisation.h

@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include "BinaryIO.h"
 #include "TextIO.h"
 #include "XmlIO.h"
-#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
+#ifndef GRID_HIP
 #include "JSON_IO.h"
 #endif
 

Grid/simd/Grid_a64fx-2.h

@@ -501,7 +501,7 @@ struct Conj{
 struct TimesMinusI{
   // Complex
   template <typename T>
-  inline vec<T> operator()(vec<T> a, vec<T> b){
+  inline vec<T> operator()(vec<T> a){
     vec<T> out;
     const vec<typename acle<T>::uint> tbl_swap = acle<T>::tbl_swap();
     svbool_t pg1 = acle<T>::pg1();
@@ -520,7 +520,7 @@ struct TimesMinusI{
 struct TimesI{
   // Complex
   template <typename T>
-  inline vec<T> operator()(vec<T> a, vec<T> b){
+  inline vec<T> operator()(vec<T> a){
     vec<T> out;
     const vec<typename acle<T>::uint> tbl_swap = acle<T>::tbl_swap();
     svbool_t pg1 = acle<T>::pg1();

Grid/simd/Grid_a64fx-fixedsize.h

@@ -418,7 +418,7 @@ struct Conj{
 
 struct TimesMinusI{
   // Complex float
-  inline vecf operator()(vecf a, vecf b){
+  inline vecf operator()(vecf a){
     lutf tbl_swap = acle<float>::tbl_swap();
     pred pg1 = acle<float>::pg1();
     pred pg_odd = acle<float>::pg_odd();
@@ -428,7 +428,7 @@ struct TimesMinusI{
     return svneg_m(a_v, pg_odd, a_v);
   }
   // Complex double
-  inline vecd operator()(vecd a, vecd b){
+  inline vecd operator()(vecd a){
     lutd tbl_swap = acle<double>::tbl_swap();
     pred pg1 = acle<double>::pg1();
     pred pg_odd = acle<double>::pg_odd();
@@ -441,7 +441,7 @@ struct TimesMinusI{
 
 struct TimesI{
   // Complex float
-  inline vecf operator()(vecf a, vecf b){
+  inline vecf operator()(vecf a){
     lutf tbl_swap = acle<float>::tbl_swap();
     pred pg1 = acle<float>::pg1();
     pred pg_even = acle<float>::pg_even();
@@ -451,7 +451,7 @@ struct TimesI{
     return svneg_m(a_v, pg_even, a_v);
   }
   // Complex double
-  inline vecd operator()(vecd a, vecd b){
+  inline vecd operator()(vecd a){
     lutd tbl_swap = acle<double>::tbl_swap();
     pred pg1 = acle<double>::pg1();
     pred pg_even = acle<double>::pg_even();

Grid/simd/Grid_avx.h

@@ -405,12 +405,12 @@ struct Conj{
 
 struct TimesMinusI{
   //Complex single
-  inline __m256 operator()(__m256 in, __m256 ret){
+  inline __m256 operator()(__m256 in){
     __m256 tmp =_mm256_addsub_ps(_mm256_setzero_ps(),in);   // r,-i
     return _mm256_shuffle_ps(tmp,tmp,_MM_SELECT_FOUR_FOUR(2,3,0,1)); //-i,r
   }
   //Complex double
-  inline __m256d operator()(__m256d in, __m256d ret){
+  inline __m256d operator()(__m256d in){
     __m256d tmp = _mm256_addsub_pd(_mm256_setzero_pd(),in); // r,-i
     return _mm256_shuffle_pd(tmp,tmp,0x5);
   }
@@ -418,12 +418,12 @@ struct TimesMinusI{
 
 struct TimesI{
   //Complex single
-  inline __m256 operator()(__m256 in, __m256 ret){
+  inline __m256 operator()(__m256 in){
     __m256 tmp =_mm256_shuffle_ps(in,in,_MM_SELECT_FOUR_FOUR(2,3,0,1)); // i,r
     return _mm256_addsub_ps(_mm256_setzero_ps(),tmp);          // i,-r
   }
   //Complex double
-  inline __m256d operator()(__m256d in, __m256d ret){
+  inline __m256d operator()(__m256d in){
     __m256d tmp = _mm256_shuffle_pd(in,in,0x5);
     return _mm256_addsub_pd(_mm256_setzero_pd(),tmp); // i,-r
   }

Grid/simd/Grid_avx512.h

@@ -271,14 +271,14 @@ struct Conj{
 
 struct TimesMinusI{
   //Complex single
-  inline __m512 operator()(__m512 in, __m512 ret){
+  inline __m512 operator()(__m512 in){
     //__m512 tmp = _mm512_mask_sub_ps(in,0xaaaa,_mm512_setzero_ps(),in); // real -imag 
     //return _mm512_shuffle_ps(tmp,tmp,_MM_SELECT_FOUR_FOUR(2,3,1,0));   // 0x4E??
     __m512 tmp = _mm512_shuffle_ps(in,in,_MM_SELECT_FOUR_FOUR(2,3,0,1));
     return _mm512_mask_sub_ps(tmp,0xaaaa,_mm512_setzero_ps(),tmp);
   }
   //Complex double
-  inline __m512d operator()(__m512d in, __m512d ret){
+  inline __m512d operator()(__m512d in){
     //__m512d tmp = _mm512_mask_sub_pd(in,0xaa,_mm512_setzero_pd(),in); // real -imag 
     //return _mm512_shuffle_pd(tmp,tmp,0x55);
     __m512d tmp = _mm512_shuffle_pd(in,in,0x55);
@@ -288,17 +288,16 @@ struct TimesMinusI{
 
 struct TimesI{
   //Complex single
-  inline __m512 operator()(__m512 in, __m512 ret){
+  inline __m512 operator()(__m512 in){
     __m512 tmp = _mm512_shuffle_ps(in,in,_MM_SELECT_FOUR_FOUR(2,3,0,1));
     return _mm512_mask_sub_ps(tmp,0x5555,_mm512_setzero_ps(),tmp); 
   }
   //Complex double
-  inline __m512d operator()(__m512d in, __m512d ret){
+  inline __m512d operator()(__m512d in){
     __m512d tmp = _mm512_shuffle_pd(in,in,0x55);
     return _mm512_mask_sub_pd(tmp,0x55,_mm512_setzero_pd(),tmp); 
   }
 
-
 };
   
 // Gpermute utilities consider coalescing into 1 Gpermute

Grid/simd/Grid_doubled_vector.h

@@ -0,0 +1,666 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/simd/Grid_vector_types.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template <class Scalar_type, class Vector_type>
+class Grid_simd2 {
+public:
+  typedef typename RealPart<Scalar_type>::type Real;
+  typedef Vector_type vector_type;
+  typedef Scalar_type scalar_type;
+
+  typedef union conv_t_union {
+    Vector_type v;
+    Scalar_type s[sizeof(Vector_type) / sizeof(Scalar_type)];
+    accelerator_inline conv_t_union(){};
+  } conv_t;
+
+  static constexpr int nvec=2;
+  Vector_type v[nvec];
+
+  static accelerator_inline constexpr int Nsimd(void) {
+    static_assert( (sizeof(Vector_type) / sizeof(Scalar_type) >= 1), " size mismatch " );
+    
+    return nvec*sizeof(Vector_type) / sizeof(Scalar_type);
+  }
+  
+  accelerator_inline Grid_simd2 &operator=(const Grid_simd2 &&rhs) {
+    for(int n=0;n<nvec;n++) v[n] = rhs.v[n];
+    return *this;
+  };
+  accelerator_inline Grid_simd2 &operator=(const Grid_simd2 &rhs) {
+    for(int n=0;n<nvec;n++) v[n] = rhs.v[n];
+    return *this;
+  };  // faster than not declaring it and leaving to the compiler
+
+  accelerator Grid_simd2() = default;
+  accelerator_inline Grid_simd2(const Grid_simd2 &rhs) {    for(int n=0;n<nvec;n++) v[n] = rhs.v[n];  };
+  accelerator_inline Grid_simd2(const Grid_simd2 &&rhs){    for(int n=0;n<nvec;n++) v[n] = rhs.v[n];  };
+  accelerator_inline Grid_simd2(const Real a) { vsplat(*this, Scalar_type(a)); };
+  // Enable if complex type
+  template <typename S = Scalar_type> accelerator_inline
+  Grid_simd2(const typename std::enable_if<is_complex<S>::value, S>::type a) {
+      vsplat(*this, a);
+  };
+
+  /////////////////////////////
+  // Constructors
+  /////////////////////////////
+  accelerator_inline Grid_simd2 &  operator=(const Zero &z) {
+    vzero(*this);
+    return (*this);
+  }
+
+  ///////////////////////////////////////////////
+  // mac, mult, sub, add, adj
+  ///////////////////////////////////////////////
+
+  friend accelerator_inline void mac(Grid_simd2 *__restrict__ y,
+				     const Grid_simd2 *__restrict__ a,
+				     const Grid_simd2 *__restrict__ x) {
+    *y = (*a) * (*x) + (*y);
+  };
+
+  friend accelerator_inline void mult(Grid_simd2 *__restrict__ y,
+				      const Grid_simd2 *__restrict__ l,
+				      const Grid_simd2 *__restrict__ r) {
+    *y = (*l) * (*r);
+  }
+
+  friend accelerator_inline void sub(Grid_simd2 *__restrict__ y,
+				     const Grid_simd2 *__restrict__ l,
+				     const Grid_simd2 *__restrict__ r) {
+    *y = (*l) - (*r);
+  }
+  friend accelerator_inline void add(Grid_simd2 *__restrict__ y,
+				     const Grid_simd2 *__restrict__ l,
+				     const Grid_simd2 *__restrict__ r) {
+    *y = (*l) + (*r);
+  }
+  friend accelerator_inline void mac(Grid_simd2 *__restrict__ y,
+				     const Scalar_type *__restrict__ a,
+				     const Grid_simd2 *__restrict__ x) {
+    *y = (*a) * (*x) + (*y);
+  };
+  friend accelerator_inline void mult(Grid_simd2 *__restrict__ y,
+				      const Scalar_type *__restrict__ l,
+				      const Grid_simd2 *__restrict__ r) {
+    *y = (*l) * (*r);
+  }
+  friend accelerator_inline void sub(Grid_simd2 *__restrict__ y,
+				     const Scalar_type *__restrict__ l,
+				     const Grid_simd2 *__restrict__ r) {
+    *y = (*l) - (*r);
+  }
+  friend accelerator_inline void add(Grid_simd2 *__restrict__ y,
+				     const Scalar_type *__restrict__ l,
+				     const Grid_simd2 *__restrict__ r) {
+    *y = (*l) + (*r);
+  }
+
+  friend accelerator_inline void mac(Grid_simd2 *__restrict__ y,
+				     const Grid_simd2 *__restrict__ a,
+				     const Scalar_type *__restrict__ x) {
+    *y = (*a) * (*x) + (*y);
+  };
+  friend accelerator_inline void mult(Grid_simd2 *__restrict__ y,
+				      const Grid_simd2 *__restrict__ l,
+				      const Scalar_type *__restrict__ r) {
+    *y = (*l) * (*r);
+  }
+  friend accelerator_inline void sub(Grid_simd2 *__restrict__ y,
+				     const Grid_simd2 *__restrict__ l,
+				     const Scalar_type *__restrict__ r) {
+    *y = (*l) - (*r);
+  }
+  friend accelerator_inline void add(Grid_simd2 *__restrict__ y,
+				     const Grid_simd2 *__restrict__ l,
+				     const Scalar_type *__restrict__ r) {
+    *y = (*l) + (*r);
+  }
+
+  ////////////////////////////////////////////////////////////////////////
+  // FIXME:  gonna remove these load/store, get, set, prefetch
+  ////////////////////////////////////////////////////////////////////////
+  friend accelerator_inline void vset(Grid_simd2 &ret, Scalar_type *a) {
+    for(int n=0;n<nvec;n++) vset(ret.v[n],a);
+  }
+
+  ///////////////////////
+  // Vstore
+  ///////////////////////
+  friend accelerator_inline void vstore(const Grid_simd2 &ret, Scalar_type *a) {
+    for(int n=0;n<nvec;n++) vstore(ret.v[n],a);
+  }
+
+  ///////////////////////
+  // Vprefetch
+  ///////////////////////
+  friend accelerator_inline void vprefetch(const Grid_simd2 &v) {
+    vprefetch(v.v[0]);
+  }
+
+  ///////////////////////
+  // Reduce
+  ///////////////////////
+  friend accelerator_inline Scalar_type Reduce(const Grid_simd2 &in) {
+    return Reduce(in.v[0])+ Reduce(in.v[1]);
+  }
+
+  ////////////////////////////
+  // operator scalar * simd
+  ////////////////////////////
+  friend accelerator_inline Grid_simd2 operator*(const Scalar_type &a, Grid_simd2 b) {
+    Grid_simd2 va;
+    vsplat(va, a);
+    return va * b;
+  }
+  friend accelerator_inline Grid_simd2 operator*(Grid_simd2 b, const Scalar_type &a) {
+    return a * b;
+  }
+
+  //////////////////////////////////
+  // Divides
+  //////////////////////////////////
+  friend accelerator_inline Grid_simd2 operator/(const Scalar_type &a, Grid_simd2 b) {
+    Grid_simd2 va;
+    vsplat(va, a);
+    return va / b;
+  }
+  friend accelerator_inline Grid_simd2 operator/(Grid_simd2 b, const Scalar_type &a) {
+    Grid_simd2 va;
+    vsplat(va, a);
+    return b / a;
+  }
+
+  ///////////////////////
+  // Unary negation
+  ///////////////////////
+  friend accelerator_inline Grid_simd2 operator-(const Grid_simd2 &r) {
+    Grid_simd2 ret;
+    vzero(ret);
+    ret = ret - r;
+    return ret;
+  }
+  // *=,+=,-= operators
+  accelerator_inline Grid_simd2 &operator*=(const Grid_simd2 &r) {
+    *this = (*this) * r;
+    return *this;
+  }
+  accelerator_inline Grid_simd2 &operator+=(const Grid_simd2 &r) {
+    *this = *this + r;
+    return *this;
+  }
+  accelerator_inline Grid_simd2 &operator-=(const Grid_simd2 &r) {
+    *this = *this - r;
+    return *this;
+  }
+
+  ///////////////////////////////////////
+  // Not all functions are supported
+  // through SIMD and must breakout to
+  // scalar type and back again. This
+  // provides support
+  ///////////////////////////////////////
+  template <class functor>
+  friend accelerator_inline Grid_simd2 SimdApply(const functor &func, const Grid_simd2 &v) {
+    Grid_simd2 ret;
+    for(int n=0;n<nvec;n++){
+      ret.v[n]=SimdApply(func,v.v[n]);
+    }
+    return ret;
+  }
+  template <class functor>
+  friend accelerator_inline Grid_simd2 SimdApplyBinop(const functor &func,
+                                         const Grid_simd2 &x,
+                                         const Grid_simd2 &y) {
+    Grid_simd2 ret;
+    for(int n=0;n<nvec;n++){
+      ret.v[n]=SimdApplyBinop(func,x.v[n],y.v[n]);
+    }
+    return ret;
+  }
+  ///////////////////////
+  // Exchange
+  // Al Ah , Bl Bh -> Al Bl Ah,Bh
+  ///////////////////////
+  friend accelerator_inline void exchange0(Grid_simd2 &out1,Grid_simd2 &out2,Grid_simd2 in1,Grid_simd2 in2){
+      out1.v[0] = in1.v[0];
+      out1.v[1] = in2.v[0];
+      out2.v[0] = in1.v[1];
+      out2.v[1] = in2.v[1];
+  }
+  friend accelerator_inline void exchange1(Grid_simd2 &out1,Grid_simd2 &out2,Grid_simd2 in1,Grid_simd2 in2){
+    exchange0(out1.v[0],out2.v[0],in1.v[0],in2.v[0]);
+    exchange0(out1.v[1],out2.v[1],in1.v[1],in2.v[1]);
+  }
+  friend accelerator_inline void exchange2(Grid_simd2 &out1,Grid_simd2 &out2,Grid_simd2 in1,Grid_simd2 in2){
+    exchange1(out1.v[0],out2.v[0],in1.v[0],in2.v[0]);
+    exchange1(out1.v[1],out2.v[1],in1.v[1],in2.v[1]);
+  }
+  friend accelerator_inline void exchange3(Grid_simd2 &out1,Grid_simd2 &out2,Grid_simd2 in1,Grid_simd2 in2){
+    exchange2(out1.v[0],out2.v[0],in1.v[0],in2.v[0]);
+    exchange2(out1.v[1],out2.v[1],in1.v[1],in2.v[1]);
+  }
+  friend accelerator_inline void exchange4(Grid_simd2 &out1,Grid_simd2 &out2,Grid_simd2 in1,Grid_simd2 in2){
+    exchange3(out1.v[0],out2.v[0],in1.v[0],in2.v[0]);
+    exchange3(out1.v[1],out2.v[1],in1.v[1],in2.v[1]);
+  }
+  friend accelerator_inline void exchange(Grid_simd2 &out1,Grid_simd2 &out2,Grid_simd2 in1,Grid_simd2 in2,int n)
+  {
+    if       (n==3) {
+      exchange3(out1,out2,in1,in2);
+    } else if(n==2) {
+      exchange2(out1,out2,in1,in2);
+    } else if(n==1) {
+      exchange1(out1,out2,in1,in2);
+    } else if(n==0) {
+      exchange0(out1,out2,in1,in2);
+    }
+  }
+  ////////////////////////////////////////////////////////////////////
+  // General permute; assumes vector length is same across
+  // all subtypes; may not be a good assumption, but could
+  // add the vector width as a template param for BG/Q for example
+  ////////////////////////////////////////////////////////////////////
+  friend accelerator_inline void permute0(Grid_simd2 &y, Grid_simd2 b) {
+    y.v[0]=b.v[1];
+    y.v[1]=b.v[0];
+  }
+  friend accelerator_inline void permute1(Grid_simd2 &y, Grid_simd2 b) {
+    permute0(y.v[0],b.v[0]);
+    permute0(y.v[1],b.v[1]);
+  }
+  friend accelerator_inline void permute2(Grid_simd2 &y, Grid_simd2 b) {
+    permute1(y.v[0],b.v[0]);
+    permute1(y.v[1],b.v[1]);
+  }
+  friend accelerator_inline void permute3(Grid_simd2 &y, Grid_simd2 b) {
+    permute2(y.v[0],b.v[0]);
+    permute2(y.v[1],b.v[1]);
+  }
+  friend accelerator_inline void permute4(Grid_simd2 &y, Grid_simd2 b) {
+    permute3(y.v[0],b.v[0]);
+    permute3(y.v[1],b.v[1]);
+  }
+  friend accelerator_inline void permute(Grid_simd2 &y, Grid_simd2 b, int perm) {
+    if(perm==3) permute3(y, b);
+    else if(perm==2) permute2(y, b);
+    else if(perm==1) permute1(y, b);
+    else if(perm==0) permute0(y, b);
+  }
+
+  ///////////////////////////////
+  // Getting single lanes
+  ///////////////////////////////
+  accelerator_inline Scalar_type getlane(int lane) const {
+    if(lane < vector_type::Nsimd() ) return v[0].getlane(lane);
+    else                             return v[1].getlane(lane%vector_type::Nsimd());
+  }
+
+  accelerator_inline void putlane(const Scalar_type &S, int lane){
+    if(lane < vector_type::Nsimd() ) v[0].putlane(S,lane);
+    else                             v[1].putlane(S,lane%vector_type::Nsimd());
+  }
+};  // end of Grid_simd2 class definition
+
+///////////////////////////////
+// Define available types
+///////////////////////////////
+
+typedef Grid_simd2<complex<double>  , vComplexD>  vComplexD2;
+typedef Grid_simd2<double           , vRealD>     vRealD2;
+
+
+
+/////////////////////////////////////////
+// Some traits to recognise the types
+/////////////////////////////////////////
+template <typename T>
+struct is_simd : public std::false_type {};
+template <> struct is_simd<vRealF>     : public std::true_type {};
+template <> struct is_simd<vRealD>     : public std::true_type {};
+template <> struct is_simd<vRealH>     : public std::true_type {};
+template <> struct is_simd<vComplexF>  : public std::true_type {};
+template <> struct is_simd<vComplexD>  : public std::true_type {};
+template <> struct is_simd<vComplexH>  : public std::true_type {};
+template <> struct is_simd<vInteger>   : public std::true_type {};
+template <> struct is_simd<vRealD2>    : public std::true_type {};
+template <> struct is_simd<vComplexD2> : public std::true_type {};
+
+template <typename T> using IfSimd    = Invoke<std::enable_if<is_simd<T>::value, int> >;
+template <typename T> using IfNotSimd = Invoke<std::enable_if<!is_simd<T>::value, unsigned> >;
+
+///////////////////////////////////////////////
+// insert / extract with complex support
+///////////////////////////////////////////////
+template <class S, class V>
+accelerator_inline S getlane(const Grid_simd<S, V> &in,int lane) {
+  return in.getlane(lane);
+}
+template <class S, class V>
+accelerator_inline void putlane(Grid_simd<S, V> &vec,const S &_S, int lane){
+  vec.putlane(_S,lane);
+}
+template <class S,IfNotSimd<S> = 0 >
+accelerator_inline S getlane(const S &in,int lane) {
+  return in;
+}
+template <class S,IfNotSimd<S> = 0 >
+accelerator_inline void putlane(S &vec,const S &_S, int lane){
+  vec = _S;
+}
+template <class S, class V>
+accelerator_inline S getlane(const Grid_simd2<S, V> &in,int lane) {
+  return in.getlane(lane);
+}
+template <class S, class V>
+accelerator_inline void putlane(Grid_simd2<S, V> &vec,const S &_S, int lane){
+  vec.putlane(_S,lane);
+}
+
+
+////////////////////////////////////////////////////////////////////
+// General rotate
+////////////////////////////////////////////////////////////////////
+
+template <class S, class V>
+accelerator_inline void vbroadcast(Grid_simd2<S,V> &ret,const Grid_simd2<S,V> &src,int lane){
+  S* typepun =(S*) &src;
+  vsplat(ret,typepun[lane]);
+}
+template <class S, class V, IfComplex<S> =0>
+accelerator_inline void rbroadcast(Grid_simd2<S,V> &ret,const Grid_simd2<S,V> &src,int lane){
+  typedef typename V::vector_type vector_type;
+  S* typepun =(S*) &src;
+  ret.v[0].v = unary<vector_type>(real(typepun[lane]), VsplatSIMD());
+  ret.v[1].v = unary<vector_type>(real(typepun[lane]), VsplatSIMD());
+}
+
+
+///////////////////////
+// Splat
+///////////////////////
+
+// this is only for the complex version
+template <class S, class V, IfComplex<S> = 0, class ABtype>
+accelerator_inline void vsplat(Grid_simd2<S, V> &ret, ABtype a, ABtype b) {
+  vsplat(ret.v[0],a,b);
+  vsplat(ret.v[1],a,b);
+}
+
+// overload if complex
+template <class S, class V>
+accelerator_inline void vsplat(Grid_simd2<S, V> &ret, EnableIf<is_complex<S>, S> c) {
+  vsplat(ret, real(c), imag(c));
+}
+template <class S, class V>
+accelerator_inline void rsplat(Grid_simd2<S, V> &ret, EnableIf<is_complex<S>, S> c) {
+  vsplat(ret, real(c), real(c));
+}
+
+// if real fill with a, if complex fill with a in the real part (first function
+// above)
+template <class S, class V>
+accelerator_inline void vsplat(Grid_simd2<S, V> &ret, NotEnableIf<is_complex<S>, S> a)
+{
+  vsplat(ret.v[0],a);
+  vsplat(ret.v[1],a);
+}
+//////////////////////////
+
+///////////////////////////////////////////////
+// Initialise to 1,0,i for the correct types
+///////////////////////////////////////////////
+// For complex types
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline void vone(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(1.0, 0.0));
+}
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline void vzero(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(0.0, 0.0));
+}  // use xor?
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline void vcomplex_i(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(0.0, 1.0));
+}
+
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline void visign(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(1.0, -1.0));
+}
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline void vrsign(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(-1.0, 1.0));
+}
+
+// if not complex overload here
+template <class S, class V, IfReal<S> = 0>
+accelerator_inline void vone(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(1.0));
+}
+template <class S, class V, IfReal<S> = 0>
+accelerator_inline void vzero(Grid_simd2<S, V> &ret) {
+  vsplat(ret, S(0.0));
+}
+
+// For integral types
+template <class S, class V, IfInteger<S> = 0>
+accelerator_inline void vone(Grid_simd2<S, V> &ret) {
+  vsplat(ret, 1);
+}
+template <class S, class V, IfInteger<S> = 0>
+accelerator_inline void vzero(Grid_simd2<S, V> &ret) {
+  vsplat(ret, 0);
+}
+template <class S, class V, IfInteger<S> = 0>
+accelerator_inline void vtrue(Grid_simd2<S, V> &ret) {
+  vsplat(ret, 0xFFFFFFFF);
+}
+template <class S, class V, IfInteger<S> = 0>
+accelerator_inline void vfalse(Grid_simd2<S, V> &ret) {
+  vsplat(ret, 0);
+}
+template <class S, class V>
+accelerator_inline void zeroit(Grid_simd2<S, V> &z) {
+  vzero(z);
+}
+
+///////////////////////
+// Vstream
+///////////////////////
+template <class S, class V, IfReal<S> = 0>
+accelerator_inline void vstream(Grid_simd2<S, V> &out, const Grid_simd2<S, V> &in) {
+  vstream(out.v[0],in.v[0]);
+  vstream(out.v[1],in.v[1]);
+}
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline void vstream(Grid_simd2<S, V> &out, const Grid_simd2<S, V> &in) {
+  vstream(out.v[0],in.v[0]);
+  vstream(out.v[1],in.v[1]);
+}
+template <class S, class V, IfInteger<S> = 0>
+accelerator_inline void vstream(Grid_simd2<S, V> &out, const Grid_simd2<S, V> &in) {
+  vstream(out.v[0],in.v[0]);
+  vstream(out.v[1],in.v[1]);
+}
+
+////////////////////////////////////
+// Arithmetic operator overloads +,-,*
+////////////////////////////////////
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> operator+(Grid_simd2<S, V> a, Grid_simd2<S, V> b) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] = a.v[0]+b.v[0];
+  ret.v[1] = a.v[1]+b.v[1];
+  return ret;
+};
+
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> operator-(Grid_simd2<S, V> a, Grid_simd2<S, V> b) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] = a.v[0]-b.v[0];
+  ret.v[1] = a.v[1]-b.v[1];
+  return ret;
+};
+
+// Distinguish between complex types and others
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline Grid_simd2<S, V> real_mult(Grid_simd2<S, V> a, Grid_simd2<S, V> b) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] =real_mult(a.v[0],b.v[0]);
+  ret.v[1] =real_mult(a.v[1],b.v[1]);
+  return ret;
+};
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline Grid_simd2<S, V> real_madd(Grid_simd2<S, V> a, Grid_simd2<S, V> b, Grid_simd2<S,V> c) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] =real_madd(a.v[0],b.v[0],c.v[0]);
+  ret.v[1] =real_madd(a.v[1],b.v[1],c.v[1]);
+  return ret;
+};
+
+
+// Distinguish between complex types and others
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> operator*(Grid_simd2<S, V> a, Grid_simd2<S, V> b) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] = a.v[0]*b.v[0];
+  ret.v[1] = a.v[1]*b.v[1];
+  return ret;
+};
+
+// Distinguish between complex types and others
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> operator/(Grid_simd2<S, V> a, Grid_simd2<S, V> b) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] = a.v[0]/b.v[0];
+  ret.v[1] = a.v[1]/b.v[1];
+  return ret;
+};
+
+///////////////////////
+// Conjugate
+///////////////////////
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> conjugate(const Grid_simd2<S, V> &in) {
+  Grid_simd2<S, V> ret;
+  ret.v[0] = conjugate(in.v[0]);
+  ret.v[1] = conjugate(in.v[1]);
+  return ret;
+}
+template <class S, class V, IfNotInteger<S> = 0>
+accelerator_inline Grid_simd2<S, V> adj(const Grid_simd2<S, V> &in) {
+  return conjugate(in);
+}
+
+///////////////////////
+// timesMinusI
+///////////////////////
+template <class S, class V>
+accelerator_inline void timesMinusI(Grid_simd2<S, V> &ret, const Grid_simd2<S, V> &in) {
+  timesMinusI(ret.v[0],in.v[0]);
+  timesMinusI(ret.v[1],in.v[1]);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> timesMinusI(const Grid_simd2<S, V> &in) {
+  Grid_simd2<S, V> ret;
+  timesMinusI(ret.v[0],in.v[0]);
+  timesMinusI(ret.v[1],in.v[1]);
+  return ret;
+}
+
+///////////////////////
+// timesI
+///////////////////////
+template <class S, class V>
+accelerator_inline void timesI(Grid_simd2<S, V> &ret, const Grid_simd2<S, V> &in) {
+  timesI(ret.v[0],in.v[0]);
+  timesI(ret.v[1],in.v[1]);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> timesI(const Grid_simd2<S, V> &in) {
+  Grid_simd2<S, V> ret;
+  timesI(ret.v[0],in.v[0]);
+  timesI(ret.v[1],in.v[1]);
+  return ret;
+}
+
+/////////////////////
+// Inner, outer
+/////////////////////
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> innerProduct(const Grid_simd2<S, V> &l,const Grid_simd2<S, V> &r) {
+  return conjugate(l) * r;
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> outerProduct(const Grid_simd2<S, V> &l,const Grid_simd2<S, V> &r) {
+  return l * conjugate(r);
+}
+
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> trace(const Grid_simd2<S, V> &arg) {
+  return arg;
+}
+
+////////////////////////////////////////////////////////////
+// copy/splat complex real parts into real;
+// insert real into complex and zero imag;
+////////////////////////////////////////////////////////////
+accelerator_inline void precisionChange(vComplexD2 &out,const vComplexF  &in){
+  Optimization::PrecisionChange::StoD(in.v,out.v[0].v,out.v[1].v);
+}
+accelerator_inline void precisionChange(vComplexF  &out,const vComplexD2 &in){
+  out.v=Optimization::PrecisionChange::DtoS(in.v[0].v,in.v[1].v);
+}
+accelerator_inline void precisionChange(vComplexD2 *out,const vComplexF  *in,int nvec){
+  for(int m=0;m<nvec;m++){ precisionChange(out[m],in[m]); }
+}
+accelerator_inline void precisionChange(vComplexF  *out,const vComplexD2 *in,int nvec){
+  for(int m=0;m<nvec;m++){ precisionChange(out[m],in[m]); }
+}
+
+accelerator_inline void precisionChange(vRealD2 &out,const vRealF  &in){
+  Optimization::PrecisionChange::StoD(in.v,out.v[0].v,out.v[1].v);
+}
+accelerator_inline void precisionChange(vRealF  &out,const vRealD2 &in){
+  out.v=Optimization::PrecisionChange::DtoS(in.v[0].v,in.v[1].v);
+}
+accelerator_inline void precisionChange(vRealD2 *out,const vRealF  *in,int nvec){
+  for(int m=0;m<nvec;m++){ precisionChange(out[m],in[m]); }
+}
+accelerator_inline void precisionChange(vRealF  *out,const vRealD2 *in,int nvec){
+  for(int m=0;m<nvec;m++){ precisionChange(out[m],in[m]); }
+}
+
+NAMESPACE_END(Grid);
+
+

Grid/simd/Grid_generic.h

@@ -244,7 +244,7 @@ struct Conj{
 struct TimesMinusI{
   // Complex
   template <typename T>
-  accelerator_inline vec<T> operator()(vec<T> a, vec<T> b){
+  accelerator_inline vec<T> operator()(vec<T> a){
     vec<T> out;
       
     VECTOR_FOR(i, W<T>::c, 1)
@@ -265,7 +265,7 @@ struct TimesMinusI{
 struct TimesI{
   // Complex
   template <typename T>
-  accelerator_inline vec<T> operator()(vec<T> a, vec<T> b){
+  accelerator_inline vec<T> operator()(vec<T> a){
     vec<T> out;
       
     VECTOR_FOR(i, W<T>::c, 1)

Grid/simd/Grid_gpu_rrii.h

@@ -0,0 +1,878 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/simd/Grid_gpu.h
+
+    Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+//----------------------------------------------------------------------
+/*! @file Grid_gpu_rrii.h*/
+//----------------------------------------------------------------------
+
+//////////////////////////////
+// fp16
+//////////////////////////////
+#ifdef GRID_CUDA
+#include <cuda_fp16.h>
+#endif
+#ifdef GRID_HIP
+#include <hip/hip_fp16.h>
+#endif
+#if !defined(GRID_HIP) && !defined(GRID_CUDA) 
+namespace Grid {
+  typedef struct { uint16_t x;} half;
+}
+#endif
+namespace Grid {
+  accelerator_inline float half2float(half h)
+  {
+    float f;
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+    f = __half2float(h);
+#else 
+    Grid_half hh; 
+    hh.x = h.x;
+    f=  sfw_half_to_float(hh);
+#endif
+    return f;
+  }
+  accelerator_inline half float2half(float f)
+  {
+    half h;
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+    h = __float2half(f);
+#else
+    Grid_half hh = sfw_float_to_half(f);
+    h.x = hh.x;
+#endif
+    return h;
+  }
+}
+
+
+#define COALESCE_GRANULARITY ( GEN_SIMD_WIDTH )
+
+namespace Grid {
+
+////////////////////////////////////////////////////////////////////////
+// Real vector
+////////////////////////////////////////////////////////////////////////  
+template<int _N, class _datum>
+struct GpuVector {
+  _datum rrrr[_N];
+  static const int N = _N;
+  typedef _datum datum;
+};
+template<int N,class datum>
+inline accelerator GpuVector<N,datum> operator*(const GpuVector<N,datum> l,const GpuVector<N,datum> r) {
+  GpuVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]*r.rrrr[i];
+  }
+  return ret;
+}
+template<int N,class datum>
+inline accelerator GpuVector<N,datum> operator-(const GpuVector<N,datum> l,const GpuVector<N,datum> r) {
+  GpuVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]-r.rrrr[i];
+  }
+  return ret;
+}
+template<int N,class datum>
+inline accelerator GpuVector<N,datum> operator+(const GpuVector<N,datum> l,const GpuVector<N,datum> r) {
+  GpuVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]+r.rrrr[i];
+  }
+  return ret;
+}
+template<int N,class datum>
+inline accelerator GpuVector<N,datum> operator/(const GpuVector<N,datum> l,const GpuVector<N,datum> r) {
+  GpuVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]/r.rrrr[i];
+  }
+  return ret;
+}
+
+////////////////////////////////////////////////////////////////////////
+// Complex vector
+////////////////////////////////////////////////////////////////////////  
+template<int _N, class _datum>
+struct GpuComplexVector {
+  _datum rrrr[_N];
+  _datum iiii[_N];
+  static const int N = _N;
+  typedef _datum datum;
+};
+template<int N,class datum>
+inline accelerator GpuComplexVector<N,datum> operator*(const GpuComplexVector<N,datum> l,const GpuComplexVector<N,datum> r) {
+  GpuComplexVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]*r.rrrr[i] - l.iiii[i]*r.iiii[i];
+    ret.iiii[i] = l.rrrr[i]*r.iiii[i] + l.iiii[i]*r.rrrr[i];
+  }
+  return ret;
+}
+template<int N,class datum>
+inline accelerator GpuComplexVector<N,datum> operator-(const GpuComplexVector<N,datum> l,const GpuComplexVector<N,datum> r) {
+  GpuComplexVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]-r.rrrr[i];
+    ret.iiii[i] = l.iiii[i]-r.iiii[i];
+  }
+  return ret;
+}
+template<int N,class datum>
+inline accelerator GpuComplexVector<N,datum> operator+(const GpuComplexVector<N,datum> l,const GpuComplexVector<N,datum> r) {
+  GpuComplexVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]+r.rrrr[i];
+    ret.iiii[i] = l.iiii[i]+r.iiii[i];
+  }
+  return ret;
+}
+template<int N,class datum>
+inline accelerator GpuComplexVector<N,datum> operator/(const GpuComplexVector<N,datum> l,const GpuComplexVector<N,datum> r) {
+  GpuComplexVector<N,datum> ret;
+  for(int i=0;i<N;i++) { 
+    ret.rrrr[i] = l.rrrr[i]/r.rrrr[i];
+    ret.iiii[i] = l.iiii[i]/r.iiii[i];
+  }
+  return ret;
+}
+
+////////////////////////////////
+// SIMD counts
+////////////////////////////////
+
+constexpr int NSIMD_RealH    = COALESCE_GRANULARITY / sizeof(half);
+constexpr int NSIMD_ComplexH = COALESCE_GRANULARITY / sizeof(half);
+constexpr int NSIMD_RealF    = COALESCE_GRANULARITY / sizeof(float);
+constexpr int NSIMD_ComplexF = COALESCE_GRANULARITY / sizeof(float);
+constexpr int NSIMD_RealD    = COALESCE_GRANULARITY / sizeof(double);
+constexpr int NSIMD_ComplexD = COALESCE_GRANULARITY / sizeof(double);
+constexpr int NSIMD_Integer  = COALESCE_GRANULARITY / sizeof(Integer);
+
+typedef GpuVector<NSIMD_RealH   , half        > GpuVectorRH;
+typedef GpuComplexVector<NSIMD_ComplexH, half > GpuVectorCH;
+typedef GpuVector<NSIMD_RealF,    float       > GpuVectorRF;
+typedef GpuComplexVector<NSIMD_ComplexF, float> GpuVectorCF;
+typedef GpuVector<NSIMD_RealD,    double      > GpuVectorRD;
+typedef GpuComplexVector<NSIMD_ComplexD,double> GpuVectorCD;
+typedef GpuVector<NSIMD_Integer,  Integer     > GpuVectorI;
+
+namespace Optimization {
+
+  struct Vsplat{
+    //Complex float
+    accelerator_inline GpuVectorCF operator()(float a, float b){
+      GpuVectorCF ret;
+      for(int i=0;i<GpuVectorCF::N;i++){
+	ret.rrrr[i] = typename GpuVectorCF::datum(a);
+	ret.iiii[i] = typename GpuVectorCF::datum(b);
+      }
+      return ret;
+    }
+    // Real float
+    accelerator_inline GpuVectorRF operator()(float a){
+      GpuVectorRF ret;
+      for(int i=0;i<GpuVectorRF::N;i++){
+	ret.rrrr[i] = typename GpuVectorRF::datum(a);
+      }
+      return ret;
+    }
+    //Complex double
+    accelerator_inline GpuVectorCD operator()(double a, double b){
+      GpuVectorCD ret;
+      for(int i=0;i<GpuVectorCD::N;i++){
+	ret.rrrr[i] = typename GpuVectorCD::datum(a);
+	ret.iiii[i] = typename GpuVectorCD::datum(b);
+      }
+      return ret;
+    }
+    //Real double
+    accelerator_inline GpuVectorRD operator()(double a){
+      GpuVectorRD ret; 
+      for(int i=0;i<GpuVectorRD::N;i++){
+	ret.rrrr[i] = typename GpuVectorRD::datum(a);
+      }
+      return ret;
+    }
+    //Integer
+    accelerator_inline GpuVectorI operator()(Integer a){
+      GpuVectorI ret;
+      for(int i=0;i<GpuVectorI::N;i++){
+	ret.rrrr[i] = typename GpuVectorI::datum(a);
+      }
+      return ret;
+    }
+  };
+
+  struct Vstore{
+    template<int N,class datum,class P>
+    accelerator_inline void operator()(GpuVector<N,datum> a, P* Fp){
+      GpuVector<N,datum> *vF = (GpuVector<N,datum> *)Fp;
+      *vF = a;
+    }
+    template<int N,class datum,class P>
+    accelerator_inline void operator()(GpuComplexVector<N,datum> a, P* Fp){
+      GpuComplexVector<N,datum> *vF = (GpuComplexVector<N,datum> *)Fp;
+      *vF = a;
+    }
+  };
+
+  struct Vstream{
+    template<int N,class datum, class P>
+    accelerator_inline void operator()(P* F,GpuVector<N,datum> a){
+      GpuVector<N,datum> *vF = (GpuVector<N,datum> *)F;
+      *vF = a;
+    }
+    template<int N,class datum, class P>
+    accelerator_inline void operator()(P* F,GpuComplexVector<N,datum> a){
+      GpuComplexVector<N,datum> *vF = (GpuComplexVector<N,datum> *)F;
+      *vF = a;
+    }
+  };
+
+  struct Vset{
+    // Complex float 
+    accelerator_inline GpuVectorCF operator()(Grid::ComplexF *a){
+      typedef GpuVectorCF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = vec::datum(a[i].real());
+	ret.iiii[i] = vec::datum(a[i].imag());
+      }
+      return ret;
+    }
+    // Complex double 
+    accelerator_inline GpuVectorCD operator()(Grid::ComplexD *a){
+      typedef GpuVectorCD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = vec::datum(a[i].real());
+	ret.iiii[i] = vec::datum(a[i].imag());
+      }
+      return ret;
+    }
+    // Real float 
+    accelerator_inline GpuVectorRF operator()(float *a){
+      typedef GpuVectorRF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = vec::datum(a[i]);
+      }
+      return ret;
+    }
+    // Real double
+    accelerator_inline GpuVectorRD operator()(double *a){
+      typedef GpuVectorRD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = vec::datum(a[i]);
+      }
+      return ret;
+    }
+    // Integer
+    accelerator_inline GpuVectorI operator()(Integer *a){
+      typedef GpuVectorI vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = vec::datum(a[i]);
+      }
+      return ret;
+    }
+  };
+
+  template <typename Out_type, typename In_type>
+  struct Reduce{
+    //Need templated class to overload output type
+    //General form must generate error if compiled
+    accelerator_inline Out_type operator()(In_type in){
+      printf("Error, using wrong Reduce function\n");
+      exit(1);
+      return 0;
+    }
+  };
+
+  /////////////////////////////////////////////////////
+  // Arithmetic operations
+  /////////////////////////////////////////////////////
+  struct Sum{
+    //Real float
+    accelerator_inline GpuVectorRF operator()(GpuVectorRF a,GpuVectorRF b){
+      return a+b;
+    }
+    accelerator_inline GpuVectorRD operator()(GpuVectorRD a,GpuVectorRD b){
+      return a+b;
+    }
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF a,GpuVectorCF b){
+      return a+b;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD a,GpuVectorCD b){
+      return a+b;
+    }
+    accelerator_inline GpuVectorI operator()(GpuVectorI a,GpuVectorI b){
+      return a+b;
+    }
+  };
+
+  struct Sub{
+    accelerator_inline GpuVectorRF operator()(GpuVectorRF a,GpuVectorRF b){
+      return a-b;
+    }
+    accelerator_inline GpuVectorRD operator()(GpuVectorRD a,GpuVectorRD b){
+      return a-b;
+    }
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF a,GpuVectorCF b){
+      return a-b;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD a,GpuVectorCD b){
+      return a-b;
+    }
+    accelerator_inline GpuVectorI operator()(GpuVectorI a,GpuVectorI b){
+      return a-b;
+    }
+  };
+
+  struct MultRealPart{
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF a,GpuVectorCF b){
+      typedef GpuVectorCF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = a.rrrr[i]*b.rrrr[i];
+	ret.iiii[i] = a.rrrr[i]*b.iiii[i];
+      }
+      return ret;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD a,GpuVectorCD b){
+      typedef GpuVectorCD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = a.rrrr[i]*b.rrrr[i];
+	ret.iiii[i] = a.rrrr[i]*b.iiii[i];
+      }
+      return ret;
+    }
+  };
+
+  struct MaddRealPart{
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF a,GpuVectorCF b,GpuVectorCF c){
+      typedef GpuVectorCF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = a.rrrr[i]*b.rrrr[i]+c.rrrr[i];
+	ret.iiii[i] = a.rrrr[i]*b.iiii[i]+c.iiii[i];
+      }
+      return ret;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD a,GpuVectorCD b,GpuVectorCD c){
+      typedef GpuVectorCD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = a.rrrr[i]*b.rrrr[i]+c.rrrr[i];
+	ret.iiii[i] = a.rrrr[i]*b.iiii[i]+c.iiii[i];
+      }
+      return ret;
+    }
+  };
+
+  struct MultComplex{
+
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF a,GpuVectorCF b){
+      return a*b;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD a,GpuVectorCD b){
+      return a*b;
+    }
+  };
+
+  struct Mult{
+    accelerator_inline void mac(GpuVectorRF &a, GpuVectorRF b, GpuVectorRF c){
+      a= a+b*c;
+    }
+    accelerator_inline void mac(GpuVectorRD &a, GpuVectorRD b, GpuVectorRD c){
+      a= a+b*c;
+    }
+    // Real float
+    accelerator_inline GpuVectorRF operator()(GpuVectorRF a, GpuVectorRF b){
+      return a*b;
+    }
+    // Real double
+    accelerator_inline GpuVectorRD operator()(GpuVectorRD a, GpuVectorRD b){
+      return a*b;
+    }
+    accelerator_inline GpuVectorI operator()(GpuVectorI a, GpuVectorI b){
+      return a*b;
+    }
+  };
+
+  struct Div{
+    // Real float
+    accelerator_inline GpuVectorRF operator()(GpuVectorRF a, GpuVectorRF b){
+      return a/b;
+    }
+    accelerator_inline GpuVectorRD operator()(GpuVectorRD a, GpuVectorRD b){
+      return a/b;
+    }
+    accelerator_inline GpuVectorI operator()(GpuVectorI a, GpuVectorI b){
+      return a/b;
+    }
+
+    // Danger -- element wise divide fro complex, not complex div. 
+    // See Grid_vector_types.h lines around 735, applied after "toReal"
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF a, GpuVectorCF b){
+      return a/b;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD a, GpuVectorCD b){
+      return a/b;
+    }
+  };
+
+
+  struct Conj{
+    // Complex single
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF in){
+      typedef GpuVectorCF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = in.rrrr[i];
+	ret.iiii[i] =-in.iiii[i];
+      }
+      return ret;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD in){
+      typedef GpuVectorCD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = in.rrrr[i];
+	ret.iiii[i] =-in.iiii[i];
+      }
+      return ret;
+    }
+  };
+
+  struct TimesMinusI{
+    //Complex single
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF in){
+      typedef GpuVectorCF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = in.iiii[i];
+	ret.iiii[i] =-in.rrrr[i];
+      }
+      return ret;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD in){
+      typedef GpuVectorCD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] = in.iiii[i];
+	ret.iiii[i] =-in.rrrr[i];
+      }
+      return ret;
+    }
+  };
+
+  struct TimesI{
+    //Complex single
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF in){
+      typedef GpuVectorCF vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] =-in.iiii[i];
+	ret.iiii[i] = in.rrrr[i];
+      }
+      return ret;
+    }
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD in){
+      typedef GpuVectorCD vec;
+      vec ret;
+      for(int i=0;i<vec::N;i++){
+	ret.rrrr[i] =-in.iiii[i];
+	ret.iiii[i] = in.rrrr[i];
+      }
+      return ret;
+    }
+  };
+
+  struct Permute{
+
+    template <int n,int _N, class _datum >
+    static accelerator_inline GpuVector<_N,_datum> PermuteN(GpuVector<_N,_datum> &in) {   
+      typedef GpuVector<_N,_datum> vec;
+      vec out;					
+      unsigned int _mask = vec::N >> (n + 1);	
+      for(int i=0;i<vec::N;i++) {
+	out.rrrr[i] = in.rrrr[i^_mask];
+      }
+      return out;	
+    }
+    template <int n,int _N, class _datum >
+    static accelerator_inline GpuComplexVector<_N,_datum> PermuteN(GpuComplexVector<_N,_datum> &in) {   
+      typedef GpuComplexVector<_N,_datum> vec;
+      vec out;					
+      unsigned int _mask = vec::N >> (n + 1);	
+      for(int i=0;i<vec::N;i++) {
+	out.rrrr[i] = in.rrrr[i^_mask];
+	out.iiii[i] = in.iiii[i^_mask];
+      }
+      return out;	
+    }
+    
+    template <typename vec>  static accelerator_inline vec Permute0(vec in) { return PermuteN<0,vec::N,typename vec::datum>(in);  }
+    template <typename vec>  static accelerator_inline vec Permute1(vec in) { return PermuteN<1,vec::N,typename vec::datum>(in);  }
+    template <typename vec>  static accelerator_inline vec Permute2(vec in) { return PermuteN<2,vec::N,typename vec::datum>(in);  }
+    template <typename vec>  static accelerator_inline vec Permute3(vec in) { return PermuteN<3,vec::N,typename vec::datum>(in);  }
+    
+  };
+  
+  struct PrecisionChange {
+
+    ////////////////////////////////////////////////////////////////////////////////////
+    // Single / Half
+    ////////////////////////////////////////////////////////////////////////////////////
+     static accelerator_inline GpuVectorCH StoH (GpuVectorCF a,GpuVectorCF b) {
+      int N = GpuVectorCF::N;
+      GpuVectorCH h;
+      for(int i=0;i<N;i++) {
+        h.rrrr[i  ] = float2half(a.rrrr[i]);
+        h.iiii[i  ] = float2half(a.iiii[i]);
+	h.rrrr[i+N] = float2half(b.rrrr[i]);
+	h.iiii[i+N] = float2half(b.iiii[i]);
+      }
+      return h;
+    }
+    static accelerator_inline void  HtoS (GpuVectorCH h,GpuVectorCF &sa,GpuVectorCF &sb) {
+      int N = GpuVectorCF::N;
+      for(int i=0;i<N;i++) {
+	sa.rrrr[i] = half2float(h.rrrr[i  ]);
+	sa.iiii[i] = half2float(h.iiii[i  ]);
+	sb.rrrr[i] = half2float(h.rrrr[i+N]);
+	sb.iiii[i] = half2float(h.iiii[i+N]);
+      }
+    }
+    static accelerator_inline GpuVectorRH StoH (GpuVectorRF a,GpuVectorRF b) {
+      int N = GpuVectorRF::N;
+      GpuVectorRH h;
+      for(int i=0;i<N;i++) {
+        h.rrrr[i  ] = float2half(a.rrrr[i]);
+	h.rrrr[i+N] = float2half(b.rrrr[i]);
+      }
+      return h;
+    }
+    static accelerator_inline void  HtoS (GpuVectorRH h,GpuVectorRF &sa,GpuVectorRF &sb) {
+      int N = GpuVectorRF::N;
+      for(int i=0;i<N;i++) {
+	sa.rrrr[i] = half2float(h.rrrr[i  ]);
+	sb.rrrr[i] = half2float(h.rrrr[i+N]);
+      }
+    }
+
+    ////////////////////////////////////////////////////////////////////////////////////
+    // Double Single
+    ////////////////////////////////////////////////////////////////////////////////////
+    static accelerator_inline GpuVectorCF DtoS (GpuVectorCD a,GpuVectorCD b) {
+      int N = GpuVectorCD::N;
+      GpuVectorCF h;
+      for(int i=0;i<N;i++) {
+        h.rrrr[i  ] = a.rrrr[i];
+        h.iiii[i  ] = a.iiii[i];
+	h.rrrr[i+N] = b.rrrr[i];
+	h.iiii[i+N] = b.iiii[i];
+      }
+      return h;
+    }
+
+    static accelerator_inline void  StoD (GpuVectorCF h,GpuVectorCD &sa,GpuVectorCD &sb) {
+      int N = GpuVectorCD::N;
+      for(int i=0;i<N;i++) {
+	sa.rrrr[i] = h.rrrr[i  ];
+	sa.iiii[i] = h.iiii[i  ];
+	sb.rrrr[i] = h.rrrr[i+N];
+	sb.iiii[i] = h.iiii[i+N];
+      }
+    }
+
+    static accelerator_inline GpuVectorRF DtoS (GpuVectorRD a,GpuVectorRD b) {
+      int N = GpuVectorRD::N;
+      GpuVectorRF h;
+      for(int i=0;i<N;i++) {
+        h.rrrr[i  ] = a.rrrr[i];
+	h.rrrr[i+N] = b.rrrr[i];
+      }
+      return h;
+    }
+
+    static accelerator_inline void  StoD (GpuVectorRF h,GpuVectorRD &sa,GpuVectorRD &sb) {
+      int N = GpuVectorRD::N;
+      for(int i=0;i<N;i++) {
+	sa.rrrr[i] = h.rrrr[i  ];
+	sb.rrrr[i] = h.rrrr[i+N];
+      }
+    }
+
+    ////////////////////////////////////////////////////////////////////////////////////
+    // Double Half
+    ////////////////////////////////////////////////////////////////////////////////////
+    static accelerator_inline GpuVectorCH DtoH (GpuVectorCD a,GpuVectorCD b,GpuVectorCD c,GpuVectorCD d) {
+      GpuVectorCF sa,sb;
+      sa = DtoS(a,b);
+      sb = DtoS(c,d);
+      return StoH(sa,sb);
+    }
+    static accelerator_inline void HtoD (GpuVectorCH h,GpuVectorCD &a,GpuVectorCD &b,GpuVectorCD &c,GpuVectorCD &d) {
+      GpuVectorCF sa,sb;
+      HtoS(h,sa,sb);
+      StoD(sa,a,b);
+      StoD(sb,c,d);
+    }
+    static accelerator_inline GpuVectorRH DtoH (GpuVectorRD a,GpuVectorRD b,GpuVectorRD c,GpuVectorRD d) {
+      GpuVectorRF sa,sb;
+      sa = DtoS(a,b);
+      sb = DtoS(c,d);
+      return StoH(sa,sb);
+    }
+    static accelerator_inline void HtoD (GpuVectorRH h,GpuVectorRD &a,GpuVectorRD &b,GpuVectorRD &c,GpuVectorRD &d) {
+      GpuVectorRF sa,sb;
+      HtoS(h,sa,sb);
+      StoD(sa,a,b);
+      StoD(sb,c,d);
+    }
+  };
+
+struct Exchange{
+
+  template <int n,int _N, class _datum >
+  static accelerator_inline void ExchangeN(GpuVector<_N,_datum> &out1,
+					   GpuVector<_N,_datum> &out2,
+					   GpuVector<_N,_datum> &in1,
+					   GpuVector<_N,_datum> &in2 )
+  {   
+    typedef GpuVector<_N,_datum> vec;
+    unsigned int mask = vec::N >> (n + 1);
+    for(int i=0;i<vec::N;i++) {
+      int j1 = i&(~mask);
+      if  ( (i&mask) == 0 ) { out1.rrrr[i]=in1.rrrr[j1];}
+      else                  { out1.rrrr[i]=in2.rrrr[j1];}
+      int j2 = i|mask;
+      if  ( (i&mask) == 0 ) { out2.rrrr[i]=in1.rrrr[j2];}
+      else                  { out2.rrrr[i]=in2.rrrr[j2];}
+    }      
+  }
+  template <int n,int _N, class _datum >
+  static accelerator_inline void ExchangeN(GpuComplexVector<_N,_datum> &out1,
+					   GpuComplexVector<_N,_datum> &out2,
+					   GpuComplexVector<_N,_datum> &in1,
+					   GpuComplexVector<_N,_datum> &in2 )
+  {   
+    typedef GpuComplexVector<_N,_datum> vec;
+    unsigned int mask = vec::N >> (n + 1);
+    for(int i=0;i<vec::N;i++) {
+      int j1 = i&(~mask);
+      if  ( (i&mask) == 0 ) {
+	out1.rrrr[i]=in1.rrrr[j1];
+	out1.iiii[i]=in1.iiii[j1];
+      }
+      else                  {
+	out1.rrrr[i]=in2.rrrr[j1];
+	out1.iiii[i]=in2.iiii[j1];
+      }
+      int j2 = i|mask;
+      if  ( (i&mask) == 0 ) {
+	out2.rrrr[i]=in1.rrrr[j2];
+	out2.iiii[i]=in1.iiii[j2];
+      }
+      else                  {
+	out2.rrrr[i]=in2.rrrr[j2];
+	out2.iiii[i]=in2.iiii[j2];
+      }
+    }      
+  }
+  template <typename vec>
+  static accelerator_inline void Exchange0(vec &out1,vec &out2,vec &in1,vec &in2){
+    ExchangeN<0>(out1,out2,in1,in2);
+  };
+  template <typename vec>
+  static accelerator_inline void Exchange1(vec &out1,vec &out2,vec &in1,vec &in2){
+    ExchangeN<1>(out1,out2,in1,in2);
+  };
+  template <typename vec>
+  static accelerator_inline void Exchange2(vec &out1,vec &out2,vec &in1,vec &in2){
+    ExchangeN<2>(out1,out2,in1,in2);
+  };
+  template <typename vec>
+  static accelerator_inline void Exchange3(vec &out1,vec &out2,vec &in1,vec &in2){
+    ExchangeN<3>(out1,out2,in1,in2);
+  };
+
+};
+
+struct Rotate{
+
+  template <int n, typename vec> static accelerator_inline vec tRotate(vec in){
+    return rotate(in, n);
+  }
+    
+  template <int _N, class _datum >
+  static accelerator_inline GpuComplexVector<_N,_datum> rotate_template(GpuComplexVector<_N,_datum> &in, int n)
+  {
+    typedef GpuComplexVector<_N,_datum> vec;
+    vec out;
+    for(int i=0;i<vec::N;i++){
+      out.rrrr[i] = in.rrrr[(i + n)%vec::N];
+      out.iiii[i] = in.iiii[(i + n)%vec::N];
+    }
+    return out;
+  }
+
+  template <int _N, class _datum >
+  static accelerator_inline GpuVector<_N,_datum> rotate_template(GpuVector<_N,_datum> &in, int n)
+  {
+    typedef GpuVector<_N,_datum> vec;
+    vec out;
+    for(int i=0;i<vec::N;i++){
+      out.rrrr[i] = in.rrrr[(i + n)%vec::N];
+    }
+    return out;
+  }
+
+  typedef GpuVectorRH  SIMD_Htype; // Single precision type
+  typedef GpuVectorRF  SIMD_Ftype; // Single precision type
+  typedef GpuVectorRD  SIMD_Dtype; // Double precision type
+  typedef GpuVectorI   SIMD_Itype; // Integer type
+
+  typedef GpuVectorCH  SIMD_CHtype; // Single precision type
+  typedef GpuVectorCF  SIMD_CFtype; // Single precision type
+  typedef GpuVectorCD  SIMD_CDtype; // Double precision type
+
+  static accelerator_inline GpuVectorRH rotate(GpuVectorRH in, int n){ return rotate_template(in,n);}
+  static accelerator_inline GpuVectorRF rotate(GpuVectorRF in, int n){ return rotate_template(in,n);}
+  static accelerator_inline GpuVectorRD rotate(GpuVectorRD in, int n){ return rotate_template(in,n);}
+  static accelerator_inline GpuVectorI  rotate(GpuVectorI  in, int n){ return rotate_template(in,n);}
+  static accelerator_inline GpuVectorCH rotate(GpuVectorCH in, int n){ return rotate_template(in,n/2);} // Measure in complex not float
+  static accelerator_inline GpuVectorCF rotate(GpuVectorCF in, int n){ return rotate_template(in,n/2);}
+  static accelerator_inline GpuVectorCD rotate(GpuVectorCD in, int n){ return rotate_template(in,n/2);}
+
+};
+
+//////////////////////////////////////////////
+// Some Template specialization
+
+  //Complex float Reduce
+  template<>
+  accelerator_inline Grid::ComplexF 
+  Reduce<Grid::ComplexF, GpuVectorCF>::operator()(GpuVectorCF in)
+  {
+    Grid::ComplexF greduce(in.rrrr[0],in.iiii[0]);
+    for(int i=1;i<GpuVectorCF::N;i++) {
+      greduce = greduce+Grid::ComplexF(in.rrrr[i],in.iiii[i]);
+    }
+    return greduce;
+  }
+
+  template<>
+  accelerator_inline Grid::ComplexD
+  Reduce<Grid::ComplexD, GpuVectorCD>::operator()(GpuVectorCD in)
+  {
+    Grid::ComplexD greduce(in.rrrr[0],in.iiii[0]);
+    for(int i=1;i<GpuVectorCD::N;i++) {
+      greduce = greduce+ Grid::ComplexD(in.rrrr[i],in.iiii[i]);
+    }
+    return greduce;
+  }
+
+  // Real
+  template<>
+  accelerator_inline Grid::RealF 
+  Reduce<RealF, GpuVectorRF>::operator()(GpuVectorRF in)
+  {
+    RealF ret = in.rrrr[0];
+    for(int i=1;i<GpuVectorRF::N;i++) {
+      ret = ret+in.rrrr[i];
+    }
+    return ret;
+  }
+
+  template<>
+  accelerator_inline Grid::RealD 
+  Reduce<RealD, GpuVectorRD>::operator()(GpuVectorRD in)
+  {
+    RealD ret = in.rrrr[0];
+    for(int i=1;i<GpuVectorRD::N;i++) {
+      ret = ret+in.rrrr[i];
+    }
+    return ret;
+  }
+
+  template<>
+  accelerator_inline Integer
+  Reduce<Integer, GpuVectorI>::operator()(GpuVectorI in)
+  {
+    Integer ret = in.rrrr[0];
+    for(int i=1;i<GpuVectorI::N;i++) {
+      ret = ret+in.rrrr[i];
+    }
+    return ret;
+  }
+
+}// End optimizatoin
+
+//////////////////////////////////////////////////////////////////////////////////////
+// Here assign types 
+//////////////////////////////////////////////////////////////////////////////////////
+  typedef GpuVectorRH  SIMD_Htype; // Single precision type
+  typedef GpuVectorRF  SIMD_Ftype; // Single precision type
+  typedef GpuVectorRD  SIMD_Dtype; // Double precision type
+  typedef GpuVectorI   SIMD_Itype; // Integer type
+
+  typedef GpuVectorCH  SIMD_CHtype; // Single precision type
+  typedef GpuVectorCF  SIMD_CFtype; // Single precision type
+  typedef GpuVectorCD  SIMD_CDtype; // Double precision type
+
+  // prefetch utilities
+  accelerator_inline void v_prefetch0(int size, const char *ptr){};
+  accelerator_inline void prefetch_HINT_T0(const char *ptr){};
+
+  // Function name aliases
+  typedef Optimization::Vsplat   VsplatSIMD;
+  typedef Optimization::Vstore   VstoreSIMD;
+  typedef Optimization::Vset     VsetSIMD;
+  typedef Optimization::Vstream  VstreamSIMD;
+  template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
+
+  // Arithmetic operations
+  typedef Optimization::Sum         SumSIMD;
+  typedef Optimization::Sub         SubSIMD;
+  typedef Optimization::Div         DivSIMD;
+  typedef Optimization::Mult        MultSIMD;
+  typedef Optimization::MultComplex MultComplexSIMD;
+  typedef Optimization::MultRealPart MultRealPartSIMD;
+  typedef Optimization::MaddRealPart MaddRealPartSIMD;
+  typedef Optimization::Conj        ConjSIMD;
+  typedef Optimization::TimesMinusI TimesMinusISIMD;
+  typedef Optimization::TimesI      TimesISIMD;
+
+}

Grid/simd/Grid_gpu_vec.h

@@ -38,7 +38,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_HIP
 #include <hip/hip_fp16.h>
 #endif
-#ifdef GRID_SYCL
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
 namespace Grid {
   typedef struct { uint16_t x;} half;
   typedef struct { half   x; half   y;} half2;
@@ -486,7 +486,7 @@ namespace Optimization {
 
   struct TimesMinusI{
     //Complex single
-    accelerator_inline GpuVectorCF operator()(GpuVectorCF in,GpuVectorCF dummy){
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF in){
       typedef GpuVectorCF vec;
       vec ret;
       for(int i=0;i<vec::N;i++){
@@ -495,7 +495,7 @@ namespace Optimization {
       }
       return ret;
     }
-    accelerator_inline GpuVectorCD operator()(GpuVectorCD in,GpuVectorCD dummy){
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD in){
       typedef GpuVectorCD vec;
       vec ret;
       for(int i=0;i<vec::N;i++){
@@ -508,7 +508,7 @@ namespace Optimization {
 
   struct TimesI{
     //Complex single
-    accelerator_inline GpuVectorCF operator()(GpuVectorCF in,GpuVectorCF dummy){
+    accelerator_inline GpuVectorCF operator()(GpuVectorCF in){
       typedef GpuVectorCF vec;
       vec ret;
       for(int i=0;i<vec::N;i++){
@@ -517,7 +517,7 @@ namespace Optimization {
       }
       return ret;
     }
-    accelerator_inline GpuVectorCD operator()(GpuVectorCD in,GpuVectorCD dummy){
+    accelerator_inline GpuVectorCD operator()(GpuVectorCD in){
       typedef GpuVectorCD vec;
       vec ret;
       for(int i=0;i<vec::N;i++){

Grid/simd/Grid_qpx.h

@@ -356,7 +356,7 @@ struct Conj{
   
 struct TimesMinusI{
   //Complex double
-  inline vector4double operator()(vector4double v, vector4double ret){
+  inline vector4double operator()(vector4double v){
     return vec_xxcpnmadd(v, (vector4double){1., 1., 1., 1.},
 			 (vector4double){0., 0., 0., 0.});
   }
@@ -367,7 +367,7 @@ struct TimesMinusI{
   
 struct TimesI{
   //Complex double
-  inline vector4double operator()(vector4double v, vector4double ret){
+  inline vector4double operator()(vector4double v){
     return vec_xxcpnmadd(v, (vector4double){-1., -1., -1., -1.},
 			 (vector4double){0., 0., 0., 0.});
   }

Grid/simd/Grid_sse4.h

@@ -273,27 +273,25 @@ struct Conj{
 
 struct TimesMinusI{
   //Complex single
-  inline __m128 operator()(__m128 in, __m128 ret){
+  inline __m128 operator()(__m128 in){
     __m128 tmp =_mm_addsub_ps(_mm_setzero_ps(),in); // r,-i
     return _mm_shuffle_ps(tmp,tmp,_MM_SELECT_FOUR_FOUR(2,3,0,1));
   }
   //Complex double
-  inline __m128d operator()(__m128d in, __m128d ret){
+  inline __m128d operator()(__m128d in){
     __m128d tmp =_mm_addsub_pd(_mm_setzero_pd(),in); // r,-i
     return _mm_shuffle_pd(tmp,tmp,0x1);
   }
-
-
 };
 
 struct TimesI{
   //Complex single
-  inline __m128 operator()(__m128 in, __m128 ret){
+  inline __m128 operator()(__m128 in){
     __m128 tmp =_mm_shuffle_ps(in,in,_MM_SELECT_FOUR_FOUR(2,3,0,1));
     return _mm_addsub_ps(_mm_setzero_ps(),tmp); // r,-i
   }
   //Complex double
-  inline __m128d operator()(__m128d in, __m128d ret){
+  inline __m128d operator()(__m128d in){
     __m128d tmp = _mm_shuffle_pd(in,in,0x1);
     return _mm_addsub_pd(_mm_setzero_pd(),tmp); // r,-i
   }

Grid/simd/Grid_vector_types.h

@@ -110,11 +110,10 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
 #ifdef GPU_VEC
 #include "Grid_gpu_vec.h"
 #endif
-/*
+
-#ifdef GEN
+#ifdef GPU_RRII
-#include "Grid_generic.h"
+#include "Grid_gpu_rrii.h"
 #endif
-*/
 
 #ifdef GEN
   #if defined(A64FX) || defined(A64FXFIXEDSIZE) // breakout A64FX SVE ACLE here
@@ -131,7 +130,6 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
       #include "Grid_a64fx-fixedsize.h"
     #endif
   #else
-    //#pragma message("building GEN") // generic
     #include "Grid_generic.h"
   #endif
 #endif
@@ -150,23 +148,6 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
   #endif
 #endif
 
-/*
-#ifdef A64FXVLA
-#pragma message("building A64FX VLA")
-#if defined(ARMCLANGCOMPAT)
-  #pragma message("applying data types patch")
-#endif
-#include <arm_sve.h>
-#include "Grid_a64fx-2.h"
-#endif
-
-#ifdef A64FXVLS
-#pragma message("building A64FX VLS")
-#include <arm_sve.h>
-#include "Grid_a64fx-fixedsize.h"
-#endif
-*/
-
 #ifdef SSE4
 #include "Grid_sse4.h"
 #endif
@@ -270,11 +251,13 @@ public:
   typedef Vector_type vector_type;
   typedef Scalar_type scalar_type;
 
+  /*
   typedef union conv_t_union {
     Vector_type v;
     Scalar_type s[sizeof(Vector_type) / sizeof(Scalar_type)];
     accelerator_inline conv_t_union(){};
   } conv_t;
+  */
   
   Vector_type v;
 
@@ -555,15 +538,13 @@ public:
   template <class functor>
   friend accelerator_inline Grid_simd SimdApply(const functor &func, const Grid_simd &v) {
     Grid_simd ret;
-    Grid_simd::conv_t conv;
     Grid_simd::scalar_type s;
 
-    conv.v = v.v;
     for (int i = 0; i < Nsimd(); i++) {
-      s = conv.s[i];
+      s = v.getlane(i);
-      conv.s[i] = func(s);
+      s = func(s);
+      ret.putlane(s,i);
     }
-    ret.v = conv.v;
     return ret;
   }
   template <class functor>
@@ -571,18 +552,14 @@ public:
                                          const Grid_simd &x,
                                          const Grid_simd &y) {
     Grid_simd ret;
-    Grid_simd::conv_t cx;
-    Grid_simd::conv_t cy;
     Grid_simd::scalar_type sx,sy;
 
-    cx.v = x.v;
-    cy.v = y.v;
     for (int i = 0; i < Nsimd(); i++) {
-      sx = cx.s[i];
+      sx = x.getlane(i);
-      sy = cy.s[i];
+      sy = y.getlane(i);
-      cx.s[i] = func(sx,sy);
+      sx = func(sx,sy);
+      ret.putlane(sx,i);
     }
-    ret.v = cx.v;
     return ret;
   }
   ///////////////////////
@@ -645,15 +622,36 @@ public:
   ///////////////////////////////
   // Getting single lanes
   ///////////////////////////////
-  accelerator_inline Scalar_type getlane(int lane) {
+#ifdef GPU_RRII
+  template <class S = Scalar_type,IfComplex<S> = 0>
+  accelerator_inline Scalar_type getlane(int lane) const {
+    return Scalar_type(v.rrrr[lane],v.iiii[lane]);
+  }
+  template <class S = Scalar_type,IfComplex<S> = 0>
+  accelerator_inline void putlane(const Scalar_type &_S, int lane){
+    v.rrrr[lane] = real(_S);
+    v.iiii[lane] = imag(_S);
+  }
+  template <class S = Scalar_type,IfNotComplex<S> = 0>
+  accelerator_inline Scalar_type getlane(int lane) const {
+    return ((S*)&v)[lane];
+  }
+  template <class S = Scalar_type,IfNotComplex<S> = 0>
+  accelerator_inline void putlane(const S &_S, int lane){
+    ((Scalar_type*)&v)[lane] = _S;
+  }
+#else // Can pun to an array of complex
+  accelerator_inline Scalar_type getlane(int lane) const {
     return ((Scalar_type*)&v)[lane];
   }
-
   accelerator_inline void putlane(const Scalar_type &S, int lane){
     ((Scalar_type*)&v)[lane] = S;
   }
+#endif
+
 };  // end of Grid_simd class definition
 
+
 ///////////////////////////////
 // Define available types
 ///////////////////////////////
@@ -663,7 +661,7 @@ typedef Grid_simd<double , SIMD_Dtype> vRealD;
 typedef Grid_simd<Integer, SIMD_Itype> vInteger;
 typedef Grid_simd<uint16_t,SIMD_Htype> vRealH;
 
-#ifdef GPU_VEC
+#if defined(GPU_VEC) || defined(GPU_RRII)
 typedef Grid_simd<complex<uint16_t>, SIMD_CHtype> vComplexH;
 typedef Grid_simd<complex<float>   , SIMD_CFtype> vComplexF;
 typedef Grid_simd<complex<double>  , SIMD_CDtype> vComplexD;
@@ -763,6 +761,7 @@ accelerator_inline void vsplat(Grid_simd<S, V> &ret, NotEnableIf<is_complex<S>,
 }
 //////////////////////////
 
+
 ///////////////////////////////////////////////
 // Initialise to 1,0,i for the correct types
 ///////////////////////////////////////////////
@@ -907,34 +906,6 @@ accelerator_inline Grid_simd<S, V> fxmac(Grid_simd<S, V> a, Grid_simd<S, V> b, G
 // ----------------------------------------------
 
 
-// Distinguish between complex types and others
-template <class S, class V, IfComplex<S> = 0>
-accelerator_inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
-  typedef Grid_simd<S, V> simd;
-
-  simd ret;
-  simd den;
-  typename simd::conv_t conv;
-
-  ret = a * conjugate(b) ;
-  den = b * conjugate(b) ;
-
-  // duplicates real part
-  auto real_den  = toReal(den);
-  simd zden;
-  memcpy((void *)&zden.v,(void *)&real_den.v,sizeof(zden));
-  ret.v=binary<V>(ret.v, zden.v, DivSIMD());
-  return ret;
-};
-
-// Real/Integer types
-template <class S, class V, IfNotComplex<S> = 0>
-accelerator_inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
-  Grid_simd<S, V> ret;
-  ret.v = binary<V>(a.v, b.v, DivSIMD());
-  return ret;
-};
-
 ///////////////////////
 // Conjugate
 ///////////////////////
@@ -959,30 +930,29 @@ accelerator_inline Grid_simd<S, V> adj(const Grid_simd<S, V> &in) {
 ///////////////////////
 template <class S, class V, IfComplex<S> = 0>
 accelerator_inline void timesMinusI(Grid_simd<S, V> &ret, const Grid_simd<S, V> &in) {
-  ret.v = binary<V>(in.v, ret.v, TimesMinusISIMD());
+  ret.v = unary<V>(in.v, TimesMinusISIMD());
 }
 template <class S, class V, IfComplex<S> = 0>
 accelerator_inline Grid_simd<S, V> timesMinusI(const Grid_simd<S, V> &in) {
   Grid_simd<S, V> ret;
-  timesMinusI(ret, in);
+  ret.v=unary<V>(in.v, TimesMinusISIMD());
   return ret;
 }
 template <class S, class V, IfNotComplex<S> = 0>
 accelerator_inline Grid_simd<S, V> timesMinusI(const Grid_simd<S, V> &in) {
   return in;
 }
-
 ///////////////////////
 // timesI
 ///////////////////////
 template <class S, class V, IfComplex<S> = 0>
 accelerator_inline void timesI(Grid_simd<S, V> &ret, const Grid_simd<S, V> &in) {
-  ret.v = binary<V>(in.v, ret.v, TimesISIMD());
+  ret.v = unary<V>(in.v, TimesISIMD());
 }
 template <class S, class V, IfComplex<S> = 0>
 accelerator_inline Grid_simd<S, V> timesI(const Grid_simd<S, V> &in) {
   Grid_simd<S, V> ret;
-  timesI(ret, in);
+  ret.v= unary<V>(in.v, TimesISIMD());
   return ret;
 }
 template <class S, class V, IfNotComplex<S> = 0>
@@ -990,6 +960,35 @@ accelerator_inline Grid_simd<S, V> timesI(const Grid_simd<S, V> &in) {
   return in;
 }
 
+
+// Distinguish between complex types and others
+template <class S, class V, IfComplex<S> = 0>
+accelerator_inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
+  typedef Grid_simd<S, V> simd;
+
+  simd ret;
+  simd den;
+
+  ret = a * conjugate(b) ;
+  den = b * conjugate(b) ;
+
+  // duplicates real part
+  auto real_den  = toReal(den);
+  simd zden;
+  memcpy((void *)&zden.v,(void *)&real_den.v,sizeof(zden));
+  ret.v=binary<V>(ret.v, zden.v, DivSIMD());
+  return ret;
+};
+
+// Real/Integer types
+template <class S, class V, IfNotComplex<S> = 0>
+accelerator_inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
+  Grid_simd<S, V> ret;
+  ret.v = binary<V>(a.v, b.v, DivSIMD());
+  return ret;
+};
+
+
 /////////////////////
 // Inner, outer
 /////////////////////
@@ -1021,12 +1020,12 @@ template <class Csimd>  // must be a real arg
 accelerator_inline typename toRealMapper<Csimd>::Realified toReal(const Csimd &in) {
   typedef typename toRealMapper<Csimd>::Realified Rsimd;
   Rsimd ret;
-  typename Rsimd::conv_t conv;
+  int j=0;
-  memcpy((void *)&conv.v,(void *)&in.v,sizeof(conv.v));
   for (int i = 0; i < Rsimd::Nsimd(); i += 2) {
-    conv.s[i + 1] = conv.s[i];  // duplicate (r,r);(r,r);(r,r); etc...
+    auto s = real(in.getlane(j++));
+    ret.putlane(s,i);
+    ret.putlane(s,i+1);
   }
-  memcpy((void *)&ret.v,(void *)&conv.v,sizeof(ret.v));
   return ret;
 }
 
@@ -1039,23 +1038,24 @@ template <class Rsimd>  // must be a real arg
 accelerator_inline typename toComplexMapper<Rsimd>::Complexified toComplex(const Rsimd &in) {
 
   typedef typename toComplexMapper<Rsimd>::Complexified   Csimd;
-  typename Rsimd::conv_t conv;  // address as real
+  typedef typename Csimd::scalar_type scalar_type;
-
+  int j=0;
-  conv.v = in.v;
+  Csimd ret;
   for (int i = 0; i < Rsimd::Nsimd(); i += 2) {
-    assert(conv.s[i + 1] == conv.s[i]);
+    auto rr = in.getlane(i);
+    auto ri = in.getlane(i+1);
+    assert(rr==ri);
     // trap any cases where real was not duplicated
     // indicating the SIMD grids of real and imag assignment did not correctly
     // match
-    conv.s[i + 1] = 0.0;  // zero imaginary parts
+    scalar_type s(rr,0.0);
+    ret.putlane(s,j++);
   }
-  Csimd ret;
-  memcpy((void *)&ret.v,(void *)&conv.v,sizeof(ret.v));
   return ret;
 }
 
 
-accelerator_inline void precisionChange(vRealF    *out,vRealD    *in,int nvec)
+accelerator_inline void precisionChange(vRealF    *out,const vRealD    *in,int nvec)
 {
   assert((nvec&0x1)==0);
   for(int m=0;m*2<nvec;m++){
@@ -1063,7 +1063,7 @@ accelerator_inline void precisionChange(vRealF    *out,vRealD    *in,int nvec)
     out[m].v=Optimization::PrecisionChange::DtoS(in[n].v,in[n+1].v);
   }
 }
-accelerator_inline void precisionChange(vRealH    *out,vRealD    *in,int nvec)
+accelerator_inline void precisionChange(vRealH    *out,const vRealD    *in,int nvec)
 {
   assert((nvec&0x3)==0);
   for(int m=0;m*4<nvec;m++){
@@ -1071,7 +1071,7 @@ accelerator_inline void precisionChange(vRealH    *out,vRealD    *in,int nvec)
     out[m].v=Optimization::PrecisionChange::DtoH(in[n].v,in[n+1].v,in[n+2].v,in[n+3].v);
   }
 }
-accelerator_inline void precisionChange(vRealH    *out,vRealF    *in,int nvec)
+accelerator_inline void precisionChange(vRealH    *out,const vRealF    *in,int nvec)
 {
   assert((nvec&0x1)==0);
   for(int m=0;m*2<nvec;m++){
@@ -1079,7 +1079,7 @@ accelerator_inline void precisionChange(vRealH    *out,vRealF    *in,int nvec)
     out[m].v=Optimization::PrecisionChange::StoH(in[n].v,in[n+1].v);
   }
 }
-accelerator_inline void precisionChange(vRealD    *out,vRealF    *in,int nvec)
+accelerator_inline void precisionChange(vRealD    *out,const vRealF    *in,int nvec)
 {
   assert((nvec&0x1)==0);
   for(int m=0;m*2<nvec;m++){
@@ -1095,7 +1095,7 @@ accelerator_inline void precisionChange(vRealD    *out,vRealF    *in,int nvec)
     //  |                                                 ~~~~~~~^
   }
 }
-accelerator_inline void precisionChange(vRealD    *out,vRealH    *in,int nvec)
+accelerator_inline void precisionChange(vRealD    *out,const vRealH    *in,int nvec)
 {
   assert((nvec&0x3)==0);
   for(int m=0;m*4<nvec;m++){
@@ -1103,7 +1103,7 @@ accelerator_inline void precisionChange(vRealD    *out,vRealH    *in,int nvec)
     Optimization::PrecisionChange::HtoD(in[m].v,out[n].v,out[n+1].v,out[n+2].v,out[n+3].v);
   }
 }
-accelerator_inline void precisionChange(vRealF    *out,vRealH    *in,int nvec)
+accelerator_inline void precisionChange(vRealF    *out,const vRealH    *in,int nvec)
 {
   assert((nvec&0x1)==0);
   for(int m=0;m*2<nvec;m++){
@@ -1111,12 +1111,12 @@ accelerator_inline void precisionChange(vRealF    *out,vRealH    *in,int nvec)
     Optimization::PrecisionChange::HtoS(in[m].v,out[n].v,out[n+1].v);
   }
 }
-accelerator_inline void precisionChange(vComplexF *out,vComplexD *in,int nvec){ precisionChange((vRealF *)out,(vRealD *)in,nvec);}
+accelerator_inline void precisionChange(vComplexF *out,const vComplexD *in,int nvec){ precisionChange((vRealF *)out,(vRealD *)in,nvec);}
-accelerator_inline void precisionChange(vComplexH *out,vComplexD *in,int nvec){ precisionChange((vRealH *)out,(vRealD *)in,nvec);}
+accelerator_inline void precisionChange(vComplexH *out,const vComplexD *in,int nvec){ precisionChange((vRealH *)out,(vRealD *)in,nvec);}
-accelerator_inline void precisionChange(vComplexH *out,vComplexF *in,int nvec){ precisionChange((vRealH *)out,(vRealF *)in,nvec);}
+accelerator_inline void precisionChange(vComplexH *out,const vComplexF *in,int nvec){ precisionChange((vRealH *)out,(vRealF *)in,nvec);}
-accelerator_inline void precisionChange(vComplexD *out,vComplexF *in,int nvec){ precisionChange((vRealD *)out,(vRealF *)in,nvec);}
+accelerator_inline void precisionChange(vComplexD *out,const vComplexF *in,int nvec){ precisionChange((vRealD *)out,(vRealF *)in,nvec);}
-accelerator_inline void precisionChange(vComplexD *out,vComplexH *in,int nvec){ precisionChange((vRealD *)out,(vRealH *)in,nvec);}
+accelerator_inline void precisionChange(vComplexD *out,const vComplexH *in,int nvec){ precisionChange((vRealD *)out,(vRealH *)in,nvec);}
-accelerator_inline void precisionChange(vComplexF *out,vComplexH *in,int nvec){ precisionChange((vRealF *)out,(vRealH *)in,nvec);}
+accelerator_inline void precisionChange(vComplexF *out,const vComplexH *in,int nvec){ precisionChange((vRealF *)out,(vRealH *)in,nvec);}
 
 // Check our vector types are of an appropriate size.
 
@@ -1130,21 +1130,6 @@ static_assert(sizeof(SIMD_Ftype) == sizeof(SIMD_Itype), "SIMD vector lengths inc
 #endif
 #endif
 
-/////////////////////////////////////////
-// Some traits to recognise the types
-/////////////////////////////////////////
-template <typename T>
-struct is_simd : public std::false_type {};
-template <> struct is_simd<vRealF>     : public std::true_type {};
-template <> struct is_simd<vRealD>     : public std::true_type {};
-template <> struct is_simd<vRealH>     : public std::true_type {};
-template <> struct is_simd<vComplexF>  : public std::true_type {};
-template <> struct is_simd<vComplexD>  : public std::true_type {};
-template <> struct is_simd<vComplexH>  : public std::true_type {};
-template <> struct is_simd<vInteger>   : public std::true_type {};
-
-template <typename T> using IfSimd    = Invoke<std::enable_if<is_simd<T>::value, int> >;
-template <typename T> using IfNotSimd = Invoke<std::enable_if<!is_simd<T>::value, unsigned> >;
 
 NAMESPACE_END(Grid);
 

Grid/simd/Grid_vector_unops.h

@@ -29,8 +29,7 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 			   /*  END LEGAL */
-#ifndef GRID_VECTOR_UNOPS
+#pragma once
-#define GRID_VECTOR_UNOPS
 
 #include <cmath>
 
@@ -112,6 +111,9 @@ template <class scalar>
 struct ImagFunctor {
   accelerator scalar operator()(const scalar &a) const { return imag(a); }
 };
+/////////////
+// Unary operations
+/////////////
 template <class S, class V>
 accelerator_inline Grid_simd<S, V> real(const Grid_simd<S, V> &r) {
   return SimdApply(RealFunctor<S>(), r);
@@ -168,6 +170,65 @@ template <class S, class V>
 accelerator_inline Grid_simd<S, V> div(const Grid_simd<S, V> &r, Integer y) {
   return SimdApply(DivIntFunctor<S>(y), r);
 }
+/// Double 2 cases
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> real(const Grid_simd2<S, V> &r) {
+  return SimdApply(RealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> imag(const Grid_simd2<S, V> &r) {
+  return SimdApply(ImagFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> sqrt(const Grid_simd2<S, V> &r) {
+  return SimdApply(SqrtRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> cos(const Grid_simd2<S, V> &r) {
+  return SimdApply(CosRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> sin(const Grid_simd2<S, V> &r) {
+  return SimdApply(SinRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> acos(const Grid_simd2<S, V> &r) {
+  return SimdApply(AcosRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> asin(const Grid_simd2<S, V> &r) {
+  return SimdApply(AsinRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> log(const Grid_simd2<S, V> &r) {
+  return SimdApply(LogRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> abs(const Grid_simd2<S, V> &r) {
+  return SimdApply(AbsRealFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> exp(const Grid_simd2<S, V> &r) {
+  return SimdApply(ExpFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> Not(const Grid_simd2<S, V> &r) {
+  return SimdApply(NotFunctor<S>(), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> pow(const Grid_simd2<S, V> &r, double y) {
+  return SimdApply(PowRealFunctor<S>(y), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> mod(const Grid_simd2<S, V> &r, Integer y) {
+  return SimdApply(ModIntFunctor<S>(y), r);
+}
+template <class S, class V>
+accelerator_inline Grid_simd2<S, V> div(const Grid_simd2<S, V> &r, Integer y) {
+  return SimdApply(DivIntFunctor<S>(y), r);
+}
+
+
 ////////////////////////////////////////////////////////////////////////////
 // Allows us to assign into **conformable** real vectors from complex
 ////////////////////////////////////////////////////////////////////////////
@@ -212,4 +273,3 @@ accelerator_inline Grid_simd<S, V> operator||(const Grid_simd<S, V> &x,
   return SimdApplyBinop(OrOrFunctor<S>(), x, y);
 }
 NAMESPACE_END(Grid);
-#endif

Grid/simd/Simd.h

@@ -69,6 +69,7 @@ typedef RealF  Real;
 typedef thrust::complex<RealF> ComplexF;
 typedef thrust::complex<RealD> ComplexD;
 typedef thrust::complex<Real>  Complex;
+typedef thrust::complex<uint16_t>  ComplexH;
 template<class T> using complex = thrust::complex<T>;
 
 accelerator_inline ComplexD pow(const ComplexD& r,RealD y){ return(thrust::pow(r,(double)y)); }
@@ -77,6 +78,7 @@ accelerator_inline ComplexF pow(const ComplexF& r,RealF y){ return(thrust::pow(r
 typedef std::complex<RealF> ComplexF;
 typedef std::complex<RealD> ComplexD;
 typedef std::complex<Real>  Complex;
+typedef std::complex<uint16_t>  ComplexH; // Hack
 template<class T> using complex = std::complex<T>;
 
 accelerator_inline ComplexD pow(const ComplexD& r,RealD y){ return(std::pow(r,y)); }
@@ -224,18 +226,14 @@ accelerator_inline void Gpermute(VectorSIMD &y,const VectorSIMD &b,int perm);
 NAMESPACE_END(Grid);
 
 #include <Grid/simd/Grid_vector_types.h>
+#include <Grid/simd/Grid_doubled_vector.h>
 #include <Grid/simd/Grid_vector_unops.h>
 
 NAMESPACE_BEGIN(Grid);
-// Default precision
+
-#ifdef GRID_DEFAULT_PRECISION_DOUBLE
+// Default precision is wired to double
 typedef vRealD vReal;
 typedef vComplexD vComplex;
-#else
-typedef vRealF vReal;
-typedef vComplexF vComplex;
-#endif
-
  
 inline std::ostream& operator<< (std::ostream& stream, const vComplexF &o){
   int nn=vComplexF::Nsimd();
@@ -262,6 +260,13 @@ inline std::ostream& operator<< (std::ostream& stream, const vComplexD &o){
   stream<<">";
   return stream;
 }
+inline std::ostream& operator<< (std::ostream& stream, const vComplexD2 &o){
+  stream<<"<";
+  stream<<o.v[0];
+  stream<<o.v[1];
+  stream<<">";
+  return stream;
+}
 
 inline std::ostream& operator<< (std::ostream& stream, const vRealF &o){
   int nn=vRealF::Nsimd();

Grid/stencil/SimpleCompressor.h

@@ -3,26 +3,108 @@
 
 NAMESPACE_BEGIN(Grid);
 
-template<class vobj>
+class SimpleStencilParams{
-accelerator_inline void exchangeSIMT(vobj &mp0,vobj &mp1,const vobj &vp0,const vobj &vp1,Integer type)
+public:
+  Coordinate dirichlet;
+  int partialDirichlet;
+  SimpleStencilParams() { partialDirichlet = 0; };
+};
+
+
+// Compressors will inherit buffer management policies
+// Standard comms buffer management
+class FaceGatherSimple
 {
-    typedef decltype(coalescedRead(mp0)) sobj;
+public:
-    unsigned int Nsimd = vobj::Nsimd();
+  static int PartialCompressionFactor(GridBase *grid) {return 1;};
-    unsigned int mask = Nsimd >> (type + 1);
+  // Decompress is after merge so ok
-    int lane = acceleratorSIMTlane(Nsimd);
+  template<class vobj,class cobj,class compressor> 
-    int j0 = lane &(~mask); // inner coor zero
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
-    int j1 = lane |(mask) ; // inner coor one
+				   const Lattice<vobj> &rhs,
-    const vobj *vpa = &vp0;
+				   cobj *buffer,
-    const vobj *vpb = &vp1;
+				   compressor &compress,
-    const vobj *vp = (lane&mask) ? (vpb) : (vpa);
+				   int off,int so,int partial)
-    auto sa = coalescedRead(vp[0],j0);
+  {
-    auto sb = coalescedRead(vp[0],j1);
+    int num=table.size();
-    coalescedWrite(mp0,sa);
+    std::pair<int,int> *table_v = & table[0];
-    coalescedWrite(mp1,sb);
+    
+    auto rhs_v = rhs.View(AcceleratorRead);
+    accelerator_forNB( i,num, vobj::Nsimd(), {
+	compress.Compress(buffer[off+table_v[i].first],rhs_v[so+table_v[i].second]);
+    });
+    rhs_v.ViewClose();
+  }
+  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)
+  {
+    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];
+    auto rhs_p = &rhs_v[0];
+    accelerator_forNB(j, num, vobj::Nsimd(), {
+	compress.CompressExchange(p0[j],p1[j],
+				  rhs_p[so+tp[2*j  ].second],
+				  rhs_p[so+tp[2*j+1].second],
+				  type);
+    });
+    rhs_v.ViewClose();
   }
 
-template<class vobj>
+  template<class decompressor,class Decompression>
-class SimpleCompressor {
+  static void DecompressFace(decompressor decompress,Decompression &dd)
+  {
+    auto kp = dd.kernel_p;
+    auto mp = dd.mpi_p;
+    accelerator_forNB(o,dd.buffer_size,1,{
+      decompress.Decompress(kp[o],mp[o]);
+    });
+  }
+  template<class decompressor,class Merger>
+  static void MergeFace(decompressor decompress,Merger &mm)
+  {
+    auto mp = &mm.mpointer[0];
+    auto vp0= &mm.vpointers[0][0];
+    auto vp1= &mm.vpointers[1][0];
+    auto type= mm.type;
+    accelerator_forNB(o,mm.buffer_size/2,Merger::Nsimd,{
+	decompress.Exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+    });
+  }
+};
+
+////////////////////////////////////
+// Wilson compressor will add alternate policies for Dirichlet
+// and possibly partial Dirichlet for DWF
+////////////////////////////////////
+/*
+class FaceGatherDirichlet
+{
+  // If it's dirichlet we don't assemble comms buffers
+  //
+  // Rely on zeroes in gauge field to drive the correct result
+  // NAN propgagation: field will locally wrap, so fermion should NOT contain NAN and just permute
+  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){};
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				   Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				   compressor &compress,int type) {}
+  template<class decompressor,class Merger>
+  static void Merge(decompressor decompress,Merge &mm)  {  }
+  template<class decompressor,class Decompression>
+  static void Decompress(decompressor decompress,Decompression &dd) {}
+};
+*/
+
+template<class vobj,class FaceGather>
+class SimpleCompressorGather : public FaceGather {
 public:
   void Point(int) {};
   accelerator_inline int  CommDatumSize(void) const { return sizeof(vobj); }
@@ -30,20 +112,19 @@ public:
   accelerator_inline void Compress(vobj &buf,const vobj &in) const {
     coalescedWrite(buf,coalescedRead(in));
   }
-  accelerator_inline void Exchange(vobj *mp,vobj *vp0,vobj *vp1,Integer type,Integer o) const {
+  accelerator_inline void Exchange(vobj &mp0,vobj &mp1,vobj &vp0,vobj &vp1,Integer type) const {
 #ifdef GRID_SIMT
-    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+    exchangeSIMT(mp0,mp1,vp0,vp1,type);
 #else
-    exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+    exchange(mp0,mp1,vp0,vp1,type);
 #endif
   }
-  accelerator_inline void Decompress(vobj *out,vobj *in, int o) const { assert(0); }
+  accelerator_inline void Decompress(vobj &out,vobj &in) const {  };
-  accelerator_inline void CompressExchange(vobj *out0,vobj *out1,const vobj *in,
+  accelerator_inline void CompressExchange(vobj &out0,vobj &out1,const vobj &in0,const vobj &in1,int type) const {
-					   int j,int k, int m,int type) const {
 #ifdef GRID_SIMT
-    exchangeSIMT(out0[j],out1[j],in[k],in[m],type);
+    exchangeSIMT(out0,out1,in0,in1,type);
 #else
-    exchange(out0[j],out1[j],in[k],in[m],type);
+    exchange(out0,out1,in0,in1,type);
 #endif
   }
   // For cshift. Cshift should drop compressor coupling altogether 
@@ -52,11 +133,18 @@ public:
     return arg;
   }
 };
-class SimpleStencilParams{
+
-public:
+// Standard compressor never needs dirichlet.
-  Coordinate dirichlet;
+//
-  SimpleStencilParams() {};
+// Get away with a local period wrap and rely on dirac operator to use a zero gauge link as it is faster
-};
+//
+// Compressors that inherit Dirichlet and Non-dirichlet behaviour.
+//
+// Currently run-time behaviour through StencilParameters paramaters, p.dirichlet
+// combined with the FaceGatherSimple behaviour
+
+template <class vobj> using SimpleCompressor = SimpleCompressorGather<vobj,FaceGatherSimple>;
+//template <class vobj> using SimpleCompressorDirichlet = SimpleCompressorGather<vobj,FaceGatherDirichlet>;
 
 NAMESPACE_END(Grid);
 

Grid/stencil/Stencil.cc

@@ -29,6 +29,27 @@
 
 NAMESPACE_BEGIN(Grid);
 
+uint64_t DslashFullCount;
+uint64_t DslashPartialCount;
+uint64_t DslashDirichletCount;
+
+void DslashResetCounts(void)
+{
+  DslashFullCount=0;
+  DslashPartialCount=0;
+  DslashDirichletCount=0;
+}
+void DslashGetCounts(uint64_t &dirichlet,uint64_t &partial,uint64_t &full)
+{
+  dirichlet = DslashDirichletCount;
+  partial   = DslashPartialCount;
+  full      = DslashFullCount;
+}
+void DslashLogFull(void)     { DslashFullCount++;}
+void DslashLogPartial(void)  { DslashPartialCount++;}
+void DslashLogDirichlet(void){ DslashDirichletCount++;}
+
+
 void Gather_plane_table_compute (GridBase *grid,int dimension,int plane,int cbmask,
 				 int off,std::vector<std::pair<int,int> > & table)
 {

Grid/stencil/Stencil.h

@@ -52,6 +52,16 @@
 
 NAMESPACE_BEGIN(Grid);
 
+// These can move into a params header and be given MacroMagic serialisation
+struct DefaultImplParams {
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  int  partialDirichlet;
+  DefaultImplParams()  {
+    dirichlet.resize(0);
+    partialDirichlet=0;
+  };
+};
+
 ///////////////////////////////////////////////////////////////////
 // Gather for when there *is* need to SIMD split with compression
 ///////////////////////////////////////////////////////////////////
@@ -59,6 +69,7 @@ NAMESPACE_BEGIN(Grid);
 void Gather_plane_table_compute (GridBase *grid,int dimension,int plane,int cbmask,
 				 int off,std::vector<std::pair<int,int> > & table);
 
+/*
 template<class vobj,class cobj,class compressor>
 void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so)   __attribute__((noinline));
 
@@ -80,11 +91,14 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
 ///////////////////////////////////////////////////////////////////
 template<class cobj,class vobj,class compressor>
 void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
-				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
+				 commVector<cobj *> pointers,
+				 int dimension,int plane,
+				 int cbmask,compressor &compress,int type) __attribute__((noinline));
 
 template<class cobj,class vobj,class compressor>
-void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
-				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				 const Lattice<vobj> &rhs,
+				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -92,17 +106,25 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const La
   int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
   auto rhs_v = rhs.View(AcceleratorRead);
+  auto rhs_p = &rhs_v[0];
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
   accelerator_forNB(j, num, vobj::Nsimd(), {
-      compress.CompressExchange(p0,p1, &rhs_v[0], j,
+      compress.CompressExchange(p0,p1, rhs_p, j,
 				so+tp[2*j  ].second,
 				so+tp[2*j+1].second,
 				type);
   });
   rhs_v.ViewClose();
 }
+*/
+
+void DslashResetCounts(void);
+void DslashGetCounts(uint64_t &dirichlet,uint64_t &partial,uint64_t &full);
+void DslashLogFull(void);
+void DslashLogPartial(void);
+void DslashLogDirichlet(void);
 
 struct StencilEntry {
 #ifdef GRID_CUDA
@@ -133,8 +155,18 @@ class CartesianStencilAccelerator {
   int           _osites;
   StencilVector _directions;
   StencilVector _distances;
+  ///////////////////////////////////////////////////
+  // If true, this is FULLY communicated per face
+  // Otherwise will either be full or partial dirichlet
+  ///////////////////////////////////////////////////
   StencilVector _comms_send; 
-  StencilVector _comms_recv;
+  StencilVector _comms_recv; // this is FULLY communicated per face
+  ///////////////////////////////////////////////////
+  // If true, this is partially communicated per face
+  ///////////////////////////////////////////////////
+  StencilVector _comms_partial_send; 
+  StencilVector _comms_partial_recv;
+  //
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -181,7 +213,7 @@ class CartesianStencilAccelerator {
 template<class vobj,class cobj,class Parameters>
 class CartesianStencilView : public CartesianStencilAccelerator<vobj,cobj,Parameters>
 {
- private:
+public:
   int *closed;
   StencilEntry *cpu_ptr;
   ViewMode      mode;
@@ -216,7 +248,6 @@ class CartesianStencil : public CartesianStencilAccelerator<vobj,cobj,Parameters
 public:
 
   typedef typename cobj::vector_type vector_type;
-  typedef typename cobj::scalar_type scalar_type;
   typedef typename cobj::scalar_object scalar_object;
   typedef const CartesianStencilView<vobj,cobj,Parameters> View_type;
   typedef typename View_type::StencilVector StencilVector;
@@ -230,19 +261,26 @@ public:
     Integer from_rank;
     Integer do_send;
     Integer do_recv;
-    Integer bytes;
+    Integer xbytes;
+    Integer rbytes;
   };
   struct Merge {
+    static constexpr int Nsimd = vobj::Nsimd();
     cobj * mpointer;
-    Vector<scalar_object *> rpointers;
+    //    std::vector<scalar_object *> rpointers;
-    Vector<cobj *> vpointers;
+    std::vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
+    Integer partial; // partial dirichlet BCs
+    Coordinate dims;
   };
   struct Decompress {
+    static constexpr int Nsimd = vobj::Nsimd();
     cobj * kernel_p;
     cobj * mpi_p;
     Integer buffer_size;
+    Integer partial; // partial dirichlet BCs
+    Coordinate dims;
   };
   struct CopyReceiveBuffer {
     void * from_p;
@@ -253,7 +291,8 @@ public:
     Integer direction;
     Integer OrthogPlane;
     Integer DestProc;
-    Integer bytes;
+    Integer xbytes;
+    Integer rbytes;
     Integer lane;
     Integer cb;
     void *recv_buf;
@@ -261,9 +300,9 @@ public:
 
 protected:
   GridBase *                        _grid;
-
 public:
   GridBase *Grid(void) const { return _grid; }
+  LebesgueOrder *lo;
 
   ////////////////////////////////////////////////////////////////////////
   // Needed to conveniently communicate gparity parameters into GPU memory
@@ -278,6 +317,8 @@ public:
   }
 
   int face_table_computed;
+  int partialDirichlet;
+  int fullDirichlet;
   std::vector<commVector<std::pair<int,int> > > face_table ;
   Vector<int> surface_list;
 
@@ -307,6 +348,7 @@ public:
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
+#ifdef SHM_FAST_PATH
   inline int SameNode(int point) {
 
     int dimension    = this->_directions[point];
@@ -326,7 +368,40 @@ public:
     if ( displacement == 0 ) return 1;
     return 0;
   }
+#else
+  // fancy calculation for shm code
+  inline int SameNode(int point) {
 
+    int dimension    = this->_directions[point];
+    int displacement = this->_distances[point];
+
+    int pd              = _grid->_processors[dimension];
+    int fd              = _grid->_fdimensions[dimension];
+    int ld              = _grid->_ldimensions[dimension];
+    int rd              = _grid->_rdimensions[dimension];
+    int simd_layout     = _grid->_simd_layout[dimension];
+    int comm_dim        = _grid->_processors[dimension] >1 ;
+ 
+    int recv_from_rank;
+    int xmit_to_rank;
+
+    if ( ! comm_dim ) return 1;
+
+    int nbr_proc;
+    if (displacement>0) nbr_proc = 1;
+    else                nbr_proc = pd-1;
+
+    // FIXME  this logic needs to be sorted for three link term
+    //    assert( (displacement==1) || (displacement==-1));
+    // Present hack only works for >= 4^4 subvol per node
+    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
+
+    void *shm = (void *) _grid->ShmBufferTranslate(recv_from_rank,this->u_recv_buf_p);
+
+    if ( shm==NULL ) return 0;
+    return 1;
+  }
+#endif
   //////////////////////////////////////////
   // Comms packet queue for asynch thread
   // Use OpenMP Tasks for cleaner ???
@@ -366,13 +441,17 @@ public:
 					Packets[i].to_rank,Packets[i].do_send,
 					Packets[i].recv_buf,
 					Packets[i].from_rank,Packets[i].do_recv,
-					Packets[i].bytes,i);
+					Packets[i].xbytes,Packets[i].rbytes,i);
     }
+    _grid->StencilBarrier();// Synch shared memory on a single nodes
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
     _grid->StencilSendToRecvFromComplete(MpiReqs,0);
+    if   ( this->partialDirichlet ) DslashLogPartial();
+    else if ( this->fullDirichlet ) DslashLogDirichlet();
+    else DslashLogFull();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
@@ -502,7 +581,9 @@ public:
     }
   }
   
-  Integer CheckForDuplicate(Integer direction, Integer OrthogPlane, Integer DestProc, void *recv_buf,Integer lane,Integer bytes,Integer cb)
+  Integer CheckForDuplicate(Integer direction, Integer OrthogPlane, Integer DestProc, void *recv_buf,Integer lane,
+			    Integer xbytes,Integer rbytes,
+			    Integer cb)
   {
     CachedTransfer obj;
     obj.direction   = direction;
@@ -510,19 +591,22 @@ public:
     obj.DestProc    = DestProc;
     obj.recv_buf    = recv_buf;
     obj.lane        = lane;
-    obj.bytes       = bytes;
+    obj.xbytes      = xbytes;
+    obj.rbytes      = rbytes;
     obj.cb          = cb;
 
     for(int i=0;i<CachedTransfers.size();i++){
       if (   (CachedTransfers[i].direction  ==direction)
 	   &&(CachedTransfers[i].OrthogPlane==OrthogPlane)
 	   &&(CachedTransfers[i].DestProc   ==DestProc)
-	   &&(CachedTransfers[i].bytes      ==bytes)
+	   &&(CachedTransfers[i].xbytes      ==xbytes)
+	   &&(CachedTransfers[i].rbytes      ==rbytes)
 	   &&(CachedTransfers[i].lane       ==lane)
 	   &&(CachedTransfers[i].cb         ==cb)
 	     ){
-
+	// FIXME worry about duplicate with partial compression
-	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
+	// Wont happen as DWF has no duplicates, but...
+	AddCopy(CachedTransfers[i].recv_buf,recv_buf,rbytes);
 	return 1;
       }
     }
@@ -533,7 +617,7 @@ public:
   void AddPacket(void *xmit,void * rcv,
 		 Integer to, Integer do_send,
 		 Integer from, Integer do_recv,
-		 Integer bytes){
+		 Integer xbytes,Integer rbytes){
     Packet p;
     p.send_buf = xmit;
     p.recv_buf = rcv;
@@ -541,18 +625,25 @@ public:
     p.from_rank= from;
     p.do_send  = do_send;
     p.do_recv  = do_recv;
-    p.bytes    = bytes;
+    p.xbytes    = xbytes;
+    p.rbytes    = rbytes;
+    //    if (do_send) std::cout << GridLogMessage << " MPI packet to   "<<to<< " of size "<<xbytes<<std::endl;
+    //    if (do_recv) std::cout << GridLogMessage << " MPI packet from "<<from<< " of size "<<xbytes<<std::endl;
     Packets.push_back(p);
   }
   void AddDecompress(cobj *k_p,cobj *m_p,Integer buffer_size,std::vector<Decompress> &dv) {
     Decompress d;
+    d.partial  = this->partialDirichlet;
+    d.dims     = _grid->_fdimensions;
     d.kernel_p = k_p;
     d.mpi_p    = m_p;
     d.buffer_size = buffer_size;
     dv.push_back(d);
   }
-  void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
+  void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
     Merge m;
+    m.partial  = this->partialDirichlet;
+    m.dims     = _grid->_fdimensions;
     m.type     = type;
     m.mpointer = merge_p;
     m.vpointers= rpointers;
@@ -572,21 +663,10 @@ public:
   void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd)
   {
     for(int i=0;i<mm.size();i++){
-      auto mp = &mm[i].mpointer[0];
+      decompressor::MergeFace(decompress,mm[i]);
-      auto vp0= &mm[i].vpointers[0][0];
-      auto vp1= &mm[i].vpointers[1][0];
-      auto type= mm[i].type;
-      accelerator_forNB(o,mm[i].buffer_size/2,vobj::Nsimd(),{
-	  decompress.Exchange(mp,vp0,vp1,type,o);
-      });
     }
-
     for(int i=0;i<dd.size();i++){
-      auto kp = dd[i].kernel_p;
+      decompressor::DecompressFace(decompress,dd[i]);
-      auto mp = dd[i].mpi_p;
-      accelerator_forNB(o,dd[i].buffer_size,1,{
-	decompress.Decompress(kp,mp,o);
-      });
     }
   }
   ////////////////////////////////////////
@@ -645,6 +725,8 @@ public:
       int block = dirichlet_block[dimension];
       this->_comms_send[ii] = comm_dim;
       this->_comms_recv[ii] = comm_dim;
+      this->_comms_partial_send[ii] = 0;
+      this->_comms_partial_recv[ii] = 0;
       if ( block && comm_dim ) {
 	assert(abs(displacement) < ld );
 	// Quiesce communication across block boundaries
@@ -665,6 +747,10 @@ public:
 	  if ( ( (ld*(pc+1) ) % block ) == 0 ) this->_comms_send[ii] = 0;
 	  if ( ( (ld*pc     ) % block ) == 0 ) this->_comms_recv[ii] = 0;
 	}
+	if ( partialDirichlet ) {
+	  this->_comms_partial_send[ii] = !this->_comms_send[ii];
+	  this->_comms_partial_recv[ii] = !this->_comms_recv[ii];
+	}
       }
     }
   }
@@ -673,7 +759,7 @@ public:
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
-		   Parameters p)
+		   Parameters p=Parameters())
   {
     face_table_computed=0;
     _grid    = grid;
@@ -692,8 +778,12 @@ public:
     this->same_node.resize(npoints);
 
     if ( p.dirichlet.size() ==0 ) p.dirichlet.resize(grid->Nd(),0);
-
+    partialDirichlet = p.partialDirichlet;
     DirichletBlock(p.dirichlet); // comms send/recv set up
+    fullDirichlet=0;
+    for(int d=0;d<p.dirichlet.size();d++){
+      if (p.dirichlet[d]) fullDirichlet=1;
+    }
 
     _unified_buffer_size=0;
     surface_list.resize(0);
@@ -828,7 +918,7 @@ public:
     GridBase *grid=_grid;
     const int Nsimd = grid->Nsimd();
 
-    int comms_recv      = this->_comms_recv[point];
+    int comms_recv      = this->_comms_recv[point] || this->_comms_partial_recv[point] ;
     int fd              = _grid->_fdimensions[dimension];
     int ld              = _grid->_ldimensions[dimension];
     int rd              = _grid->_rdimensions[dimension];
@@ -1014,10 +1104,11 @@ public:
   int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx, int point)
   {
     typedef typename cobj::vector_type vector_type;
-    typedef typename cobj::scalar_type scalar_type;
 
     int comms_send   = this->_comms_send[point];
     int comms_recv   = this->_comms_recv[point];
+    int comms_partial_send   = this->_comms_partial_send[point] ;
+    int comms_partial_recv   = this->_comms_partial_recv[point] ;
     
     assert(rhs.Grid()==_grid);
     //	  conformable(_grid,rhs.Grid());
@@ -1048,6 +1139,16 @@ public:
 	if (cbmask != 0x3) words=words>>1;
 
 	int bytes =  words * compress.CommDatumSize();
+	int xbytes;
+	int rbytes;
+
+	if ( comms_send ) xbytes = bytes; // Full send
+	else if ( comms_partial_send ) xbytes = bytes/compressor::PartialCompressionFactor(_grid);
+	else xbytes = 0; // full dirichlet
+
+	if ( comms_recv ) rbytes = bytes;
+	else if ( comms_partial_recv ) rbytes = bytes/compressor::PartialCompressionFactor(_grid);
+	else rbytes = 0;
 	
 	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 	int comm_off = u_comm_offset;
@@ -1061,12 +1162,11 @@ public:
 	assert (xmit_to_rank   != _grid->ThisRank());
 	assert (recv_from_rank != _grid->ThisRank());
 
-	if( comms_send ) {
-
 	if ( !face_table_computed ) {
 	  face_table.resize(face_idx+1);
 	  std::vector<std::pair<int,int> >  face_table_host ;
 	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
+	  //	  std::cout << "bytes expect "<< bytes << " " << face_table_host.size()* compress.CommDatumSize()<<std::endl;
 	    face_table[face_idx].resize(face_table_host.size());
 	    acceleratorCopyToDevice(&face_table_host[0],
 				    &face_table[face_idx][0],
@@ -1074,36 +1174,60 @@ public:
 	}
 
 
-	  if ( compress.DecompressionStep() ) {
+	if ( (compress.DecompressionStep()&&comms_recv) || comms_partial_recv ) {
 	  recv_buf=u_simd_recv_buf[0];
 	} else {
 	  recv_buf=this->u_recv_buf_p;
 	}
 
-	  send_buf = this->u_send_buf_p; // Gather locally, must send
+	// potential SHM fast path for intranode
-	
+	int shm_send=0;
+	int shm_recv=0;
+#ifdef SHM_FAST_PATH
+	// Put directly in place if we can
+	send_buf = (cobj *)_grid->ShmBufferTranslate(xmit_to_rank,recv_buf);
+	if ( (send_buf==NULL) ) {
+	  shm_send=0;
+	  send_buf = this->u_send_buf_p;
+	} else {
+	  shm_send=1;
+	}
+	void *test_ptr = _grid->ShmBufferTranslate(recv_from_rank,recv_buf);
+	if ( test_ptr != NULL ) shm_recv = 1;
+	//	static int printed;
+	//	if (!printed){
+	  //	  std::cout << " GATHER FAST PATH SHM "<<shm_send<< " "<<shm_recv<<std::endl;
+	//	  printed = 1;
+	//	}
+#else
 	////////////////////////////////////////////////////////
 	// Gather locally
 	////////////////////////////////////////////////////////
+	send_buf = this->u_send_buf_p; // Gather locally, must send
 	assert(send_buf!=NULL);
-	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,comm_off,so);
+#endif
-	}
 
-	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,bytes,cbmask);
+	//	std::cout << " GatherPlaneSimple partial send "<< comms_partial_send<<std::endl;
-	if ( (!duplicate) ) { // Force comms for now
+	compressor::Gather_plane_simple(face_table[face_idx],rhs,send_buf,compress,comm_off,so,comms_partial_send);
+
+        int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,xbytes,rbytes,cbmask);
+	if ( !duplicate ) { // Force comms for now
 	  
 	  ///////////////////////////////////////////////////////////
 	  // Build a list of things to do after we synchronise GPUs
 	  // Start comms now???
 	  ///////////////////////////////////////////////////////////
+	  int do_send = (comms_send|comms_partial_send) && (!shm_send );
+	  int do_recv = (comms_send|comms_partial_send) && (!shm_recv );
+	  
 	  AddPacket((void *)&send_buf[comm_off],
 		    (void *)&recv_buf[comm_off],
-		    xmit_to_rank, comms_send,
+		    xmit_to_rank, do_send,
-		    recv_from_rank, comms_recv,
+		    recv_from_rank, do_recv,
-		    bytes);
+		    xbytes,rbytes);
 	}
 
-	if ( compress.DecompressionStep()  && comms_recv ) {
+	if ( (compress.DecompressionStep() && comms_recv) || comms_partial_recv ) {
 	  AddDecompress(&this->u_recv_buf_p[comm_off],
 			&recv_buf[comm_off],
 			words,Decompressions);
@@ -1111,7 +1235,6 @@ public:
 	
 	u_comm_offset+=words;
 	face_idx++;
-
       }
     }
     return 0;
@@ -1126,6 +1249,8 @@ public:
 
     int comms_send   = this->_comms_send[point];
     int comms_recv   = this->_comms_recv[point];
+    int comms_partial_send   = this->_comms_partial_send[point] ;
+    int comms_partial_recv   = this->_comms_partial_recv[point] ;
 
     int fd = _grid->_fdimensions[dimension];
     int rd = _grid->_rdimensions[dimension];
@@ -1155,10 +1280,15 @@ public:
 
     int datum_bytes = compress.CommDatumSize();
     int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
+
+    // how many bytes on wire : partial dirichlet or dirichlet may set to < bytes
+    int xbytes; 
+    int rbytes; 
+    
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    Vector<cobj *> rpointers(maxl);
+    std::vector<cobj *> rpointers(maxl);
-    Vector<cobj *> spointers(maxl);
+    std::vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where
@@ -1190,16 +1320,31 @@ public:
 	  acceleratorCopyToDevice(&face_table_host[0],
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
+
 	}
 
-	if ( comms_send || comms_recv ) {
+	
-	  Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
+	if ( comms_send ) xbytes = bytes;
+	else if ( comms_partial_send ) xbytes = bytes/compressor::PartialCompressionFactor(_grid);
+	else xbytes = 0;
+
+	if ( comms_recv ) rbytes = bytes;
+	else if ( comms_partial_recv ) rbytes = bytes/compressor::PartialCompressionFactor(_grid);
+	else rbytes = 0;
+
+	// Gathers SIMD lanes for send and merge
+	// Different faces can be full comms or partial comms with  multiple ranks per node
+	if ( comms_send || comms_recv||comms_partial_send||comms_partial_recv ) {
+
+	  int partial = partialDirichlet;
+	  compressor::Gather_plane_exchange(face_table[face_idx],rhs,
+					    spointers,dimension,sx,cbmask,
+					    compress,permute_type,partial );
 	}
 	face_idx++;
 
-	//spointers[0] -- low
+	//spointers[0] -- low simd coor
-	//spointers[1] -- high
+	//spointers[1] -- high simd coor
-
 	for(int i=0;i<maxl;i++){
 
 	  int my_coor  = rd*i + x;            // self explanatory
@@ -1220,17 +1365,48 @@ public:
 
 	    int recv_from_rank;
 	    int xmit_to_rank;
-
+	    int shm_send=0;
+	    int shm_recv=0;
 	    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
-
+#ifdef SHM_FAST_PATH
+  #warning STENCIL SHM FAST PATH SELECTED
+	    // shm == receive pointer         if offnode
+	    // shm == Translate[send pointer] if on node -- my view of his send pointer
+	    cobj *shm = (cobj *) _grid->ShmBufferTranslate(recv_from_rank,sp);
+	    if (shm==NULL) {
+	      shm = rp;
+	      // we found a packet that comes from MPI and contributes to this shift.
+	      // is_same_node is only used in the WilsonStencil, and gets set for this point in the stencil.
+	      // Kernel will add the exterior_terms except if is_same_node.
+	      // leg of stencil
+	      shm_recv=0;
+	    } else {
+	      shm_recv=1;
+	    }
+	    rpointers[i] = shm;
+	    // Test send side
+	    void *test_ptr = (void *) _grid->ShmBufferTranslate(xmit_to_rank,sp);
+	    if ( test_ptr != NULL ) shm_send = 1;
+	    //	    static int printed;
+	    //	    if (!printed){
+	    //	      std::cout << " GATHERSIMD FAST PATH SHM "<<shm_send<< " "<<shm_recv<<std::endl;
+	    //	      printed = 1;
+	    //	    }
+#else
 	    rpointers[i] = rp;
+#endif
 	    
-	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,bytes,cbmask);
+	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,xbytes,rbytes,cbmask);
 	    if ( !duplicate  ) { 
+	      if ( (bytes != rbytes) && (rbytes!=0) ){
+		acceleratorMemSet(rp,0,bytes); // Zero prefill comms buffer to zero
+	      }
+	      int do_send = (comms_send|comms_partial_send) && (!shm_send );
+	      int do_recv = (comms_send|comms_partial_send) && (!shm_recv );
 	      AddPacket((void *)sp,(void *)rp,
-			xmit_to_rank,comms_send,
+			xmit_to_rank,do_send,
-			recv_from_rank,comms_recv,
+			recv_from_rank,do_send,
-			bytes);
+			xbytes,rbytes);
 	    }
 
 	  } else {
@@ -1239,8 +1415,8 @@ public:
 
 	  }
 	}
-
+	// rpointer may be doing a remote read in the gather over SHM
-	if ( comms_recv ) {
+	if ( comms_recv|comms_partial_recv ) {
 	  AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
 	}
 

Grid/tensors/Tensor_SIMT.h

@@ -31,6 +31,27 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+////////////////////////////////////////////////
+// Inside a GPU thread
+////////////////////////////////////////////////
+template<class vobj>
+accelerator_inline void exchangeSIMT(vobj &mp0,vobj &mp1,const vobj &vp0,const vobj &vp1,Integer type)
+{
+    typedef decltype(coalescedRead(mp0)) sobj;
+    unsigned int Nsimd = vobj::Nsimd();
+    unsigned int mask = Nsimd >> (type + 1);
+    int lane = acceleratorSIMTlane(Nsimd);
+    int j0 = lane &(~mask); // inner coor zero
+    int j1 = lane |(mask) ; // inner coor one
+    const vobj *vpa = &vp0;
+    const vobj *vpb = &vp1;
+    const vobj *vp = (lane&mask) ? (vpb) : (vpa);
+    auto sa = coalescedRead(vp[0],j0);
+    auto sb = coalescedRead(vp[0],j1);
+    coalescedWrite(mp0,sa);
+    coalescedWrite(mp1,sb);
+}
+
 
 #ifndef GRID_SIMT
 //////////////////////////////////////////

Grid/tensors/Tensor_class.h

@@ -178,6 +178,7 @@ public:
     stream << "S {" << o._internal << "}";
     return stream;
   };
+  // FIXME These will break with change of data layout
   strong_inline const scalar_type * begin() const { return reinterpret_cast<const scalar_type *>(&_internal); }
   strong_inline       scalar_type * begin()       { return reinterpret_cast<      scalar_type *>(&_internal); }
   strong_inline const scalar_type * end()   const { return begin() + Traits::count; }
@@ -288,6 +289,7 @@ public:
   //      return _internal[i];
   //    }
 
+  // FIXME These will break with change of data layout
   strong_inline const scalar_type * begin() const { return reinterpret_cast<const scalar_type *>(_internal); }
   strong_inline       scalar_type * begin()       { return reinterpret_cast<      scalar_type *>(_internal); }
   strong_inline const scalar_type * end()   const { return begin() + Traits::count; }
@@ -430,6 +432,7 @@ public:
   //    return _internal[i][j];
   //  }
 
+  // FIXME These will break with change of data layout
   strong_inline const scalar_type * begin() const { return reinterpret_cast<const scalar_type *>(_internal[0]); }
   strong_inline       scalar_type * begin()       { return reinterpret_cast<      scalar_type *>(_internal[0]); }
   strong_inline const scalar_type * end()   const { return begin() + Traits::count; }

Grid/tensors/Tensor_extract_merge.h

@@ -1,5 +1,5 @@
 /*************************************************************************************
-n
+
     Grid physics library, www.github.com/paboyle/Grid 
 
     Source file: ./lib/tensors/Tensor_extract_merge.h
@@ -62,8 +62,18 @@ void extract(const vobj &vec,ExtractBuffer<sobj> &extracted)
   const int words=sizeof(vobj)/sizeof(vector_type);
   const int Nsimd=vector_type::Nsimd();
   const int Nextr=extracted.size();
+  vector_type * vp = (vector_type *)&vec;
   const int s=Nsimd/Nextr;
   sobj_scalar_type *sp = (sobj_scalar_type *) &extracted[0];
+  sobj_scalar_type stmp;
+  for(int w=0;w<words;w++){
+    for(int i=0;i<Nextr;i++){
+      stmp = vp[w].getlane(i*s);
+      sp[i*words+w] =stmp;
+      //      memcpy((char *)&sp[i*words+w],(char *)&stmp,sizeof(stmp));
+    }
+  }
+  /*
   scalar_type *vp = (scalar_type *)&vec;
   scalar_type      vtmp;
   sobj_scalar_type stmp;
@@ -74,6 +84,8 @@ void extract(const vobj &vec,ExtractBuffer<sobj> &extracted)
       memcpy((char *)&sp[i*words+w],(char *)&stmp,sizeof(stmp));
     }
   }
+  */
+  
   return;
 }
 
@@ -93,7 +105,7 @@ void   merge(vobj &vec,ExtractBuffer<sobj> &extracted)
   const int s=Nsimd/Nextr;
 
   sobj_scalar_type *sp = (sobj_scalar_type *)&extracted[0];
-  scalar_type *vp = (scalar_type *)&vec;
+  vector_type *vp = (vector_type *)&vec;
   scalar_type      vtmp;
   sobj_scalar_type stmp;
   for(int w=0;w<words;w++){
@@ -101,7 +113,8 @@ void   merge(vobj &vec,ExtractBuffer<sobj> &extracted)
       for(int ii=0;ii<s;ii++){
 	memcpy((char *)&stmp,(char *)&sp[i*words+w],sizeof(stmp));
 	vtmp = stmp;
-	memcpy((char *)&vp[w*Nsimd+i*s+ii],(char *)&vtmp,sizeof(vtmp));
+	vp[w].putlane(vtmp,i*s+ii);
+	//	memcpy((char *)&vp[w*Nsimd+i*s+ii],(char *)&vtmp,sizeof(vtmp));
       }
     }
   }
@@ -117,16 +130,16 @@ typename vobj::scalar_object extractLane(int lane, const vobj & __restrict__ vec
   typedef typename vobj::scalar_object scalar_object;
   typedef typename vobj::vector_type   vector_type;
   typedef typename ExtractTypeMap<scalar_type>::extract_type extract_type;
-  typedef extract_type * pointer;
+  typedef scalar_type * pointer;
 
   constexpr int words=sizeof(vobj)/sizeof(vector_type);
   constexpr int Nsimd=vector_type::Nsimd();
 
   scalar_object extracted;
   pointer __restrict__  sp = (pointer)&extracted; // Type pun
-  pointer __restrict__  vp = (pointer)&vec;
+  vector_type *vp = (vector_type *)&vec;
   for(int w=0;w<words;w++){
-    sp[w]=vp[w*Nsimd+lane];
+    sp[w]=vp[w].getlane(lane);
   }
   return extracted;
 }
@@ -137,15 +150,15 @@ void insertLane(int lane, vobj & __restrict__ vec,const typename vobj::scalar_ob
   typedef typename vobj::vector_type vector_type;
   typedef typename vector_type::scalar_type scalar_type;
   typedef typename ExtractTypeMap<scalar_type>::extract_type extract_type;
-  typedef extract_type * pointer;
+  typedef scalar_type * pointer;
 
   constexpr int words=sizeof(vobj)/sizeof(vector_type);
   constexpr int Nsimd=vector_type::Nsimd();
 
   pointer __restrict__ sp = (pointer)&extracted;
-  pointer __restrict__ vp = (pointer)&vec;
+  vector_type *vp = (vector_type *)&vec;
   for(int w=0;w<words;w++){
-    vp[w*Nsimd+lane]=sp[w];
+    vp[w].putlane(sp[w],lane);
   }
 }
 
@@ -164,15 +177,13 @@ void extract(const vobj &vec,const ExtractPointerArray<sobj> &extracted, int off
   const int Nextr=extracted.size();
   const int s = Nsimd/Nextr;
 
-  scalar_type * vp = (scalar_type *)&vec;
+  vector_type * vp = (vector_type *)&vec;
   scalar_type      vtmp;
   sobj_scalar_type stmp;
   for(int w=0;w<words;w++){
     for(int i=0;i<Nextr;i++){
       sobj_scalar_type * pointer = (sobj_scalar_type *)& extracted[i][offset];
-      memcpy((char *)&vtmp,(char *)&vp[w*Nsimd+i*s],sizeof(vtmp));
+      pointer[w] = vp[w].getlane(i*s);
-      stmp = vtmp;
-      memcpy((char *)&pointer[w],(char *)&stmp,sizeof(stmp)); // may do a precision conversion
     }
   }
 }
@@ -192,23 +203,21 @@ void merge(vobj &vec,const ExtractPointerArray<sobj> &extracted, int offset)
   const int Nextr=extracted.size();
   const int s = Nsimd/Nextr;
 
-  scalar_type * vp = (scalar_type *)&vec;
+  vector_type * vp = (vector_type *)&vec;
   scalar_type      vtmp;
   sobj_scalar_type stmp;
   for(int w=0;w<words;w++){
     for(int i=0;i<Nextr;i++){
       sobj_scalar_type * pointer = (sobj_scalar_type *)& extracted[i][offset];
       for(int ii=0;ii<s;ii++){
-	memcpy((char *)&stmp,(char *)&pointer[w],sizeof(stmp));
+	vtmp=pointer[w];
-	vtmp=stmp;
+	vp[w].putlane(vtmp,i*s+ii);
-	memcpy((char *)&vp[w*Nsimd+i*s+ii],(char *)&vtmp,sizeof(vtmp));
       }
     }
   }
 }
 
 
-
 //////////////////////////////////////////////////////////////////////////////////
 //Copy a single lane of a SIMD tensor type from one object to another
 //Output object must be of the same tensor type but may be of a different precision (i.e. it can have a different root data type)
@@ -217,7 +226,7 @@ template<class vobjOut, class vobjIn>
 accelerator_inline 
 void copyLane(vobjOut & __restrict__ vecOut, int lane_out, const vobjIn & __restrict__ vecIn, int lane_in)
 {
-  static_assert( std::is_same<typename vobjOut::DoublePrecision, typename vobjIn::DoublePrecision>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+  static_assert( std::is_same<typename vobjOut::scalar_typeD, typename vobjIn::scalar_typeD>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
 
   typedef typename vobjOut::vector_type ovector_type;  
   typedef typename vobjIn::vector_type ivector_type;  
@@ -239,12 +248,12 @@ void copyLane(vobjOut & __restrict__ vecOut, int lane_out, const vobjIn & __rest
   iscalar_type itmp;
   oscalar_type otmp;
 
-  opointer __restrict__  op = (opointer)&vecOut;
+  ovector_type * __restrict__ op = (ovector_type *)&vecOut;
-  ipointer __restrict__  ip = (ipointer)&vecIn;
+  ivector_type * __restrict__ ip = (ivector_type *)&vecIn;
   for(int w=0;w<owords;w++){
-    memcpy( (char*)&itmp, (char*)(ip + lane_in + iNsimd*w), sizeof(iscalar_type) );
+    itmp = ip[w].getlane(lane_in);
     otmp = itmp; //potential precision change
-    memcpy( (char*)(op + lane_out + oNsimd*w), (char*)&otmp, sizeof(oscalar_type) );
+    op[w].putlane(otmp,lane_out);
   }
 }
 

Grid/tensors/Tensor_inner.h

@@ -214,24 +214,20 @@ accelerator_inline vRealD    innerProductD2(const vRealD    &l,const vRealD    &
 
 accelerator_inline vComplexD2 innerProductD2(const vComplexF &l,const vComplexF &r)
 {  
-  vComplexD la,lb;
+  vComplexD2 dl,dr;
-  vComplexD ra,rb;
-  Optimization::PrecisionChange::StoD(l.v,la.v,lb.v);
-  Optimization::PrecisionChange::StoD(r.v,ra.v,rb.v);
   vComplexD2 ret;
-  ret._internal[0] = innerProduct(la,ra);
+  precisionChange(dl,l);
-  ret._internal[1] = innerProduct(lb,rb);
+  precisionChange(dr,r);
+  ret = innerProduct(dl,dr);
   return ret;
 }
 accelerator_inline vRealD2 innerProductD2(const vRealF &l,const vRealF &r)
 {  
-  vRealD la,lb;
+  vRealD2 dl,dr;
-  vRealD ra,rb;
-  Optimization::PrecisionChange::StoD(l.v,la.v,lb.v);
-  Optimization::PrecisionChange::StoD(r.v,ra.v,rb.v);
   vRealD2 ret;
-  ret._internal[0]=innerProduct(la,ra);
+  precisionChange(dl,l);
-  ret._internal[1]=innerProduct(lb,rb); 
+  precisionChange(dr,r);
+  ret=innerProduct(dl,dr);
   return ret;
 }