Run at physical mass

Feature/paddedcellgauge

Grid/Makefile.am

@@ -66,6 +66,10 @@ if BUILD_FERMION_REPS
   extra_sources+=$(ADJ_FERMION_FILES)
   extra_sources+=$(TWOIND_FERMION_FILES)
 endif
+if BUILD_SP
+    extra_sources+=$(SP_FERMION_FILES)
+    extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
 
 lib_LIBRARIES = libGrid.a
 

Grid/algorithms/Algorithms.h

@@ -69,7 +69,8 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/PowerMethod.h>
 
 NAMESPACE_CHECK(PowerMethod);
-#include <Grid/algorithms/CoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/MultiGrid.h>
+
 NAMESPACE_CHECK(CoarsendMatrix);
 #include <Grid/algorithms/FFT.h>
 

Grid/algorithms/approx/Chebyshev.h

@@ -90,9 +90,8 @@ public:
     order=_order;
       
     if(order < 2) exit(-1);
-    Coeffs.resize(order);
-    Coeffs.assign(0.,order);
-    Coeffs[order-1] = 1.;
+    Coeffs.resize(order,0.0);
+    Coeffs[order-1] = 1.0;
   };
   
   // PB - more efficient low pass drops high modes above the low as 1/x uses all Chebyshev's.

Grid/algorithms/iterative/AdefGeneric.h

@@ -33,218 +33,254 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    * Script A = SolverMatrix 
    * Script P = Preconditioner
    *
-   * Deflation methods considered
-   *      -- Solve P A x = P b        [ like Luscher ]
-   * DEF-1        M P A x = M P b     [i.e. left precon]
-   * DEF-2        P^T M A x = P^T M b
-   * ADEF-1       Preconditioner = M P + Q      [ Q + M + M A Q]
-   * ADEF-2       Preconditioner = P^T M + Q
-   * BNN          Preconditioner = P^T M P + Q
-   * BNN2         Preconditioner = M P + P^TM +Q - M P A M 
-   * 
    * Implement ADEF-2
    *
    * Vstart = P^Tx + Qb
    * M1 = P^TM + Q
    * M2=M3=1
-   * Vout = x
    */
+NAMESPACE_BEGIN(Grid);
 
-// abstract base
-template<class Field, class CoarseField>
-class TwoLevelFlexiblePcg : public LinearFunction<Field>
+template<class Field>
+class TwoLevelCG : public LinearFunction<Field>
 {
  public:
-  int verbose;
   RealD   Tolerance;
   Integer MaxIterations;
-  const int mmax = 5;
   GridBase *grid;
-  GridBase *coarsegrid;
 
-  LinearOperatorBase<Field>   *_Linop
-  OperatorFunction<Field>     *_Smoother,
-  LinearFunction<CoarseField> *_CoarseSolver;
-
-  // Need somthing that knows how to get from Coarse to fine and back again
+  // Fine operator, Smoother, CoarseSolver
+  LinearOperatorBase<Field>   &_FineLinop;
+  LinearFunction<Field>   &_Smoother;
   
   // more most opertor functions
-  TwoLevelFlexiblePcg(RealD tol,
-		     Integer maxit,
-		     LinearOperatorBase<Field> *Linop,
-		     LinearOperatorBase<Field> *SmootherLinop,
-		     OperatorFunction<Field>   *Smoother,
-		     OperatorFunction<CoarseField>  CoarseLinop
-		     ) : 
+  TwoLevelCG(RealD tol,
+	     Integer maxit,
+	     LinearOperatorBase<Field>   &FineLinop,
+	     LinearFunction<Field>       &Smoother,
+	     GridBase *fine) : 
       Tolerance(tol), 
       MaxIterations(maxit),
-      _Linop(Linop),
-      _PreconditionerLinop(PrecLinop),
-      _Preconditioner(Preconditioner)
+      _FineLinop(FineLinop),
+      _Smoother(Smoother)
   {
-    verbose=0;
+    grid       = fine;
   };
   
-  // The Pcg routine is common to all, but the various matrices differ from derived 
-  // implementation to derived implmentation
-  void operator() (const Field &src, Field &psi){
-  void operator() (const Field &src, Field &psi){
-
-    psi.Checkerboard() = src.Checkerboard();
-    grid             = src.Grid();
-
+  virtual void operator() (const Field &src, Field &x)
+  {
+    Field resid(grid);
     RealD f;
     RealD rtzp,rtz,a,d,b;
     RealD rptzp;
-    RealD tn;
-    RealD guess = norm2(psi);
-    RealD ssq   = norm2(src);
-    RealD rsq   = ssq*Tolerance*Tolerance;
     
-    /////////////////////////////
-    // Set up history vectors
-    /////////////////////////////
-    std::vector<Field> p  (mmax,grid);
-    std::vector<Field> mmp(mmax,grid);
-    std::vector<RealD> pAp(mmax);
-
-    Field x  (grid); x = psi;
-    Field z  (grid);
+    Field p(grid);
+    Field z(grid);
     Field tmp(grid);
+    Field mmp(grid);
     Field r  (grid);
     Field mu (grid);
+    Field rp (grid);
     
+    //Initial residual computation & set up
+    double tn;
+
+    GridStopWatch HDCGTimer;
+    HDCGTimer.Start();
     //////////////////////////
     // x0 = Vstart -- possibly modify guess
     //////////////////////////
-    x=src;
+    x=Zero();
     Vstart(x,src);
 
     // r0 = b -A x0
-    HermOp(x,mmp); // Shouldn't this be something else?
-    axpy (r, -1.0,mmp[0], src);    // Recomputes r=src-Ax0
+    _FineLinop.HermOp(x,mmp);
+
+    axpy(r, -1.0, mmp, src);    // Recomputes r=src-x0
+    rp=r;
 
     //////////////////////////////////
     // Compute z = M1 x
     //////////////////////////////////
-    M1(r,z,tmp,mp,SmootherMirs);
+    PcgM1(r,z);
     rtzp =real(innerProduct(r,z));
 
     ///////////////////////////////////////
-    // Solve for Mss mu = P A z and set p = z-mu
-    // Def2: p = 1 - Q Az = Pright z 
-    // Other algos M2 is trivial
+    // Except Def2, M2 is trivial
     ///////////////////////////////////////
-    M2(z,p[0]);
+    p=z;
 
-    for (int k=0;k<=MaxIterations;k++){
+    RealD ssq =  norm2(src);
+    RealD rsq =  ssq*Tolerance*Tolerance;
 
-      int peri_k  = k % mmax;
-      int peri_kp = (k+1) % mmax;
+    std::cout<<GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" target rsq "<<rsq<<" ssq "<<ssq<<std::endl;
+    
+    for (int k=1;k<=MaxIterations;k++){
 
       rtz=rtzp;
-      d= M3(p[peri_k],mp,mmp[peri_k],tmp);
+      d= PcgM3(p,mmp);
       a = rtz/d;
 
-      // Memorise this
-      pAp[peri_k] = d;
+      axpy(x,a,p,x);
+      RealD rn = axpy_norm(r,-a,mmp,r);
 
-      axpy(x,a,p[peri_k],x);
-      RealD rn = axpy_norm(r,-a,mmp[peri_k],r);
-
-      // Compute z = M x
-      M1(r,z,tmp,mp);
+      PcgM1(r,z);
 
       rtzp =real(innerProduct(r,z));
 
-      M2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
-
-      p[peri_kp]=p[peri_k];
-
-      // Standard search direction  p -> z + b p    ; b = 
-      b = (rtzp)/rtz;
-
-      int northog;
-      //    northog     = (peri_kp==0)?1:peri_kp; // This is the fCG(mmax) algorithm
-      northog     = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
-    
-      for(int back=0; back < northog; back++){
-	int peri_back = (k-back)%mmax;
-	RealD pbApk= real(innerProduct(mmp[peri_back],p[peri_kp]));
-	RealD beta = -pbApk/pAp[peri_back];
-	axpy(p[peri_kp],beta,p[peri_back],p[peri_kp]);
+      int ipcg=1; // almost free inexact preconditioned CG
+      if (ipcg) {
+	rptzp =real(innerProduct(rp,z));
+      } else {
+	rptzp =0;
       }
+      b = (rtzp-rptzp)/rtz;
+
+      PcgM2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+
+      axpy(p,b,p,mu);  // mu = A r
 
       RealD rrn=sqrt(rn/ssq);
-      std::cout<<GridLogMessage<<"TwoLevelfPcg: k= "<<k<<" residual = "<<rrn<<std::endl;
+      RealD rtn=sqrt(rtz/ssq);
+      std::cout<<GridLogMessage<<"HDCG: Pcg k= "<<k<<" residual = "<<rrn<<std::endl;
+
+      if ( ipcg ) {
+	axpy(rp,0.0,r,r);
+      }
 
       // Stopping condition
       if ( rn <= rsq ) { 
 
-	HermOp(x,mmp); // Shouldn't this be something else?
-	axpy(tmp,-1.0,src,mmp[0]);
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: Pcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
 
-	RealD psinorm = sqrt(norm2(x));
-	RealD srcnorm = sqrt(norm2(src));
-	RealD tmpnorm = sqrt(norm2(tmp));
-	RealD true_residual = tmpnorm/srcnorm;
-	std::cout<<GridLogMessage<<"TwoLevelfPcg:   true residual is "<<true_residual<<std::endl;
-	std::cout<<GridLogMessage<<"TwoLevelfPcg: target residual was"<<Tolerance<<std::endl;
-	return k;
+	_FineLinop.HermOp(x,mmp);			  
+	axpy(tmp,-1.0,src,mmp);
+
+	RealD  mmpnorm = sqrt(norm2(mmp));
+	RealD  xnorm   = sqrt(norm2(x));
+	RealD  srcnorm = sqrt(norm2(src));
+	RealD  tmpnorm = sqrt(norm2(tmp));
+	RealD  true_residual = tmpnorm/srcnorm;
+	std::cout<<GridLogMessage
+		 <<"HDCG: true residual is "<<true_residual
+		 <<" solution "<<xnorm
+		 <<" source "<<srcnorm
+		 <<" mmp "<<mmpnorm	  
+		 <<std::endl;
+
+	return;
       }
+
     }
-    // Non-convergence
-    assert(0);
+    std::cout<<GridLogMessage<<"HDCG: not converged"<<std::endl;
+    RealD  xnorm   = sqrt(norm2(x));
+    RealD  srcnorm = sqrt(norm2(src));
+    std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+    
+    return ;
   }
   
+
  public:
 
-  virtual void M(Field & in,Field & out,Field & tmp) {
+  virtual void PcgM1(Field & in, Field & out)     =0;
+  virtual void Vstart(Field & x,const Field & src)=0;
 
+  virtual void PcgM2(const Field & in, Field & out) {
+    out=in;
   }
 
-  virtual void M1(Field & in, Field & out) {// the smoother
+  virtual RealD PcgM3(const Field & p, Field & mmp){
+    RealD dd;
+    _FineLinop.HermOp(p,mmp);
+    ComplexD dot = innerProduct(p,mmp);
+    dd=real(dot);
+    return dd;
+  }
 
+  /////////////////////////////////////////////////////////////////////
+  // Only Def1 has non-trivial Vout.
+  /////////////////////////////////////////////////////////////////////
+
+};
+  
+template<class Field, class CoarseField, class Aggregation>
+class TwoLevelADEF2 : public TwoLevelCG<Field>
+{
+ public:
+  ///////////////////////////////////////////////////////////////////////////////////
+  // Need something that knows how to get from Coarse to fine and back again
+  //  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
+  //  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+  ///////////////////////////////////////////////////////////////////////////////////
+  GridBase *coarsegrid;
+  Aggregation &_Aggregates;                    
+  LinearFunction<CoarseField> &_CoarseSolver;
+  LinearFunction<CoarseField> &_CoarseSolverPrecise;
+  ///////////////////////////////////////////////////////////////////////////////////
+  
+  // more most opertor functions
+  TwoLevelADEF2(RealD tol,
+		Integer maxit,
+		LinearOperatorBase<Field>    &FineLinop,
+		LinearFunction<Field>        &Smoother,
+		LinearFunction<CoarseField>  &CoarseSolver,
+		LinearFunction<CoarseField>  &CoarseSolverPrecise,
+		Aggregation &Aggregates
+		) :
+      TwoLevelCG<Field>(tol,maxit,FineLinop,Smoother,Aggregates.FineGrid),
+      _CoarseSolver(CoarseSolver),
+      _CoarseSolverPrecise(CoarseSolverPrecise),
+      _Aggregates(Aggregates)
+  {
+    coarsegrid = Aggregates.CoarseGrid;
+  };
+
+  virtual void PcgM1(Field & in, Field & out)
+  {
     // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
-    Field tmp(grid);
-    Field Min(grid);
 
-    PcgM(in,Min); // Smoother call
+    Field tmp(this->grid);
+    Field Min(this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
 
-    HermOp(Min,out);
+    GridStopWatch SmootherTimer;
+    GridStopWatch MatrixTimer;
+    SmootherTimer.Start();
+    this->_Smoother(in,Min);
+    SmootherTimer.Stop();
+
+    MatrixTimer.Start();
+    this->_FineLinop.HermOp(Min,out);
+    MatrixTimer.Stop();
     axpy(tmp,-1.0,out,in);          // tmp  = in - A Min
 
-    ProjectToSubspace(tmp,PleftProj);     
-    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
-    PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    GridStopWatch ProjTimer;
+    GridStopWatch CoarseTimer;
+    GridStopWatch PromTimer;
+    ProjTimer.Start();
+    this->_Aggregates.ProjectToSubspace(PleftProj,tmp);     
+    ProjTimer.Stop();
+    CoarseTimer.Start();
+    this->_CoarseSolver(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
+    CoarseTimer.Stop();
+    PromTimer.Start();
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    PromTimer.Stop();
+    std::cout << GridLogPerformance << "PcgM1 breakdown "<<std::endl;
+    std::cout << GridLogPerformance << "\tSmoother   " << SmootherTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tProj       " << ProjTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tCoarse     " << CoarseTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tProm       " << PromTimer.Elapsed() <<std::endl;
+
     axpy(out,1.0,Min,tmp); // Min+tmp
   }
 
-  virtual void M2(const Field & in, Field & out) {
-    out=in;
-    // Must override for Def2 only
-    //  case PcgDef2:
-    //    Pright(in,out);
-    //    break;
-  }
-
-  virtual RealD M3(const Field & p, Field & mmp){
-    double d,dd;
-    HermOpAndNorm(p,mmp,d,dd);
-    return dd;
-    // Must override for Def1 only
-    //  case PcgDef1:
-    //    d=linop_d->Mprec(p,mmp,tmp,0,1);// Dag no
-    //      linop_d->Mprec(mmp,mp,tmp,1);// Dag yes
-    //    Pleft(mp,mmp);
-    //    d=real(linop_d->inner(p,mmp));
-  }
-
-  virtual void VstartDef2(Field & xconst Field & src){
-    //case PcgDef2:
-    //case PcgAdef2: 
-    //case PcgAdef2f:
-    //case PcgV11f:
+  virtual void Vstart(Field & x,const Field & src)
+  {
     ///////////////////////////////////
     // Choose x_0 such that 
     // x_0 = guess +  (A_ss^inv) r_s = guess + Ass_inv [src -Aguess]
@@ -256,142 +292,73 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
     //                   = src_s - (A guess)_s - src_s  + (A guess)_s 
     //                   = 0 
     ///////////////////////////////////
-    Field r(grid);
-    Field mmp(grid);
+    Field r(this->grid);
+    Field mmp(this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
 
-    HermOp(x,mmp);
-    axpy (r, -1.0, mmp, src);        // r_{-1} = src - A x
-    ProjectToSubspace(r,PleftProj);     
-    ApplyInverseCG(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    PromoteFromSubspace(PleftMss_proj,mmp);  
-    x=x+mmp;
+    this->_Aggregates.ProjectToSubspace(PleftProj,src);     
+    this->_CoarseSolverPrecise(PleftProj,PleftMss_proj); // Ass^{-1} r_s
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,x);  
 
   }
 
+};
+
+template<class Field>
+class TwoLevelADEF1defl : public TwoLevelCG<Field>
+{
+public:
+  const std::vector<Field> &evec;
+  const std::vector<RealD> &eval;
+  
+  TwoLevelADEF1defl(RealD tol,
+		   Integer maxit,
+		   LinearOperatorBase<Field>   &FineLinop,
+		   LinearFunction<Field>   &Smoother,
+		   std::vector<Field> &_evec,
+		   std::vector<RealD> &_eval) : 
+    TwoLevelCG<Field>(tol,maxit,FineLinop,Smoother,_evec[0].Grid()),
+    evec(_evec),
+    eval(_eval)
+  {};
+
+  // Can just inherit existing M2
+  // Can just inherit existing M3
+
+  // Simple vstart - do nothing
   virtual void Vstart(Field & x,const Field & src){
-    return;
+    x=src; // Could apply Q
+  };
+
+  // Override PcgM1
+  virtual void PcgM1(Field & in, Field & out)
+  {
+    int N=evec.size();
+    Field Pin(this->grid);
+    Field Qin(this->grid);
+
+    //MP  + Q = M(1-AQ) + Q = M
+    // // If we are eigenvector deflating in coarse space
+    // // Q   = Sum_i |phi_i> 1/lambda_i <phi_i|
+    // // A Q = Sum_i |phi_i> <phi_i|
+    // // M(1-AQ) = M(1-proj) + Q
+    Qin.Checkerboard()=in.Checkerboard();
+    Qin = Zero();
+    Pin = in;
+    for (int i=0;i<N;i++) {
+      const Field& tmp = evec[i];
+      auto ip = TensorRemove(innerProduct(tmp,in));
+      axpy(Qin, ip / eval[i],tmp,Qin);
+      axpy(Pin, -ip ,tmp,Pin);
+    }
+
+    this->_Smoother(Pin,out);
+
+    out = out + Qin;
   }
+};
 
-  /////////////////////////////////////////////////////////////////////
-  // Only Def1 has non-trivial Vout. Override in Def1
-  /////////////////////////////////////////////////////////////////////
-  virtual void   Vout  (Field & in, Field & out,Field & src){
-    out = in;
-    //case PcgDef1:
-    //    //Qb + PT x
-    //    ProjectToSubspace(src,PleftProj);     
-    //    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    //    PromoteFromSubspace(PleftMss_proj,tmp);  
-    //    
-    //    Pright(in,out);
-    //    
-    //    linop_d->axpy(out,tmp,out,1.0);
-    //    break;
-  }
+NAMESPACE_END(Grid);
 
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // Pright and Pleft are common to all implementations
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  virtual void Pright(Field & in,Field & out){
-    // P_R  = [ 1              0 ] 
-    //        [ -Mss^-1 Msb    0 ] 
-    Field in_sbar(grid);
-
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
-    axpy(in_sbar,-1.0,out,in);       // in_sbar = in - in_s 
-
-    HermOp(in_sbar,out);
-    ProjectToSubspace(out,PleftProj);           // Mssbar in_sbar  (project)
-
-    ApplyInverse     (PleftProj,PleftMss_proj); // Mss^{-1} Mssbar 
-    PromoteFromSubspace(PleftMss_proj,out);     // 
-
-    axpy(out,-1.0,out,in_sbar);     // in_sbar - Mss^{-1} Mssbar in_sbar
-  }
-  virtual void Pleft (Field & in,Field & out){
-    // P_L  = [ 1  -Mbs Mss^-1] 
-    //        [ 0   0         ] 
-    Field in_sbar(grid);
-    Field    tmp2(grid);
-    Field    Mtmp(grid);
-
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
-    axpy(in_sbar,-1.0,out,in);      // in_sbar = in - in_s
-
-    ApplyInverse(PleftProj,PleftMss_proj); // Mss^{-1} in_s
-    PromoteFromSubspace(PleftMss_proj,out);
-
-    HermOp(out,Mtmp);
-
-    ProjectToSubspace(Mtmp,PleftProj);      // Msbar s Mss^{-1}
-    PromoteFromSubspace(PleftProj,tmp2);
-
-    axpy(out,-1.0,tmp2,Mtmp);
-    axpy(out,-1.0,out,in_sbar);     // in_sbar - Msbars Mss^{-1} in_s
-  }
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgADef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp){
-
-  } 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp){
-
-  }
-  virtual void M2(Field & in, Field & out){
-
-  }
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp){
-
-  }
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp){
-
-  }
-}
-/*
-template<class Field>
-class TwoLevelFlexiblePcgAD : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef1 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-  virtual void   Vout  (Field & in, Field & out,Field & src,Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgV11: public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-*/
 #endif

Grid/algorithms/iterative/ConjugateGradient.h

@@ -183,13 +183,13 @@ public:
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
 	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 << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "Time breakdown "<<std::endl;
+	std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
 	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -457,7 +457,7 @@ until convergence
 	    std::vector<Field>& evec,
 	    Field& w,int Nm,int k)
   {
-    std::cout<<GridLogIRL << "Lanczos step " <<k<<std::endl;
+    std::cout<<GridLogDebug << "Lanczos step " <<k<<std::endl;
     const RealD tiny = 1.0e-20;
     assert( k< Nm );
 
@@ -465,7 +465,7 @@ until convergence
 
     Field& evec_k = evec[k];
 
-    _PolyOp(evec_k,w);    std::cout<<GridLogIRL << "PolyOp" <<std::endl;
+    _PolyOp(evec_k,w);    std::cout<<GridLogDebug << "PolyOp" <<std::endl;
 
     if(k>0) w -= lme[k-1] * evec[k-1];
 
@@ -480,18 +480,18 @@ until convergence
     lme[k] = beta;
 
     if ( (k>0) && ( (k % orth_period) == 0 )) {
-      std::cout<<GridLogIRL << "Orthogonalising " <<k<<std::endl;
+      std::cout<<GridLogDebug << "Orthogonalising " <<k<<std::endl;
       orthogonalize(w,evec,k); // orthonormalise
-      std::cout<<GridLogIRL << "Orthogonalised " <<k<<std::endl;
+      std::cout<<GridLogDebug << "Orthogonalised " <<k<<std::endl;
     }
 
     if(k < Nm-1) evec[k+1] = w;
 
-    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+    std::cout<<GridLogIRL << "Lanczos step alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
     if ( beta < tiny ) 
       std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
 
-    std::cout<<GridLogIRL << "Lanczos step complete " <<k<<std::endl;
+    std::cout<<GridLogDebug << "Lanczos step complete " <<k<<std::endl;
   }
 
   void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 

Grid/algorithms/iterative/NormalEquations.h

@@ -33,7 +33,7 @@ NAMESPACE_BEGIN(Grid);
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
 // Take a matrix and form an NE solver calling a Herm solver
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class NormalEquations {
+template<class Field> class NormalEquations : public LinearFunction<Field>{
 private:
   SparseMatrixBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;
@@ -60,7 +60,7 @@ public:
   }     
 };
 
-template<class Field> class HPDSolver {
+template<class Field> class HPDSolver : public LinearFunction<Field> {
 private:
   LinearOperatorBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;
@@ -78,13 +78,13 @@ public:
   void operator() (const Field &in, Field &out){
  
     _Guess(in,out);
-    _HermitianSolver(_Matrix,in,out);  // Mdag M out = Mdag in
+    _HermitianSolver(_Matrix,in,out);  //M out = in
 
   }     
 };
 
 
-template<class Field> class MdagMSolver {
+template<class Field> class MdagMSolver : public LinearFunction<Field> {
 private:
   SparseMatrixBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;

Grid/algorithms/iterative/PowerMethod.h

@@ -20,7 +20,7 @@ template<class Field> class PowerMethod
     RealD evalMaxApprox = 0.0; 
     auto src_n = src; 
     auto tmp = src; 
-    const int _MAX_ITER_EST_ = 50; 
+    const int _MAX_ITER_EST_ = 100; 
 
     for (int i=0;i<_MAX_ITER_EST_;i++) { 
       

Grid/algorithms/multigrid/Aggregates.h

@@ -0,0 +1,262 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/Aggregates.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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 Fobj,class CComplex,int nbasis>
+class Aggregation {
+public:
+  typedef iVector<CComplex,nbasis >             siteVector;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+
+  GridBase *CoarseGrid;
+  GridBase *FineGrid;
+  std::vector<Lattice<Fobj> > subspace;
+  int checkerboard;
+  int Checkerboard(void){return checkerboard;}
+  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
+    CoarseGrid(_CoarseGrid),
+    FineGrid(_FineGrid),
+    subspace(nbasis,_FineGrid),
+    checkerboard(_checkerboard)
+  {
+  };
+  
+  
+  void Orthogonalise(void){
+    CoarseScalar InnerProd(CoarseGrid); 
+    //    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
+    blockOrthogonalise(InnerProd,subspace);
+  } 
+  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
+    blockProject(CoarseVec,FineVec,subspace);
+  }
+  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+    FineVec.Checkerboard() = subspace[0].Checkerboard();
+    blockPromote(CoarseVec,FineVec,subspace);
+  }
+
+  virtual void CreateSubspaceRandom(GridParallelRNG  &RNG) {
+    int nn=nbasis;
+    RealD scale;
+    FineField noise(FineGrid);
+    for(int b=0;b<nn;b++){
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      subspace[b] = noise;
+    }
+  }
+  virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis)
+  {
+
+    RealD scale;
+
+    ConjugateGradient<FineField> CG(1.0e-2,100,false);
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+      for(int i=0;i<1;i++){
+
+	CG(hermop,noise,subspace[b]);
+
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
+  // and this is the best I found
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+
+  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo,
+				       int orderfilter,
+				       int ordermin,
+				       int orderstep,
+				       double filterlo
+				       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" Chebyshev subspace pass-1 : ord "<<orderfilter<<" ["<<lo<<","<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pass-2 : nbasis"<<nn<<" min "
+	      <<ordermin<<" step "<<orderstep
+	      <<" lo"<<filterlo<<std::endl;
+
+    // Initial matrix element
+    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      // Filter
+      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      b++;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    {
+      lo=filterlo;
+      noise=Mn;
+
+      FineField T0(FineGrid); T0 = noise;  
+      FineField T1(FineGrid); 
+      FineField T2(FineGrid);
+      FineField y(FineGrid);
+      
+      FineField *Tnm = &T0;
+      FineField *Tn  = &T1;
+      FineField *Tnp = &T2;
+
+      // Tn=T1 = (xscale M + mscale)in
+      RealD xscale = 2.0/(hi-lo);
+      RealD mscale = -(hi+lo)/(hi-lo);
+      hermop.HermOp(T0,y);
+      T1=y*xscale+noise*mscale;
+
+      for(int n=2;n<=ordermin+orderstep*(nn-2);n++){
+	
+	hermop.HermOp(*Tn,y);
+
+	autoView( y_v , y, AcceleratorWrite);
+	autoView( Tn_v , (*Tn), AcceleratorWrite);
+	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
+	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
+	const int Nsimd = CComplex::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));
+        });
+
+	// Possible more fine grained control is needed than a linear sweep,
+	// but huge productivity gain if this is simple algorithm and not a tunable
+	int m =1;
+	if ( n>=ordermin ) m=n-ordermin;
+	if ( (m%orderstep)==0 ) { 
+	  Mn=*Tnp;
+	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
+	  subspace[b] = Mn;
+	  hermop.Op(Mn,tmp); 
+	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+	  b++;
+	}
+
+	// Cycle pointers to avoid copies
+	FineField *swizzle = Tnm;
+	Tnm    =Tn;
+	Tn     =Tnp;
+	Tnp    =swizzle;
+	  
+      }
+    }
+    assert(b==nn);
+  }
+  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo,
+				       int orderfilter
+				       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise : ord "<<orderfilter<<" ["<<lo<<","<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise  : nbasis "<<nn<<std::endl;
+
+
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+      // Filter
+      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+
+  }
+
+};
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/CoarsenedMatrix.h

@@ -56,243 +56,6 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
   blockSum(CoarseInner,fine_inner_msk);
 }
 
-
-class Geometry {
-public:
-  int npoint;
-  int base;
-  std::vector<int> directions   ;
-  std::vector<int> displacements;
-  std::vector<int> points_dagger;
-
-  Geometry(int _d)  {
-    
-    base = (_d==5) ? 1:0;
-
-    // make coarse grid stencil for 4d , not 5d
-    if ( _d==5 ) _d=4;
-
-    npoint = 2*_d+1;
-    directions.resize(npoint);
-    displacements.resize(npoint);
-    points_dagger.resize(npoint);
-    for(int d=0;d<_d;d++){
-      directions[d   ] = d+base;
-      directions[d+_d] = d+base;
-      displacements[d  ] = +1;
-      displacements[d+_d]= -1;
-      points_dagger[d   ] = d+_d;
-      points_dagger[d+_d] = d;
-    }
-    directions   [2*_d]=0;
-    displacements[2*_d]=0;
-    points_dagger[2*_d]=2*_d;
-  }
-
-  int point(int dir, int disp) {
-    assert(disp == -1 || disp == 0 || disp == 1);
-    assert(base+0 <= dir && dir < base+4);
-
-    // directions faster index = new indexing
-    // 4d (base = 0):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   0  1  2  3  0  1  2  3  0
-    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
-    // 5d (base = 1):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   1  2  3  4  1  2  3  4  0
-    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
-
-    // displacements faster index = old indexing
-    // 4d (base = 0):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   0  0  1  1  2  2  3  3  0
-    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
-    // 5d (base = 1):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   1  1  2  2  3  3  4  4  0
-    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
-
-    if(dir == 0 and disp == 0)
-      return 8;
-    else // New indexing
-      return (1 - disp) / 2 * 4 + dir - base;
-    // else // Old indexing
-    //   return (4 * (dir - base) + 1 - disp) / 2;
-  }
-};
-  
-template<class Fobj,class CComplex,int nbasis>
-class Aggregation   {
-public:
-  typedef iVector<CComplex,nbasis >             siteVector;
-  typedef Lattice<siteVector>                 CoarseVector;
-  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
-
-  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
-  typedef Lattice<Fobj >        FineField;
-
-  GridBase *CoarseGrid;
-  GridBase *FineGrid;
-  std::vector<Lattice<Fobj> > subspace;
-  int checkerboard;
-  int Checkerboard(void){return checkerboard;}
-  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
-    CoarseGrid(_CoarseGrid),
-    FineGrid(_FineGrid),
-    subspace(nbasis,_FineGrid),
-    checkerboard(_checkerboard)
-  {
-  };
-  
-  void Orthogonalise(void){
-    CoarseScalar InnerProd(CoarseGrid); 
-    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
-    blockOrthogonalise(InnerProd,subspace);
-  } 
-  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
-    blockProject(CoarseVec,FineVec,subspace);
-  }
-  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
-    FineVec.Checkerboard() = subspace[0].Checkerboard();
-    blockPromote(CoarseVec,FineVec,subspace);
-  }
-
-  virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
-
-    RealD scale;
-
-    ConjugateGradient<FineField> CG(1.0e-2,100,false);
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-
-    for(int b=0;b<nn;b++){
-      
-      subspace[b] = Zero();
-      gaussian(RNG,noise);
-      scale = std::pow(norm2(noise),-0.5); 
-      noise=noise*scale;
-      
-      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-      for(int i=0;i<1;i++){
-
-	CG(hermop,noise,subspace[b]);
-
-	noise = subspace[b];
-	scale = std::pow(norm2(noise),-0.5); 
-	noise=noise*scale;
-
-      }
-
-      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
-      subspace[b]   = noise;
-
-    }
-  }
-
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
-  // and this is the best I found
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-    {
-      // Filter
-      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
-      Cheb(hermop,noise,Mn);
-      // normalise
-      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
-      subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
-      b++;
-    }
-
-    // Generate a full sequence of Chebyshevs
-    {
-      lo=filterlo;
-      noise=Mn;
-
-      FineField T0(FineGrid); T0 = noise;  
-      FineField T1(FineGrid); 
-      FineField T2(FineGrid);
-      FineField y(FineGrid);
-      
-      FineField *Tnm = &T0;
-      FineField *Tn  = &T1;
-      FineField *Tnp = &T2;
-
-      // Tn=T1 = (xscale M + mscale)in
-      RealD xscale = 2.0/(hi-lo);
-      RealD mscale = -(hi+lo)/(hi-lo);
-      hermop.HermOp(T0,y);
-      T1=y*xscale+noise*mscale;
-
-      for(int n=2;n<=ordermin+orderstep*(nn-2);n++){
-	
-	hermop.HermOp(*Tn,y);
-
-	autoView( y_v , y, AcceleratorWrite);
-	autoView( Tn_v , (*Tn), AcceleratorWrite);
-	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
-	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
-	const int Nsimd = CComplex::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));
-        });
-
-	// Possible more fine grained control is needed than a linear sweep,
-	// but huge productivity gain if this is simple algorithm and not a tunable
-	int m =1;
-	if ( n>=ordermin ) m=n-ordermin;
-	if ( (m%orderstep)==0 ) { 
-	  Mn=*Tnp;
-	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
-	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
-	  b++;
-	}
-
-	// Cycle pointers to avoid copies
-	FineField *swizzle = Tnm;
-	Tnm    =Tn;
-	Tn     =Tnp;
-	Tnp    =swizzle;
-	  
-      }
-    }
-    assert(b==nn);
-  }
-
-};
-
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -0,0 +1,419 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrix.h
+
+    Copyright (C) 2015
+
+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
+
+#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
+
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+NAMESPACE_BEGIN(Grid);
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class GeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  typedef Lattice<iScalar<CComplex> >         CoarseComplexField;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+  typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+  typedef CoarseVector Field;
+  ////////////////////
+  // Data members
+  ////////////////////
+  int hermitian;
+  GridBase      *       _FineGrid; 
+  GridCartesian *       _CoarseGrid; 
+  NonLocalStencilGeometry &geom;
+  PaddedCell Cell;
+  GeneralLocalStencil Stencil;
+  
+  std::vector<CoarseMatrix> _A;
+  std::vector<CoarseMatrix> _Adag;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase      * Grid(void)           { return _FineGrid; };   // this is all the linalg routines need to know
+  GridBase      * FineGrid(void)       { return _FineGrid; };   // this is all the linalg routines need to know
+  GridCartesian * CoarseGrid(void)     { return _CoarseGrid; };   // this is all the linalg routines need to know
+
+  void ProjectNearestNeighbour(RealD shift, GeneralCoarseOp &CopyMe)
+  {
+    int nfound=0;
+    std::cout << GridLogMessage <<"GeneralCoarsenedMatrix::ProjectNearestNeighbour "<< CopyMe._A[0].Grid()<<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      for(int pp=0;pp<CopyMe.geom.npoint;pp++){
+ 	// Search for the same relative shift
+	// Avoids brutal handling of Grid pointers
+	if ( CopyMe.geom.shifts[pp]==geom.shifts[p] ) {
+	  _A[p] = CopyMe.Cell.Extract(CopyMe._A[pp]);
+	  _Adag[p] = CopyMe.Cell.Extract(CopyMe._Adag[pp]);
+	  nfound++;
+	}
+      }
+    }
+    assert(nfound==geom.npoint);
+    ExchangeCoarseLinks();
+  }
+  
+  GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid)
+    : geom(_geom),
+      _FineGrid(FineGrid),
+      _CoarseGrid(CoarseGrid),
+      hermitian(1),
+      Cell(_geom.Depth(),_CoarseGrid),
+      Stencil(Cell.grids.back(),geom.shifts)
+  {
+    {
+      int npoint = _geom.npoint;
+      autoView( Stencil_v  , Stencil, AcceleratorRead);
+      int osites=Stencil.Grid()->oSites();
+      for(int ss=0;ss<osites;ss++){
+	for(int point=0;point<npoint;point++){
+	  auto SE = Stencil_v.GetEntry(point,ss);
+	  int o = SE->_offset;
+	  assert( o< osites);
+	}
+      }    
+    }
+
+    _A.resize(geom.npoint,CoarseGrid);
+    _Adag.resize(geom.npoint,CoarseGrid);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    Mult(_A,in,out);
+  }
+  void Mdag (const CoarseVector &in, CoarseVector &out)
+  {
+    if ( hermitian ) M(in,out);
+    else Mult(_Adag,in,out);
+  }
+  void Mult (std::vector<CoarseMatrix> &A,const CoarseVector &in, CoarseVector &out)
+  {
+    RealD tviews=0;
+    RealD ttot=0;
+    RealD tmult=0;
+    RealD texch=0;
+    RealD text=0;
+    ttot=-usecond();
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+    CoarseVector tin=in;
+
+    texch-=usecond();
+    CoarseVector pin  = Cell.Exchange(tin);
+    texch+=usecond();
+
+    CoarseVector pout(pin.Grid()); pout=Zero();
+
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+      
+    const int Nsimd = CComplex::Nsimd();
+    
+    int osites=pin.Grid()->oSites();
+    //    int gsites=pin.Grid()->gSites();
+
+    RealD flops = 1.0* npoint * nbasis * nbasis * 8 * osites;
+    RealD bytes = (1.0*osites*sizeof(siteMatrix)*npoint+2.0*osites*sizeof(siteVector))*npoint;
+      
+    //    for(int point=0;point<npoint;point++){
+    //      conformable(A[point],pin);
+    //    }
+
+    {
+      tviews-=usecond();
+      autoView( in_v , pin, AcceleratorRead);
+      autoView( out_v , pout, AcceleratorWrite);
+      autoView( Stencil_v  , Stencil, AcceleratorRead);
+      tviews+=usecond();
+      
+      for(int point=0;point<npoint;point++){
+	tviews-=usecond();
+	autoView( A_v, A[point],AcceleratorRead);
+	tviews+=usecond();
+	tmult-=usecond();
+	accelerator_for(sss, osites*nbasis, Nsimd, {
+
+	    typedef decltype(coalescedRead(in_v[0]))    calcVector;
+
+	    int ss = sss/nbasis;
+	    int b  = sss%nbasis;
+
+	    auto SE  = Stencil_v.GetEntry(point,ss);
+	    auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
+	    auto res = out_v(ss)(b);
+	    for(int bb=0;bb<nbasis;bb++) {
+	      res = res + coalescedRead(A_v[ss](b,bb))*nbr(bb);
+	    }
+	    coalescedWrite(out_v[ss](b),res);
+	});
+
+	tmult+=usecond();
+      }
+    }
+    text-=usecond();
+    out = Cell.Extract(pout);
+    text+=usecond();
+    ttot+=usecond();
+    
+    std::cout << GridLogDebug<<"Coarse Mult Aviews "<<tviews<<" us"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse Mult exch "<<texch<<" us"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse Mult ext  "<<text<<" us"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse Mult tot  "<<ttot<<" us"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse Kernel bytes/s"<< bytes/tmult<<" MB/s"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;
+    std::cout << GridLogDebug<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+
+  };
+
+  void PopulateAdag(void)
+  {
+    for(int64_t bidx=0;bidx<CoarseGrid()->gSites() ;bidx++){
+      Coordinate bcoor;
+      CoarseGrid()->GlobalIndexToGlobalCoor(bidx,bcoor);
+      
+      for(int p=0;p<geom.npoint;p++){
+	Coordinate scoor = bcoor;
+	for(int mu=0;mu<bcoor.size();mu++){
+	  int L = CoarseGrid()->GlobalDimensions()[mu];
+	  scoor[mu] = (bcoor[mu] - geom.shifts[p][mu] + L) % L; // Modulo arithmetic
+	}
+	// Flip to poke/peekLocalSite and not too bad
+	auto link = peekSite(_A[p],scoor);
+	int pp = geom.Reverse(p);
+	pokeSite(adj(link),_Adag[pp],bcoor);
+      }
+    }
+  }
+  /////////////////////////////////////////////////////////////
+  // 
+  // A) Only reduced flops option is to use a padded cell of depth 4
+  // and apply MpcDagMpc in the padded cell.
+  //
+  // Makes for ONE application of MpcDagMpc per vector instead of 30 or 80.
+  // With the effective cell size around (B+8)^4 perhaps 12^4/4^4 ratio
+  // Cost is 81x more, same as stencil size.
+  //
+  // But: can eliminate comms and do as local dirichlet.
+  //
+  // Local exchange gauge field once.
+  // Apply to all vectors, local only computation.
+  // Must exchange ghost subcells in reverse process of PaddedCell to take inner products
+  //
+  // B) Can reduce cost: pad by 1, apply Deo      (4^4+6^4+8^4+8^4 )/ (4x 4^4)
+  //                     pad by 2, apply Doe
+  //                     pad by 3, apply Deo
+  //                     then break out 8x directions; cost is ~10x MpcDagMpc per vector
+  //
+  // => almost factor of 10 in setup cost, excluding data rearrangement
+  //
+  // Intermediates -- ignore the corner terms, leave approximate and force Hermitian
+  // Intermediates -- pad by 2 and apply 1+8+24 = 33 times.
+  /////////////////////////////////////////////////////////////
+
+    //////////////////////////////////////////////////////////
+    // BFM HDCG style approach: Solve a system of equations to get Aij
+    //////////////////////////////////////////////////////////
+    /*
+     *     Here, k,l index which possible shift within the 3^Nd "ball" connected by MdagM.
+     *
+     *     conj(phases[block]) proj[k][ block*Nvec+j ] =  \sum_ball  e^{i q_k . delta} < phi_{block,j} | MdagM | phi_{(block+delta),i} > 
+     *                                                 =  \sum_ball e^{iqk.delta} A_ji
+     *
+     *     Must invert matrix M_k,l = e^[i q_k . delta_l]
+     *
+     *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+     */
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
+    GridBase *grid = FineGrid();
+
+    RealD tproj=0.0;
+    RealD teigen=0.0;
+    RealD tmat=0.0;
+    RealD tphase=0.0;
+    RealD tinv=0.0;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid()); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    const int npoint = geom.npoint;
+      
+    Coordinate clatt = CoarseGrid()->GlobalDimensions();
+    int Nd = CoarseGrid()->Nd();
+
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    teigen-=usecond();
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom.shifts[k][mu]*geom.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+    teigen+=usecond();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    CoarseVector coarseInner(CoarseGrid());
+
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid());
+    std::vector<CoarseVector>          FT(npoint,CoarseGrid());
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+	/////////////////////////////////////////////////////
+	// Stick a phase on every block
+	/////////////////////////////////////////////////////
+	tphase-=usecond();
+	CoarseComplexField coor(CoarseGrid());
+	CoarseComplexField pha(CoarseGrid());	pha=Zero();
+	for(int mu=0;mu<Nd;mu++){
+	  LatticeCoordinate(coor,mu);
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  pha = pha + (TwoPiL * geom.shifts[p][mu]) * coor;
+	}
+	pha  =exp(pha*ci);
+	phaV=Zero();
+	blockZAXPY(phaV,pha,Subspace.subspace[i],phaV);
+	tphase+=usecond();
+
+	/////////////////////////////////////////////////////////////////////
+	// Multiple phased subspace vector by matrix and project to subspace
+	// Remove local bulk phase to leave relative phases
+	/////////////////////////////////////////////////////////////////////
+	tmat-=usecond();
+	linop.Op(phaV,MphaV);
+	tmat+=usecond();
+
+	tproj-=usecond();
+	blockProject(coarseInner,MphaV,Subspace.subspace);
+	coarseInner = conjugate(pha) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+	tproj+=usecond();
+
+      }
+
+      tinv-=usecond();
+      for(int k=0;k<npoint;k++){
+	FT[k] = Zero();
+	for(int l=0;l<npoint;l++){
+	  FT[k]= FT[k]+ invMkl(l,k)*ComputeProj[l];
+	}
+      
+	int osites=CoarseGrid()->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT[k], AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](j,i) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    for(int p=0;p<geom.npoint;p++){
+      Coordinate coor({0,0,0,0,0});
+      auto sval = peekSite(_A[p],coor);
+    }
+
+    // Only needed if nonhermitian
+    if ( ! hermitian ) {
+      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      PopulateAdag();
+    }
+
+    // Need to write something to populate Adag from A
+    ExchangeCoarseLinks();
+    std::cout << GridLogMessage<<"CoarsenOperator eigen  "<<teigen<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+  }
+  void ExchangeCoarseLinks(void){
+    for(int p=0;p<geom.npoint;p++){
+      _A[p] = Cell.Exchange(_A[p]);
+      _Adag[p]= Cell.Exchange(_Adag[p]);
+    }
+  }
+  virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
+  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
+  virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
+};
+
+
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/Geometry.h

@@ -0,0 +1,243 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrix.h
+
+    Copyright (C) 2015
+
+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);
+
+
+/////////////////////////////////////////////////////////////////
+// Geometry class in cartesian case
+/////////////////////////////////////////////////////////////////
+
+class Geometry {
+public:
+  int npoint;
+  int base;
+  std::vector<int> directions   ;
+  std::vector<int> displacements;
+  std::vector<int> points_dagger;
+
+  Geometry(int _d)  {
+    
+    base = (_d==5) ? 1:0;
+
+    // make coarse grid stencil for 4d , not 5d
+    if ( _d==5 ) _d=4;
+
+    npoint = 2*_d+1;
+    directions.resize(npoint);
+    displacements.resize(npoint);
+    points_dagger.resize(npoint);
+    for(int d=0;d<_d;d++){
+      directions[d   ] = d+base;
+      directions[d+_d] = d+base;
+      displacements[d  ] = +1;
+      displacements[d+_d]= -1;
+      points_dagger[d   ] = d+_d;
+      points_dagger[d+_d] = d;
+    }
+    directions   [2*_d]=0;
+    displacements[2*_d]=0;
+    points_dagger[2*_d]=2*_d;
+  }
+
+  int point(int dir, int disp) {
+    assert(disp == -1 || disp == 0 || disp == 1);
+    assert(base+0 <= dir && dir < base+4);
+
+    // directions faster index = new indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  1  2  3  0  1  2  3  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  2  3  4  1  2  3  4  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+
+    // displacements faster index = old indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  0  1  1  2  2  3  3  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  1  2  2  3  3  4  4  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+
+    if(dir == 0 and disp == 0)
+      return 8;
+    else // New indexing
+      return (1 - disp) / 2 * 4 + dir - base;
+    // else // Old indexing
+    //   return (4 * (dir - base) + 1 - disp) / 2;
+  }
+};
+
+/////////////////////////////////////////////////////////////////
+// Less local equivalent of Geometry class in cartesian case
+/////////////////////////////////////////////////////////////////
+class NonLocalStencilGeometry {
+public:
+  int depth;
+  int hops;
+  int npoint;
+  std::vector<Coordinate> shifts;
+  Coordinate stencil_size;
+  Coordinate stencil_lo;
+  Coordinate stencil_hi;
+  GridCartesian *grid;
+  GridCartesian *Grid() {return grid;};
+  int Depth(void){return 1;};   // Ghost zone depth
+  int Hops(void){return hops;}; // # of hops=> level of corner fill in in stencil
+
+  virtual int DimSkip(void) =0;
+
+  virtual ~NonLocalStencilGeometry() {};
+
+  int  Reverse(int point)
+  {
+    int Nd = Grid()->Nd();
+    Coordinate shft = shifts[point];
+    Coordinate rev(Nd);
+    for(int mu=0;mu<Nd;mu++) rev[mu]= -shft[mu];
+    for(int p=0;p<npoint;p++){
+      if(rev==shifts[p]){
+	return p;
+      }
+    }
+    assert(0);
+    return -1;
+  }
+  void BuildShifts(void)
+  {
+    this->shifts.resize(0);
+    int Nd = this->grid->Nd();
+
+    int dd = this->DimSkip();
+    for(int s0=this->stencil_lo[dd+0];s0<=this->stencil_hi[dd+0];s0++){
+    for(int s1=this->stencil_lo[dd+1];s1<=this->stencil_hi[dd+1];s1++){
+    for(int s2=this->stencil_lo[dd+2];s2<=this->stencil_hi[dd+2];s2++){
+    for(int s3=this->stencil_lo[dd+3];s3<=this->stencil_hi[dd+3];s3++){
+      Coordinate sft(Nd,0);
+      sft[dd+0] = s0;
+      sft[dd+1] = s1;
+      sft[dd+2] = s2;
+      sft[dd+3] = s3;
+      int nhops = abs(s0)+abs(s1)+abs(s2)+abs(s3);
+      if(nhops<=this->hops) this->shifts.push_back(sft);
+    }}}}
+    this->npoint = this->shifts.size();
+    std::cout << GridLogMessage << "NonLocalStencilGeometry has "<< this->npoint << " terms in stencil "<<std::endl;
+  }
+  
+  NonLocalStencilGeometry(GridCartesian *_coarse_grid,int _hops) : grid(_coarse_grid), hops(_hops)
+  {
+    Coordinate latt = grid->GlobalDimensions();
+    stencil_size.resize(grid->Nd());
+    stencil_lo.resize(grid->Nd());
+    stencil_hi.resize(grid->Nd());
+    for(int d=0;d<grid->Nd();d++){
+     if ( latt[d] == 1 ) {
+      stencil_lo[d] = 0;
+      stencil_hi[d] = 0;
+      stencil_size[d]= 1;
+     } else if ( latt[d] == 2 ) {
+      stencil_lo[d] = -1;
+      stencil_hi[d] = 0;
+      stencil_size[d]= 2;
+     } else if ( latt[d] > 2 ) {
+       stencil_lo[d] = -1;
+       stencil_hi[d] =  1;
+       stencil_size[d]= 3;
+     }
+    }
+  };
+
+};
+
+// Need to worry about red-black now
+class NonLocalStencilGeometry4D : public NonLocalStencilGeometry {
+public:
+  virtual int DimSkip(void) { return 0;};
+  NonLocalStencilGeometry4D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops) { };
+  virtual ~NonLocalStencilGeometry4D() {};
+};
+class NonLocalStencilGeometry5D : public NonLocalStencilGeometry {
+public:
+  virtual int DimSkip(void) { return 1; }; 
+  NonLocalStencilGeometry5D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops)  { };
+  virtual ~NonLocalStencilGeometry5D() {};
+};
+/*
+ * Bunch of different options classes
+ */
+class NextToNextToNextToNearestStencilGeometry4D : public NonLocalStencilGeometry4D {
+public:
+  NextToNextToNextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,4)
+  {
+    this->BuildShifts();
+  };
+};
+class NextToNextToNextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNextToNextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,4)
+  {
+    this->BuildShifts();
+  };
+};
+class NextToNearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,2)
+  {
+    this->BuildShifts();
+  };
+};
+class NextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,2)
+  {
+    this->BuildShifts();
+  };
+};
+class NearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,1)
+  {
+    this->BuildShifts();
+  };
+};
+class NearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,1)
+  {
+    this->BuildShifts();
+  };
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/MultiGrid.h

@@ -0,0 +1,33 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Grid/algorithms/multigrid/MultiGrid.h
+
+    Copyright (C) 2023
+
+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
+
+#include <Grid/algorithms/multigrid/Aggregates.h>
+#include <Grid/algorithms/multigrid/Geometry.h>
+#include <Grid/algorithms/multigrid/CoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h>

Grid/cartesian/Cartesian_base.h

@@ -70,8 +70,8 @@ public:
   Coordinate _istride;    // Inner stride i.e. within simd lane
   int _osites;                  // _isites*_osites = product(dimensions).
   int _isites;
-  int _fsites;                  // _isites*_osites = product(dimensions).
-  int _gsites;
+  int64_t _fsites;                  // _isites*_osites = product(dimensions).
+  int64_t _gsites;
   Coordinate _slice_block;// subslice information
   Coordinate _slice_stride;
   Coordinate _slice_nblock;
@@ -183,7 +183,7 @@ public:
   inline int Nsimd(void)  const { return _isites; };// Synonymous with iSites
   inline int oSites(void) const { return _osites; };
   inline int lSites(void) const { return _isites*_osites; }; 
-  inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
+  inline int64_t gSites(void) const { return (int64_t)_isites*(int64_t)_osites*(int64_t)_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
   inline const Coordinate LocalStarts(void)             { return _lstart;    };
@@ -214,7 +214,7 @@ public:
   ////////////////////////////////////////////////////////////////
   // Global addressing
   ////////////////////////////////////////////////////////////////
-  void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
+  void GlobalIndexToGlobalCoor(int64_t gidx,Coordinate &gcoor){
     assert(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
@@ -222,7 +222,7 @@ public:
     assert(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
-  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
+  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int64_t & gidx){
     gidx=0;
     int mult=1;
     for(int mu=0;mu<_ndimension;mu++) {

Grid/communicator/SharedMemoryMPI.cc

@@ -604,8 +604,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;

Grid/lattice/Lattice.h

@@ -47,3 +47,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
 #include <Grid/lattice/Lattice_crc.h>
+#include <Grid/lattice/PaddedCell.h>

Grid/lattice/Lattice_ET.h

@@ -345,7 +345,9 @@ GridUnopClass(UnaryNot, Not(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
+GridUnopClass(UnarySpTa, SpTa(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
+GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
@@ -456,7 +458,9 @@ GRID_DEF_UNOP(operator!, UnaryNot);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
+GRID_DEF_UNOP(SpTa, UnarySpTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
+GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup);
 GRID_DEF_UNOP(timesI, UnaryTimesI);
 GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
 GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the

Grid/lattice/Lattice_base.h

@@ -360,7 +360,7 @@ public:
 
 template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
   typedef typename vobj::scalar_object sobj;
-  for(int g=0;g<o.Grid()->_gsites;g++){
+  for(int64_t g=0;g<o.Grid()->_gsites;g++){
 
     Coordinate gcoor;
     o.Grid()->GlobalIndexToGlobalCoor(g,gcoor);

Grid/lattice/Lattice_rng.h

@@ -361,9 +361,14 @@ public:
     _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
   }
-
-  template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
-
+  template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist)
+  {
+    if ( l.Grid()->_isCheckerBoarded ) {
+      Lattice<vobj> tmp(_grid);
+      fill(tmp,dist);
+      pickCheckerboard(l.Checkerboard(),l,tmp);
+      return;
+    }
     typedef typename vobj::scalar_object scalar_object;
     typedef typename vobj::scalar_type scalar_type;
     typedef typename vobj::vector_type vector_type;
@@ -427,7 +432,7 @@ public:
 #if 1
     thread_for( lidx, _grid->lSites(), {
 
-	int gidx;
+	int64_t gidx;
 	int o_idx;
 	int i_idx;
 	int rank;

Grid/lattice/Lattice_trace.h

@@ -66,6 +66,65 @@ inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<
   return ret;
 };
 
+template<int N, class Vec>
+Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
+  typedef typename Vec::scalar_type scalar;
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<scalar, N> > > Us;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	scalar tmp= Us()()(i,j);
+	ComplexD ztmp(real(tmp),imag(tmp));
+	EigenU(i,j)=ztmp;
+      }}
+    ComplexD detD  = EigenU.determinant();
+    typename Vec::scalar_type det(detD.real(),detD.imag());
+    pokeLocalSite(det,ret_v,lcoor);
+  });
+  return ret;
+}
+
+template<int N>
+Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
+  
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us()()(i,j);
+      }}
+    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	Ui()()(i,j) = EigenUinv(i,j);
+      }}
+    pokeLocalSite(Ui,ret_v,lcoor);
+  });
+  return ret;
+}
+
+
 NAMESPACE_END(Grid);
 #endif
 

Grid/lattice/Lattice_transfer.h

@@ -471,13 +471,13 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 
   vobj zz = Zero();
   
-  accelerator_for(sc,coarse->oSites(),1,{
+  accelerator_for(sc,coarse->oSites(),vobj::Nsimd(),{
 
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
 
-      vobj cd = zz;
+      auto cd = coalescedRead(zz);
       
       for(int sb=0;sb<blockVol;sb++){
 
@@ -488,10 +488,10 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
 	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
 
-	cd=cd+fineData_p[sf];
+	cd=cd+coalescedRead(fineData_p[sf]);
       }
 
-      coarseData_p[sc] = cd;
+      coalescedWrite(coarseData_p[sc],cd);
 
     });
   return;
@@ -697,8 +697,68 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
-
   // the above should guarantee that the operations are local
+  
+#if 1
+
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
+  
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+ 
+  thread_for(idx, nsite, {
+      Coordinate from_coor, to_coor;
+      size_t rem = idx;
+      for(int i=0;i<nd;i++){
+	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i];
+	from_coor[i] = base_i + FromLowerLeft[i];
+	to_coor[i] = base_i + ToLowerLeft[i];
+      }
+      
+      int foidx = Fg->oIndex(from_coor);
+      int fiidx = Fg->iIndex(from_coor);
+      int toidx = Tg->oIndex(to_coor);
+      int tiidx = Tg->iIndex(to_coor);
+      int* tt = table + 4*idx;
+      tt[0] = foidx;
+      tt[1] = fiidx;
+      tt[2] = toidx;
+      tt[3] = tiidx;
+    });
+  
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+
+#else  
   Coordinate ldf = Fg->_ldimensions;
   Coordinate rdf = Fg->_rdimensions;
   Coordinate isf = Fg->_istride;
@@ -738,6 +798,8 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
 #endif
     }
   });
+
+#endif
 }
 
 
@@ -830,6 +892,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 }
 
 
+//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
+//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -851,6 +915,65 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
     }
   }
 
+#if 1
+  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+  
+  thread_for(idx,nsite,{
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
+    lcoor[orthog] = slice_lo;
+    hcoor[orthog] = slice_hi;
+    size_t rem = idx;
+    for(int mu=0;mu<nl;mu++){
+      if(mu != orthog){
+	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
+	lcoor[mu] = hcoor[mu] = xmu;
+      }
+    }
+    int loidx = lg->oIndex(lcoor);
+    int liidx = lg->iIndex(lcoor);
+    int hoidx = hg->oIndex(hcoor);
+    int hiidx = hg->iIndex(hcoor);
+    int* tt = table + 4*idx;
+    tt[0] = loidx;
+    tt[1] = liidx;
+    tt[2] = hoidx;
+    tt[3] = hiidx;
+    });
+   
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(lowDim_v,lowDim,AcceleratorRead);
+  autoView(higherDim_v,higherDim,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
+      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+#else
   // the above should guarantee that the operations are local
   autoView(lowDimv,lowDim,CpuRead);
   autoView(higherDimv,higherDim,CpuWrite);
@@ -866,6 +989,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       pokeLocalSite(s,higherDimv,hcoor);
     }
   });
+#endif
 }
 
 
@@ -930,7 +1054,7 @@ void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
   Coordinate fcoor(nd);
   Coordinate ccoor(nd);
-  for(int g=0;g<fg->gSites();g++){
+  for(int64_t g=0;g<fg->gSites();g++){
 
     fg->GlobalIndexToGlobalCoor(g,fcoor);
     for(int d=0;d<nd;d++){

Grid/lattice/PaddedCell.h

@@ -26,14 +26,32 @@ Author: Peter Boyle pboyle@bnl.gov
 /*  END LEGAL */
 #pragma once
 
+#include<Grid/cshift/Cshift.h>
+
 NAMESPACE_BEGIN(Grid);
 
+//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
+template<typename vobj>
+struct CshiftImplBase{
+  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
+  virtual ~CshiftImplBase(){}
+};
+template<typename vobj>
+struct CshiftImplDefault: public CshiftImplBase<vobj>{
+  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
+};
+template<typename Gimpl>
+struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
+  typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
+};  
+
 class PaddedCell {
 public:
   GridCartesian * unpadded_grid;
   int dims;
   int depth;
   std::vector<GridCartesian *> grids;
+
   ~PaddedCell()
   {
     DeleteGrids();
@@ -45,8 +63,9 @@ public:
     dims=_grid->Nd();
     AllocateGrids();
     Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate procs     =unpadded_grid->ProcessorGrid();
     for(int d=0;d<dims;d++){
-      assert(local[d]>=depth);
+      if ( procs[d] > 1 ) assert(local[d]>=depth);
     }
   }
   void DeleteGrids(void)
@@ -67,7 +86,9 @@ public:
     // expand up one dim at a time
     for(int d=0;d<dims;d++){
 
-      plocal[d] += 2*depth; 
+      if ( processors[d] > 1 ) { 
+	plocal[d] += 2*depth; 
+      }
       
       for(int d=0;d<dims;d++){
 	global[d] = plocal[d]*processors[d];
@@ -77,31 +98,38 @@ public:
     }
   };
   template<class vobj>
-  inline Lattice<vobj> Extract(Lattice<vobj> &in)
+  inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
   {
+    Coordinate processors=unpadded_grid->_processors;
+
     Lattice<vobj> out(unpadded_grid);
 
     Coordinate local     =unpadded_grid->LocalDimensions();
-    Coordinate fll(dims,depth); // depends on the MPI spread
+    // depends on the MPI spread      
+    Coordinate fll(dims,depth);
     Coordinate tll(dims,0); // depends on the MPI spread
+    for(int d=0;d<dims;d++){
+      if( processors[d]==1 ) fll[d]=0;
+    }
     localCopyRegion(in,out,fll,tll,local);
     return out;
   }
   template<class vobj>
-  inline Lattice<vobj> Exchange(Lattice<vobj> &in)
+  inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
   {
     GridBase *old_grid = in.Grid();
     int dims = old_grid->Nd();
     Lattice<vobj> tmp = in;
     for(int d=0;d<dims;d++){
-      tmp = Expand(d,tmp); // rvalue && assignment
+      tmp = Expand(d,tmp,cshift); // rvalue && assignment
     }
     return tmp;
   }
   // expand up one dim at a time
   template<class vobj>
-  inline Lattice<vobj> Expand(int dim,Lattice<vobj> &in)
+  inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
   {
+    Coordinate processors=unpadded_grid->_processors;
     GridBase *old_grid = in.Grid();
     GridCartesian *new_grid = grids[dim];//These are new grids
     Lattice<vobj>  padded(new_grid);
@@ -111,21 +139,53 @@ public:
     if(dim==0) conformable(old_grid,unpadded_grid);
     else       conformable(old_grid,grids[dim-1]);
 
-    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
-    // Middle bit
-    for(int x=0;x<local[dim];x++){
-      InsertSliceLocal(in,padded,x,depth+x,dim);
-    }
-    // High bit
-    shifted = Cshift(in,dim,depth);
-    for(int x=0;x<depth;x++){
-      InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
-    }
-    // Low bit
-    shifted = Cshift(in,dim,-depth);
-    for(int x=0;x<depth;x++){
-      InsertSliceLocal(shifted,padded,x,x,dim);
+    //    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
+
+    double tins=0, tshift=0;
+
+    int islocal = 0 ;
+    if ( processors[dim] == 1 ) islocal = 1;
+
+    if ( islocal ) {
+      
+      double t = usecond();
+      for(int x=0;x<local[dim];x++){
+	InsertSliceLocal(in,padded,x,x,dim);
+      }
+      tins += usecond() - t;
+      
+    } else { 
+      // Middle bit
+      double t = usecond();
+      for(int x=0;x<local[dim];x++){
+	InsertSliceLocal(in,padded,x,depth+x,dim);
+      }
+      tins += usecond() - t;
+    
+      // High bit
+      t = usecond();
+      shifted = cshift.Cshift(in,dim,depth);
+      tshift += usecond() - t;
+
+      t=usecond();
+      for(int x=0;x<depth;x++){
+	InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
+      }
+      tins += usecond() - t;
+    
+      // Low bit
+      t = usecond();
+      shifted = cshift.Cshift(in,dim,-depth);
+      tshift += usecond() - t;
+    
+      t = usecond();
+      for(int x=0;x<depth;x++){
+	InsertSliceLocal(shifted,padded,x,x,dim);
+      }
+      tins += usecond() - t;
     }
+    std::cout << GridLogDebug << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
+    
     return padded;
   }
 

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -124,11 +124,6 @@ public:
   RealD                _b;
   RealD                _c;
 
-  // possible boost
-  std::vector<ComplexD> qmu;
-  void set_qmu(std::vector<ComplexD> _qmu) { qmu=_qmu; assert(qmu.size()==Nd);};
-  void addQmu(const FermionField &in, FermionField &out, int dag);
-  
   // Cayley form Moebius (tanh and zolotarev)
   Vector<Coeff_t> omega;
   Vector<Coeff_t> bs;    // S dependent coeffs

Grid/qcd/action/fermion/ContinuedFractionFermion5D.h

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

Grid/qcd/action/fermion/Fermion.h

@@ -126,6 +126,16 @@ typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermi
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD;
 
+// Sp(2n)
+typedef WilsonFermion<SpWilsonImplF> SpWilsonFermionF;
+typedef WilsonFermion<SpWilsonImplD> SpWilsonFermionD;
+
+typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplF> SpWilsonTwoIndexAntiSymmetricFermionF;
+typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplD> SpWilsonTwoIndexAntiSymmetricFermionD;
+
+typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplF> SpWilsonTwoIndexSymmetricFermionF;
+typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplD> SpWilsonTwoIndexSymmetricFermionD;
+
 // Twisted mass fermion
 typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;

Grid/qcd/action/fermion/PartialFractionFermion5D.h

@@ -39,7 +39,7 @@ class PartialFractionFermion5D : public WilsonFermion5D<Impl>
 public:
   INHERIT_IMPL_TYPES(Impl);
 
-  const int part_frac_chroma_convention=0;
+  const int part_frac_chroma_convention=1;
 
   void   Meooe_internal(const FermionField &in, FermionField &out,int dag);
   void   Mooee_internal(const FermionField &in, FermionField &out,int dag);
@@ -83,63 +83,12 @@ public:
 			   GridRedBlackCartesian &FourDimRedBlackGrid,
 			   RealD _mass,RealD M5,const ImplParams &p= ImplParams());
 
-  PartialFractionFermion5D(GaugeField &_Umu,
-			   GridCartesian         &FiveDimGrid,
-			   GridRedBlackCartesian &FiveDimRedBlackGrid,
-			   GridCartesian         &FourDimGrid,
-			   GridRedBlackCartesian &FourDimRedBlackGrid,
-			   RealD _mass,RealD M5,std::vector<RealD> &_qmu,const ImplParams &p= ImplParams());
-
-  void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary, std::vector<double> twist)
-  {
-    std::cout << "Free Propagator for PartialFraction"<<std::endl;
-    FermionField in_k(in.Grid());
-    FermionField prop_k(in.Grid());
-    
-    FFT theFFT((GridCartesian *) in.Grid());
-
-    //phase for boundary condition
-    ComplexField coor(in.Grid());
-    ComplexField ph(in.Grid());  ph = Zero();
-    FermionField in_buf(in.Grid()); in_buf = Zero();
-    typedef typename Simd::scalar_type Scalar;
-    Scalar ci(0.0,1.0);
-    assert(twist.size() == Nd);//check that twist is Nd
-    assert(boundary.size() == Nd);//check that boundary conditions is Nd
-    int shift = 0;
-    for(unsigned int nu = 0; nu < Nd; nu++)
-      {
-	// Shift coordinate lattice index by 1 to account for 5th dimension.
-	LatticeCoordinate(coor, nu + shift);
-	double boundary_phase = ::acos(real(boundary[nu]));
-	ph = ph + boundary_phase*coor*((1./(in.Grid()->_fdimensions[nu+shift])));
-	//momenta for propagator shifted by twist+boundary
-	twist[nu] = twist[nu] + boundary_phase/((2.0*M_PI));
-      }
-    in_buf = exp(ci*ph*(-1.0))*in;
-
-    theFFT.FFT_all_dim(in_k,in,FFT::forward);
-    this->MomentumSpacePropagatorHw(prop_k,in_k,mass,twist);
-    theFFT.FFT_all_dim(out,prop_k,FFT::backward);
-    
-    //phase for boundary condition
-    out = out * exp(ci*ph);
-  };
-
-  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass) {
-    std::vector<double> twist(Nd,0.0); //default: periodic boundarys in all directions
-    std::vector<Complex> boundary;
-    for(int i=0;i<Nd;i++) boundary.push_back(1);//default: periodic boundary conditions
-    FreePropagator(in,out,mass,boundary,twist);
-  };
-  
 protected:
 
   virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD scale);
   virtual void SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata);
 
   // Part frac
-  std::vector<RealD> qmu;
   RealD mass;
   RealD dw_diag;
   RealD R;

Grid/qcd/action/fermion/WilsonImpl.h

@@ -261,6 +261,22 @@ typedef WilsonImpl<vComplex,  TwoIndexAntiSymmetricRepresentation, CoeffReal > W
 typedef WilsonImpl<vComplexF, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplF;  // Float
 typedef WilsonImpl<vComplexD, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplD;  // Double
 
+//sp 2n
+
+typedef WilsonImpl<vComplex,  SpFundamentalRepresentation, CoeffReal > SpWilsonImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpFundamentalRepresentation, CoeffReal > SpWilsonImplF;  // Float
+typedef WilsonImpl<vComplexD, SpFundamentalRepresentation, CoeffReal > SpWilsonImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplF;  // Float
+typedef WilsonImpl<vComplexD, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplF;  // Float
+typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplR;  // Real.. whichever prec    // adj = 2indx symmetric for Sp(2N)
+typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplF;  // Float     // adj = 2indx symmetric for Sp(2N)
+typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplD;  // Double    // adj = 2indx symmetric for Sp(2N)
 
 NAMESPACE_END(Grid);
-

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

@@ -49,7 +49,6 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FourDimRedBlackGrid,_M5,p),
   mass_plus(_mass), mass_minus(_mass)
 { 
-  // qmu defaults to zero size;
 }
 
 ///////////////////////////////////////////////////////////////
@@ -271,34 +270,6 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   M5Ddag(psi,psi,Din,lower,diag,upper);
 }
 
-template<class Impl>
-void CayleyFermion5D<Impl>::addQmu(const FermionField &psi,FermionField &chi, int dag)
-{
-  if ( qmu.size() ) {
-
-    Gamma::Algebra Gmu [] = {
-      Gamma::Algebra::GammaX,
-      Gamma::Algebra::GammaY,
-      Gamma::Algebra::GammaZ,
-      Gamma::Algebra::GammaT
-    };
-    std::vector<ComplexD> coeff(Nd);
-    ComplexD ci(0,1);
-
-    assert(qmu.size()==Nd);
-
-    for(int mu=0;mu<Nd;mu++){
-       coeff[mu] = ci*qmu[mu];
-       if ( dag ) coeff[mu] = conjugate(coeff[mu]);
-    }
-
-    chi = chi + Gamma(Gmu[0])*psi*coeff[0];
-    for(int mu=1;mu<Nd;mu++){
-      chi = chi + Gamma(Gmu[mu])*psi*coeff[mu];
-    }
-  }
-}
-
 template<class Impl>
 void CayleyFermion5D<Impl>::M    (const FermionField &psi, FermionField &chi)
 {
@@ -308,10 +279,6 @@ void CayleyFermion5D<Impl>::M    (const FermionField &psi, FermionField &chi)
   Meooe5D(psi,Din);
   
   this->DW(Din,chi,DaggerNo);
-
-  // add i q_mu gamma_mu here
-  addQmu(Din,chi,DaggerNo);
-  
   // ((b D_W + D_w hop terms +1) on s-diag
   axpby(chi,1.0,1.0,chi,psi); 
   
@@ -329,9 +296,6 @@ void CayleyFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
   // Apply Dw
   this->DW(psi,Din,DaggerYes); 
   
-  // add -i conj(q_mu) gamma_mu here ... if qmu is real, gammm_5 hermitian, otherwise not.
-  addQmu(psi,Din,DaggerYes);
-  
   MeooeDag5D(Din,chi);
   
   M5Ddag(psi,chi);

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

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

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

@@ -255,45 +255,6 @@ void   PartialFractionFermion5D<Impl>::M_internal(const FermionField &psi, Fermi
   }
 	
   {
-    // The 'conventional' Cayley overlap operator is
-    //
-    // Dov = (1+m)/2 + (1-m)/2 g5 sgn Hw
-    //
-    //
-    // With massless limit 1/2(1+g5 sgnHw)
-    //
-    // Luscher shows quite neatly that 1+g5 sgn Hw has tree level propagator i qslash +O(a^2)
-    //
-    // However, the conventional normalisation has both a leading order factor of 2 in Zq
-    // at tree level AND a mass dependent (1-m) that are convenient to absorb.
-    //
-    // In WilsonFermion5DImplementation.h, the tree level propagator for Hw is
-    //
-    // num = -i sin kmu gmu
-    //
-    // denom ( sqrt(sk^2 + (2shk^2 - 1)^2
-    //    b_k = sk2 - M5;
-    //     
-    //    w_k = sqrt(sk + b_k*b_k);
-    //
-    //    denom= ( w_k + b_k + mass*mass) ;
-    //
-    //    denom= one/denom;
-    //    out = num*denom;
-    //
-    // Chroma, and Grid define partial fraction via 4d operator
-    //
-    //   Dpf = 2/(1-m) x Dov = (1+m)/(1-m) + g5 sgn Hw
-    //
-    // Now since:
-    //
-    //      (1+m)/(1-m) = (1-m)/(1-m) + 2m/(1-m) = 1 + 2m/(1-m)
-    //
-    // This corresponds to a modified mass parameter
-    //
-    // It has an annoying 
-    //
-    // 
     double R=(1+this->mass)/(1-this->mass);
     //R g5 psi[Ls] + p[0] H
     ag5xpbg5y_ssp(chi,R*scale,psi,p[nblock]*scale/amax,D,Ls-1,Ls-1);
@@ -303,28 +264,6 @@ void   PartialFractionFermion5D<Impl>::M_internal(const FermionField &psi, Fermi
       double pp = p[nblock-1-b];
       axpby_ssp(chi,1.0,chi,-sqrt(amax*pp)*scale*sign,psi,Ls-1,s);
     }
-
-    if ( qmu.size() ) {
-
-      FermionField qslash_psi(psi.Grid());
-      
-      Gamma::Algebra Gmu [] = {
-			 Gamma::Algebra::GammaX,
-			 Gamma::Algebra::GammaY,
-			 Gamma::Algebra::GammaZ,
-			 Gamma::Algebra::GammaT
-      };
-      ComplexD ci(0,1);
-      assert(qmu.size()==Nd);
-      qslash_psi = Gamma(Gmu[0])*psi;
-      for(int mu=1;mu<Nd;mu++){
-	qslash_psi = Gamma(Gmu[mu])*psi;
-      }
-      //      RealD coeff = 1.0;
-      qslash_psi = Gamma(Gamma::Algebra::Gamma5)*qslash_psi*ci ; // i g5 qslash -- 1-m factor???
-      axpby_ssp(chi,1.0,chi,1.0, qslash_psi,Ls-1,Ls-1);
-    }
-    
   }
 
 }
@@ -472,7 +411,7 @@ void  PartialFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,App
       int Ls = this->Ls;
       conformable(solution5d.Grid(),this->FermionGrid());
       conformable(exported4d.Grid(),this->GaugeGrid());
-      ExtractSlice(exported4d, solution5d, Ls-1, 0);
+      ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
     }
     template<class Impl>
     void PartialFractionFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
@@ -482,8 +421,7 @@ void  PartialFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,App
       conformable(input4d.Grid()   ,this->GaugeGrid());
       FermionField tmp(this->FermionGrid());
       tmp=Zero();
-      std::cout << " importing to slice " << Ls-1 <<std::endl;
-      InsertSlice(input4d, tmp, Ls-1, 0);
+      InsertSlice(input4d, tmp, Ls-1, Ls-1);
       tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
       this->Dminus(tmp,imported5d);
     }
@@ -504,7 +442,7 @@ PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
 
 {
   int Ls = this->Ls;
-  qmu.resize(0);
+
   assert((Ls&0x1)==1); // Odd Ls required
   int nrational=Ls-1;
 
@@ -522,22 +460,6 @@ PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
   Approx::zolotarev_free(zdata);
 
 }
-template<class Impl>
-PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
-							 GridCartesian         &FiveDimGrid,
-							 GridRedBlackCartesian &FiveDimRedBlackGrid,
-							 GridCartesian         &FourDimGrid,
-							 GridRedBlackCartesian &FourDimRedBlackGrid,
-							 RealD _mass,RealD M5,
-							 std::vector<RealD> &_qmu,
-							 const ImplParams &p)
-  : PartialFractionFermion5D<Impl>(_Umu,
-			     FiveDimGrid,FiveDimRedBlackGrid,
-			     FourDimGrid,FourDimRedBlackGrid,
-			     _mass,M5,p)
-{
-  qmu=_qmu;
-}
 
 NAMESPACE_END(Grid);
 

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonCloverFermionInstantiationSpWilsonImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonFermionInstantiationSpWilsonImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonKernelsInstantiationSpWilsonImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonTMFermionInstantiationSpWilsonImplD.cc

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

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

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonCloverFermionInstantiationSpWilsonImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonFermionInstantiationSpWilsonImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonKernelsInstantiationSpWilsonImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonTMFermionInstantiationSpWilsonImplF.cc

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

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

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonCloverFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonTMFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

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

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

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonCloverFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonTMFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

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

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

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonCloverFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationSpWilsonTwoIndexSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonTMFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc

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

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

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonCloverFermionInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonFermionInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonTMFermionInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

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

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

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

@@ -10,12 +10,18 @@ WILSON_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
 	   WilsonImplD2 \
+	   SpWilsonImplF \
+	   SpWilsonImplD \
 	   WilsonAdjImplF \
 	   WilsonAdjImplD \
 	   WilsonTwoIndexSymmetricImplF \
 	   WilsonTwoIndexSymmetricImplD \
 	   WilsonTwoIndexAntiSymmetricImplF \
 	   WilsonTwoIndexAntiSymmetricImplD \
+	   SpWilsonTwoIndexAntiSymmetricImplF \
+	   SpWilsonTwoIndexAntiSymmetricImplD \
+	   SpWilsonTwoIndexSymmetricImplF \
+	   SpWilsonTwoIndexSymmetricImplD \
 	   GparityWilsonImplF \
 	   GparityWilsonImplD "
 

Grid/qcd/action/gauge/Gauge.h

@@ -39,6 +39,9 @@ NAMESPACE_BEGIN(Grid);
 typedef WilsonGaugeAction<PeriodicGimplR>          WilsonGaugeActionR;
 typedef WilsonGaugeAction<PeriodicGimplF>          WilsonGaugeActionF;
 typedef WilsonGaugeAction<PeriodicGimplD>          WilsonGaugeActionD;
+typedef WilsonGaugeAction<SpPeriodicGimplR>        SpWilsonGaugeActionR;
+typedef WilsonGaugeAction<SpPeriodicGimplF>        SpWilsonGaugeActionF;
+typedef WilsonGaugeAction<SpPeriodicGimplD>        SpWilsonGaugeActionD;
 typedef PlaqPlusRectangleAction<PeriodicGimplR>    PlaqPlusRectangleActionR;
 typedef PlaqPlusRectangleAction<PeriodicGimplF>    PlaqPlusRectangleActionF;
 typedef PlaqPlusRectangleAction<PeriodicGimplD>    PlaqPlusRectangleActionD;

Grid/qcd/action/gauge/GaugeImplTypes.h

@@ -61,7 +61,7 @@ NAMESPACE_BEGIN(Grid);
   typedef typename Impl::Field Field;
 
 // hardcodes the exponential approximation in the template
-template <class S, int Nrepresentation = Nc, int Nexp = 12 > class GaugeImplTypes {
+template <class S, int Nrepresentation = Nc, int Nexp = 12, class Group = SU<Nc> > class GaugeImplTypes {
 public:
   typedef S Simd;
   typedef typename Simd::scalar_type scalar_type;
@@ -78,8 +78,6 @@ public:
   typedef Lattice<SiteLink>    LinkField; 
   typedef Lattice<SiteField>   Field;
 
-  typedef SU<Nrepresentation> Group;
-
   // Guido: we can probably separate the types from the HMC functions
   // this will create 2 kind of implementations
   // probably confusing the users
@@ -119,6 +117,7 @@ public:
     //
     LinkField Pmu(P.Grid());
     Pmu = Zero();
+
     for (int mu = 0; mu < Nd; mu++) {
       Group::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
       RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR) ;
@@ -127,7 +126,11 @@ public:
     }
   }
     
-  static inline Field projectForce(Field &P) { return Ta(P); }
+  static inline Field projectForce(Field &P) {
+      Field ret(P.Grid());
+      Group::taProj(P, ret);
+      return ret;
+    }
 
   static inline void update_field(Field& P, Field& U, double ep){
     //static std::chrono::duration<double> diff;
@@ -137,7 +140,8 @@ public:
     autoView(P_v,P,AcceleratorRead);
     accelerator_for(ss, P.Grid()->oSites(),1,{
       for (int mu = 0; mu < Nd; mu++) {
-        U_v[ss](mu) = ProjectOnGroup(Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu));
+          U_v[ss](mu) = Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu);
+          U_v[ss](mu) = Group::ProjectOnGeneralGroup(U_v[ss](mu));
       }
     });
    //auto end = std::chrono::high_resolution_clock::now();
@@ -157,7 +161,7 @@ public:
   }
 
   static inline void Project(Field &U) {
-    ProjectSUn(U);
+    Group::ProjectOnSpecialGroup(U);
   }
 
   static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
@@ -171,6 +175,7 @@ public:
   static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
     Group::ColdConfiguration(pRNG, U);
   }
+
 };
 
 
@@ -178,10 +183,17 @@ typedef GaugeImplTypes<vComplex, Nc> GimplTypesR;
 typedef GaugeImplTypes<vComplexF, Nc> GimplTypesF;
 typedef GaugeImplTypes<vComplexD, Nc> GimplTypesD;
 
+typedef GaugeImplTypes<vComplex, Nc, 12, Sp<Nc> > SpGimplTypesR;
+typedef GaugeImplTypes<vComplexF, Nc, 12, Sp<Nc> > SpGimplTypesF;
+typedef GaugeImplTypes<vComplexD, Nc, 12, Sp<Nc> > SpGimplTypesD;
+
 typedef GaugeImplTypes<vComplex, SU<Nc>::AdjointDimension> GimplAdjointTypesR;
 typedef GaugeImplTypes<vComplexF, SU<Nc>::AdjointDimension> GimplAdjointTypesF;
 typedef GaugeImplTypes<vComplexD, SU<Nc>::AdjointDimension> GimplAdjointTypesD;
 
+
+
+
 NAMESPACE_END(Grid);
 
 #endif // GRID_GAUGE_IMPL_TYPES_H

Grid/qcd/action/gauge/GaugeImplementations.h

@@ -176,7 +176,7 @@ public:
       return PeriodicBC::CshiftLink(Link,mu,shift);
   }
 
-  static inline void       setDirections(std::vector<int> &conjDirs) { _conjDirs=conjDirs; }
+  static inline void       setDirections(const std::vector<int> &conjDirs) { _conjDirs=conjDirs; }
   static inline std::vector<int> getDirections(void) { return _conjDirs; }
   static inline bool isPeriodicGaugeField(void) { return false; }
 };
@@ -193,6 +193,11 @@ typedef ConjugateGaugeImpl<GimplTypesR> ConjugateGimplR; // Real.. whichever pre
 typedef ConjugateGaugeImpl<GimplTypesF> ConjugateGimplF; // Float
 typedef ConjugateGaugeImpl<GimplTypesD> ConjugateGimplD; // Double
 
+typedef PeriodicGaugeImpl<SpGimplTypesR> SpPeriodicGimplR; // Real.. whichever prec
+typedef PeriodicGaugeImpl<SpGimplTypesF> SpPeriodicGimplF; // Float
+typedef PeriodicGaugeImpl<SpGimplTypesD> SpPeriodicGimplD; // Double
+
+
 NAMESPACE_END(Grid);
 
 #endif

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

@@ -43,7 +43,7 @@ public:
 private:
   RealD c_plaq;
   RealD c_rect;
-
+  typename WilsonLoops<Gimpl>::StapleAndRectStapleAllWorkspace workspace;
 public:
   PlaqPlusRectangleAction(RealD b,RealD c): c_plaq(b),c_rect(c){};
 
@@ -79,27 +79,18 @@ public:
     GridBase *grid = Umu.Grid();
 
     std::vector<GaugeLinkField> U (Nd,grid);
-    std::vector<GaugeLinkField> U2(Nd,grid);
-
     for(int mu=0;mu<Nd;mu++){
       U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
-      WilsonLoops<Gimpl>::RectStapleDouble(U2[mu],U[mu],mu);
     }
+    std::vector<GaugeLinkField> RectStaple(Nd,grid), Staple(Nd,grid);
+    WilsonLoops<Gimpl>::StapleAndRectStapleAll(Staple, RectStaple, U, workspace);
 
     GaugeLinkField dSdU_mu(grid);
     GaugeLinkField staple(grid);
 
     for (int mu=0; mu < Nd; mu++){
-
-      // Staple in direction mu
-
-      WilsonLoops<Gimpl>::Staple(staple,Umu,mu);
-
-      dSdU_mu = Ta(U[mu]*staple)*factor_p;
-
-      WilsonLoops<Gimpl>::RectStaple(Umu,staple,U2,U,mu);
-
-      dSdU_mu = dSdU_mu + Ta(U[mu]*staple)*factor_r;
+      dSdU_mu = Ta(U[mu]*Staple[mu])*factor_p;
+      dSdU_mu = dSdU_mu + Ta(U[mu]*RectStaple[mu])*factor_r;
 	  
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);
     }

Grid/qcd/hmc/GenericHMCrunner.h

@@ -225,6 +225,18 @@ template <class RepresentationsPolicy,
 using GenericHMCRunnerHirep =
 				     HMCWrapperTemplate<PeriodicGimplR, Integrator, RepresentationsPolicy>;
 
+// sp2n
+
+template <template <typename, typename, typename> class Integrator>
+using GenericSpHMCRunner = HMCWrapperTemplate<SpPeriodicGimplR, Integrator>;
+
+template <class RepresentationsPolicy,
+          template <typename, typename, typename> class Integrator>
+using GenericSpHMCRunnerHirep =
+                     HMCWrapperTemplate<SpPeriodicGimplR, Integrator, RepresentationsPolicy>;
+
+
+
 template <class Implementation, class RepresentationsPolicy, 
           template <typename, typename, typename> class Integrator>
 using GenericHMCRunnerTemplate = HMCWrapperTemplate<Implementation, Integrator, RepresentationsPolicy>;

Grid/qcd/representations/fundamental.h

@@ -13,7 +13,7 @@ NAMESPACE_BEGIN(Grid);
  * Empty since HMC updates already the fundamental representation 
  */
 
-template <int ncolour>
+template <int ncolour, class group_name>
 class FundamentalRep {
 public:
   static const int Dimension = ncolour;
@@ -21,7 +21,7 @@ public:
 
   // typdef to be used by the Representations class in HMC to get the
   // types for the higher representation fields
-  typedef typename SU<ncolour>::LatticeMatrix LatticeMatrix;
+  typedef typename GaugeGroup<ncolour,group_name>::LatticeMatrix LatticeMatrix;
   typedef LatticeGaugeField LatticeField;
   
   explicit FundamentalRep(GridBase* grid) {} //do nothing
@@ -45,7 +45,8 @@ public:
     
 
   
-typedef	 FundamentalRep<Nc> FundamentalRepresentation;
+typedef	 FundamentalRep<Nc,GroupName::SU> FundamentalRepresentation;
+typedef	 FundamentalRep<Nc,GroupName::Sp> SpFundamentalRepresentation;
 
 NAMESPACE_END(Grid);  
 

Grid/qcd/representations/two_index.h

@@ -20,14 +20,14 @@ NAMESPACE_BEGIN(Grid);
  * in the SUnTwoIndex.h file
  */
 
-template <int ncolour, TwoIndexSymmetry S>
+template <int ncolour, TwoIndexSymmetry S, class group_name = GroupName::SU>
 class TwoIndexRep {
 public:
   // typdef to be used by the Representations class in HMC to get the
   // types for the higher representation fields
-  typedef typename SU_TwoIndex<ncolour, S>::LatticeTwoIndexMatrix LatticeMatrix;
-  typedef typename SU_TwoIndex<ncolour, S>::LatticeTwoIndexField LatticeField;
-  static const int Dimension = ncolour * (ncolour + S) / 2;
+  typedef typename GaugeGroupTwoIndex<ncolour, S, group_name>::LatticeTwoIndexMatrix LatticeMatrix;
+  typedef typename GaugeGroupTwoIndex<ncolour, S, group_name>::LatticeTwoIndexField LatticeField;
+  static const int Dimension = GaugeGroupTwoIndex<ncolour,S,group_name>::Dimension;
   static const bool isFundamental = false;
 
   LatticeField U;
@@ -43,10 +43,10 @@ public:
     U = Zero();
     LatticeColourMatrix tmp(Uin.Grid());
 
-    Vector<typename SU<ncolour>::Matrix> eij(Dimension);
+    Vector<typename GaugeGroup<ncolour,group_name>::Matrix> eij(Dimension);
 
     for (int a = 0; a < Dimension; a++)
-      SU_TwoIndex<ncolour, S>::base(a, eij[a]);
+      GaugeGroupTwoIndex<ncolour, S, group_name>::base(a, eij[a]);
 
     for (int mu = 0; mu < Nd; mu++) {
       auto Uin_mu = peekLorentz(Uin, mu);
@@ -71,7 +71,7 @@ public:
 
       out_mu = Zero();
 
-      typename SU<ncolour>::LatticeAlgebraVector h(in.Grid());
+      typename GaugeGroup<ncolour, group_name>::LatticeAlgebraVector h(in.Grid());
       projectOnAlgebra(h, in_mu, double(Nc + 2 * S));  // factor T(r)/T(fund)
       FundamentalLieAlgebraMatrix(h, out_mu);          // apply scale only once
       pokeLorentz(out, out_mu, mu);
@@ -80,20 +80,23 @@ public:
   }
 
 private:
-  void projectOnAlgebra(typename SU<ncolour>::LatticeAlgebraVector &h_out,
+  void projectOnAlgebra(typename GaugeGroup<ncolour, group_name>::LatticeAlgebraVector &h_out,
                         const LatticeMatrix &in, Real scale = 1.0) const {
-    SU_TwoIndex<ncolour, S>::projectOnAlgebra(h_out, in, scale);
+    GaugeGroupTwoIndex<ncolour, S,group_name>::projectOnAlgebra(h_out, in, scale);
   }
 
   void FundamentalLieAlgebraMatrix(
-				   typename SU<ncolour>::LatticeAlgebraVector &h,
-				   typename SU<ncolour>::LatticeMatrix &out, Real scale = 1.0) const {
-    SU<ncolour>::FundamentalLieAlgebraMatrix(h, out, scale);
+				   typename GaugeGroup<ncolour, group_name>::LatticeAlgebraVector &h,
+				   typename GaugeGroup<ncolour, group_name>::LatticeMatrix &out, Real scale = 1.0) const {
+    GaugeGroup<ncolour,group_name>::FundamentalLieAlgebraMatrix(h, out, scale);
   }
 };
 
-typedef TwoIndexRep<Nc, Symmetric> TwoIndexSymmetricRepresentation;
-typedef TwoIndexRep<Nc, AntiSymmetric> TwoIndexAntiSymmetricRepresentation;
+typedef TwoIndexRep<Nc, Symmetric, GroupName::SU> TwoIndexSymmetricRepresentation;
+typedef TwoIndexRep<Nc, AntiSymmetric, GroupName::SU> TwoIndexAntiSymmetricRepresentation;
+
+typedef TwoIndexRep<Nc, Symmetric, GroupName::Sp> SpTwoIndexSymmetricRepresentation;
+typedef TwoIndexRep<Nc, AntiSymmetric, GroupName::Sp> SpTwoIndexAntiSymmetricRepresentation;
 
 NAMESPACE_END(Grid);
 

Grid/qcd/utils/CovariantCshift.h

@@ -37,13 +37,14 @@ NAMESPACE_BEGIN(Grid);
 // Make these members of an Impl class for BC's.
 
 namespace PeriodicBC { 
-
+  //Out(x) = Link(x)*field(x+mu)
   template<class covariant,class gauge> Lattice<covariant> CovShiftForward(const Lattice<gauge> &Link, 
 									   int mu,
 									   const Lattice<covariant> &field)
   {
     return Link*Cshift(field,mu,1);// moves towards negative mu
   }
+  //Out(x) = Link^dag(x-mu)*field(x-mu)
   template<class covariant,class gauge> Lattice<covariant> CovShiftBackward(const Lattice<gauge> &Link, 
 									    int mu,
 									    const Lattice<covariant> &field)
@@ -52,19 +53,19 @@ namespace PeriodicBC {
     tmp = adj(Link)*field;
     return Cshift(tmp,mu,-1);// moves towards positive mu
   }
-
+  //Out(x) = Link^dag(x-mu)
   template<class gauge> Lattice<gauge>
   CovShiftIdentityBackward(const Lattice<gauge> &Link, int mu) 
   {
     return Cshift(adj(Link), mu, -1);
   }
-
+  //Out(x) = Link(x)
   template<class gauge> Lattice<gauge>
   CovShiftIdentityForward(const Lattice<gauge> &Link, int mu)
   {
     return Link;
   }
-
+  //Link(x) = Link(x+mu)
   template<class gauge> Lattice<gauge>
   ShiftStaple(const Lattice<gauge> &Link, int mu)
   {

Grid/qcd/utils/GaugeGroup.h

@@ -0,0 +1,470 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/utils/GaugeGroup.h
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: neo <cossu@post.kek.jp>
+Author: paboyle <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 */
+#ifndef QCD_UTIL_GAUGEGROUP_H
+#define QCD_UTIL_GAUGEGROUP_H
+
+// Important detail: nvcc requires all template parameters to have names.
+// This is the only reason why the second template parameter has a name.
+#define ONLY_IF_SU                                                       \
+  typename dummy_name = group_name,                                      \
+           typename named_dummy = std::enable_if_t <                                 \
+                          std::is_same<dummy_name, group_name>::value && \
+                      is_su<dummy_name>::value >
+
+#define ONLY_IF_Sp                                                       \
+  typename dummy_name = group_name,                                      \
+           typename named_dummy = std::enable_if_t <                                 \
+                          std::is_same<dummy_name, group_name>::value && \
+                      is_sp<dummy_name>::value >
+
+NAMESPACE_BEGIN(Grid);
+namespace GroupName {
+class SU {};
+class Sp {};
+}  // namespace GroupName
+
+template <typename group_name>
+struct is_su {
+  static const bool value = false;
+};
+
+template <>
+struct is_su<GroupName::SU> {
+  static const bool value = true;
+};
+
+template <typename group_name>
+struct is_sp {
+  static const bool value = false;
+};
+
+template <>
+struct is_sp<GroupName::Sp> {
+  static const bool value = true;
+};
+
+template <typename group_name>
+constexpr int compute_adjoint_dimension(int ncolour);
+
+template <>
+constexpr int compute_adjoint_dimension<GroupName::SU>(int ncolour) {
+  return ncolour * ncolour - 1;
+}
+
+template <>
+constexpr int compute_adjoint_dimension<GroupName::Sp>(int ncolour) {
+  return ncolour / 2 * (ncolour + 1);
+}
+
+template <int ncolour, class group_name>
+class GaugeGroup {
+ public:
+  static const int Dimension = ncolour;
+  static const int AdjointDimension =
+      compute_adjoint_dimension<group_name>(ncolour);
+  static const int AlgebraDimension =
+      compute_adjoint_dimension<group_name>(ncolour);
+
+  template <typename vtype>
+  using iSU2Matrix = iScalar<iScalar<iMatrix<vtype, 2> > >;
+  template <typename vtype>
+  using iGroupMatrix = iScalar<iScalar<iMatrix<vtype, ncolour> > >;
+  template <typename vtype>
+  using iAlgebraVector = iScalar<iScalar<iVector<vtype, AdjointDimension> > >;
+  static int su2subgroups(void) { return su2subgroups(group_name()); }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////
+  // Types can be accessed as SU<2>::Matrix , SU<2>::vSUnMatrix,
+  // SU<2>::LatticeMatrix etc...
+  //////////////////////////////////////////////////////////////////////////////////////////////////
+  typedef iGroupMatrix<Complex> Matrix;
+  typedef iGroupMatrix<ComplexF> MatrixF;
+  typedef iGroupMatrix<ComplexD> MatrixD;
+
+  typedef iGroupMatrix<vComplex> vMatrix;
+  typedef iGroupMatrix<vComplexF> vMatrixF;
+  typedef iGroupMatrix<vComplexD> vMatrixD;
+
+  // For the projectors to the algebra
+  // these should be real...
+  // keeping complex for consistency with the SIMD vector types
+  typedef iAlgebraVector<Complex> AlgebraVector;
+  typedef iAlgebraVector<ComplexF> AlgebraVectorF;
+  typedef iAlgebraVector<ComplexD> AlgebraVectorD;
+
+  typedef iAlgebraVector<vComplex> vAlgebraVector;
+  typedef iAlgebraVector<vComplexF> vAlgebraVectorF;
+  typedef iAlgebraVector<vComplexD> vAlgebraVectorD;
+
+  typedef Lattice<vMatrix> LatticeMatrix;
+  typedef Lattice<vMatrixF> LatticeMatrixF;
+  typedef Lattice<vMatrixD> LatticeMatrixD;
+
+  typedef Lattice<vAlgebraVector> LatticeAlgebraVector;
+  typedef Lattice<vAlgebraVectorF> LatticeAlgebraVectorF;
+  typedef Lattice<vAlgebraVectorD> LatticeAlgebraVectorD;
+
+  typedef iSU2Matrix<Complex> SU2Matrix;
+  typedef iSU2Matrix<ComplexF> SU2MatrixF;
+  typedef iSU2Matrix<ComplexD> SU2MatrixD;
+
+  typedef iSU2Matrix<vComplex> vSU2Matrix;
+  typedef iSU2Matrix<vComplexF> vSU2MatrixF;
+  typedef iSU2Matrix<vComplexD> vSU2MatrixD;
+
+  typedef Lattice<vSU2Matrix> LatticeSU2Matrix;
+  typedef Lattice<vSU2MatrixF> LatticeSU2MatrixF;
+  typedef Lattice<vSU2MatrixD> LatticeSU2MatrixD;
+
+  // Private implementation details are specified in the following files:
+  // Grid/qcd/utils/SUn.impl
+  // Grid/qcd/utils/SUn.impl
+  // The public part of the interface follows below and refers to these
+  // private member functions.
+
+#include <Grid/qcd/utils/SUn.impl.h>
+#include <Grid/qcd/utils/Sp2n.impl.h>
+
+ public:
+  template <class cplx>
+  static void generator(int lieIndex, iGroupMatrix<cplx> &ta) {
+    return generator(lieIndex, ta, group_name());
+  }
+
+  static void su2SubGroupIndex(int &i1, int &i2, int su2_index) {
+    return su2SubGroupIndex(i1, i2, su2_index, group_name());
+  }
+
+  static void testGenerators(void) { testGenerators(group_name()); }
+
+  static void printGenerators(void) {
+    for (int gen = 0; gen < AlgebraDimension; gen++) {
+      Matrix ta;
+      generator(gen, ta);
+      std::cout << GridLogMessage << "Nc = " << ncolour << " t_" << gen
+                << std::endl;
+      std::cout << GridLogMessage << ta << std::endl;
+    }
+  }
+
+  template <typename LatticeMatrixType>
+  static void LieRandomize(GridParallelRNG &pRNG, LatticeMatrixType &out,
+                           double scale = 1.0) {
+    GridBase *grid = out.Grid();
+
+    typedef typename LatticeMatrixType::vector_type vector_type;
+
+    typedef iSinglet<vector_type> vTComplexType;
+
+    typedef Lattice<vTComplexType> LatticeComplexType;
+    typedef typename GridTypeMapper<
+        typename LatticeMatrixType::vector_object>::scalar_object MatrixType;
+
+    LatticeComplexType ca(grid);
+    LatticeMatrixType lie(grid);
+    LatticeMatrixType la(grid);
+    ComplexD ci(0.0, scale);
+    MatrixType ta;
+
+    lie = Zero();
+
+    for (int a = 0; a < AlgebraDimension; a++) {
+      random(pRNG, ca);
+
+      ca = (ca + conjugate(ca)) * 0.5;
+      ca = ca - 0.5;
+
+      generator(a, ta);
+
+      la = ci * ca * ta;
+
+      lie = lie + la;  // e^{i la ta}
+    }
+    taExp(lie, out);
+  }
+
+  static void GaussianFundamentalLieAlgebraMatrix(GridParallelRNG &pRNG,
+                                                  LatticeMatrix &out,
+                                                  Real scale = 1.0) {
+    GridBase *grid = out.Grid();
+    LatticeReal ca(grid);
+    LatticeMatrix la(grid);
+    Complex ci(0.0, scale);
+    Matrix ta;
+
+    out = Zero();
+    for (int a = 0; a < AlgebraDimension; a++) {
+      gaussian(pRNG, ca);
+      generator(a, ta);
+      la = toComplex(ca) * ta;
+      out += la;
+    }
+    out *= ci;
+  }
+
+  static void FundamentalLieAlgebraMatrix(const LatticeAlgebraVector &h,
+                                          LatticeMatrix &out,
+                                          Real scale = 1.0) {
+    conformable(h, out);
+    GridBase *grid = out.Grid();
+    LatticeMatrix la(grid);
+    Matrix ta;
+
+    out = Zero();
+    for (int a = 0; a < AlgebraDimension; a++) {
+      generator(a, ta);
+      la = peekColour(h, a) * timesI(ta) * scale;
+      out += la;
+    }
+  }
+
+  // Projects the algebra components a lattice matrix (of dimension ncol*ncol -1
+  // ) inverse operation: FundamentalLieAlgebraMatrix
+  static void projectOnAlgebra(LatticeAlgebraVector &h_out,
+                               const LatticeMatrix &in, Real scale = 1.0) {
+    conformable(h_out, in);
+    h_out = Zero();
+    Matrix Ta;
+
+    for (int a = 0; a < AlgebraDimension; a++) {
+      generator(a, Ta);
+      pokeColour(h_out, -2.0 * (trace(timesI(Ta) * in)) * scale, a);
+    }
+  }
+
+   
+  template <class vtype>
+  accelerator_inline static iScalar<vtype> ProjectOnGeneralGroup(const iScalar<vtype> &r) {
+    return ProjectOnGeneralGroup(r, group_name());
+  }
+
+  template <class vtype, int N>
+  accelerator_inline static iVector<vtype,N> ProjectOnGeneralGroup(const iVector<vtype,N> &r) {
+    return ProjectOnGeneralGroup(r, group_name());
+  }
+
+  template <class vtype,int N, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0 >::type * =nullptr>
+  accelerator_inline static iMatrix<vtype,N> ProjectOnGeneralGroup(const iMatrix<vtype,N> &arg) {
+    return ProjectOnGeneralGroup(arg, group_name());
+  }
+
+  template <int N,class vComplex_t>                  // Projects on the general groups U(N), Sp(2N)xZ2 i.e. determinant is allowed a complex phase.
+  static void ProjectOnGeneralGroup(Lattice<iVector<iScalar<iMatrix<vComplex_t, N> >, Nd> > &U) {
+    for (int mu = 0; mu < Nd; mu++) {
+      auto Umu = PeekIndex<LorentzIndex>(U, mu);
+      Umu = ProjectOnGeneralGroup(Umu);
+    }
+  }
+       
+
+  
+  template <int N,class vComplex_t>
+  static Lattice<iScalar<iScalar<iMatrix<vComplex_t, N> > > > ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vComplex_t, N> > > > &Umu) {
+    return ProjectOnGeneralGroup(Umu, group_name());
+  }
+
+  template <int N,class vComplex_t>       // Projects on SU(N), Sp(2N), with unit determinant, by first projecting on general group and then enforcing unit determinant
+  static void ProjectOnSpecialGroup(Lattice<iScalar<iScalar<iMatrix<vComplex_t, N> > > > &Umu) {
+       Umu = ProjectOnGeneralGroup(Umu);
+       auto det = Determinant(Umu);
+
+       det = conjugate(det);
+
+       for (int i = 0; i < N; i++) {
+           auto element = PeekIndex<ColourIndex>(Umu, N - 1, i);
+           element = element * det;
+           PokeIndex<ColourIndex>(Umu, element, Nc - 1, i);
+       }
+   }
+
+  template <int N,class vComplex_t>    // reunitarise, resimplectify... previously ProjectSUn
+    static void ProjectOnSpecialGroup(Lattice<iVector<iScalar<iMatrix<vComplex_t, N> >, Nd> > &U) {
+      // Reunitarise
+      for (int mu = 0; mu < Nd; mu++) {
+        auto Umu = PeekIndex<LorentzIndex>(U, mu);
+        ProjectOnSpecialGroup(Umu);
+        PokeIndex<LorentzIndex>(U, Umu, mu);
+      }
+    }
+    
+  template <typename GaugeField>
+  static void HotConfiguration(GridParallelRNG &pRNG, GaugeField &out) {
+    typedef typename GaugeField::vector_type vector_type;
+    typedef iGroupMatrix<vector_type> vMatrixType;
+    typedef Lattice<vMatrixType> LatticeMatrixType;
+
+    LatticeMatrixType Umu(out.Grid());
+    LatticeMatrixType tmp(out.Grid());
+    for (int mu = 0; mu < Nd; mu++) {
+      //      LieRandomize(pRNG, Umu, 1.0);
+      //      PokeIndex<LorentzIndex>(out, Umu, mu);
+      gaussian(pRNG,Umu);
+      tmp = Ta(Umu);
+      taExp(tmp,Umu);
+      ProjectOnSpecialGroup(Umu);
+      //      ProjectSUn(Umu);
+      PokeIndex<LorentzIndex>(out, Umu, mu);
+    }
+  }
+  template <typename GaugeField>
+  static void TepidConfiguration(GridParallelRNG &pRNG, GaugeField &out) {
+    typedef typename GaugeField::vector_type vector_type;
+    typedef iGroupMatrix<vector_type> vMatrixType;
+    typedef Lattice<vMatrixType> LatticeMatrixType;
+
+    LatticeMatrixType Umu(out.Grid());
+    for (int mu = 0; mu < Nd; mu++) {
+      LieRandomize(pRNG, Umu, 0.01);
+      PokeIndex<LorentzIndex>(out, Umu, mu);
+    }
+  }
+    
+  template <typename GaugeField>
+  static void ColdConfiguration(GaugeField &out) {
+    typedef typename GaugeField::vector_type vector_type;
+    typedef iGroupMatrix<vector_type> vMatrixType;
+    typedef Lattice<vMatrixType> LatticeMatrixType;
+
+    LatticeMatrixType Umu(out.Grid());
+    Umu = 1.0;
+    for (int mu = 0; mu < Nd; mu++) {
+      PokeIndex<LorentzIndex>(out, Umu, mu);
+    }
+  }
+    
+  template <typename GaugeField>
+  static void ColdConfiguration(GridParallelRNG &pRNG, GaugeField &out) {
+    ColdConfiguration(out);
+  }
+
+  template <typename LatticeMatrixType>
+  static void taProj(const LatticeMatrixType &in, LatticeMatrixType &out) {
+    taProj(in, out, group_name());
+  }
+    
+  template <typename LatticeMatrixType>
+  static void taExp(const LatticeMatrixType &x, LatticeMatrixType &ex) {
+    typedef typename LatticeMatrixType::scalar_type ComplexType;
+
+    LatticeMatrixType xn(x.Grid());
+    RealD nfac = 1.0;
+
+    xn = x;
+    ex = xn + ComplexType(1.0);  // 1+x
+
+    // Do a 12th order exponentiation
+    for (int i = 2; i <= 12; ++i) {
+      nfac = nfac / RealD(i);  // 1/2, 1/2.3 ...
+      xn = xn * x;             // x2, x3,x4....
+      ex = ex + xn * nfac;     // x2/2!, x3/3!....
+    }
+  }
+
+};
+    
+template <int ncolour>
+using SU = GaugeGroup<ncolour, GroupName::SU>;
+
+template <int ncolour>
+using Sp = GaugeGroup<ncolour, GroupName::Sp>;
+
+typedef SU<2> SU2;
+typedef SU<3> SU3;
+typedef SU<4> SU4;
+typedef SU<5> SU5;
+
+typedef SU<Nc> FundamentalMatrices;
+    
+typedef Sp<2> Sp2;
+typedef Sp<4> Sp4;
+typedef Sp<6> Sp6;
+typedef Sp<8> Sp8;
+
+template <int N,class vComplex_t>
+static void ProjectSUn(Lattice<iScalar<iScalar<iMatrix<vComplex_t, N> > > > &Umu)
+{
+    GaugeGroup<N,GroupName::SU>::ProjectOnSpecialGroup(Umu);
+}
+  
+template <int N,class vComplex_t>
+static void ProjectSUn(Lattice<iVector<iScalar<iMatrix<vComplex_t, N> >,Nd> > &U)
+{
+    GaugeGroup<N,GroupName::SU>::ProjectOnSpecialGroup(U);
+}
+    
+template <int N,class vComplex_t>
+static void ProjectSpn(Lattice<iScalar<iScalar<iMatrix<vComplex_t, N> > > > &Umu)
+{
+    GaugeGroup<N,GroupName::Sp>::ProjectOnSpecialGroup(Umu);
+}
+    
+template <int N,class vComplex_t>
+static void ProjectSpn(Lattice<iVector<iScalar<iMatrix<vComplex_t, N> >,Nd> > &U)
+{
+    GaugeGroup<N,GroupName::Sp>::ProjectOnSpecialGroup(U);
+}
+
+// Explicit specialisation for SU(3).
+static void ProjectSU3(Lattice<iScalar<iScalar<iMatrix<vComplexD, 3> > > > &Umu)
+{
+  GridBase *grid = Umu.Grid();
+  const int x = 0;
+  const int y = 1;
+  const int z = 2;
+  // Reunitarise
+  Umu = ProjectOnGroup(Umu);
+  autoView(Umu_v, Umu, CpuWrite);
+  thread_for(ss, grid->oSites(), {
+    auto cm = Umu_v[ss];
+    cm()()(2, x) = adj(cm()()(0, y) * cm()()(1, z) -
+                       cm()()(0, z) * cm()()(1, y));  // x= yz-zy
+    cm()()(2, y) = adj(cm()()(0, z) * cm()()(1, x) -
+                       cm()()(0, x) * cm()()(1, z));  // y= zx-xz
+    cm()()(2, z) = adj(cm()()(0, x) * cm()()(1, y) -
+                       cm()()(0, y) * cm()()(1, x));  // z= xy-yx
+    Umu_v[ss] = cm;
+  });
+}
+static void ProjectSU3(Lattice<iVector<iScalar<iMatrix<vComplexD, 3> >, Nd> > &U)
+{
+  GridBase *grid = U.Grid();
+  // Reunitarise
+  for (int mu = 0; mu < Nd; mu++) {
+    auto Umu = PeekIndex<LorentzIndex>(U, mu);
+    Umu = ProjectOnGroup(Umu);
+    ProjectSU3(Umu);
+    PokeIndex<LorentzIndex>(U, Umu, mu);
+  }
+}
+
+NAMESPACE_END(Grid);
+#endif

Grid/qcd/utils/GaugeGroupTwoIndex.h

@@ -0,0 +1,371 @@
+////////////////////////////////////////////////////////////////////////
+//
+// * Two index representation generators
+//
+// * Normalisation for the fundamental generators:
+//   trace ta tb = 1/2 delta_ab = T_F delta_ab
+//   T_F = 1/2  for SU(N) groups
+//
+//
+//   base for NxN two index (anti-symmetric) matrices
+//   normalized to 1 (d_ij is the kroenecker delta)
+//
+//   (e^(ij)_{kl} = 1 / sqrt(2) (d_ik d_jl +/- d_jk d_il)
+//
+//   Then the generators are written as
+//
+//   (iT_a)^(ij)(lk) = i * ( tr[e^(ij)^dag e^(lk) T^trasp_a] +
+//   tr[e^(lk)e^(ij)^dag T_a] )  //
+//
+//
+////////////////////////////////////////////////////////////////////////
+
+// Authors: David Preti, Guido Cossu
+
+#ifndef QCD_UTIL_GAUGEGROUPTWOINDEX_H
+#define QCD_UTIL_GAUGEGROUPTWOINDEX_H
+
+NAMESPACE_BEGIN(Grid);
+
+enum TwoIndexSymmetry { Symmetric = 1, AntiSymmetric = -1 };
+
+constexpr inline Real delta(int a, int b) { return (a == b) ? 1.0 : 0.0; }
+
+namespace detail {
+
+template <class cplx, int nc, TwoIndexSymmetry S>
+struct baseOffDiagonalSpHelper;
+
+template <class cplx, int nc>
+struct baseOffDiagonalSpHelper<cplx, nc, AntiSymmetric> {
+  static const int ngroup = nc / 2;
+  static void baseOffDiagonalSp(int i, int j, iScalar<iScalar<iMatrix<cplx, nc> > > &eij) {
+    eij = Zero();
+    RealD tmp;
+
+    if ((i == ngroup + j) && (1 <= j) && (j < ngroup)) {
+      for (int k = 0; k < j+1; k++) {
+        if (k < j) {
+          tmp = 1 / sqrt(j * (j + 1));
+          eij()()(k, k + ngroup) = tmp;
+          eij()()(k + ngroup, k) = -tmp;
+        }
+        if (k == j) {
+          tmp = -j / sqrt(j * (j + 1));
+          eij()()(k, k + ngroup) = tmp;
+          eij()()(k + ngroup, k) = -tmp;
+        }
+      }
+
+    }
+
+    else if (i != ngroup + j) {
+      for (int k = 0; k < nc; k++)
+        for (int l = 0; l < nc; l++) {
+          eij()()(l, k) =
+              delta(i, k) * delta(j, l) - delta(j, k) * delta(i, l);
+        }
+    }
+    RealD nrm = 1. / std::sqrt(2.0);
+    eij = eij * nrm;
+  }
+};
+
+template <class cplx, int nc>
+struct baseOffDiagonalSpHelper<cplx, nc, Symmetric> {
+  static void baseOffDiagonalSp(int i, int j, iScalar<iScalar<iMatrix<cplx, nc> > > &eij) {
+    eij = Zero();
+    for (int k = 0; k < nc; k++)
+      for (int l = 0; l < nc; l++)
+        eij()()(l, k) =
+            delta(i, k) * delta(j, l) + delta(j, k) * delta(i, l);
+
+    RealD nrm = 1. / std::sqrt(2.0);
+    eij = eij * nrm;
+  }
+};
+
+}   // closing detail namespace
+
+template <int ncolour, TwoIndexSymmetry S, class group_name>
+class GaugeGroupTwoIndex : public GaugeGroup<ncolour, group_name> {
+ public:
+  // The chosen convention is that we are taking ncolour to be N in SU<N> but 2N
+  // in Sp(2N). ngroup is equal to N for SU but 2N/2 = N for Sp(2N).
+  static_assert(std::is_same<group_name, GroupName::SU>::value or
+                    std::is_same<group_name, GroupName::Sp>::value,
+                "ngroup is only implemented for SU and Sp currently.");
+  static const int ngroup =
+      std::is_same<group_name, GroupName::SU>::value ? ncolour : ncolour / 2;
+  static const int Dimension =
+      (ncolour * (ncolour + S) / 2) + (std::is_same<group_name, GroupName::Sp>::value ? (S - 1) / 2 : 0);
+  static const int DimensionAS =
+      (ncolour * (ncolour - 1) / 2) + (std::is_same<group_name, GroupName::Sp>::value ? (- 1) : 0);
+  static const int DimensionS =
+      ncolour * (ncolour + 1) / 2;
+  static const int NumGenerators =
+      GaugeGroup<ncolour, group_name>::AlgebraDimension;
+
+  template <typename vtype>
+  using iGroupTwoIndexMatrix = iScalar<iScalar<iMatrix<vtype, Dimension> > >;
+
+  typedef iGroupTwoIndexMatrix<Complex> TIMatrix;
+  typedef iGroupTwoIndexMatrix<ComplexF> TIMatrixF;
+  typedef iGroupTwoIndexMatrix<ComplexD> TIMatrixD;
+
+  typedef iGroupTwoIndexMatrix<vComplex> vTIMatrix;
+  typedef iGroupTwoIndexMatrix<vComplexF> vTIMatrixF;
+  typedef iGroupTwoIndexMatrix<vComplexD> vTIMatrixD;
+
+  typedef Lattice<vTIMatrix> LatticeTwoIndexMatrix;
+  typedef Lattice<vTIMatrixF> LatticeTwoIndexMatrixF;
+  typedef Lattice<vTIMatrixD> LatticeTwoIndexMatrixD;
+
+  typedef Lattice<iVector<iScalar<iMatrix<vComplex, Dimension> >, Nd> >
+      LatticeTwoIndexField;
+  typedef Lattice<iVector<iScalar<iMatrix<vComplexF, Dimension> >, Nd> >
+      LatticeTwoIndexFieldF;
+  typedef Lattice<iVector<iScalar<iMatrix<vComplexD, Dimension> >, Nd> >
+      LatticeTwoIndexFieldD;
+
+  template <typename vtype>
+  using iGroupMatrix = iScalar<iScalar<iMatrix<vtype, ncolour> > >;
+
+  typedef iGroupMatrix<Complex> Matrix;
+  typedef iGroupMatrix<ComplexF> MatrixF;
+  typedef iGroupMatrix<ComplexD> MatrixD;
+    
+private:
+  template <class cplx>
+  static void baseDiagonal(int Index, iGroupMatrix<cplx> &eij) {
+    eij = Zero();
+    eij()()(Index - ncolour * (ncolour - 1) / 2,
+            Index - ncolour * (ncolour - 1) / 2) = 1.0;
+  }
+    
+  template <class cplx>
+  static void baseOffDiagonal(int i, int j, iGroupMatrix<cplx> &eij, GroupName::SU) {
+    eij = Zero();
+    for (int k = 0; k < ncolour; k++)
+      for (int l = 0; l < ncolour; l++)
+        eij()()(l, k) =
+            delta(i, k) * delta(j, l) + S * delta(j, k) * delta(i, l);
+
+    RealD nrm = 1. / std::sqrt(2.0);
+    eij = eij * nrm;
+  }
+    
+  template <class cplx>
+  static void baseOffDiagonal(int i, int j, iGroupMatrix<cplx> &eij, GroupName::Sp) {
+    detail::baseOffDiagonalSpHelper<cplx, ncolour, S>::baseOffDiagonalSp(i, j, eij);
+  }
+
+public:
+    
+  template <class cplx>
+  static void base(int Index, iGroupMatrix<cplx> &eij) {
+  // returns (e)^(ij)_{kl} necessary for change of base U_F -> U_R
+    assert(Index < Dimension);
+    eij = Zero();
+  // for the linearisation of the 2 indexes
+    static int a[ncolour * (ncolour - 1) / 2][2];  // store the a <-> i,j
+    static bool filled = false;
+    if (!filled) {
+      int counter = 0;
+      for (int i = 1; i < ncolour; i++) {
+      for (int j = 0; j < i; j++) {
+        if (std::is_same<group_name, GroupName::Sp>::value)
+          {
+            if (j==0 && i==ngroup+j && S==-1) {
+            //std::cout << "skipping" << std::endl; // for Sp2n this vanishes identically.
+              j = j+1;
+            }
+          }
+          a[counter][0] = i;
+          a[counter][1] = j;
+          counter++;
+          }
+      }
+      filled = true;
+    }
+    if (Index < ncolour*ncolour - DimensionS)
+    {
+      baseOffDiagonal(a[Index][0], a[Index][1], eij, group_name());
+    } else {
+      baseDiagonal(Index, eij);
+    }
+  }
+    
+  static void printBase(void) {
+    for (int gen = 0; gen < Dimension; gen++) {
+      Matrix tmp;
+      base(gen, tmp);
+      std::cout << GridLogMessage << "Nc = " << ncolour << " t_" << gen
+                << std::endl;
+      std::cout << GridLogMessage << tmp << std::endl;
+    }
+  }
+
+  template <class cplx>
+  static void generator(int Index, iGroupTwoIndexMatrix<cplx> &i2indTa) {
+    Vector<iGroupMatrix<cplx> > ta(NumGenerators);
+    Vector<iGroupMatrix<cplx> > eij(Dimension);
+    iGroupMatrix<cplx> tmp;
+
+    for (int a = 0; a < NumGenerators; a++)
+      GaugeGroup<ncolour, group_name>::generator(a, ta[a]);
+
+    for (int a = 0; a < Dimension; a++) base(a, eij[a]);
+
+    for (int a = 0; a < Dimension; a++) {
+      tmp = transpose(eij[a]*ta[Index]) + transpose(eij[a]) * ta[Index];
+      for (int b = 0; b < Dimension; b++) {
+        Complex iTr = TensorRemove(timesI(trace(tmp * eij[b])));
+        i2indTa()()(a, b) = iTr;
+      }
+    }
+  }
+
+  static void printGenerators(void) {
+    for (int gen = 0; gen < NumGenerators; gen++) {
+      TIMatrix i2indTa;
+      generator(gen, i2indTa);
+      std::cout << GridLogMessage << "Nc = " << ncolour << " t_" << gen
+                << std::endl;
+      std::cout << GridLogMessage << i2indTa << std::endl;
+    }
+  }
+
+  static void testGenerators(void) {
+    TIMatrix i2indTa, i2indTb;
+    std::cout << GridLogMessage << "2IndexRep - Checking if traceless"
+              << std::endl;
+    for (int a = 0; a < NumGenerators; a++) {
+      generator(a, i2indTa);
+      std::cout << GridLogMessage << a << std::endl;
+      assert(norm2(trace(i2indTa)) < 1.0e-6);
+    }
+    std::cout << GridLogMessage << std::endl;
+
+    std::cout << GridLogMessage << "2IndexRep - Checking if antihermitean"
+              << std::endl;
+    for (int a = 0; a < NumGenerators; a++) {
+      generator(a, i2indTa);
+      std::cout << GridLogMessage << a << std::endl;
+      assert(norm2(adj(i2indTa) + i2indTa) < 1.0e-6);
+    }
+
+    std::cout << GridLogMessage << std::endl;
+    std::cout << GridLogMessage
+              << "2IndexRep - Checking Tr[Ta*Tb]=delta(a,b)*(N +- 2)/2"
+              << std::endl;
+    for (int a = 0; a < NumGenerators; a++) {
+      for (int b = 0; b < NumGenerators; b++) {
+        generator(a, i2indTa);
+        generator(b, i2indTb);
+
+        // generator returns iTa, so we need a minus sign here
+        Complex Tr = -TensorRemove(trace(i2indTa * i2indTb));
+        std::cout << GridLogMessage << "a=" << a << "b=" << b << "Tr=" << Tr
+                  << std::endl;
+        if (a == b) {
+          assert(real(Tr) - ((ncolour + S * 2) * 0.5) < 1e-8);
+        } else {
+          assert(real(Tr) < 1e-8);
+        }
+        assert(imag(Tr) < 1e-8);
+      }
+    }
+    std::cout << GridLogMessage << std::endl;
+  }
+
+  static void TwoIndexLieAlgebraMatrix(
+      const typename GaugeGroup<ncolour, group_name>::LatticeAlgebraVector &h,
+      LatticeTwoIndexMatrix &out, Real scale = 1.0) {
+    conformable(h, out);
+    GridBase *grid = out.Grid();
+    LatticeTwoIndexMatrix la(grid);
+    TIMatrix i2indTa;
+
+    out = Zero();
+    for (int a = 0; a < NumGenerators; a++) {
+      generator(a, i2indTa);
+      la = peekColour(h, a) * i2indTa;
+      out += la;
+    }
+    out *= scale;
+  }
+
+  // Projects the algebra components
+  // of a lattice matrix ( of dimension ncol*ncol -1 )
+  static void projectOnAlgebra(
+      typename GaugeGroup<ncolour, group_name>::LatticeAlgebraVector &h_out,
+      const LatticeTwoIndexMatrix &in, Real scale = 1.0) {
+    conformable(h_out, in);
+    h_out = Zero();
+    TIMatrix i2indTa;
+    Real coefficient = -2.0 / (ncolour + 2 * S) * scale;
+    // 2/(Nc +/- 2) for the normalization of the trace in the two index rep
+    for (int a = 0; a < NumGenerators; a++) {
+      generator(a, i2indTa);
+      pokeColour(h_out, real(trace(i2indTa * in)) * coefficient, a);
+    }
+  }
+
+  // a projector that keeps the generators stored to avoid the overhead of
+  // recomputing them
+  static void projector(
+      typename GaugeGroup<ncolour, group_name>::LatticeAlgebraVector &h_out,
+      const LatticeTwoIndexMatrix &in, Real scale = 1.0) {
+    conformable(h_out, in);
+    // to store the generators
+    static std::vector<TIMatrix> i2indTa(NumGenerators);
+    h_out = Zero();
+    static bool precalculated = false;
+    if (!precalculated) {
+      precalculated = true;
+      for (int a = 0; a < NumGenerators; a++) generator(a, i2indTa[a]);
+    }
+
+    Real coefficient =
+        -2.0 / (ncolour + 2 * S) * scale;  // 2/(Nc +/- 2) for the normalization
+    // of the trace in the two index rep
+
+    for (int a = 0; a < NumGenerators; a++) {
+      auto tmp = real(trace(i2indTa[a] * in)) * coefficient;
+      pokeColour(h_out, tmp, a);
+    }
+  }
+};
+
+template <int ncolour, TwoIndexSymmetry S>
+using SU_TwoIndex = GaugeGroupTwoIndex<ncolour, S, GroupName::SU>;
+
+// Some useful type names
+typedef SU_TwoIndex<Nc, Symmetric> TwoIndexSymmMatrices;
+typedef SU_TwoIndex<Nc, AntiSymmetric> TwoIndexAntiSymmMatrices;
+
+typedef SU_TwoIndex<2, Symmetric> SU2TwoIndexSymm;
+typedef SU_TwoIndex<3, Symmetric> SU3TwoIndexSymm;
+typedef SU_TwoIndex<4, Symmetric> SU4TwoIndexSymm;
+typedef SU_TwoIndex<5, Symmetric> SU5TwoIndexSymm;
+
+typedef SU_TwoIndex<2, AntiSymmetric> SU2TwoIndexAntiSymm;
+typedef SU_TwoIndex<3, AntiSymmetric> SU3TwoIndexAntiSymm;
+typedef SU_TwoIndex<4, AntiSymmetric> SU4TwoIndexAntiSymm;
+typedef SU_TwoIndex<5, AntiSymmetric> SU5TwoIndexAntiSymm;
+
+template <int ncolour, TwoIndexSymmetry S>
+using Sp_TwoIndex = GaugeGroupTwoIndex<ncolour, S, GroupName::Sp>;
+
+typedef Sp_TwoIndex<Nc, Symmetric> SpTwoIndexSymmMatrices;
+typedef Sp_TwoIndex<Nc, AntiSymmetric> SpTwoIndexAntiSymmMatrices;
+
+typedef Sp_TwoIndex<2, Symmetric> Sp2TwoIndexSymm;
+typedef Sp_TwoIndex<4, Symmetric> Sp4TwoIndexSymm;
+
+typedef Sp_TwoIndex<4, AntiSymmetric> Sp4TwoIndexAntiSymm;
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/qcd/utils/SUn.h

@@ -1,932 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/utils/SUn.h
-
-Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-Author: paboyle <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 */
-#ifndef QCD_UTIL_SUN_H
-#define QCD_UTIL_SUN_H
-
-NAMESPACE_BEGIN(Grid);
-
-template<int N, class Vec>
-Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
-{
-  GridBase *grid=Umu.Grid();
-  auto lvol = grid->lSites();
-  Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
-  typedef typename Vec::scalar_type scalar;
-  autoView(Umu_v,Umu,CpuRead);
-  autoView(ret_v,ret,CpuWrite);
-  thread_for(site,lvol,{
-    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
-    Coordinate lcoor;
-    grid->LocalIndexToLocalCoor(site, lcoor);
-    iScalar<iScalar<iMatrix<scalar, N> > > Us;
-    peekLocalSite(Us, Umu_v, lcoor);
-    for(int i=0;i<N;i++){
-      for(int j=0;j<N;j++){
-	scalar tmp= Us()()(i,j);
-	ComplexD ztmp(real(tmp),imag(tmp));
-	EigenU(i,j)=ztmp;
-      }}
-    ComplexD detD  = EigenU.determinant();
-    typename Vec::scalar_type det(detD.real(),detD.imag());
-    pokeLocalSite(det,ret_v,lcoor);
-  });
-  return ret;
-}
-
-template<int N, class Vec>
-static void ProjectSUn(Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
-{
-  Umu      = ProjectOnGroup(Umu);
-  auto det = Determinant(Umu);
-
-  det = conjugate(det);
-
-  for(int i=0;i<N;i++){
-    auto element = PeekIndex<ColourIndex>(Umu,N-1,i);
-    element = element * det;
-    PokeIndex<ColourIndex>(Umu,element,Nc-1,i);
-  }
-}
-template<int N,class Vec>
-static void ProjectSUn(Lattice<iVector<iScalar<iMatrix<Vec, N> >,Nd> > &U)
-{
-  GridBase *grid=U.Grid();
-  // Reunitarise
-  for(int mu=0;mu<Nd;mu++){
-    auto Umu = PeekIndex<LorentzIndex>(U,mu);
-    Umu      = ProjectOnGroup(Umu);
-    ProjectSUn(Umu);
-    PokeIndex<LorentzIndex>(U,Umu,mu);
-  }
-}
-
-template <int ncolour>
-class SU {
-public:
-  static const int Dimension = ncolour;
-  static const int AdjointDimension = ncolour * ncolour - 1;
-  static int su2subgroups(void) { return (ncolour * (ncolour - 1)) / 2; }
-
-  template <typename vtype>
-  using iSUnMatrix = iScalar<iScalar<iMatrix<vtype, ncolour> > >;
-  template <typename vtype>
-  using iSU2Matrix = iScalar<iScalar<iMatrix<vtype, 2> > >;
-  template <typename vtype>
-  using iSUnAlgebraVector =
-    iScalar<iScalar<iVector<vtype, AdjointDimension> > >;
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////
-  // Types can be accessed as SU<2>::Matrix , SU<2>::vSUnMatrix,
-  // SU<2>::LatticeMatrix etc...
-  //////////////////////////////////////////////////////////////////////////////////////////////////
-  typedef iSUnMatrix<Complex> Matrix;
-  typedef iSUnMatrix<ComplexF> MatrixF;
-  typedef iSUnMatrix<ComplexD> MatrixD;
-
-  typedef iSUnMatrix<vComplex> vMatrix;
-  typedef iSUnMatrix<vComplexF> vMatrixF;
-  typedef iSUnMatrix<vComplexD> vMatrixD;
-
-  // For the projectors to the algebra
-  // these should be real...
-  // keeping complex for consistency with the SIMD vector types
-  typedef iSUnAlgebraVector<Complex> AlgebraVector;
-  typedef iSUnAlgebraVector<ComplexF> AlgebraVectorF;
-  typedef iSUnAlgebraVector<ComplexD> AlgebraVectorD;
-
-  typedef iSUnAlgebraVector<vComplex> vAlgebraVector;
-  typedef iSUnAlgebraVector<vComplexF> vAlgebraVectorF;
-  typedef iSUnAlgebraVector<vComplexD> vAlgebraVectorD;
-
-  typedef Lattice<vMatrix> LatticeMatrix;
-  typedef Lattice<vMatrixF> LatticeMatrixF;
-  typedef Lattice<vMatrixD> LatticeMatrixD;
-
-  typedef Lattice<vAlgebraVector> LatticeAlgebraVector;
-  typedef Lattice<vAlgebraVectorF> LatticeAlgebraVectorF;
-  typedef Lattice<vAlgebraVectorD> LatticeAlgebraVectorD;
-
-  typedef iSU2Matrix<Complex> SU2Matrix;
-  typedef iSU2Matrix<ComplexF> SU2MatrixF;
-  typedef iSU2Matrix<ComplexD> SU2MatrixD;
-
-  typedef iSU2Matrix<vComplex> vSU2Matrix;
-  typedef iSU2Matrix<vComplexF> vSU2MatrixF;
-  typedef iSU2Matrix<vComplexD> vSU2MatrixD;
-
-  typedef Lattice<vSU2Matrix> LatticeSU2Matrix;
-  typedef Lattice<vSU2MatrixF> LatticeSU2MatrixF;
-  typedef Lattice<vSU2MatrixD> LatticeSU2MatrixD;
-
-  ////////////////////////////////////////////////////////////////////////
-  // There are N^2-1 generators for SU(N).
-  //
-  // We take a traceless hermitian generator basis as follows
-  //
-  // * Normalisation: trace ta tb = 1/2 delta_ab = T_F delta_ab
-  //   T_F = 1/2  for SU(N) groups
-  //
-  // * Off diagonal
-  //    - pairs of rows i1,i2 behaving like pauli matrices signma_x, sigma_y
-  //
-  //    - there are (Nc-1-i1) slots for i2 on each row [ x  0  x ]
-  //      direct count off each row
-  //
-  //    - Sum of all pairs is Nc(Nc-1)/2: proof arithmetic series
-  //
-  //      (Nc-1) + (Nc-2)+...  1      ==> Nc*(Nc-1)/2
-  //      1+ 2+          +   + Nc-1
-  //
-  //    - There are 2 x Nc (Nc-1)/ 2 of these = Nc^2 - Nc
-  //
-  //    - We enumerate the row-col pairs.
-  //    - for each row col pair there is a (sigma_x) and a (sigma_y) like
-  //    generator
-  //
-  //
-  //   t^a_ij = { in 0.. Nc(Nc-1)/2 -1} =>  1/2(delta_{i,i1} delta_{j,i2} +
-  //   delta_{i,i1} delta_{j,i2})
-  //   t^a_ij = { in Nc(Nc-1)/2 ... Nc(Nc-1) - 1} =>  i/2( delta_{i,i1}
-  //   delta_{j,i2} - i delta_{i,i1} delta_{j,i2})
-  //
-  // * Diagonal; must be traceless and normalised
-  //   - Sequence is
-  //   N  (1,-1,0,0...)
-  //   N  (1, 1,-2,0...)
-  //   N  (1, 1, 1,-3,0...)
-  //   N  (1, 1, 1, 1,-4,0...)
-  //
-  //   where 1/2 = N^2 (1+.. m^2)etc.... for the m-th diagonal generator
-  //   NB this gives the famous SU3 result for su2 index 8
-  //
-  //   N= sqrt(1/2 . 1/6 ) = 1/2 . 1/sqrt(3)
-  //
-  //   ( 1      )
-  //   (    1   ) / sqrt(3) /2  = 1/2 lambda_8
-  //   (      -2)
-  //
-  ////////////////////////////////////////////////////////////////////////
-  template <class cplx>
-  static void generator(int lieIndex, iSUnMatrix<cplx> &ta) {
-    // map lie index to which type of generator
-    int diagIndex;
-    int su2Index;
-    int sigxy;
-    int NNm1 = ncolour * (ncolour - 1);
-    if (lieIndex >= NNm1) {
-      diagIndex = lieIndex - NNm1;
-      generatorDiagonal(diagIndex, ta);
-      return;
-    }
-    sigxy = lieIndex & 0x1;  // even or odd
-    su2Index = lieIndex >> 1;
-    if (sigxy)
-      generatorSigmaY(su2Index, ta);
-    else
-      generatorSigmaX(su2Index, ta);
-  }
-  
-  template <class cplx>
-  static void generatorSigmaY(int su2Index, iSUnMatrix<cplx> &ta) {
-    ta = Zero();
-    int i1, i2;
-    su2SubGroupIndex(i1, i2, su2Index);
-    ta()()(i1, i2) = 1.0;
-    ta()()(i2, i1) = 1.0;
-    ta = ta * 0.5;
-  }
-  
-  template <class cplx>
-  static void generatorSigmaX(int su2Index, iSUnMatrix<cplx> &ta) {
-    ta = Zero();
-    cplx i(0.0, 1.0);
-    int i1, i2;
-    su2SubGroupIndex(i1, i2, su2Index);
-    ta()()(i1, i2) = i;
-    ta()()(i2, i1) = -i;
-    ta = ta * 0.5;
-  }
-
-  template <class cplx>
-  static void generatorDiagonal(int diagIndex, iSUnMatrix<cplx> &ta) {
-    // diag ({1, 1, ..., 1}(k-times), -k, 0, 0, ...)
-    ta = Zero();
-    int k = diagIndex + 1;                  // diagIndex starts from 0
-    for (int i = 0; i <= diagIndex; i++) {  // k iterations
-      ta()()(i, i) = 1.0;
-    }
-    ta()()(k, k) = -k;  // indexing starts from 0
-    RealD nrm = 1.0 / std::sqrt(2.0 * k * (k + 1));
-    ta = ta * nrm;
-  }
-
-
-
-  ////////////////////////////////////////////////////////////////////////
-  // Map a su2 subgroup number to the pair of rows that are non zero
-  ////////////////////////////////////////////////////////////////////////
-  static void su2SubGroupIndex(int &i1, int &i2, int su2_index) {
-    assert((su2_index >= 0) && (su2_index < (ncolour * (ncolour - 1)) / 2));
-
-    int spare = su2_index;
-    for (i1 = 0; spare >= (ncolour - 1 - i1); i1++) {
-      spare = spare - (ncolour - 1 - i1);  // remove the Nc-1-i1 terms
-    }
-    i2 = i1 + 1 + spare;
-  }
-
-  //////////////////////////////////////////////////////////////////////////////////////////
-  // Pull out a subgroup and project on to real coeffs x pauli basis
-  //////////////////////////////////////////////////////////////////////////////////////////
-  template <class vcplx>
-  static void su2Extract(Lattice<iSinglet<vcplx> > &Determinant,
-                         Lattice<iSU2Matrix<vcplx> > &subgroup,
-                         const Lattice<iSUnMatrix<vcplx> > &source,
-                         int su2_index) {
-    GridBase *grid(source.Grid());
-    conformable(subgroup, source);
-    conformable(subgroup, Determinant);
-    int i0, i1;
-    su2SubGroupIndex(i0, i1, su2_index);
-
-    autoView( subgroup_v , subgroup,AcceleratorWrite);
-    autoView( source_v   , source,AcceleratorRead);
-    autoView( Determinant_v , Determinant,AcceleratorWrite);
-    accelerator_for(ss, grid->oSites(), 1, {
-
-      subgroup_v[ss]()()(0, 0) = source_v[ss]()()(i0, i0);
-      subgroup_v[ss]()()(0, 1) = source_v[ss]()()(i0, i1);
-      subgroup_v[ss]()()(1, 0) = source_v[ss]()()(i1, i0);
-      subgroup_v[ss]()()(1, 1) = source_v[ss]()()(i1, i1);
-
-      iSU2Matrix<vcplx> Sigma = subgroup_v[ss];
-
-      Sigma = Sigma - adj(Sigma) + trace(adj(Sigma));
-
-      subgroup_v[ss] = Sigma;
-
-      // this should be purely real
-      Determinant_v[ss] =
-	Sigma()()(0, 0) * Sigma()()(1, 1) - Sigma()()(0, 1) * Sigma()()(1, 0);
-    });
-  }
-
-  //////////////////////////////////////////////////////////////////////////////////////////
-  // Set matrix to one and insert a pauli subgroup
-  //////////////////////////////////////////////////////////////////////////////////////////
-  template <class vcplx>
-  static void su2Insert(const Lattice<iSU2Matrix<vcplx> > &subgroup,
-                        Lattice<iSUnMatrix<vcplx> > &dest, int su2_index) {
-    GridBase *grid(dest.Grid());
-    conformable(subgroup, dest);
-    int i0, i1;
-    su2SubGroupIndex(i0, i1, su2_index);
-
-    dest = 1.0;  // start out with identity
-    autoView( dest_v , dest, AcceleratorWrite);
-    autoView( subgroup_v, subgroup, AcceleratorRead);
-    accelerator_for(ss, grid->oSites(),1,
-    {
-      dest_v[ss]()()(i0, i0) = subgroup_v[ss]()()(0, 0);
-      dest_v[ss]()()(i0, i1) = subgroup_v[ss]()()(0, 1);
-      dest_v[ss]()()(i1, i0) = subgroup_v[ss]()()(1, 0);
-      dest_v[ss]()()(i1, i1) = subgroup_v[ss]()()(1, 1);
-    });
-
-  }
-
-  ///////////////////////////////////////////////
-  // Generate e^{ Re Tr Staple Link} dlink
-  //
-  // *** Note Staple should be appropriate linear compbination between all
-  // staples.
-  // *** If already by beta pass coefficient 1.0.
-  // *** This routine applies the additional 1/Nc factor that comes after trace
-  // in action.
-  //
-  ///////////////////////////////////////////////
-  static void SubGroupHeatBath(GridSerialRNG &sRNG, GridParallelRNG &pRNG,
-			       RealD beta,  // coeff multiplying staple in action (with no 1/Nc)
-			       LatticeMatrix &link,
-			       const LatticeMatrix &barestaple,  // multiplied by action coeffs so th
-			       int su2_subgroup, int nheatbath, LatticeInteger &wheremask) 
-  {
-    GridBase *grid = link.Grid();
-
-    const RealD twopi = 2.0 * M_PI;
-
-    LatticeMatrix staple(grid);
-
-    staple = barestaple * (beta / ncolour);
-
-    LatticeMatrix V(grid);
-    V = link * staple;
-
-    // Subgroup manipulation in the lie algebra space
-    LatticeSU2Matrix u(grid);  // Kennedy pendleton "u" real projected normalised Sigma
-    LatticeSU2Matrix uinv(grid);
-    LatticeSU2Matrix ua(grid);  // a in pauli form
-    LatticeSU2Matrix b(grid);   // rotated matrix after hb
-
-    // Some handy constant fields
-    LatticeComplex ones(grid);
-    ones = 1.0;
-    LatticeComplex zeros(grid);
-    zeros = Zero();
-    LatticeReal rones(grid);
-    rones = 1.0;
-    LatticeReal rzeros(grid);
-    rzeros = Zero();
-    LatticeComplex udet(grid);  // determinant of real(staple)
-    LatticeInteger mask_true(grid);
-    mask_true = 1;
-    LatticeInteger mask_false(grid);
-    mask_false = 0;
-
-    /*
-      PLB 156 P393 (1985) (Kennedy and Pendleton)
-
-      Note: absorb "beta" into the def of sigma compared to KP paper; staple
-      passed to this routine has "beta" already multiplied in
-
-      Action linear in links h and of form:
-
-      beta S = beta  Sum_p (1 - 1/Nc Re Tr Plaq )
-
-      Writing Sigma = 1/Nc (beta Sigma') where sum over staples is "Sigma' "
-
-      beta S = const - beta/Nc Re Tr h Sigma'
-      = const - Re Tr h Sigma
-
-      Decompose h and Sigma into (1, sigma_j) ; h_i real, h^2=1, Sigma_i complex
-      arbitrary.
-
-      Tr h Sigma = h_i Sigma_j Tr (sigma_i sigma_j)  = h_i Sigma_j 2 delta_ij
-      Re Tr h Sigma = 2 h_j Re Sigma_j
-
-      Normalised re Sigma_j = xi u_j
-
-      With u_j a unit vector and U can be in SU(2);
-
-      Re Tr h Sigma = 2 h_j Re Sigma_j = 2 xi (h.u)
-
-      4xi^2 = Det [ Sig - Sig^dag  + 1 Tr Sigdag]
-      u   = 1/2xi [ Sig - Sig^dag  + 1 Tr Sigdag]
-
-      xi = sqrt(Det)/2;
-
-      Write a= u h in SU(2); a has pauli decomp a_j;
-
-      Note: Product b' xi is unvariant because scaling Sigma leaves
-      normalised vector "u" fixed; Can rescale Sigma so b' = 1.
-    */
-
-    ////////////////////////////////////////////////////////
-    // Real part of Pauli decomposition
-    // Note a subgroup can project to zero in cold start
-    ////////////////////////////////////////////////////////
-    su2Extract(udet, u, V, su2_subgroup);
-
-    //////////////////////////////////////////////////////
-    // Normalising this vector if possible; else identity
-    //////////////////////////////////////////////////////
-    LatticeComplex xi(grid);
-
-    LatticeSU2Matrix lident(grid);
-
-    SU2Matrix ident = Complex(1.0);
-    SU2Matrix pauli1;
-    SU<2>::generator(0, pauli1);
-    SU2Matrix pauli2;
-    SU<2>::generator(1, pauli2);
-    SU2Matrix pauli3;
-    SU<2>::generator(2, pauli3);
-    pauli1 = timesI(pauli1) * 2.0;
-    pauli2 = timesI(pauli2) * 2.0;
-    pauli3 = timesI(pauli3) * 2.0;
-
-    LatticeComplex cone(grid);
-    LatticeReal adet(grid);
-    adet = abs(toReal(udet));
-    lident = Complex(1.0);
-    cone = Complex(1.0);
-    Real machine_epsilon = 1.0e-7;
-    u = where(adet > machine_epsilon, u, lident);
-    udet = where(adet > machine_epsilon, udet, cone);
-
-    xi = 0.5 * sqrt(udet);  // 4xi^2 = Det [ Sig - Sig^dag  + 1 Tr Sigdag]
-    u = 0.5 * u *
-      pow(xi, -1.0);  //  u   = 1/2xi [ Sig - Sig^dag  + 1 Tr Sigdag]
-
-    // Debug test for sanity
-    uinv = adj(u);
-    b = u * uinv - 1.0;
-    assert(norm2(b) < 1.0e-4);
-
-    /*
-      Measure: Haar measure dh has d^4a delta(1-|a^2|)
-      In polars:
-      da = da0 r^2 sin theta dr dtheta dphi delta( 1 - r^2 -a0^2)
-      = da0 r^2 sin theta dr dtheta dphi delta( (sqrt(1-a0^) - r)(sqrt(1-a0^) +
-      r) )
-      = da0 r/2 sin theta dr dtheta dphi delta( (sqrt(1-a0^) - r) )
-
-      Action factor Q(h) dh  = e^-S[h]  dh =  e^{  xi Tr uh} dh    // beta enters
-      through xi
-      =  e^{2 xi (h.u)} dh
-      =  e^{2 xi h0u0}.e^{2 xi h1u1}.e^{2 xi
-      h2u2}.e^{2 xi h3u3} dh
-
-      Therefore for each site, take xi for that site
-      i) generate  |a0|<1 with dist
-      (1-a0^2)^0.5 e^{2 xi a0 } da0
-
-      Take alpha = 2 xi  = 2 xi [ recall 2 beta/Nc unmod staple norm]; hence 2.0/Nc
-      factor in Chroma ]
-      A. Generate two uniformly distributed pseudo-random numbers R and R', R'',
-      R''' in the unit interval;
-      B. Set X = -(ln R)/alpha, X' =-(ln R')/alpha;
-      C. Set C = cos^2(2pi R"), with R" another uniform random number in [0,1] ;
-      D. Set A = XC;
-      E. Let d  = X'+A;
-      F. If R'''^2 :> 1 - 0.5 d,  go back to A;
-      G. Set a0 = 1 - d;
-
-      Note that in step D setting B ~ X - A and using B in place of A in step E will
-      generate a second independent a 0 value.
-    */
-
-    /////////////////////////////////////////////////////////
-    // count the number of sites by picking "1"'s out of hat
-    /////////////////////////////////////////////////////////
-    Integer hit = 0;
-    LatticeReal rtmp(grid);
-    rtmp = where(wheremask, rones, rzeros);
-    RealD numSites = sum(rtmp);
-    RealD numAccepted;
-    LatticeInteger Accepted(grid);
-    Accepted = Zero();
-    LatticeInteger newlyAccepted(grid);
-
-    std::vector<LatticeReal> xr(4, grid);
-    std::vector<LatticeReal> a(4, grid);
-    LatticeReal d(grid);
-    d = Zero();
-    LatticeReal alpha(grid);
-
-    //    std::cout<<GridLogMessage<<"xi "<<xi <<std::endl;
-    xi = 2.0 *xi;
-    alpha = toReal(xi);
-
-    do {
-      // A. Generate two uniformly distributed pseudo-random numbers R and R',
-      // R'', R''' in the unit interval;
-      random(pRNG, xr[0]);
-      random(pRNG, xr[1]);
-      random(pRNG, xr[2]);
-      random(pRNG, xr[3]);
-
-      // B. Set X = - ln R/alpha, X' = -ln R'/alpha
-      xr[1] = -log(xr[1]) / alpha;
-      xr[2] = -log(xr[2]) / alpha;
-
-      // C. Set C = cos^2(2piR'')
-      xr[3] = cos(xr[3] * twopi);
-      xr[3] = xr[3] * xr[3];
-
-      LatticeReal xrsq(grid);
-
-      // D. Set A = XC;
-      // E. Let d  = X'+A;
-      xrsq = xr[2] + xr[1] * xr[3];
-
-      d = where(Accepted, d, xr[2] + xr[1] * xr[3]);
-
-      // F. If R'''^2 :> 1 - 0.5 d,  go back to A;
-      LatticeReal thresh(grid);
-      thresh = 1.0 - d * 0.5;
-      xrsq = xr[0] * xr[0];
-      LatticeInteger ione(grid);
-      ione = 1;
-      LatticeInteger izero(grid);
-      izero = Zero();
-
-      newlyAccepted = where(xrsq < thresh, ione, izero);
-      Accepted = where(newlyAccepted, newlyAccepted, Accepted);
-      Accepted = where(wheremask, Accepted, izero);
-
-      // FIXME need an iSum for integer to avoid overload on return type??
-      rtmp = where(Accepted, rones, rzeros);
-      numAccepted = sum(rtmp);
-
-      hit++;
-
-    } while ((numAccepted < numSites) && (hit < nheatbath));
-
-    // G. Set a0 = 1 - d;
-    a[0] = Zero();
-    a[0] = where(wheremask, 1.0 - d, a[0]);
-
-    //////////////////////////////////////////
-    //    ii) generate a_i uniform on two sphere radius (1-a0^2)^0.5
-    //////////////////////////////////////////
-
-    LatticeReal a123mag(grid);
-    a123mag = sqrt(abs(1.0 - a[0] * a[0]));
-
-    LatticeReal cos_theta(grid);
-    LatticeReal sin_theta(grid);
-    LatticeReal phi(grid);
-
-    random(pRNG, phi);
-    phi = phi * twopi;  // uniform in [0,2pi]
-    random(pRNG, cos_theta);
-    cos_theta = (cos_theta * 2.0) - 1.0;  // uniform in [-1,1]
-    sin_theta = sqrt(abs(1.0 - cos_theta * cos_theta));
-
-    a[1] = a123mag * sin_theta * cos(phi);
-    a[2] = a123mag * sin_theta * sin(phi);
-    a[3] = a123mag * cos_theta;
-
-    ua = toComplex(a[0]) * ident  + toComplex(a[1]) * pauli1 +
-         toComplex(a[2]) * pauli2 + toComplex(a[3]) * pauli3;
-
-    b = 1.0;
-    b = where(wheremask, uinv * ua, b);
-    su2Insert(b, V, su2_subgroup);
-
-    // mask the assignment back based on Accptance
-    link = where(Accepted, V * link, link);
-
-    //////////////////////////////
-    // Debug Checks
-    // SU2 check
-    LatticeSU2Matrix check(grid);  // rotated matrix after hb
-    u = Zero();
-    check = ua * adj(ua) - 1.0;
-    check = where(Accepted, check, u);
-    assert(norm2(check) < 1.0e-4);
-
-    check = b * adj(b) - 1.0;
-    check = where(Accepted, check, u);
-    assert(norm2(check) < 1.0e-4);
-
-    LatticeMatrix Vcheck(grid);
-    Vcheck = Zero();
-    Vcheck = where(Accepted, V * adj(V) - 1.0, Vcheck);
-    //    std::cout<<GridLogMessage << "SU3 check " <<norm2(Vcheck)<<std::endl;
-    assert(norm2(Vcheck) < 1.0e-4);
-
-    // Verify the link stays in SU(3)
-    //    std::cout<<GridLogMessage <<"Checking the modified link"<<std::endl;
-    Vcheck = link * adj(link) - 1.0;
-    assert(norm2(Vcheck) < 1.0e-4);
-    /////////////////////////////////
-  }
-
-  static void printGenerators(void) {
-    for (int gen = 0; gen < AdjointDimension; gen++) {
-      Matrix ta;
-      generator(gen, ta);
-      std::cout << GridLogMessage << "Nc = " << ncolour << " t_" << gen
-                << std::endl;
-      std::cout << GridLogMessage << ta << std::endl;
-    }
-  }
-
-
-
-  static void testGenerators(void) {
-    Matrix ta;
-    Matrix tb;
-    std::cout << GridLogMessage
-              << "Fundamental - Checking trace ta tb is 0.5 delta_ab"
-              << std::endl;
-    for (int a = 0; a < AdjointDimension; a++) {
-      for (int b = 0; b < AdjointDimension; b++) {
-        generator(a, ta);
-        generator(b, tb);
-        Complex tr = TensorRemove(trace(ta * tb));
-        std::cout << GridLogMessage << "(" << a << "," << b << ") =  " << tr
-                  << std::endl;
-        if (a == b) assert(abs(tr - Complex(0.5)) < 1.0e-6);
-        if (a != b) assert(abs(tr) < 1.0e-6);
-      }
-      std::cout << GridLogMessage << std::endl;
-    }
-    std::cout << GridLogMessage << "Fundamental - Checking if hermitian"
-              << std::endl;
-    for (int a = 0; a < AdjointDimension; a++) {
-      generator(a, ta);
-      std::cout << GridLogMessage << a << std::endl;
-      assert(norm2(ta - adj(ta)) < 1.0e-6);
-    }
-    std::cout << GridLogMessage << std::endl;
-
-    std::cout << GridLogMessage << "Fundamental - Checking if traceless"
-              << std::endl;
-    for (int a = 0; a < AdjointDimension; a++) {
-      generator(a, ta);
-      Complex tr = TensorRemove(trace(ta));
-      std::cout << GridLogMessage << a << " " << std::endl;
-      assert(abs(tr) < 1.0e-6);
-    }
-    std::cout << GridLogMessage << std::endl;
-  }
-
-  // reunitarise??
-  template <typename LatticeMatrixType>
-  static void LieRandomize(GridParallelRNG &pRNG, LatticeMatrixType &out, double scale = 1.0) 
-  {
-    GridBase *grid = out.Grid();
-
-    typedef typename LatticeMatrixType::vector_type vector_type;
-
-    typedef iSinglet<vector_type> vTComplexType;
-
-    typedef Lattice<vTComplexType> LatticeComplexType;
-    typedef typename GridTypeMapper<typename LatticeMatrixType::vector_object>::scalar_object MatrixType;
-
-    LatticeComplexType ca(grid);
-    LatticeMatrixType lie(grid);
-    LatticeMatrixType la(grid);
-    ComplexD ci(0.0, scale);
-    //    ComplexD cone(1.0, 0.0);
-    MatrixType ta;
-
-    lie = Zero();
-
-    for (int a = 0; a < AdjointDimension; a++) {
-      random(pRNG, ca);
-
-      ca = (ca + conjugate(ca)) * 0.5;
-      ca = ca - 0.5;
-
-      generator(a, ta);
-
-      la = ci * ca * ta;
-
-      lie = lie + la;  // e^{i la ta}
-
-    }
-    taExp(lie, out);
-  }
-
-  static void GaussianFundamentalLieAlgebraMatrix(GridParallelRNG &pRNG,
-                                                  LatticeMatrix &out,
-                                                  Real scale = 1.0) {
-    GridBase *grid = out.Grid();
-    LatticeReal ca(grid);
-    LatticeMatrix la(grid);
-    Complex ci(0.0, scale);
-    Matrix ta;
-
-    out = Zero();
-    for (int a = 0; a < AdjointDimension; a++) {
-      gaussian(pRNG, ca);
-      generator(a, ta);
-      la = toComplex(ca) * ta;
-      out += la;
-    }
-    out *= ci;
-  }
-
-  static void FundamentalLieAlgebraMatrix(const LatticeAlgebraVector &h,
-                                          LatticeMatrix &out,
-                                          Real scale = 1.0) {
-    conformable(h, out);
-    GridBase *grid = out.Grid();
-    LatticeMatrix la(grid);
-    Matrix ta;
-
-    out = Zero();
-    for (int a = 0; a < AdjointDimension; a++) {
-      generator(a, ta);
-      la = peekColour(h, a) * timesI(ta) * scale;
-      out += la;
-    }
-  }
-/*
- * Fundamental rep gauge xform
- */
-  template<typename Fundamental,typename GaugeMat>
-  static void GaugeTransformFundamental( Fundamental &ferm, GaugeMat &g){
-    GridBase *grid = ferm._grid;
-    conformable(grid,g._grid);
-    ferm = g*ferm;
-  }
-/*
- * Adjoint rep gauge xform
- */
-
-  template<typename Gimpl>
-  static void GaugeTransform(typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
-    GridBase *grid = Umu.Grid();
-    conformable(grid,g.Grid());
-
-    typename Gimpl::GaugeLinkField U(grid);
-    typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
-
-    for(int mu=0;mu<Nd;mu++){
-      U= PeekIndex<LorentzIndex>(Umu,mu);
-      U = g*U*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
-      PokeIndex<LorentzIndex>(Umu,U,mu);
-    }
-  }
-  template<typename Gimpl>
-  static void GaugeTransform( std::vector<typename Gimpl::GaugeLinkField> &U, typename Gimpl::GaugeLinkField &g){
-    GridBase *grid = g.Grid();
-    typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
-    for(int mu=0;mu<Nd;mu++){
-      U[mu] = g*U[mu]*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
-    }
-  }
-  template<typename Gimpl>
-  static void RandomGaugeTransform(GridParallelRNG &pRNG, typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
-    LieRandomize(pRNG,g,1.0);
-    GaugeTransform<Gimpl>(Umu,g);
-  }
-
-  // Projects the algebra components a lattice matrix (of dimension ncol*ncol -1 )
-  // inverse operation: FundamentalLieAlgebraMatrix
-  static void projectOnAlgebra(LatticeAlgebraVector &h_out, const LatticeMatrix &in, Real scale = 1.0) {
-    conformable(h_out, in);
-    h_out = Zero();
-    Matrix Ta;
-
-    for (int a = 0; a < AdjointDimension; a++) {
-      generator(a, Ta);
-      pokeColour(h_out, - 2.0 * (trace(timesI(Ta) * in)) * scale, a);
-    }
-  }
-
-  template <typename GaugeField>
-  static void HotConfiguration(GridParallelRNG &pRNG, GaugeField &out) {
-    typedef typename GaugeField::vector_type vector_type;
-    typedef iSUnMatrix<vector_type> vMatrixType;
-    typedef Lattice<vMatrixType> LatticeMatrixType;
-
-    LatticeMatrixType Umu(out.Grid());
-    LatticeMatrixType tmp(out.Grid());
-    for (int mu = 0; mu < Nd; mu++) {
-      //      LieRandomize(pRNG, Umu, 1.0);
-      //      PokeIndex<LorentzIndex>(out, Umu, mu);
-      gaussian(pRNG,Umu);
-      tmp = Ta(Umu);
-      taExp(tmp,Umu);
-      ProjectSUn(Umu);
-      PokeIndex<LorentzIndex>(out, Umu, mu);
-    }
-  }
-  template<typename GaugeField>
-  static void TepidConfiguration(GridParallelRNG &pRNG,GaugeField &out){
-    typedef typename GaugeField::vector_type vector_type;
-    typedef iSUnMatrix<vector_type> vMatrixType;
-    typedef Lattice<vMatrixType> LatticeMatrixType;
-
-    LatticeMatrixType Umu(out.Grid());
-    for(int mu=0;mu<Nd;mu++){
-      LieRandomize(pRNG,Umu,0.01);
-      PokeIndex<LorentzIndex>(out,Umu,mu);
-    }
-  }
-  template<typename GaugeField>
-  static void ColdConfiguration(GaugeField &out){
-    typedef typename GaugeField::vector_type vector_type;
-    typedef iSUnMatrix<vector_type> vMatrixType;
-    typedef Lattice<vMatrixType> LatticeMatrixType;
-
-    LatticeMatrixType Umu(out.Grid());
-    Umu=1.0;
-    for(int mu=0;mu<Nd;mu++){
-      PokeIndex<LorentzIndex>(out,Umu,mu);
-    }
-  }
-  template<typename GaugeField>
-  static void ColdConfiguration(GridParallelRNG &pRNG,GaugeField &out){
-    ColdConfiguration(out);
-  }
-
-  template<typename LatticeMatrixType>
-  static void taProj( const LatticeMatrixType &in,  LatticeMatrixType &out){
-    out = Ta(in);
-  }
-  template <typename LatticeMatrixType>
-  static void taExp(const LatticeMatrixType &x, LatticeMatrixType &ex) {
-    typedef typename LatticeMatrixType::scalar_type ComplexType;
-
-    LatticeMatrixType xn(x.Grid());
-    RealD nfac = 1.0;
-
-    xn = x;
-    ex = xn + ComplexType(1.0);  // 1+x
-
-    // Do a 12th order exponentiation
-    for (int i = 2; i <= 12; ++i) {
-      nfac = nfac / RealD(i);  // 1/2, 1/2.3 ...
-      xn = xn * x;             // x2, x3,x4....
-      ex = ex + xn * nfac;     // x2/2!, x3/3!....
-    }
-  }
-};
-
-template<int N>
-Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
-{
-  GridBase *grid=Umu.Grid();
-  auto lvol = grid->lSites();
-  Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
-  
-  autoView(Umu_v,Umu,CpuRead);
-  autoView(ret_v,ret,CpuWrite);
-  thread_for(site,lvol,{
-    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
-    Coordinate lcoor;
-    grid->LocalIndexToLocalCoor(site, lcoor);
-    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
-    iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
-    peekLocalSite(Us, Umu_v, lcoor);
-    for(int i=0;i<N;i++){
-      for(int j=0;j<N;j++){
-	EigenU(i,j) = Us()()(i,j);
-      }}
-    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
-    for(int i=0;i<N;i++){
-      for(int j=0;j<N;j++){
-	Ui()()(i,j) = EigenUinv(i,j);
-      }}
-    pokeLocalSite(Ui,ret_v,lcoor);
-  });
-  return ret;
-}
-// Explicit specialisation for SU(3).
-// Explicit specialisation for SU(3).
-static void
-ProjectSU3 (Lattice<iScalar<iScalar<iMatrix<vComplexD, 3> > > > &Umu)
-{
-  GridBase *grid=Umu.Grid();
-  const int x=0;
-  const int y=1;
-  const int z=2;
-  // Reunitarise
-  Umu = ProjectOnGroup(Umu);
-  autoView(Umu_v,Umu,CpuWrite);
-  thread_for(ss,grid->oSites(),{
-      auto cm = Umu_v[ss];
-      cm()()(2,x) = adj(cm()()(0,y)*cm()()(1,z)-cm()()(0,z)*cm()()(1,y)); //x= yz-zy
-      cm()()(2,y) = adj(cm()()(0,z)*cm()()(1,x)-cm()()(0,x)*cm()()(1,z)); //y= zx-xz
-      cm()()(2,z) = adj(cm()()(0,x)*cm()()(1,y)-cm()()(0,y)*cm()()(1,x)); //z= xy-yx
-      Umu_v[ss]=cm;
-  });
-}
-static void ProjectSU3(Lattice<iVector<iScalar<iMatrix<vComplexD, 3> >,Nd> > &U)
-{
-  GridBase *grid=U.Grid();
-  // Reunitarise
-  for(int mu=0;mu<Nd;mu++){
-    auto Umu = PeekIndex<LorentzIndex>(U,mu);
-    Umu      = ProjectOnGroup(Umu);
-    ProjectSU3(Umu);
-    PokeIndex<LorentzIndex>(U,Umu,mu);
-  }
-}
-
-typedef SU<2> SU2;
-typedef SU<3> SU3;
-typedef SU<4> SU4;
-typedef SU<5> SU5;
-
-
-typedef SU<Nc> FundamentalMatrices;
-
-NAMESPACE_END(Grid);
-#endif

Grid/qcd/utils/SUn.impl.h

@@ -0,0 +1,578 @@
+// This file is #included into the body of the class template definition of
+// GaugeGroup. So, image there to be
+//
+// template <int ncolour, class group_name>
+// class GaugeGroup {
+//
+// around it.
+//
+// Please note that the unconventional file extension makes sure that it
+// doesn't get found by the scripts/filelist during bootstrapping.
+
+private:
+template <ONLY_IF_SU>
+static int su2subgroups(GroupName::SU) { return (ncolour * (ncolour - 1)) / 2; }
+////////////////////////////////////////////////////////////////////////
+// There are N^2-1 generators for SU(N).
+//
+// We take a traceless hermitian generator basis as follows
+//
+// * Normalisation: trace ta tb = 1/2 delta_ab = T_F delta_ab
+//   T_F = 1/2  for SU(N) groups
+//
+// * Off diagonal
+//    - pairs of rows i1,i2 behaving like pauli matrices signma_x, sigma_y
+//
+//    - there are (Nc-1-i1) slots for i2 on each row [ x  0  x ]
+//      direct count off each row
+//
+//    - Sum of all pairs is Nc(Nc-1)/2: proof arithmetic series
+//
+//      (Nc-1) + (Nc-2)+...  1      ==> Nc*(Nc-1)/2
+//      1+ 2+          +   + Nc-1
+//
+//    - There are 2 x Nc (Nc-1)/ 2 of these = Nc^2 - Nc
+//
+//    - We enumerate the row-col pairs.
+//    - for each row col pair there is a (sigma_x) and a (sigma_y) like
+//    generator
+//
+//
+//   t^a_ij = { in 0.. Nc(Nc-1)/2 -1} =>  1/2(delta_{i,i1} delta_{j,i2} +
+//   delta_{i,i1} delta_{j,i2})
+//   t^a_ij = { in Nc(Nc-1)/2 ... Nc(Nc-1) - 1} =>  i/2( delta_{i,i1}
+//   delta_{j,i2} - i delta_{i,i1} delta_{j,i2})
+//
+// * Diagonal; must be traceless and normalised
+//   - Sequence is
+//   N  (1,-1,0,0...)
+//   N  (1, 1,-2,0...)
+//   N  (1, 1, 1,-3,0...)
+//   N  (1, 1, 1, 1,-4,0...)
+//
+//   where 1/2 = N^2 (1+.. m^2)etc.... for the m-th diagonal generator
+//   NB this gives the famous SU3 result for su2 index 8
+//
+//   N= sqrt(1/2 . 1/6 ) = 1/2 . 1/sqrt(3)
+//
+//   ( 1      )
+//   (    1   ) / sqrt(3) /2  = 1/2 lambda_8
+//   (      -2)
+//
+////////////////////////////////////////////////////////////////////////
+template <class cplx, ONLY_IF_SU>
+static void generator(int lieIndex, iGroupMatrix<cplx> &ta, GroupName::SU) {
+  // map lie index to which type of generator
+  int diagIndex;
+  int su2Index;
+  int sigxy;
+  int NNm1 = ncolour * (ncolour - 1);
+  if (lieIndex >= NNm1) {
+    diagIndex = lieIndex - NNm1;
+    generatorDiagonal(diagIndex, ta);
+    return;
+  }
+  sigxy = lieIndex & 0x1;  // even or odd
+  su2Index = lieIndex >> 1;
+  if (sigxy)
+    generatorSigmaY(su2Index, ta);
+  else
+    generatorSigmaX(su2Index, ta);
+}
+
+template <class cplx, ONLY_IF_SU>
+static void generatorSigmaY(int su2Index, iGroupMatrix<cplx> &ta) {
+  ta = Zero();
+  int i1, i2;
+  su2SubGroupIndex(i1, i2, su2Index);
+  ta()()(i1, i2) = 1.0;
+  ta()()(i2, i1) = 1.0;
+  ta = ta * 0.5;
+}
+
+template <class cplx, ONLY_IF_SU>
+static void generatorSigmaX(int su2Index, iGroupMatrix<cplx> &ta) {
+  ta = Zero();
+  cplx i(0.0, 1.0);
+  int i1, i2;
+  su2SubGroupIndex(i1, i2, su2Index);
+  ta()()(i1, i2) = i;
+  ta()()(i2, i1) = -i;
+  ta = ta * 0.5;
+}
+
+template <class cplx, ONLY_IF_SU>
+static void generatorDiagonal(int diagIndex, iGroupMatrix<cplx> &ta) {
+  // diag ({1, 1, ..., 1}(k-times), -k, 0, 0, ...)
+  ta = Zero();
+  int k = diagIndex + 1;                  // diagIndex starts from 0
+  for (int i = 0; i <= diagIndex; i++) {  // k iterations
+    ta()()(i, i) = 1.0;
+  }
+  ta()()(k, k) = -k;  // indexing starts from 0
+  RealD nrm = 1.0 / std::sqrt(2.0 * k * (k + 1));
+  ta = ta * nrm;
+}
+
+////////////////////////////////////////////////////////////////////////
+// Map a su2 subgroup number to the pair of rows that are non zero
+////////////////////////////////////////////////////////////////////////
+static void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::SU) {
+  assert((su2_index >= 0) && (su2_index < (ncolour * (ncolour - 1)) / 2));
+
+  int spare = su2_index;
+  for (i1 = 0; spare >= (ncolour - 1 - i1); i1++) {
+    spare = spare - (ncolour - 1 - i1);  // remove the Nc-1-i1 terms
+  }
+  i2 = i1 + 1 + spare;
+}
+
+public:
+//////////////////////////////////////////////////////////////////////////////////////////
+// Pull out a subgroup and project on to real coeffs x pauli basis
+//////////////////////////////////////////////////////////////////////////////////////////
+template <class vcplx, ONLY_IF_SU>
+static void su2Extract(Lattice<iSinglet<vcplx> > &Determinant,
+                       Lattice<iSU2Matrix<vcplx> > &subgroup,
+                       const Lattice<iGroupMatrix<vcplx> > &source,
+                       int su2_index) {
+  GridBase *grid(source.Grid());
+  conformable(subgroup, source);
+  conformable(subgroup, Determinant);
+  int i0, i1;
+  su2SubGroupIndex(i0, i1, su2_index);
+
+  autoView(subgroup_v, subgroup, AcceleratorWrite);
+  autoView(source_v, source, AcceleratorRead);
+  autoView(Determinant_v, Determinant, AcceleratorWrite);
+  accelerator_for(ss, grid->oSites(), 1, {
+    subgroup_v[ss]()()(0, 0) = source_v[ss]()()(i0, i0);
+    subgroup_v[ss]()()(0, 1) = source_v[ss]()()(i0, i1);
+    subgroup_v[ss]()()(1, 0) = source_v[ss]()()(i1, i0);
+    subgroup_v[ss]()()(1, 1) = source_v[ss]()()(i1, i1);
+
+    iSU2Matrix<vcplx> Sigma = subgroup_v[ss];
+
+    Sigma = Sigma - adj(Sigma) + trace(adj(Sigma));
+
+    subgroup_v[ss] = Sigma;
+
+    // this should be purely real
+    Determinant_v[ss] =
+        Sigma()()(0, 0) * Sigma()()(1, 1) - Sigma()()(0, 1) * Sigma()()(1, 0);
+  });
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// Set matrix to one and insert a pauli subgroup
+//////////////////////////////////////////////////////////////////////////////////////////
+template <class vcplx, ONLY_IF_SU>
+static void su2Insert(const Lattice<iSU2Matrix<vcplx> > &subgroup,
+                      Lattice<iGroupMatrix<vcplx> > &dest, int su2_index) {
+  GridBase *grid(dest.Grid());
+  conformable(subgroup, dest);
+  int i0, i1;
+  su2SubGroupIndex(i0, i1, su2_index);
+
+  dest = 1.0;  // start out with identity
+  autoView(dest_v, dest, AcceleratorWrite);
+  autoView(subgroup_v, subgroup, AcceleratorRead);
+  accelerator_for(ss, grid->oSites(), 1, {
+    dest_v[ss]()()(i0, i0) = subgroup_v[ss]()()(0, 0);
+    dest_v[ss]()()(i0, i1) = subgroup_v[ss]()()(0, 1);
+    dest_v[ss]()()(i1, i0) = subgroup_v[ss]()()(1, 0);
+    dest_v[ss]()()(i1, i1) = subgroup_v[ss]()()(1, 1);
+  });
+}
+
+///////////////////////////////////////////////
+// Generate e^{ Re Tr Staple Link} dlink
+//
+// *** Note Staple should be appropriate linear compbination between all
+// staples.
+// *** If already by beta pass coefficient 1.0.
+// *** This routine applies the additional 1/Nc factor that comes after trace
+// in action.
+//
+///////////////////////////////////////////////
+template <ONLY_IF_SU>
+static void SubGroupHeatBath(
+    GridSerialRNG &sRNG, GridParallelRNG &pRNG,
+    RealD beta,  // coeff multiplying staple in action (with no 1/Nc)
+    LatticeMatrix &link,
+    const LatticeMatrix &barestaple,  // multiplied by action coeffs so th
+    int su2_subgroup, int nheatbath, LatticeInteger &wheremask) {
+  GridBase *grid = link.Grid();
+
+  const RealD twopi = 2.0 * M_PI;
+
+  LatticeMatrix staple(grid);
+
+  staple = barestaple * (beta / ncolour);
+
+  LatticeMatrix V(grid);
+  V = link * staple;
+
+  // Subgroup manipulation in the lie algebra space
+  LatticeSU2Matrix u(
+      grid);  // Kennedy pendleton "u" real projected normalised Sigma
+  LatticeSU2Matrix uinv(grid);
+  LatticeSU2Matrix ua(grid);  // a in pauli form
+  LatticeSU2Matrix b(grid);   // rotated matrix after hb
+
+  // Some handy constant fields
+  LatticeComplex ones(grid);
+  ones = 1.0;
+  LatticeComplex zeros(grid);
+  zeros = Zero();
+  LatticeReal rones(grid);
+  rones = 1.0;
+  LatticeReal rzeros(grid);
+  rzeros = Zero();
+  LatticeComplex udet(grid);  // determinant of real(staple)
+  LatticeInteger mask_true(grid);
+  mask_true = 1;
+  LatticeInteger mask_false(grid);
+  mask_false = 0;
+
+  /*
+    PLB 156 P393 (1985) (Kennedy and Pendleton)
+
+    Note: absorb "beta" into the def of sigma compared to KP paper; staple
+    passed to this routine has "beta" already multiplied in
+
+    Action linear in links h and of form:
+
+    beta S = beta  Sum_p (1 - 1/Nc Re Tr Plaq )
+
+    Writing Sigma = 1/Nc (beta Sigma') where sum over staples is "Sigma' "
+
+    beta S = const - beta/Nc Re Tr h Sigma'
+    = const - Re Tr h Sigma
+
+    Decompose h and Sigma into (1, sigma_j) ; h_i real, h^2=1, Sigma_i complex
+    arbitrary.
+
+    Tr h Sigma = h_i Sigma_j Tr (sigma_i sigma_j)  = h_i Sigma_j 2 delta_ij
+    Re Tr h Sigma = 2 h_j Re Sigma_j
+
+    Normalised re Sigma_j = xi u_j
+
+    With u_j a unit vector and U can be in SU(2);
+
+    Re Tr h Sigma = 2 h_j Re Sigma_j = 2 xi (h.u)
+
+    4xi^2 = Det [ Sig - Sig^dag  + 1 Tr Sigdag]
+    u   = 1/2xi [ Sig - Sig^dag  + 1 Tr Sigdag]
+
+    xi = sqrt(Det)/2;
+
+    Write a= u h in SU(2); a has pauli decomp a_j;
+
+    Note: Product b' xi is unvariant because scaling Sigma leaves
+    normalised vector "u" fixed; Can rescale Sigma so b' = 1.
+  */
+
+  ////////////////////////////////////////////////////////
+  // Real part of Pauli decomposition
+  // Note a subgroup can project to zero in cold start
+  ////////////////////////////////////////////////////////
+  su2Extract(udet, u, V, su2_subgroup);
+
+  //////////////////////////////////////////////////////
+  // Normalising this vector if possible; else identity
+  //////////////////////////////////////////////////////
+  LatticeComplex xi(grid);
+
+  LatticeSU2Matrix lident(grid);
+
+  SU2Matrix ident = Complex(1.0);
+  SU2Matrix pauli1;
+  GaugeGroup<2, GroupName::SU>::generator(0, pauli1);
+  SU2Matrix pauli2;
+  GaugeGroup<2, GroupName::SU>::generator(1, pauli2);
+  SU2Matrix pauli3;
+  GaugeGroup<2, GroupName::SU>::generator(2, pauli3);
+  pauli1 = timesI(pauli1) * 2.0;
+  pauli2 = timesI(pauli2) * 2.0;
+  pauli3 = timesI(pauli3) * 2.0;
+
+  LatticeComplex cone(grid);
+  LatticeReal adet(grid);
+  adet = abs(toReal(udet));
+  lident = Complex(1.0);
+  cone = Complex(1.0);
+  Real machine_epsilon = 1.0e-7;
+  u = where(adet > machine_epsilon, u, lident);
+  udet = where(adet > machine_epsilon, udet, cone);
+
+  xi = 0.5 * sqrt(udet);        // 4xi^2 = Det [ Sig - Sig^dag  + 1 Tr Sigdag]
+  u = 0.5 * u * pow(xi, -1.0);  //  u   = 1/2xi [ Sig - Sig^dag  + 1 Tr Sigdag]
+
+  // Debug test for sanity
+  uinv = adj(u);
+  b = u * uinv - 1.0;
+  assert(norm2(b) < 1.0e-4);
+
+  /*
+    Measure: Haar measure dh has d^4a delta(1-|a^2|)
+    In polars:
+    da = da0 r^2 sin theta dr dtheta dphi delta( 1 - r^2 -a0^2)
+    = da0 r^2 sin theta dr dtheta dphi delta( (sqrt(1-a0^) - r)(sqrt(1-a0^) +
+    r) )
+    = da0 r/2 sin theta dr dtheta dphi delta( (sqrt(1-a0^) - r) )
+
+    Action factor Q(h) dh  = e^-S[h]  dh =  e^{  xi Tr uh} dh    // beta
+    enters through xi =  e^{2 xi (h.u)} dh =  e^{2 xi h0u0}.e^{2 xi h1u1}.e^{2
+    xi h2u2}.e^{2 xi h3u3} dh
+
+    Therefore for each site, take xi for that site
+    i) generate  |a0|<1 with dist
+    (1-a0^2)^0.5 e^{2 xi a0 } da0
+
+    Take alpha = 2 xi  = 2 xi [ recall 2 beta/Nc unmod staple norm];
+    hence 2.0/Nc factor in Chroma ] A. Generate two uniformly distributed
+    pseudo-random numbers R and R', R'', R''' in the unit interval; B. Set X =
+    -(ln R)/alpha, X' =-(ln R')/alpha; C. Set C = cos^2(2pi R"), with R"
+    another uniform random number in [0,1] ; D. Set A = XC; E. Let d  = X'+A;
+    F. If R'''^2 :> 1 - 0.5 d,  go back to A;
+    G. Set a0 = 1 - d;
+
+    Note that in step D setting B ~ X - A and using B in place of A in step E
+    will generate a second independent a 0 value.
+  */
+
+  /////////////////////////////////////////////////////////
+  // count the number of sites by picking "1"'s out of hat
+  /////////////////////////////////////////////////////////
+  Integer hit = 0;
+  LatticeReal rtmp(grid);
+  rtmp = where(wheremask, rones, rzeros);
+  RealD numSites = sum(rtmp);
+  RealD numAccepted;
+  LatticeInteger Accepted(grid);
+  Accepted = Zero();
+  LatticeInteger newlyAccepted(grid);
+
+  std::vector<LatticeReal> xr(4, grid);
+  std::vector<LatticeReal> a(4, grid);
+  LatticeReal d(grid);
+  d = Zero();
+  LatticeReal alpha(grid);
+
+  //    std::cout<<GridLogMessage<<"xi "<<xi <<std::endl;
+  xi = 2.0 * xi;
+  alpha = toReal(xi);
+
+  do {
+    // A. Generate two uniformly distributed pseudo-random numbers R and R',
+    // R'', R''' in the unit interval;
+    random(pRNG, xr[0]);
+    random(pRNG, xr[1]);
+    random(pRNG, xr[2]);
+    random(pRNG, xr[3]);
+
+    // B. Set X = - ln R/alpha, X' = -ln R'/alpha
+    xr[1] = -log(xr[1]) / alpha;
+    xr[2] = -log(xr[2]) / alpha;
+
+    // C. Set C = cos^2(2piR'')
+    xr[3] = cos(xr[3] * twopi);
+    xr[3] = xr[3] * xr[3];
+
+    LatticeReal xrsq(grid);
+
+    // D. Set A = XC;
+    // E. Let d  = X'+A;
+    xrsq = xr[2] + xr[1] * xr[3];
+
+    d = where(Accepted, d, xr[2] + xr[1] * xr[3]);
+
+    // F. If R'''^2 :> 1 - 0.5 d,  go back to A;
+    LatticeReal thresh(grid);
+    thresh = 1.0 - d * 0.5;
+    xrsq = xr[0] * xr[0];
+    LatticeInteger ione(grid);
+    ione = 1;
+    LatticeInteger izero(grid);
+    izero = Zero();
+
+    newlyAccepted = where(xrsq < thresh, ione, izero);
+    Accepted = where(newlyAccepted, newlyAccepted, Accepted);
+    Accepted = where(wheremask, Accepted, izero);
+
+    // FIXME need an iSum for integer to avoid overload on return type??
+    rtmp = where(Accepted, rones, rzeros);
+    numAccepted = sum(rtmp);
+
+    hit++;
+
+  } while ((numAccepted < numSites) && (hit < nheatbath));
+
+  // G. Set a0 = 1 - d;
+  a[0] = Zero();
+  a[0] = where(wheremask, 1.0 - d, a[0]);
+
+  //////////////////////////////////////////
+  //    ii) generate a_i uniform on two sphere radius (1-a0^2)^0.5
+  //////////////////////////////////////////
+
+  LatticeReal a123mag(grid);
+  a123mag = sqrt(abs(1.0 - a[0] * a[0]));
+
+  LatticeReal cos_theta(grid);
+  LatticeReal sin_theta(grid);
+  LatticeReal phi(grid);
+
+  random(pRNG, phi);
+  phi = phi * twopi;  // uniform in [0,2pi]
+  random(pRNG, cos_theta);
+  cos_theta = (cos_theta * 2.0) - 1.0;  // uniform in [-1,1]
+  sin_theta = sqrt(abs(1.0 - cos_theta * cos_theta));
+
+  a[1] = a123mag * sin_theta * cos(phi);
+  a[2] = a123mag * sin_theta * sin(phi);
+  a[3] = a123mag * cos_theta;
+
+  ua = toComplex(a[0]) * ident + toComplex(a[1]) * pauli1 +
+       toComplex(a[2]) * pauli2 + toComplex(a[3]) * pauli3;
+
+  b = 1.0;
+  b = where(wheremask, uinv * ua, b);
+  su2Insert(b, V, su2_subgroup);
+
+  // mask the assignment back based on Accptance
+  link = where(Accepted, V * link, link);
+
+  //////////////////////////////
+  // Debug Checks
+  // SU2 check
+  LatticeSU2Matrix check(grid);  // rotated matrix after hb
+  u = Zero();
+  check = ua * adj(ua) - 1.0;
+  check = where(Accepted, check, u);
+  assert(norm2(check) < 1.0e-4);
+
+  check = b * adj(b) - 1.0;
+  check = where(Accepted, check, u);
+  assert(norm2(check) < 1.0e-4);
+
+  LatticeMatrix Vcheck(grid);
+  Vcheck = Zero();
+  Vcheck = where(Accepted, V * adj(V) - 1.0, Vcheck);
+  //    std::cout<<GridLogMessage << "SU3 check " <<norm2(Vcheck)<<std::endl;
+  assert(norm2(Vcheck) < 1.0e-4);
+
+  // Verify the link stays in SU(3)
+  //    std::cout<<GridLogMessage <<"Checking the modified link"<<std::endl;
+  Vcheck = link * adj(link) - 1.0;
+  assert(norm2(Vcheck) < 1.0e-4);
+  /////////////////////////////////
+}
+
+template <ONLY_IF_SU>
+static void testGenerators(GroupName::SU) {
+  Matrix ta;
+  Matrix tb;
+  std::cout << GridLogMessage
+            << "Fundamental - Checking trace ta tb is 0.5 delta_ab"
+            << std::endl;
+  for (int a = 0; a < AdjointDimension; a++) {
+    for (int b = 0; b < AdjointDimension; b++) {
+      generator(a, ta);
+      generator(b, tb);
+      Complex tr = TensorRemove(trace(ta * tb));
+      std::cout << GridLogMessage << "(" << a << "," << b << ") =  " << tr
+                << std::endl;
+      if (a == b) assert(abs(tr - Complex(0.5)) < 1.0e-6);
+      if (a != b) assert(abs(tr) < 1.0e-6);
+    }
+    std::cout << GridLogMessage << std::endl;
+  }
+  std::cout << GridLogMessage << "Fundamental - Checking if hermitian"
+            << std::endl;
+  for (int a = 0; a < AdjointDimension; a++) {
+    generator(a, ta);
+    std::cout << GridLogMessage << a << std::endl;
+    assert(norm2(ta - adj(ta)) < 1.0e-6);
+  }
+  std::cout << GridLogMessage << std::endl;
+
+  std::cout << GridLogMessage << "Fundamental - Checking if traceless"
+            << std::endl;
+  for (int a = 0; a < AdjointDimension; a++) {
+    generator(a, ta);
+    Complex tr = TensorRemove(trace(ta));
+    std::cout << GridLogMessage << a << " " << std::endl;
+    assert(abs(tr) < 1.0e-6);
+  }
+  std::cout << GridLogMessage << std::endl;
+}
+
+
+template <int N, class vtype>
+static Lattice<iScalar<iScalar<iMatrix<vtype, N> > > >
+ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vtype, N> > > > &Umu, GroupName::SU) {
+  return ProjectOnGroup(Umu);
+}
+
+template <class vtype>
+accelerator_inline static iScalar<vtype> ProjectOnGeneralGroup(const iScalar<vtype> &r, GroupName::SU) {
+  return ProjectOnGroup(r);
+}
+
+template <class vtype, int N>
+accelerator_inline static iVector<vtype,N> ProjectOnGeneralGroup(const iVector<vtype,N> &r, GroupName::SU) {
+  return ProjectOnGroup(r);
+}
+
+template <class vtype,int N, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0 >::type * =nullptr>
+accelerator_inline static iMatrix<vtype,N> ProjectOnGeneralGroup(const iMatrix<vtype,N> &arg, GroupName::SU) {
+  return ProjectOnGroup(arg);
+}
+
+template <typename LatticeMatrixType>
+static void taProj(const LatticeMatrixType &in, LatticeMatrixType &out, GroupName::SU) {
+  out = Ta(in);
+}
+
+/*
+ * Fundamental rep gauge xform
+ */
+template<typename Fundamental,typename GaugeMat>
+static void GaugeTransformFundamental( Fundamental &ferm, GaugeMat &g){
+  GridBase *grid = ferm._grid;
+  conformable(grid,g._grid);
+  ferm = g*ferm;
+}
+/*
+ * Adjoint rep gauge xform
+ */
+
+template<typename Gimpl>
+static void GaugeTransform(typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
+  GridBase *grid = Umu.Grid();
+  conformable(grid,g.Grid());
+
+  typename Gimpl::GaugeLinkField U(grid);
+  typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
+
+  for(int mu=0;mu<Nd;mu++){
+    U= PeekIndex<LorentzIndex>(Umu,mu);
+    U = g*U*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
+    PokeIndex<LorentzIndex>(Umu,U,mu);
+  }
+}
+template<typename Gimpl>
+static void GaugeTransform( std::vector<typename Gimpl::GaugeLinkField> &U, typename Gimpl::GaugeLinkField &g){
+  GridBase *grid = g.Grid();
+  typename Gimpl::GaugeLinkField ag(grid); ag = adj(g);
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = g*U[mu]*Gimpl::CshiftLink(ag, mu, 1); //BC-aware
+  }
+}
+template<typename Gimpl>
+static void RandomGaugeTransform(GridParallelRNG &pRNG, typename Gimpl::GaugeField &Umu, typename Gimpl::GaugeLinkField &g){
+  LieRandomize(pRNG,g,1.0);
+  GaugeTransform<Gimpl>(Umu,g);
+}

Grid/qcd/utils/SUnAdjoint.h

@@ -51,6 +51,10 @@ public:
   typedef Lattice<iVector<iScalar<iMatrix<vComplexF, Dimension> >, Nd> > LatticeAdjFieldF;
   typedef Lattice<iVector<iScalar<iMatrix<vComplexD, Dimension> >, Nd> > LatticeAdjFieldD;
 
+
+  template <typename vtype>
+  using iSUnMatrix = iScalar<iScalar<iMatrix<vtype, ncolour> > >;
+
   typedef Lattice<iScalar<iScalar<iVector<vComplex, Dimension> > > >  LatticeAdjVector;
 
   template <class cplx>
@@ -58,8 +62,8 @@ public:
     // returns i(T_Adj)^index necessary for the projectors
     // see definitions above
     iAdjTa = Zero();
-    Vector<typename SU<ncolour>::template iSUnMatrix<cplx> > ta(ncolour * ncolour - 1);
-    typename SU<ncolour>::template iSUnMatrix<cplx> tmp;
+    Vector<iSUnMatrix<cplx> > ta(ncolour * ncolour - 1);
+    iSUnMatrix<cplx> tmp;
 
     // FIXME not very efficient to get all the generators everytime
     for (int a = 0; a < Dimension; a++) SU<ncolour>::generator(a, ta[a]);
@@ -67,8 +71,7 @@ public:
     for (int a = 0; a < Dimension; a++) {
       tmp = ta[a] * ta[Index] - ta[Index] * ta[a];
       for (int b = 0; b < (ncolour * ncolour - 1); b++) {
-        typename SU<ncolour>::template iSUnMatrix<cplx> tmp1 =
-	  2.0 * tmp * ta[b];  // 2.0 from the normalization
+        iSUnMatrix<cplx> tmp1 = 2.0 * tmp * ta[b];  // 2.0 from the normalization
         Complex iTr = TensorRemove(timesI(trace(tmp1)));
         //iAdjTa()()(b, a) = iTr;
         iAdjTa()()(a, b) = iTr;
@@ -134,8 +137,7 @@ public:
 
     for (int a = 0; a < Dimension; a++) {
       generator(a, iTa);
-      LatticeComplex tmp = real(trace(iTa * in)) * coefficient;
-      pokeColour(h_out, tmp, a);
+      pokeColour(h_out, real(trace(iTa * in)) * coefficient, a);
     }
   }
 

Grid/qcd/utils/SUnTwoIndex.h

@@ -1,273 +0,0 @@
-////////////////////////////////////////////////////////////////////////
-//
-// * Two index representation generators
-//
-// * Normalisation for the fundamental generators:
-//   trace ta tb = 1/2 delta_ab = T_F delta_ab
-//   T_F = 1/2  for SU(N) groups
-//
-//
-//   base for NxN two index (anti-symmetric) matrices
-//   normalized to 1 (d_ij is the kroenecker delta)
-//
-//   (e^(ij)_{kl} = 1 / sqrt(2) (d_ik d_jl +/- d_jk d_il)
-//
-//   Then the generators are written as
-//
-//   (iT_a)^(ij)(lk) = i * ( tr[e^(ij)^dag e^(lk) T^trasp_a] +
-//   tr[e^(lk)e^(ij)^dag T_a] )  //
-//   
-//
-////////////////////////////////////////////////////////////////////////
-
-// Authors: David Preti, Guido Cossu
-
-#ifndef QCD_UTIL_SUN2INDEX_H
-#define QCD_UTIL_SUN2INDEX_H
-
-
-NAMESPACE_BEGIN(Grid);
-
-enum TwoIndexSymmetry { Symmetric = 1, AntiSymmetric = -1 };
-
-inline Real delta(int a, int b) { return (a == b) ? 1.0 : 0.0; }
-
-template <int ncolour, TwoIndexSymmetry S>
-class SU_TwoIndex : public SU<ncolour> {
-public:
-  static const int Dimension = ncolour * (ncolour + S) / 2;
-  static const int NumGenerators = SU<ncolour>::AdjointDimension;
-
-  template <typename vtype>
-  using iSUnTwoIndexMatrix = iScalar<iScalar<iMatrix<vtype, Dimension> > >;
-
-  typedef iSUnTwoIndexMatrix<Complex> TIMatrix;
-  typedef iSUnTwoIndexMatrix<ComplexF> TIMatrixF;
-  typedef iSUnTwoIndexMatrix<ComplexD> TIMatrixD;
-
-  typedef iSUnTwoIndexMatrix<vComplex> vTIMatrix;
-  typedef iSUnTwoIndexMatrix<vComplexF> vTIMatrixF;
-  typedef iSUnTwoIndexMatrix<vComplexD> vTIMatrixD;
-
-  typedef Lattice<vTIMatrix> LatticeTwoIndexMatrix;
-  typedef Lattice<vTIMatrixF> LatticeTwoIndexMatrixF;
-  typedef Lattice<vTIMatrixD> LatticeTwoIndexMatrixD;
-
-  typedef Lattice<iVector<iScalar<iMatrix<vComplex, Dimension> >, Nd> >
-  LatticeTwoIndexField;
-  typedef Lattice<iVector<iScalar<iMatrix<vComplexF, Dimension> >, Nd> >
-  LatticeTwoIndexFieldF;
-  typedef Lattice<iVector<iScalar<iMatrix<vComplexD, Dimension> >, Nd> >
-  LatticeTwoIndexFieldD;
-
-  template <typename vtype>
-  using iSUnMatrix = iScalar<iScalar<iMatrix<vtype, ncolour> > >;
-
-  typedef iSUnMatrix<Complex> Matrix;
-  typedef iSUnMatrix<ComplexF> MatrixF;
-  typedef iSUnMatrix<ComplexD> MatrixD;
-
-  template <class cplx>
-  static void base(int Index, iSUnMatrix<cplx> &eij) {
-    // returns (e)^(ij)_{kl} necessary for change of base U_F -> U_R
-    assert(Index < NumGenerators);
-    eij = Zero();
-
-    // for the linearisation of the 2 indexes 
-    static int a[ncolour * (ncolour - 1) / 2][2]; // store the a <-> i,j
-    static bool filled = false;
-    if (!filled) {
-      int counter = 0;
-      for (int i = 1; i < ncolour; i++) {
-        for (int j = 0; j < i; j++) {
-          a[counter][0] = i;
-          a[counter][1] = j;
-          counter++;
-        }
-      }
-      filled = true;
-    }
-
-    if (Index < ncolour * (ncolour - 1) / 2) {
-      baseOffDiagonal(a[Index][0], a[Index][1], eij);
-    } else {
-      baseDiagonal(Index, eij);
-    }
-  }
-
-  template <class cplx>
-  static void baseDiagonal(int Index, iSUnMatrix<cplx> &eij) {
-    eij = Zero();
-    eij()()(Index - ncolour * (ncolour - 1) / 2,
-            Index - ncolour * (ncolour - 1) / 2) = 1.0;
-  }
-
-  template <class cplx>
-  static void baseOffDiagonal(int i, int j, iSUnMatrix<cplx> &eij) {
-    eij = Zero();
-    for (int k = 0; k < ncolour; k++)
-      for (int l = 0; l < ncolour; l++)
-        eij()()(l, k) = delta(i, k) * delta(j, l) +
-	  S * delta(j, k) * delta(i, l);
-
-    RealD nrm = 1. / std::sqrt(2.0);
-    eij = eij * nrm;
-  }
-
-  static void printBase(void) {
-    for (int gen = 0; gen < Dimension; gen++) {
-      Matrix tmp;
-      base(gen, tmp);
-      std::cout << GridLogMessage << "Nc = " << ncolour << " t_" << gen
-                << std::endl;
-      std::cout << GridLogMessage << tmp << std::endl;
-    }
-  }
-
-  template <class cplx>
-  static void generator(int Index, iSUnTwoIndexMatrix<cplx> &i2indTa) {
-    Vector<typename SU<ncolour>::template iSUnMatrix<cplx> > ta(
-								ncolour * ncolour - 1);
-    Vector<typename SU<ncolour>::template iSUnMatrix<cplx> > eij(Dimension);
-    typename SU<ncolour>::template iSUnMatrix<cplx> tmp;
-    i2indTa = Zero();
-    
-    for (int a = 0; a < ncolour * ncolour - 1; a++)
-      SU<ncolour>::generator(a, ta[a]);
-    
-    for (int a = 0; a < Dimension; a++) base(a, eij[a]);
-
-    for (int a = 0; a < Dimension; a++) {
-      tmp = transpose(ta[Index]) * adj(eij[a]) + adj(eij[a]) * ta[Index];
-      for (int b = 0; b < Dimension; b++) {
-        typename SU<ncolour>::template iSUnMatrix<cplx> tmp1 =
-	  tmp * eij[b]; 
-        Complex iTr = TensorRemove(timesI(trace(tmp1)));
-        i2indTa()()(a, b) = iTr;
-      }
-    }
-  }
-
-  static void printGenerators(void) {
-    for (int gen = 0; gen < ncolour * ncolour - 1; gen++) {
-      TIMatrix i2indTa;
-      generator(gen, i2indTa);
-      std::cout << GridLogMessage << "Nc = " << ncolour << " t_" << gen
-                << std::endl;
-      std::cout << GridLogMessage << i2indTa << std::endl;
-    }
-  }
-
-  static void testGenerators(void) {
-    TIMatrix i2indTa, i2indTb;
-    std::cout << GridLogMessage << "2IndexRep - Checking if traceless"
-              << std::endl;
-    for (int a = 0; a < ncolour * ncolour - 1; a++) {
-      generator(a, i2indTa);
-      std::cout << GridLogMessage << a << std::endl;
-      assert(norm2(trace(i2indTa)) < 1.0e-6);
-    }
-    std::cout << GridLogMessage << std::endl;
-
-    std::cout << GridLogMessage << "2IndexRep - Checking if antihermitean"
-              << std::endl;
-    for (int a = 0; a < ncolour * ncolour - 1; a++) {
-      generator(a, i2indTa);
-      std::cout << GridLogMessage << a << std::endl;
-      assert(norm2(adj(i2indTa) + i2indTa) < 1.0e-6);
-    }
-
-    std::cout << GridLogMessage << std::endl;
-    std::cout << GridLogMessage
-              << "2IndexRep - Checking Tr[Ta*Tb]=delta(a,b)*(N +- 2)/2"
-              << std::endl;
-    for (int a = 0; a < ncolour * ncolour - 1; a++) {
-      for (int b = 0; b < ncolour * ncolour - 1; b++) {
-        generator(a, i2indTa);
-        generator(b, i2indTb);
-
-        // generator returns iTa, so we need a minus sign here
-        Complex Tr = -TensorRemove(trace(i2indTa * i2indTb));
-        std::cout << GridLogMessage << "a=" << a << "b=" << b << "Tr=" << Tr
-                  << std::endl;
-      }
-    }
-    std::cout << GridLogMessage << std::endl;
-  }
-
-  static void TwoIndexLieAlgebraMatrix(
-				       const typename SU<ncolour>::LatticeAlgebraVector &h,
-				       LatticeTwoIndexMatrix &out, Real scale = 1.0) {
-    conformable(h, out);
-    GridBase *grid = out.Grid();
-    LatticeTwoIndexMatrix la(grid);
-    TIMatrix i2indTa;
-
-    out = Zero();
-    for (int a = 0; a < ncolour * ncolour - 1; a++) {
-      generator(a, i2indTa);
-      la = peekColour(h, a) * i2indTa;
-      out += la;
-    }
-    out *= scale;
-  }
-
-  // Projects the algebra components 
-  // of a lattice matrix ( of dimension ncol*ncol -1 )
-  static void projectOnAlgebra(
-			       typename SU<ncolour>::LatticeAlgebraVector &h_out,
-			       const LatticeTwoIndexMatrix &in, Real scale = 1.0) {
-    conformable(h_out, in);
-    h_out = Zero();
-    TIMatrix i2indTa;
-    Real coefficient = -2.0 / (ncolour + 2 * S) * scale;
-    // 2/(Nc +/- 2) for the normalization of the trace in the two index rep
-    for (int a = 0; a < ncolour * ncolour - 1; a++) {
-      generator(a, i2indTa);
-      auto tmp = real(trace(i2indTa * in)) * coefficient;
-      pokeColour(h_out, tmp, a);
-    }
-  }
-
-  // a projector that keeps the generators stored to avoid the overhead of
-  // recomputing them
-  static void projector(typename SU<ncolour>::LatticeAlgebraVector &h_out,
-                        const LatticeTwoIndexMatrix &in, Real scale = 1.0) {
-    conformable(h_out, in);
-    // to store the generators
-    static std::vector<TIMatrix> i2indTa(ncolour * ncolour -1); 
-    h_out = Zero();
-    static bool precalculated = false;
-    if (!precalculated) {
-      precalculated = true;
-      for (int a = 0; a < ncolour * ncolour - 1; a++) generator(a, i2indTa[a]);
-    }
-
-    Real coefficient =
-      -2.0 / (ncolour + 2 * S) * scale;  // 2/(Nc +/- 2) for the normalization
-    // of the trace in the two index rep
-
-    for (int a = 0; a < ncolour * ncolour - 1; a++) {
-      auto tmp = real(trace(i2indTa[a] * in)) * coefficient;
-      pokeColour(h_out, tmp, a);
-    }
-  }
-};
-
-// Some useful type names
-typedef SU_TwoIndex<Nc, Symmetric> TwoIndexSymmMatrices;
-typedef SU_TwoIndex<Nc, AntiSymmetric> TwoIndexAntiSymmMatrices;
-
-typedef SU_TwoIndex<2, Symmetric> SU2TwoIndexSymm;
-typedef SU_TwoIndex<3, Symmetric> SU3TwoIndexSymm;
-typedef SU_TwoIndex<4, Symmetric> SU4TwoIndexSymm;
-typedef SU_TwoIndex<5, Symmetric> SU5TwoIndexSymm;
-
-typedef SU_TwoIndex<2, AntiSymmetric> SU2TwoIndexAntiSymm;
-typedef SU_TwoIndex<3, AntiSymmetric> SU3TwoIndexAntiSymm;
-typedef SU_TwoIndex<4, AntiSymmetric> SU4TwoIndexAntiSymm;
-typedef SU_TwoIndex<5, AntiSymmetric> SU5TwoIndexAntiSymm;
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/qcd/utils/Sp2n.impl.h

@@ -0,0 +1,317 @@
+// This file is #included into the body of the class template definition of
+// GaugeGroup. So, image there to be
+//
+// template <int ncolour, class group_name>
+// class GaugeGroup {
+//
+// around it.
+//
+// Please note that the unconventional file extension makes sure that it
+// doesn't get found by the scripts/filelist during bootstrapping.
+
+private:
+template <ONLY_IF_Sp>
+static int su2subgroups(GroupName::Sp) { return (ncolour/2 * (ncolour/2 - 1)) / 2; }
+
+// Sp(2N) has N(2N+1) = 2N^2+N generators
+//
+// normalise the generators such that
+// Trace ( Ta Tb) = 1/2 delta_ab
+//
+// N generators in the cartan, 2N^2 off
+// off diagonal:
+//     there are 6 types named a,b,c,d and w,z
+//     abcd are N(N-1)/2 each while wz are N each
+
+template <class cplx, ONLY_IF_Sp>
+static void generator(int lieIndex, iGroupMatrix<cplx> &ta, GroupName::Sp) {
+  // map lie index into type of generators: diagonal, abcd type, wz type
+
+  const int nsp = ncolour/2;
+  int diagIndex;
+  int aIndex, bIndex, cIndex, dIndex;
+  int wIndex, zIndex;  // a,b,c,d are N(N-1)/2 and w,z are N
+  const int mod = nsp * (nsp - 1) * 0.5;
+  const int offdiag =
+      2 * nsp * nsp;  // number of generators not in the cartan subalgebra
+  const int wmod = 4 * mod;
+  const int zmod = wmod + nsp;
+  if (lieIndex >= offdiag) {
+    diagIndex = lieIndex - offdiag;  // 0, ... ,N-1
+    // std::cout << GridLogMessage << "diag type " << std::endl;
+    generatorDiagtype(diagIndex, ta);
+    return;
+  }
+  if ((lieIndex >= wmod) && (lieIndex < zmod)) {
+    // std::cout << GridLogMessage << "w type " << std::endl;
+    wIndex = lieIndex - wmod;  // 0, ... ,N-1
+    generatorWtype(wIndex, ta);
+    return;
+  }
+  if ((lieIndex >= zmod) && (lieIndex < offdiag)) {
+    // std::cout << GridLogMessage << "z type " << std::endl;
+    // std::cout << GridLogMessage << "lie index " << lieIndex << std::endl;
+    // std::cout << GridLogMessage << "z mod " << zmod << std::endl;
+    zIndex = lieIndex - zmod;  // 0, ... ,N-1
+    generatorZtype(zIndex, ta);
+    return;
+  }
+  if (lieIndex < mod) {  // atype 0, ... , N(N-1)/2=mod
+    // std::cout << GridLogMessage << "a type " << std::endl;
+    aIndex = lieIndex;
+    // std::cout << GridLogMessage << "a indx " << aIndex << std::endl;
+    generatorAtype(aIndex, ta);
+    return;
+  }
+  if ((lieIndex >= mod) && lieIndex < 2 * mod) {  // btype mod, ... , 2mod-1
+    // std::cout << GridLogMessage << "b type " << std::endl;
+    bIndex = lieIndex - mod;
+    generatorBtype(bIndex, ta);
+    return;
+  }
+  if ((lieIndex >= 2 * mod) &&
+      lieIndex < 3 * mod) {  // ctype 2mod, ... , 3mod-1
+    // std::cout << GridLogMessage << "c type " << std::endl;
+    cIndex = lieIndex - 2 * mod;
+    generatorCtype(cIndex, ta);
+    return;
+  }
+  if ((lieIndex >= 3 * mod) &&
+      lieIndex < wmod) {  // ctype 3mod, ... , 4mod-1 = wmod-1
+    // std::cout << GridLogMessage << "d type " << std::endl;
+    dIndex = lieIndex - 3 * mod;
+    generatorDtype(dIndex, ta);
+    return;
+  }
+
+}  // end of generator
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorDiagtype(int diagIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,i) = - ta(i+N,i+N) = 1/2 for each i index of the cartan subalgebra
+
+  const int nsp=ncolour/2;
+  ta = Zero();
+  RealD nrm = 1.0 / 2;
+
+  ta()()(diagIndex, diagIndex) = nrm;
+  ta()()(diagIndex + nsp, diagIndex + nsp) = -nrm;
+}
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorAtype(int aIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,j) = ta(j,i) = -ta(i+N,j+N) = -ta(j+N,i+N) = 1 / 2 sqrt(2)
+  // with i<j and i=0,...,N-2
+  // follows that j=i+1, ... , N
+  int i1, i2;
+  const int nsp=ncolour/2;
+  ta = Zero();
+  RealD nrm = 1 / (2 * std::sqrt(2));
+
+  su2SubGroupIndex(i1, i2, aIndex);
+  ta()()(i1, i2) = 1;
+  ta()()(i2, i1) = 1;
+  ta()()(i1 + nsp, i2 + nsp) = -1;
+  ta()()(i2 + nsp, i1 + nsp) = -1;
+
+  ta = ta * nrm;
+}
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorBtype(int bIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,j) = -ta(j,i) = ta(i+N,j+N) = -ta(j+N,i+N) = i / 1/ 2 sqrt(2)
+  // with i<j and i=0,...,N-2
+  // follows that j=i+1, ... , N-1
+
+  const int nsp=ncolour/2;
+  int i1, i2;
+  ta = Zero();
+  cplx i(0.0, 1.0);
+  RealD nrm = 1 / (2 * std::sqrt(2));
+  su2SubGroupIndex(i1, i2, bIndex);
+
+  ta()()(i1, i2) = i;
+  ta()()(i2, i1) = -i;
+  ta()()(i1 + nsp, i2 + nsp) = i;
+  ta()()(i2 + nsp, i1 + nsp) = -i;
+
+  ta = ta * nrm;
+}
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorCtype(int cIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,j+N) = ta(j,i+N) = ta(i+N,j) = ta(j+N,i) = 1 / 2 sqrt(2)
+
+  const int nsp=ncolour/2;
+  int i1, i2;
+  ta = Zero();
+  RealD nrm = 1 / (2 * std::sqrt(2));
+  su2SubGroupIndex(i1, i2, cIndex);
+
+  ta()()(i1, i2 + nsp) = 1;
+  ta()()(i2, i1 + nsp) = 1;
+  ta()()(i1 + nsp, i2) = 1;
+  ta()()(i2 + nsp, i1) = 1;
+
+  ta = ta * nrm;
+}
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorDtype(int dIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,j+N) = ta(j,i+N) = -ta(i+N,j) = -ta(j+N,i) = i /  2 sqrt(2)
+
+  const int nsp=ncolour/2;
+  int i1, i2;
+  ta = Zero();
+  cplx i(0.0, 1.0);
+  RealD nrm = 1 / (2 * std::sqrt(2));
+  su2SubGroupIndex(i1, i2, dIndex);
+
+  ta()()(i1, i2 + nsp) = i;
+  ta()()(i2, i1 + nsp) = i;
+  ta()()(i1 + nsp, i2) = -i;
+  ta()()(i2 + nsp, i1) = -i;
+
+  ta = ta * nrm;
+}
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorWtype(int wIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,i+N) =  ta(i+N,i) = 1/2
+
+  const int nsp=ncolour/2;
+  ta = Zero();
+  RealD nrm = 1.0 / 2;  // check
+
+  ta()()(wIndex, wIndex + nsp) = 1;
+  ta()()(wIndex + nsp, wIndex) = 1;
+
+  ta = ta * nrm;
+}
+
+template <class cplx, ONLY_IF_Sp>
+static void generatorZtype(int zIndex, iGroupMatrix<cplx> &ta) {
+  // ta(i,i+N) = - ta(i+N,i) = i/2
+
+  const int nsp=ncolour/2;
+  ta = Zero();
+  RealD nrm = 1.0 / 2;  // check
+  cplx i(0.0, 1.0);
+  ta()()(zIndex, zIndex + nsp) = i;
+  ta()()(zIndex + nsp, zIndex) = -i;
+
+  ta = ta * nrm;
+}
+
+////////////////////////////////////////////////////////////////////////
+// Map a su2 subgroup number to the pair of rows that are non zero
+////////////////////////////////////////////////////////////////////////
+template <ONLY_IF_Sp>
+static void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::Sp) {
+  const int nsp=ncolour/2;
+  assert((su2_index >= 0) && (su2_index < (nsp * (nsp - 1)) / 2));
+
+  int spare = su2_index;
+  for (i1 = 0; spare >= (nsp - 1 - i1); i1++) {
+    spare = spare - (nsp - 1 - i1);  // remove the Nc-1-i1 terms
+  }
+  i2 = i1 + 1 + spare;
+}
+
+static void testGenerators(GroupName::Sp) {
+  Matrix ta;
+  Matrix tb;
+  std::cout << GridLogMessage
+            << "Fundamental - Checking trace ta tb is 0.5 delta_ab "
+            << std::endl;
+  for (int a = 0; a < AlgebraDimension; a++) {
+    for (int b = 0; b < AlgebraDimension; b++) {
+      generator(a, ta);
+      generator(b, tb);
+      Complex tr = TensorRemove(trace(ta * tb));
+      std::cout << GridLogMessage << "(" << a << "," << b << ") =  " << tr
+                << std::endl;
+      if (a == b) assert(abs(tr - Complex(0.5)) < 1.0e-6);
+      if (a != b) assert(abs(tr) < 1.0e-6);
+    }
+  }
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "Fundamental - Checking if hermitian"
+            << std::endl;
+  for (int a = 0; a < AlgebraDimension; a++) {
+    generator(a, ta);
+    std::cout << GridLogMessage << a << std::endl;
+    assert(norm2(ta - adj(ta)) < 1.0e-6);
+  }
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "Fundamental - Checking if traceless"
+            << std::endl;
+  for (int a = 0; a < AlgebraDimension; a++) {
+    generator(a, ta);
+    Complex tr = TensorRemove(trace(ta));
+    std::cout << GridLogMessage << a << std::endl;
+    assert(abs(tr) < 1.0e-6);
+  }
+}
+
+template <int N>
+static Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > >
+ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu, GroupName::Sp) {
+  return ProjectOnSpGroup(Umu);
+}
+
+template <class vtype>
+accelerator_inline static iScalar<vtype> ProjectOnGeneralGroup(const iScalar<vtype> &r, GroupName::Sp) {
+  return ProjectOnSpGroup(r);
+}
+
+template <class vtype, int N>
+accelerator_inline static iVector<vtype,N> ProjectOnGeneralGroup(const iVector<vtype,N> &r, GroupName::Sp) {
+  return ProjectOnSpGroup(r);
+}
+
+template <class vtype,int N, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0 >::type * =nullptr>
+accelerator_inline static iMatrix<vtype,N> ProjectOnGeneralGroup(const iMatrix<vtype,N> &arg, GroupName::Sp) {
+  return ProjectOnSpGroup(arg);
+}
+
+template <typename LatticeMatrixType>   
+static void taProj(const LatticeMatrixType &in, LatticeMatrixType &out, GroupName::Sp) {
+  out = SpTa(in);
+}
+
+public:
+
+template <ONLY_IF_Sp>
+static void Omega(LatticeColourMatrixD &in) {
+  const int nsp=ncolour/2;
+  LatticeColourMatrixD OmegaLatt(in.Grid());
+  LatticeColourMatrixD identity(in.Grid());
+  ColourMatrix Omega;
+
+  OmegaLatt = Zero();
+  Omega = Zero();
+  identity = 1.;
+
+  for (int i = 0; i < nsp; i++) {
+    Omega()()(i, nsp + i) = 1.;
+    Omega()()(nsp + i, i) = -1;
+  }
+  OmegaLatt = OmegaLatt + (identity * Omega);
+  in = OmegaLatt;
+}
+
+template <ONLY_IF_Sp, class vtype, int N>
+static void Omega(iScalar<iScalar<iMatrix<vtype, N> > > &in) {
+  const int nsp=ncolour/2;
+    
+  iScalar<iScalar<iMatrix<vtype, N> > > Omega;
+  Omega = Zero();
+
+  for (int i = 0; i < nsp; i++) {
+    Omega()()(i, nsp + i) = 1.;
+    Omega()()(nsp + i, i) = -1;
+  }
+    
+  in = Omega;
+}

Grid/qcd/utils/Utils.h

@@ -8,9 +8,9 @@
 #include <Grid/qcd/utils/ScalarObjs.h>
 
 // Include representations
-#include <Grid/qcd/utils/SUn.h>
+#include <Grid/qcd/utils/GaugeGroup.h>
 #include <Grid/qcd/utils/SUnAdjoint.h>
-#include <Grid/qcd/utils/SUnTwoIndex.h>
+#include <Grid/qcd/utils/GaugeGroupTwoIndex.h>
 
 // All-to-all contraction kernels that touch the 
 // internal lattice structure

Grid/qcd/utils/WilsonLoops.h

@@ -290,7 +290,7 @@ public:
   }
 */
   //////////////////////////////////////////////////
-  // the sum over all staples on each site
+  // the sum over all nu-oriented staples for nu != mu on each site
   //////////////////////////////////////////////////
   static void Staple(GaugeMat &staple, const GaugeLorentz &Umu, int mu) {
 
@@ -300,6 +300,10 @@ public:
     for (int d = 0; d < Nd; d++) {
       U[d] = PeekIndex<LorentzIndex>(Umu, d);
     }
+    Staple(staple, U, mu);
+  }
+
+  static void Staple(GaugeMat &staple, const std::vector<GaugeMat> &U, int mu) {
     staple = Zero();
 
     for (int nu = 0; nu < Nd; nu++) {
@@ -335,6 +339,202 @@ public:
     }
   }
 
+  /////////////
+  //Staples for each direction mu, summed over nu != mu
+  //staple: output staples for each mu (Nd)
+  //U: link array (Nd)
+  /////////////
+  static void StapleAll(std::vector<GaugeMat> &staple, const std::vector<GaugeMat> &U) {
+    assert(staple.size() == Nd); assert(U.size() == Nd);
+    for(int mu=0;mu<Nd;mu++) Staple(staple[mu], U, mu);
+  }
+
+
+  //A workspace class allowing reuse of the stencil
+  class WilsonLoopPaddedStencilWorkspace{
+    std::unique_ptr<GeneralLocalStencil> stencil;
+    size_t nshift;
+
+    void generateStencil(GridBase* padded_grid){
+      double t0 = usecond();
+      
+      //Generate shift arrays
+      std::vector<Coordinate> shifts = this->getShifts();
+      nshift = shifts.size();
+      
+      double t1 = usecond();
+      //Generate local stencil
+      stencil.reset(new GeneralLocalStencil(padded_grid,shifts));
+      double t2 = usecond();
+      std::cout << GridLogPerformance << " WilsonLoopPaddedWorkspace timings: coord:" << (t1-t0)/1000 << "ms, stencil:" << (t2-t1)/1000 << "ms" << std::endl;   
+    }
+  public:
+    //Get the stencil. If not already generated, or if generated using a different Grid than in PaddedCell, it will be created on-the-fly
+    const GeneralLocalStencil & getStencil(const PaddedCell &pcell){
+      assert(pcell.depth >= this->paddingDepth());
+      if(!stencil || stencil->Grid() != (GridBase*)pcell.grids.back() ) generateStencil((GridBase*)pcell.grids.back());
+      return *stencil;
+    }
+    size_t Nshift() const{ return nshift; }
+    
+    virtual std::vector<Coordinate> getShifts() const = 0;
+    virtual int paddingDepth() const = 0; //padding depth required
+    
+    virtual ~WilsonLoopPaddedStencilWorkspace(){}
+  };
+
+  //This workspace allows the sharing of a common PaddedCell object between multiple stencil workspaces
+  class WilsonLoopPaddedWorkspace{
+    std::vector<WilsonLoopPaddedStencilWorkspace*> stencil_wk;
+    std::unique_ptr<PaddedCell> pcell;
+
+    void generatePcell(GridBase* unpadded_grid){
+      assert(stencil_wk.size());
+      int max_depth = 0;
+      for(auto const &s : stencil_wk) max_depth=std::max(max_depth, s->paddingDepth());
+      
+      pcell.reset(new PaddedCell(max_depth, dynamic_cast<GridCartesian*>(unpadded_grid)));
+    }
+    
+  public:
+    //Add a stencil definition. This should be done before the first call to retrieve a stencil object.
+    //Takes ownership of the pointer
+    void addStencil(WilsonLoopPaddedStencilWorkspace *stencil){
+      assert(!pcell);
+      stencil_wk.push_back(stencil);
+    }
+
+    const GeneralLocalStencil & getStencil(const size_t stencil_idx, GridBase* unpadded_grid){
+      if(!pcell || pcell->unpadded_grid != unpadded_grid) generatePcell(unpadded_grid);
+      return stencil_wk[stencil_idx]->getStencil(*pcell);
+    }      
+    const PaddedCell & getPaddedCell(GridBase* unpadded_grid){
+      if(!pcell || pcell->unpadded_grid != unpadded_grid) generatePcell(unpadded_grid);
+      return *pcell;
+    }
+    
+    ~WilsonLoopPaddedWorkspace(){
+      for(auto &s : stencil_wk) delete s;
+    }
+  };
+
+  //A workspace class allowing reuse of the stencil
+  class StaplePaddedAllWorkspace: public WilsonLoopPaddedStencilWorkspace{
+  public:
+    std::vector<Coordinate> getShifts() const override{
+      std::vector<Coordinate> shifts;
+      for(int mu=0;mu<Nd;mu++){
+	for(int nu=0;nu<Nd;nu++){
+	  if(nu != mu){
+	    Coordinate shift_0(Nd,0);
+	    Coordinate shift_mu(Nd,0); shift_mu[mu]=1;
+	    Coordinate shift_nu(Nd,0); shift_nu[nu]=1;
+	    Coordinate shift_mnu(Nd,0); shift_mnu[nu]=-1;
+	    Coordinate shift_mnu_pmu(Nd,0); shift_mnu_pmu[nu]=-1; shift_mnu_pmu[mu]=1;
+      
+	    //U_nu(x+mu)U^dag_mu(x+nu) U^dag_nu(x)
+	    shifts.push_back(shift_0);
+	    shifts.push_back(shift_nu);
+	    shifts.push_back(shift_mu);
+      
+	    //U_nu^dag(x-nu+mu) U_mu^dag(x-nu) U_nu(x-nu)
+	    shifts.push_back(shift_mnu);
+	    shifts.push_back(shift_mnu);
+	    shifts.push_back(shift_mnu_pmu);
+	  }
+	}
+      }
+      return shifts;
+    }
+
+    int paddingDepth() const override{ return 1; }
+  }; 
+
+  //Padded cell implementation of the staple method for all mu, summed over nu != mu
+  //staple: output staple for each mu, summed over nu != mu (Nd)
+  //U_padded: the gauge link fields padded out using the PaddedCell class
+  //Cell: the padded cell class
+  static void StaplePaddedAll(std::vector<GaugeMat> &staple, const std::vector<GaugeMat> &U_padded, const PaddedCell &Cell) {
+    StaplePaddedAllWorkspace wk;
+    StaplePaddedAll(staple,U_padded,Cell,wk.getStencil(Cell));
+  }
+  
+  //Padded cell implementation of the staple method for all mu, summed over nu != mu
+  //staple: output staple for each mu, summed over nu != mu (Nd)
+  //U_padded: the gauge link fields padded out using the PaddedCell class
+  //Cell: the padded cell class
+  //gStencil: the precomputed generalized local stencil for the staple
+  static void StaplePaddedAll(std::vector<GaugeMat> &staple, const std::vector<GaugeMat> &U_padded, const PaddedCell &Cell, const GeneralLocalStencil &gStencil) {
+    double t0 = usecond();
+    assert(U_padded.size() == Nd); assert(staple.size() == Nd);
+    assert(U_padded[0].Grid() == (GridBase*)Cell.grids.back());
+    assert(Cell.depth >= 1);
+    GridBase *ggrid = U_padded[0].Grid(); //padded cell grid
+
+    int shift_mu_off = gStencil._npoints/Nd;
+    
+    //Open views to padded gauge links and keep open over mu loop
+    typedef LatticeView<typename GaugeMat::vector_object> GaugeViewType;
+    size_t vsize = Nd*sizeof(GaugeViewType);
+    GaugeViewType* Ug_dirs_v_host = (GaugeViewType*)malloc(vsize);
+    for(int i=0;i<Nd;i++) Ug_dirs_v_host[i] = U_padded[i].View(AcceleratorRead);
+    GaugeViewType* Ug_dirs_v = (GaugeViewType*)acceleratorAllocDevice(vsize);
+    acceleratorCopyToDevice(Ug_dirs_v_host,Ug_dirs_v,vsize);
+    
+    GaugeMat gStaple(ggrid);
+
+    int outer_off = 0;
+    for(int mu=0;mu<Nd;mu++){
+      { //view scope
+	autoView( gStaple_v , gStaple, AcceleratorWrite);
+	auto gStencil_v = gStencil.View(AcceleratorRead);
+	
+	accelerator_for(ss, ggrid->oSites(), ggrid->Nsimd(), {
+	    decltype(coalescedRead(Ug_dirs_v[0][0])) stencil_ss;
+	    stencil_ss = Zero();
+	    int off = outer_off;
+	    
+	    for(int nu=0;nu<Nd;nu++){
+	      if(nu != mu){	  
+		GeneralStencilEntry const* e = gStencil_v.GetEntry(off++,ss);
+		auto U0 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(off++,ss);
+		auto U1 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(off++,ss);
+		auto U2 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+      
+		stencil_ss = stencil_ss + U2 * U1 * U0;
+
+		e = gStencil_v.GetEntry(off++,ss);
+		U0 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(off++,ss);
+		U1 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(off++,ss);
+		U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+
+		stencil_ss = stencil_ss + U2 * U1 * U0;
+	      }
+	    }
+		
+	    coalescedWrite(gStaple_v[ss],stencil_ss);
+	  }
+	  );
+      } //ensure views are all closed!
+      
+      staple[mu] = Cell.Extract(gStaple);
+      outer_off += shift_mu_off;
+    }//mu loop
+
+    for(int i=0;i<Nd;i++) Ug_dirs_v_host[i].ViewClose();
+    free(Ug_dirs_v_host);
+    acceleratorFreeDevice(Ug_dirs_v);
+    
+    double t1=usecond();
+    
+    std::cout << GridLogPerformance << "StaplePaddedAll timing:" << (t1-t0)/1000 << "ms" << std::endl;   
+  }
+
+   
   //////////////////////////////////////////////////
   // the sum over all staples on each site in direction mu,nu, upper part
   //////////////////////////////////////////////////
@@ -707,18 +907,14 @@ public:
   // the sum over all staples on each site
   //////////////////////////////////////////////////
   static void RectStapleDouble(GaugeMat &U2, const GaugeMat &U, int mu) {
-    U2 = U * Cshift(U, mu, 1);
+    U2 = U * Gimpl::CshiftLink(U, mu, 1);
   }
 
   ////////////////////////////////////////////////////////////////////////////
-  // Hop by two optimisation strategy does not work nicely with Gparity. (could
-  // do,
-  // but need to track two deep where cross boundary and apply a conjugation).
-  // Must differentiate this in Gimpl, and use Gimpl::isPeriodicGaugeField to do
-  // so .
+  // Hop by two optimisation strategy. Use RectStapleDouble to obtain 'U2'
   ////////////////////////////////////////////////////////////////////////////
-  static void RectStapleOptimised(GaugeMat &Stap, std::vector<GaugeMat> &U2,
-                                  std::vector<GaugeMat> &U, int mu) {
+  static void RectStapleOptimised(GaugeMat &Stap, const std::vector<GaugeMat> &U2,
+                                  const std::vector<GaugeMat> &U, int mu) {
 
     Stap = Zero();
 
@@ -732,9 +928,9 @@ public:
 
         // Up staple    ___ ___
         //             |       |
-        tmp = Cshift(adj(U[nu]), nu, -1);
+        tmp = Gimpl::CshiftLink(adj(U[nu]), nu, -1);
         tmp = adj(U2[mu]) * tmp;
-        tmp = Cshift(tmp, mu, -2);
+        tmp = Gimpl::CshiftLink(tmp, mu, -2);
 
         Staple2x1 = Gimpl::CovShiftForward(U[nu], nu, tmp);
 
@@ -742,14 +938,14 @@ public:
         //             |___ ___|
         //
         tmp = adj(U2[mu]) * U[nu];
-        Staple2x1 += Gimpl::CovShiftBackward(U[nu], nu, Cshift(tmp, mu, -2));
+        Staple2x1 += Gimpl::CovShiftBackward(U[nu], nu, Gimpl::CshiftLink(tmp, mu, -2));
 
         //              ___ ___
         //             |    ___|
         //             |___ ___|
         //
 
-        Stap += Cshift(Gimpl::CovShiftForward(U[mu], mu, Staple2x1), mu, 1);
+        Stap += Gimpl::CshiftLink(Gimpl::CovShiftForward(U[mu], mu, Staple2x1), mu, 1);
 
         //              ___ ___
         //             |___    |
@@ -758,7 +954,7 @@ public:
 
         //  tmp= Staple2x1* Cshift(U[mu],mu,-2);
         //  Stap+= Cshift(tmp,mu,1) ;
-        Stap += Cshift(Staple2x1, mu, 1) * Cshift(U[mu], mu, -1);
+        Stap += Gimpl::CshiftLink(Staple2x1, mu, 1) * Gimpl::CshiftLink(U[mu], mu, -1);
         ;
 
         //       --
@@ -766,10 +962,10 @@ public:
         //
         //      |  |
 
-        tmp = Cshift(adj(U2[nu]), nu, -2);
+        tmp = Gimpl::CshiftLink(adj(U2[nu]), nu, -2);
         tmp = Gimpl::CovShiftBackward(U[mu], mu, tmp);
-        tmp = U2[nu] * Cshift(tmp, nu, 2);
-        Stap += Cshift(tmp, mu, 1);
+        tmp = U2[nu] * Gimpl::CshiftLink(tmp, nu, 2);
+        Stap += Gimpl::CshiftLink(tmp, mu, 1);
 
         //      |  |
         //
@@ -778,25 +974,12 @@ public:
 
         tmp = Gimpl::CovShiftBackward(U[mu], mu, U2[nu]);
         tmp = adj(U2[nu]) * tmp;
-        tmp = Cshift(tmp, nu, -2);
-        Stap += Cshift(tmp, mu, 1);
+        tmp = Gimpl::CshiftLink(tmp, nu, -2);
+        Stap += Gimpl::CshiftLink(tmp, mu, 1);
       }
     }
   }
 
-  static void RectStaple(GaugeMat &Stap, const GaugeLorentz &Umu, int mu) {
-    RectStapleUnoptimised(Stap, Umu, mu);
-  }
-  static void RectStaple(const GaugeLorentz &Umu, GaugeMat &Stap,
-                         std::vector<GaugeMat> &U2, std::vector<GaugeMat> &U,
-                         int mu) {
-    if (Gimpl::isPeriodicGaugeField()) {
-      RectStapleOptimised(Stap, U2, U, mu);
-    } else {
-      RectStapleUnoptimised(Stap, Umu, mu);
-    }
-  }
-
   static void RectStapleUnoptimised(GaugeMat &Stap, const GaugeLorentz &Umu,
                                     int mu) {
     GridBase *grid = Umu.Grid();
@@ -895,6 +1078,288 @@ public:
     }
   }
 
+  static void RectStaple(GaugeMat &Stap, const GaugeLorentz &Umu, int mu) {
+    RectStapleUnoptimised(Stap, Umu, mu);
+  }
+  static void RectStaple(const GaugeLorentz &Umu, GaugeMat &Stap,
+                         std::vector<GaugeMat> &U2, std::vector<GaugeMat> &U,
+                         int mu) {
+    RectStapleOptimised(Stap, U2, U, mu);
+  }
+  //////////////////////////////////////////////////////
+  //Compute the rectangular staples for all orientations
+  //Stap : Array of staples (Nd)
+  //U: Gauge links in each direction (Nd)
+  /////////////////////////////////////////////////////
+  static void RectStapleAll(std::vector<GaugeMat> &Stap, const std::vector<GaugeMat> &U){
+    assert(Stap.size() == Nd); assert(U.size() == Nd);
+    std::vector<GaugeMat> U2(Nd,U[0].Grid());
+    for(int mu=0;mu<Nd;mu++) RectStapleDouble(U2[mu], U[mu], mu);
+    for(int mu=0;mu<Nd;mu++) RectStapleOptimised(Stap[mu], U2, U, mu);
+  }
+
+  //A workspace class allowing reuse of the stencil
+  class RectStaplePaddedAllWorkspace: public WilsonLoopPaddedStencilWorkspace{
+  public:
+    std::vector<Coordinate> getShifts() const override{
+      std::vector<Coordinate> shifts;
+      for (int mu = 0; mu < Nd; mu++){
+	for (int nu = 0; nu < Nd; nu++) {
+	  if (nu != mu) {
+	    auto genShift = [&](int mushift,int nushift){
+	      Coordinate out(Nd,0); out[mu]=mushift; out[nu]=nushift; return out;
+	    };
+
+	    //tmp6 = tmp5(x+mu) = U_mu(x+mu)U_nu(x+2mu)U_mu^dag(x+nu+mu) U_mu^dag(x+nu) U_nu^dag(x)
+	    shifts.push_back(genShift(0,0));
+	    shifts.push_back(genShift(0,+1));
+	    shifts.push_back(genShift(+1,+1));
+	    shifts.push_back(genShift(+2,0));
+	    shifts.push_back(genShift(+1,0));
+
+	    //tmp5 = tmp4(x+mu) = U_mu(x+mu)U^dag_nu(x-nu+2mu)U^dag_mu(x-nu+mu)U^dag_mu(x-nu)U_nu(x-nu)
+	    shifts.push_back(genShift(0,-1));
+	    shifts.push_back(genShift(0,-1));
+	    shifts.push_back(genShift(+1,-1));
+	    shifts.push_back(genShift(+2,-1));
+	    shifts.push_back(genShift(+1,0));
+
+	    //tmp5 = tmp4(x+mu) = U^dag_nu(x-nu+mu)U^dag_mu(x-nu)U^dag_mu(x-mu-nu)U_nu(x-mu-nu)U_mu(x-mu)
+	    shifts.push_back(genShift(-1,0));
+	    shifts.push_back(genShift(-1,-1));
+	    shifts.push_back(genShift(-1,-1));
+	    shifts.push_back(genShift(0,-1));
+	    shifts.push_back(genShift(+1,-1));
+
+	    //tmp5 = tmp4(x+mu) = U_nu(x+mu)U_mu^dag(x+nu)U_mu^dag(x-mu+nu)U_nu^dag(x-mu)U_mu(x-mu)
+	    shifts.push_back(genShift(-1,0));
+	    shifts.push_back(genShift(-1,0));
+	    shifts.push_back(genShift(-1,+1));
+	    shifts.push_back(genShift(0,+1));
+	    shifts.push_back(genShift(+1,0));
+
+	    //tmp6 = tmp5(x+mu) = U_nu(x+mu)U_nu(x+mu+nu)U_mu^dag(x+2nu)U_nu^dag(x+nu)U_nu^dag(x)
+	    shifts.push_back(genShift(0,0));
+	    shifts.push_back(genShift(0,+1));
+	    shifts.push_back(genShift(0,+2));
+	    shifts.push_back(genShift(+1,+1));
+	    shifts.push_back(genShift(+1,0));
+
+	    //tmp5 = tmp4(x+mu) = U_nu^dag(x+mu-nu)U_nu^dag(x+mu-2nu)U_mu^dag(x-2nu)U_nu(x-2nu)U_nu(x-nu)
+	    shifts.push_back(genShift(0,-1));
+	    shifts.push_back(genShift(0,-2));
+	    shifts.push_back(genShift(0,-2));
+	    shifts.push_back(genShift(+1,-2));
+	    shifts.push_back(genShift(+1,-1));
+	  }
+	}
+      }
+      return shifts;
+    }
+
+    int paddingDepth() const override{ return 2; }
+  }; 
+
+  //Padded cell implementation of the rectangular staple method for all mu, summed over nu != mu
+  //staple: output staple for each mu, summed over nu != mu (Nd)
+  //U_padded: the gauge link fields padded out using the PaddedCell class
+  //Cell: the padded cell class
+  static void RectStaplePaddedAll(std::vector<GaugeMat> &staple, const std::vector<GaugeMat> &U_padded, const PaddedCell &Cell) {
+    RectStaplePaddedAllWorkspace wk;
+    RectStaplePaddedAll(staple,U_padded,Cell,wk.getStencil(Cell));
+  }
+  
+  //Padded cell implementation of the rectangular staple method for all mu, summed over nu != mu
+  //staple: output staple for each mu, summed over nu != mu (Nd)
+  //U_padded: the gauge link fields padded out using the PaddedCell class
+  //Cell: the padded cell class
+  //gStencil: the stencil
+  static void RectStaplePaddedAll(std::vector<GaugeMat> &staple, const std::vector<GaugeMat> &U_padded, const PaddedCell &Cell, const GeneralLocalStencil &gStencil) {
+    double t0 = usecond();
+    assert(U_padded.size() == Nd); assert(staple.size() == Nd);
+    assert(U_padded[0].Grid() == (GridBase*)Cell.grids.back());
+    assert(Cell.depth >= 2);
+    GridBase *ggrid = U_padded[0].Grid(); //padded cell grid
+
+    size_t nshift = gStencil._npoints;
+    int mu_off_delta = nshift / Nd;
+    
+    //Open views to padded gauge links and keep open over mu loop
+    typedef LatticeView<typename GaugeMat::vector_object> GaugeViewType;
+    size_t vsize = Nd*sizeof(GaugeViewType);
+    GaugeViewType* Ug_dirs_v_host = (GaugeViewType*)malloc(vsize);
+    for(int i=0;i<Nd;i++) Ug_dirs_v_host[i] = U_padded[i].View(AcceleratorRead);
+    GaugeViewType* Ug_dirs_v = (GaugeViewType*)acceleratorAllocDevice(vsize);
+    acceleratorCopyToDevice(Ug_dirs_v_host,Ug_dirs_v,vsize);
+
+    GaugeMat gStaple(ggrid); //temp staple object on padded grid
+
+    int offset = 0;
+    for(int mu=0; mu<Nd; mu++){
+
+      { //view scope
+	autoView( gStaple_v , gStaple, AcceleratorWrite);
+	auto gStencil_v = gStencil.View(AcceleratorRead);
+
+	accelerator_for(ss, ggrid->oSites(), ggrid->Nsimd(), {
+	    decltype(coalescedRead(Ug_dirs_v[0][0])) stencil_ss;
+	    stencil_ss = Zero();
+	    int s=offset;
+	    for(int nu=0;nu<Nd;nu++){
+	      if(nu != mu){
+		//tmp6 = tmp5(x+mu) = U_mu(x+mu)U_nu(x+2mu)U_mu^dag(x+nu+mu) U_mu^dag(x+nu) U_nu^dag(x)
+		GeneralStencilEntry const* e = gStencil_v.GetEntry(s++,ss);
+		auto U0 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		auto U1 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		auto U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		auto U3 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		auto U4 = coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd);
+	    
+		stencil_ss = stencil_ss + U4*U3*U2*U1*U0;
+
+		//tmp5 = tmp4(x+mu) = U_mu(x+mu)U^dag_nu(x-nu+2mu)U^dag_mu(x-nu+mu)U^dag_mu(x-nu)U_nu(x-nu)
+		e = gStencil_v.GetEntry(s++,ss);
+		U0 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U1 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U3 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U4 = coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd);
+
+		stencil_ss = stencil_ss + U4*U3*U2*U1*U0;
+
+		//tmp5 = tmp4(x+mu) = U^dag_nu(x-nu+mu)U^dag_mu(x-nu)U^dag_mu(x-mu-nu)U_nu(x-mu-nu)U_mu(x-mu)
+		e = gStencil_v.GetEntry(s++,ss);
+		U0 = coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U1 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U3 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U4 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+
+		stencil_ss = stencil_ss + U4*U3*U2*U1*U0;
+
+		//tmp5 = tmp4(x+mu) = U_nu(x+mu)U_mu^dag(x+nu)U_mu^dag(x-mu+nu)U_nu^dag(x-mu)U_mu(x-mu)
+		e = gStencil_v.GetEntry(s++,ss);
+		U0 = coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U1 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U3 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U4 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+
+		stencil_ss = stencil_ss + U4*U3*U2*U1*U0;
+
+		//tmp6 = tmp5(x+mu) = U_nu(x+mu)U_nu(x+mu+nu)U_mu^dag(x+2nu)U_nu^dag(x+nu)U_nu^dag(x)
+		e = gStencil_v.GetEntry(s++,ss);
+		U0 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U1 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U3 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U4 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+
+		stencil_ss = stencil_ss + U4*U3*U2*U1*U0;   
+
+		//tmp5 = tmp4(x+mu) = U_nu^dag(x+mu-nu)U_nu^dag(x+mu-2nu)U_mu^dag(x-2nu)U_nu(x-2nu)U_nu(x-nu)
+		e = gStencil_v.GetEntry(s++,ss);
+		U0 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U1 = coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd);
+		e = gStencil_v.GetEntry(s++,ss);
+		U2 = adj(coalescedReadGeneralPermute(Ug_dirs_v[mu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U3 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+		e = gStencil_v.GetEntry(s++,ss);
+		U4 = adj(coalescedReadGeneralPermute(Ug_dirs_v[nu][e->_offset], e->_permute, Nd));
+
+		stencil_ss = stencil_ss + U4*U3*U2*U1*U0;   
+
+	      }
+	    }
+	    coalescedWrite(gStaple_v[ss],stencil_ss);
+	  }
+	  );
+	offset += mu_off_delta;
+      }//kernel/view scope
+
+      staple[mu] = Cell.Extract(gStaple);    
+    }//mu loop
+  
+    for(int i=0;i<Nd;i++) Ug_dirs_v_host[i].ViewClose();
+    free(Ug_dirs_v_host);
+    acceleratorFreeDevice(Ug_dirs_v);
+    
+    double t1 = usecond();
+    
+    std::cout << GridLogPerformance << "RectStaplePaddedAll timings:" << (t1-t0)/1000 << "ms" << std::endl;   
+  }
+
+  //A workspace for reusing the PaddedCell and GeneralLocalStencil objects
+  class StapleAndRectStapleAllWorkspace: public WilsonLoopPaddedWorkspace{
+  public:
+    StapleAndRectStapleAllWorkspace(){
+      this->addStencil(new StaplePaddedAllWorkspace);
+      this->addStencil(new RectStaplePaddedAllWorkspace);
+    }
+  };     
+    
+  //////////////////////////////////////////////////////
+  //Compute the 1x1 and 1x2 staples for all orientations
+  //Stap : Array of staples (Nd)
+  //RectStap: Array of rectangular staples (Nd)
+  //U: Gauge links in each direction (Nd)
+  /////////////////////////////////////////////////////
+  static void StapleAndRectStapleAll(std::vector<GaugeMat> &Stap, std::vector<GaugeMat> &RectStap, const std::vector<GaugeMat> &U){
+    StapleAndRectStapleAllWorkspace wk;
+    StapleAndRectStapleAll(Stap,RectStap,U,wk);
+  }
+  
+  //////////////////////////////////////////////////////
+  //Compute the 1x1 and 1x2 staples for all orientations
+  //Stap : Array of staples (Nd)
+  //RectStap: Array of rectangular staples (Nd)
+  //U: Gauge links in each direction (Nd)
+  //wk: a workspace containing stored PaddedCell and GeneralLocalStencil objects to maximize reuse
+  /////////////////////////////////////////////////////
+  static void StapleAndRectStapleAll(std::vector<GaugeMat> &Stap, std::vector<GaugeMat> &RectStap, const std::vector<GaugeMat> &U, StapleAndRectStapleAllWorkspace &wk){
+#if 0
+    StapleAll(Stap, U);
+    RectStapleAll(RectStap, U);
+#else
+    double t0 = usecond();
+
+    GridCartesian* unpadded_grid = dynamic_cast<GridCartesian*>(U[0].Grid());
+    const PaddedCell &Ghost = wk.getPaddedCell(unpadded_grid);
+        
+    CshiftImplGauge<Gimpl> cshift_impl;
+    std::vector<GaugeMat> U_pad(Nd, Ghost.grids.back());
+    for(int mu=0;mu<Nd;mu++) U_pad[mu] = Ghost.Exchange(U[mu], cshift_impl);
+    double t1 = usecond();
+    StaplePaddedAll(Stap, U_pad, Ghost, wk.getStencil(0,unpadded_grid) );
+    double t2 = usecond();
+    RectStaplePaddedAll(RectStap, U_pad, Ghost, wk.getStencil(1,unpadded_grid));
+    double t3 = usecond();
+    std::cout << GridLogPerformance << "StapleAndRectStapleAll timings: pad:" << (t1-t0)/1000 << "ms, staple:" << (t2-t1)/1000 << "ms, rect-staple:" << (t3-t2)/1000 << "ms" << std::endl;
+#endif
+  }
+
   //////////////////////////////////////////////////
   // Wilson loop of size (R1, R2), oriented in mu,nu plane
   //////////////////////////////////////////////////

Grid/simd/Grid_vector_types.h

@@ -1130,6 +1130,14 @@ static_assert(sizeof(SIMD_Ftype) == sizeof(SIMD_Itype), "SIMD vector lengths inc
 #endif
 #endif
 
+// Fixme need coalesced read gpermute
+template<class vobj> void gpermute(vobj & inout,int perm){
+  vobj tmp=inout;
+  if (perm & 0x1 ) { permute(inout,tmp,0); tmp=inout;}
+  if (perm & 0x2 ) { permute(inout,tmp,1); tmp=inout;}
+  if (perm & 0x4 ) { permute(inout,tmp,2); tmp=inout;}
+  if (perm & 0x8 ) { permute(inout,tmp,3); tmp=inout;}
+}
 
 NAMESPACE_END(Grid);
 

Grid/stencil/GeneralLocalStencil.h

@@ -46,7 +46,7 @@ class GeneralLocalStencilView {
   accelerator_inline GeneralStencilEntry * GetEntry(int point,int osite) { 
     return & this->_entries_p[point+this->_npoints*osite]; 
   }
-
+  void ViewClose(void){};
 };
 ////////////////////////////////////////
 // The Stencil Class itself
@@ -61,7 +61,7 @@ protected:
 public: 
   GridBase *Grid(void) const { return _grid; }
 
-  View_type View(void) const {
+  View_type View(int mode) const {
     View_type accessor(*( (View_type *) this));
     return accessor;
   }
@@ -79,60 +79,60 @@ public:
     this->_entries.resize(npoints* osites);
     this->_entries_p = &_entries[0];
 
+    thread_for(site, osites, {
+	Coordinate Coor;
+	Coordinate NbrCoor;
 
-    Coordinate Coor;
-    Coordinate NbrCoor;
-    for(Integer site=0;site<osites;site++){
-      for(Integer ii=0;ii<npoints;ii++){
-	Integer lex = site*npoints+ii;
-	GeneralStencilEntry SE;
-	////////////////////////////////////////////////
-	// Outer index of neighbour Offset calculation
-	////////////////////////////////////////////////
-	grid->oCoorFromOindex(Coor,site);
-	for(int d=0;d<Coor.size();d++){
-	  int rd = grid->_rdimensions[d];
-	  NbrCoor[d] = (Coor[d] + shifts[ii][d] + rd )%rd;
-	}
-	SE._offset      = grid->oIndexReduced(NbrCoor);
-
-	////////////////////////////////////////////////
-	// Inner index permute calculation
-	// Simpler version using icoor calculation
-	////////////////////////////////////////////////
-	SE._permute =0;
-	for(int d=0;d<Coor.size();d++){
-
-	  int fd = grid->_fdimensions[d];
-	  int rd = grid->_rdimensions[d];
-	  int ly = grid->_simd_layout[d];
-
-	  assert((ly==1)||(ly==2));
-
-	  int shift = (shifts[ii][d]+fd)%fd;  // make it strictly positive 0.. L-1
-	  int x = Coor[d];                // x in [0... rd-1] as an oSite 
-
-	  int permute_dim  = grid->PermuteDim(d);
-	  int permute_slice=0;
-	  if(permute_dim){    
-	    int  num = shift%rd; // Slice within dest osite cell of slice zero
-	    int wrap = shift/rd; // Number of osite local volume cells crossed through
-                                  // x+num < rd dictates whether we are in same permute state as slice 0
-	    if ( x< rd-num ) permute_slice=wrap;
-	    else             permute_slice=(wrap+1)%ly;
+	for(Integer ii=0;ii<npoints;ii++){
+	  Integer lex = site*npoints+ii;
+	  GeneralStencilEntry SE;
+	  ////////////////////////////////////////////////
+	  // Outer index of neighbour Offset calculation
+	  ////////////////////////////////////////////////
+	  grid->oCoorFromOindex(Coor,site);
+	  for(int d=0;d<Coor.size();d++){
+	    int rd = grid->_rdimensions[d];
+	    NbrCoor[d] = (Coor[d] + shifts[ii][d] + rd )%rd;
 	  }
-	  if ( permute_slice ) {
-	    int ptype       =grid->PermuteType(d);
-	    uint8_t mask    =0x1<<ptype;
-	    SE._permute    |= mask;
+	  SE._offset      = grid->oIndexReduced(NbrCoor);
+
+	  ////////////////////////////////////////////////
+	  // Inner index permute calculation
+	  // Simpler version using icoor calculation
+	  ////////////////////////////////////////////////
+	  SE._permute =0;
+	  for(int d=0;d<Coor.size();d++){
+
+	    int fd = grid->_fdimensions[d];
+	    int rd = grid->_rdimensions[d];
+	    int ly = grid->_simd_layout[d];
+
+	    assert((ly==1)||(ly==2));
+
+	    int shift = (shifts[ii][d]+fd)%fd;  // make it strictly positive 0.. L-1
+	    int x = Coor[d];                // x in [0... rd-1] as an oSite 
+
+	    int permute_dim  = grid->PermuteDim(d);
+	    int permute_slice=0;
+	    if(permute_dim){    
+	      int  num = shift%rd; // Slice within dest osite cell of slice zero
+	      int wrap = shift/rd; // Number of osite local volume cells crossed through
+	      // x+num < rd dictates whether we are in same permute state as slice 0
+	      if ( x< rd-num ) permute_slice=wrap;
+	      else             permute_slice=(wrap+1)%ly;
+	    }
+	    if ( permute_slice ) {
+	      int ptype       =grid->PermuteType(d);
+	      uint8_t mask    =0x1<<ptype;
+	      SE._permute    |= mask;
+	    }
 	  }	
+	  ////////////////////////////////////////////////
+	  // Store in look up table
+	  ////////////////////////////////////////////////
+	  this->_entries[lex] = SE;
 	}
-	////////////////////////////////////////////////
-	// Store in look up table
-	////////////////////////////////////////////////
-	this->_entries[lex] = SE;
-      }
-    }      
+      });
   }
   
 };

Grid/stencil/Stencil.h

@@ -32,6 +32,7 @@
 
 #include <Grid/stencil/SimpleCompressor.h>   // subdir aggregate
 #include <Grid/stencil/Lebesgue.h>   // subdir aggregate
+#include <Grid/stencil/GeneralLocalStencil.h>
 
 //////////////////////////////////////////////////////////////////////////////////////////
 // Must not lose sight that goal is to be able to construct really efficient

Grid/tensors/Tensor_SIMT.h

@@ -73,6 +73,16 @@ vobj coalescedReadPermute(const vobj & __restrict__ vec,int ptype,int doperm,int
     return vec;
   }
 }
+//'perm_mask' acts as a bitmask
+template<class vobj> accelerator_inline
+vobj coalescedReadGeneralPermute(const vobj & __restrict__ vec,int perm_mask,int nd,int lane=0)
+{
+  auto obj = vec, tmp = vec;
+  for (int d=0;d<nd;d++)
+    if (perm_mask & (0x1 << d)) { permute(obj,tmp,d); tmp=obj;}
+  return obj;
+}
+
 template<class vobj> accelerator_inline
 void coalescedWrite(vobj & __restrict__ vec,const vobj & __restrict__ extracted,int lane=0)
 {
@@ -83,7 +93,7 @@ void coalescedWriteNonTemporal(vobj & __restrict__ vec,const vobj & __restrict__
 {
   vstream(vec, extracted);
 }
-#else
+#else //==GRID_SIMT
 
 
 //#ifndef GRID_SYCL
@@ -166,6 +176,14 @@ typename vobj::scalar_object coalescedReadPermute(const vobj & __restrict__ vec,
   return extractLane(plane,vec);
 }
 template<class vobj> accelerator_inline
+typename vobj::scalar_object coalescedReadGeneralPermute(const vobj & __restrict__ vec,int perm_mask,int nd,int lane=acceleratorSIMTlane(vobj::Nsimd()))
+{
+  int plane = lane;
+  for (int d=0;d<nd;d++)
+    plane = (perm_mask & (0x1 << d)) ? plane ^ (vobj::Nsimd() >> (d + 1)) : plane;
+  return extractLane(plane,vec);
+}
+template<class vobj> accelerator_inline
 void coalescedWrite(vobj & __restrict__ vec,const typename vobj::scalar_object & __restrict__ extracted,int lane=acceleratorSIMTlane(vobj::Nsimd()))
 {
   insertLane(lane,vec,extracted);

Grid/tensors/Tensor_Ta.h

@@ -66,13 +66,61 @@ template<class vtype,int N> accelerator_inline iMatrix<vtype,N> Ta(const iMatrix
   return ret;
 }
 
+template<class vtype> accelerator_inline iScalar<vtype> SpTa(const iScalar<vtype>&r)
+{
+  iScalar<vtype> ret;
+  ret._internal = SpTa(r._internal);
+  return ret;
+}
+template<class vtype,int N> accelerator_inline iVector<vtype,N> SpTa(const iVector<vtype,N>&r)
+{
+  iVector<vtype,N> ret;
+  for(int i=0;i<N;i++){
+    ret._internal[i] = SpTa(r._internal[i]);
+  }
+  return ret;
+}
+template<class vtype,int N, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0 >::type * =nullptr>
+accelerator_inline iMatrix<vtype,N> SpTa(const iMatrix<vtype,N> &arg)
+{
+  // Generalises Ta to Sp2n
+  // Applies the following projections
+  // P_{antihermitian} P_{antihermitian-Sp-algebra} P_{traceless}
+  // where the ordering matters
+  // P_{traceless} subtracts the trace
+  // P_{antihermitian-Sp-algebra} provides the block structure of the algebra based on U = exp(T) i.e. anti-hermitian generators
+  // P_{antihermitian} does in-adj(in) / 2
+  iMatrix<vtype,N> ret(arg);
+  double factor = (1.0/(double)N);
+  vtype nrm;
+  nrm = 0.5;
+    
+  ret = arg - (trace(arg)*factor);
+    
+  for(int c1=0;c1<N/2;c1++)
+  {
+      for(int c2=0;c2<N/2;c2++)
+      {
+          ret._internal[c1][c2] = nrm*(conjugate(ret._internal[c1+N/2][c2+N/2]) + ret._internal[c1][c2]); // new[up-left] = old[up-left]+old*[down-right]
+          ret._internal[c1][c2+N/2] = nrm*(ret._internal[c1][c2+N/2] - conjugate(ret._internal[c1+N/2][c2])); // new[up-right] = old[up-right]-old*[down-left]
+      }
+      for(int c2=N/2;c2<N;c2++)
+      {
+          ret._internal[c1+N/2][c2-N/2] = -conjugate(ret._internal[c1][c2]);  //  reconstructs lower blocks
+          ret._internal[c1+N/2][c2] = conjugate(ret._internal[c1][c2-N/2]);   //  from upper blocks
+      }
+  }
+    
+  ret = (ret - adj(ret))*0.5;
+
+  return ret;
+}
 
 /////////////////////////////////////////////// 
 // ProjectOnGroup function for scalar, vector, matrix 
 // Projects on orthogonal, unitary group
 /////////////////////////////////////////////// 
 
-
 template<class vtype> accelerator_inline iScalar<vtype> ProjectOnGroup(const iScalar<vtype>&r)
 {
   iScalar<vtype> ret;
@@ -90,10 +138,12 @@ template<class vtype,int N> accelerator_inline iVector<vtype,N> ProjectOnGroup(c
 template<class vtype,int N, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0 >::type * =nullptr> 
 accelerator_inline iMatrix<vtype,N> ProjectOnGroup(const iMatrix<vtype,N> &arg)
 {
+  typedef typename iMatrix<vtype,N>::scalar_type scalar;
   // need a check for the group type?
   iMatrix<vtype,N> ret(arg);
   vtype nrm;
   vtype inner;
+  scalar one(1.0);
   for(int c1=0;c1<N;c1++){
 
     // Normalises row c1
@@ -102,7 +152,7 @@ accelerator_inline iMatrix<vtype,N> ProjectOnGroup(const iMatrix<vtype,N> &arg)
       inner += innerProduct(ret._internal[c1][c2],ret._internal[c1][c2]);
 
     nrm = sqrt(inner);
-    nrm = 1.0/nrm;
+    nrm = one/nrm;
     for(int c2=0;c2<N;c2++)
       ret._internal[c1][c2]*= nrm;
       
@@ -127,7 +177,7 @@ accelerator_inline iMatrix<vtype,N> ProjectOnGroup(const iMatrix<vtype,N> &arg)
       inner += innerProduct(ret._internal[c1][c2],ret._internal[c1][c2]);
 
     nrm = sqrt(inner);
-    nrm = 1.0/nrm;
+    nrm = one/nrm;
     for(int c2=0;c2<N;c2++)
       ret._internal[c1][c2]*= nrm;
   }
@@ -135,6 +185,85 @@ accelerator_inline iMatrix<vtype,N> ProjectOnGroup(const iMatrix<vtype,N> &arg)
   return ret;
 }
 
+// re-do for sp2n
+
+// Ta cannot be defined here for Sp2n because I need the generators from the Sp class
+// It is defined in gauge impl types
+
+template<class vtype> accelerator_inline iScalar<vtype> ProjectOnSpGroup(const iScalar<vtype>&r)
+{
+  iScalar<vtype> ret;
+  ret._internal = ProjectOnSpGroup(r._internal);
+  return ret;
+}
+template<class vtype,int N> accelerator_inline iVector<vtype,N> ProjectOnSpGroup(const iVector<vtype,N>&r)
+{
+  iVector<vtype,N> ret;
+  for(int i=0;i<N;i++){
+    ret._internal[i] = ProjectOnSpGroup(r._internal[i]);
+  }
+  return ret;
+}
+
+
+// int N is 2n in Sp(2n)
+template<class vtype,int N, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0 >::type * =nullptr>
+accelerator_inline iMatrix<vtype,N> ProjectOnSpGroup(const iMatrix<vtype,N> &arg)
+{
+  // need a check for the group type?
+  iMatrix<vtype,N> ret(arg);
+  vtype nrm;
+  vtype inner;
+  
+  for(int c1=0;c1<N/2;c1++)
+  {
+      
+    for (int b=0; b<c1; b++)                  // remove the b-rows from U_c1
+    {
+      decltype(ret._internal[b][b]*ret._internal[b][b]) pr;
+      decltype(ret._internal[b][b]*ret._internal[b][b]) prn;
+      zeroit(pr);
+      zeroit(prn);
+          
+      for(int c=0; c<N; c++)
+      {
+        pr += conjugate(ret._internal[c1][c])*ret._internal[b][c];        // <U_c1 | U_b >
+        prn += conjugate(ret._internal[c1][c])*ret._internal[b+N/2][c];   // <U_c1 | U_{b+N} >
+      }
+       
+
+      for(int c=0; c<N; c++)
+      {
+        ret._internal[c1][c] -= (conjugate(pr) * ret._internal[b][c] + conjugate(prn) * ret._internal[b+N/2][c] );    //  U_c1 -= (  <U_c1 | U_b > U_b + <U_c1 | U_{b+N} > U_{b+N}  )
+      }
+    }
+    
+    zeroit(inner);
+    for(int c2=0;c2<N;c2++)
+    {
+      inner += innerProduct(ret._internal[c1][c2],ret._internal[c1][c2]);
+    }
+      
+    nrm = sqrt(inner);
+    nrm = 1.0/nrm;
+    for(int c2=0;c2<N;c2++)
+    {
+      ret._internal[c1][c2]*= nrm;
+    }
+      
+    for(int c2=0;c2<N/2;c2++)
+    {
+      ret._internal[c1+N/2][c2+N/2] = conjugate(ret._internal[c1][c2]);          // down right in the new matrix = (up-left)* of the old matrix
+    }
+      
+    for(int c2=N/2;c2<N;c2++)
+    {
+      ret._internal[c1+N/2][c2-N/2] = -conjugate(ret._internal[c1][c2]);;     // down left in the new matrix = -(up-right)* of the old
+    }
+  }
+  return ret;
+}
+
 NAMESPACE_END(Grid);
 
 #endif

Grid/tensors/Tensor_exp.h

@@ -53,7 +53,6 @@ template<class vtype, int N> accelerator_inline iVector<vtype, N> Exponentiate(c
 }
 
 
-
 // Specialisation: Cayley-Hamilton exponential for SU(3)
 #if 0
 template<class vtype, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0>::type * =nullptr> 

Grid/threads/Accelerator.h

@@ -137,6 +137,18 @@ inline void cuda_mem(void)
     dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
     LambdaApply<<<cu_blocks,cu_threads,0,computeStream>>>(num1,num2,nsimd,lambda);	\
   }
+#define prof_accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
+  {									\
+    int nt=acceleratorThreads();					\
+    typedef uint64_t Iterator;						\
+    auto lambda = [=] accelerator					\
+      (Iterator iter1,Iterator iter2,Iterator lane) mutable {		\
+      __VA_ARGS__;							\
+    };									\
+    dim3 cu_threads(nsimd,acceleratorThreads(),1);			\
+    dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
+    ProfileLambdaApply<<<cu_blocks,cu_threads,0,computeStream>>>(num1,num2,nsimd,lambda); \
+  }
 
 #define accelerator_for6dNB(iter1, num1,				\
                             iter2, num2,				\
@@ -157,6 +169,20 @@ inline void cuda_mem(void)
     Lambda6Apply<<<cu_blocks,cu_threads,0,computeStream>>>(num1,num2,num3,num4,num5,num6,lambda); \
   }
 
+
+#define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
+  {									\
+    int nt=acceleratorThreads();					\
+    typedef uint64_t Iterator;						\
+    auto lambda = [=] accelerator					\
+      (Iterator iter1,Iterator iter2,Iterator lane) mutable {		\
+      __VA_ARGS__;							\
+    };									\
+    dim3 cu_threads(nsimd,acceleratorThreads(),1);			\
+    dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
+    LambdaApply<<<cu_blocks,cu_threads,0,computeStream>>>(num1,num2,nsimd,lambda);	\
+  }
+
 template<typename lambda>  __global__
 void LambdaApply(uint64_t num1, uint64_t num2, uint64_t num3, lambda Lambda)
 {
@@ -168,6 +194,17 @@ void LambdaApply(uint64_t num1, uint64_t num2, uint64_t num3, lambda Lambda)
     Lambda(x,y,z);
   }
 }
+template<typename lambda>  __global__
+void ProfileLambdaApply(uint64_t num1, uint64_t num2, uint64_t num3, lambda Lambda)
+{
+  // Weird permute is to make lane coalesce for large blocks
+  uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
+  uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
+  uint64_t z = threadIdx.x;
+  if ( (x < num1) && (y<num2) && (z<num3) ) {
+    Lambda(x,y,z);
+  }
+}
 
 template<typename lambda>  __global__
 void Lambda6Apply(uint64_t num1, uint64_t num2, uint64_t num3,
@@ -208,6 +245,7 @@ inline void *acceleratorAllocShared(size_t bytes)
   if( err != cudaSuccess ) {
     ptr = (void *) NULL;
     printf(" cudaMallocManaged failed for %d %s \n",bytes,cudaGetErrorString(err));
+    assert(0);
   }
   return ptr;
 };
@@ -460,6 +498,9 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
 #if defined(GRID_SYCL) || defined(GRID_CUDA) || defined(GRID_HIP)
 // FIXME -- the non-blocking nature got broken March 30 2023 by PAB
 #define accelerator_forNB( iter1, num1, nsimd, ... ) accelerator_for2dNB( iter1, num1, iter2, 1, nsimd, {__VA_ARGS__} );  
+#define prof_accelerator_for( iter1, num1, nsimd, ... ) \
+  prof_accelerator_for2dNB( iter1, num1, iter2, 1, nsimd, {__VA_ARGS__} );\
+  accelerator_barrier(dummy);
 
 #define accelerator_for( iter, num, nsimd, ... )		\
   accelerator_forNB(iter, num, nsimd, { __VA_ARGS__ } );	\

Grid/util/Coordinate.h

@@ -94,6 +94,13 @@ static constexpr int MaxDims = GRID_MAX_LATTICE_DIMENSION;
 
 typedef AcceleratorVector<int,MaxDims> Coordinate;
 
+template<class T,int _ndim>
+inline bool operator==(const AcceleratorVector<T,_ndim> &v,const AcceleratorVector<T,_ndim> &w)
+{
+  if (v.size()!=w.size()) return false;
+  for(int i=0;i<v.size();i++) if ( v[i]!=w[i] ) return false;
+  return true;
+}
 template<class T,int _ndim>
 inline std::ostream & operator<<(std::ostream &os, const AcceleratorVector<T,_ndim> &v)
 {

Grid/util/Lexicographic.h

@@ -8,7 +8,7 @@ namespace Grid{
   public:
 
     template<class coor_t>
-    static accelerator_inline void CoorFromIndex (coor_t& coor,int index,const coor_t &dims){
+    static accelerator_inline void CoorFromIndex (coor_t& coor,int64_t index,const coor_t &dims){
       int nd= dims.size();
       coor.resize(nd);
       for(int d=0;d<nd;d++){
@@ -18,28 +18,45 @@ namespace Grid{
     }
 
     template<class coor_t>
-    static accelerator_inline void IndexFromCoor (const coor_t& coor,int &index,const coor_t &dims){
+    static accelerator_inline void IndexFromCoor (const coor_t& coor,int64_t &index,const coor_t &dims){
       int nd=dims.size();
       int stride=1;
       index=0;
       for(int d=0;d<nd;d++){
-	index = index+stride*coor[d];
+	index = index+(int64_t)stride*coor[d];
 	stride=stride*dims[d];
       }
     }
+    template<class coor_t>
+    static accelerator_inline void IndexFromCoor (const coor_t& coor,int &index,const coor_t &dims){
+      int64_t index64;
+      IndexFromCoor(coor,index64,dims);
+      assert(index64<2*1024*1024*1024LL);
+      index = (int) index64;
+    }
 
     template<class coor_t>
-    static inline void IndexFromCoorReversed (const coor_t& coor,int &index,const coor_t &dims){
+    static inline void IndexFromCoorReversed (const coor_t& coor,int64_t &index,const coor_t &dims){
       int nd=dims.size();
       int stride=1;
       index=0;
       for(int d=nd-1;d>=0;d--){
-	index = index+stride*coor[d];
+	index = index+(int64_t)stride*coor[d];
 	stride=stride*dims[d];
       }
     }
     template<class coor_t>
-    static inline void CoorFromIndexReversed (coor_t& coor,int index,const coor_t &dims){
+    static inline void IndexFromCoorReversed (const coor_t& coor,int &index,const coor_t &dims){
+      int64_t index64;
+      IndexFromCoorReversed(coor,index64,dims);
+      if ( index64>=2*1024*1024*1024LL ){
+	std::cout << " IndexFromCoorReversed " << coor<<" index " << index64<< " dims "<<dims<<std::endl;
+      }
+      assert(index64<2*1024*1024*1024LL);
+      index = (int) index64;
+    }
+    template<class coor_t>
+    static inline void CoorFromIndexReversed (coor_t& coor,int64_t index,const coor_t &dims){
       int nd= dims.size();
       coor.resize(nd);
       for(int d=nd-1;d>=0;d--){

configure.ac

@@ -41,7 +41,7 @@ AC_PROG_RANLIB
 
 ############### Get compiler informations
 AC_LANG([C++])
-AX_CXX_COMPILE_STDCXX_11([noext],[mandatory])
+AX_CXX_COMPILE_STDCXX(14,noext,mandatory)
 AX_COMPILER_VENDOR
 AC_DEFINE_UNQUOTED([CXX_COMP_VENDOR],["$ax_cv_cxx_compiler_vendor"],
       [vendor of C++ compiler that will compile the code])
@@ -191,10 +191,28 @@ case ${ac_Nc} in
         AC_DEFINE([Config_Nc],[4],[Gauge group Nc]);;
     5)
         AC_DEFINE([Config_Nc],[5],[Gauge group Nc]);;
+    8)
+        AC_DEFINE([Config_Nc],[8],[Gauge group Nc]);;
     *)
       AC_MSG_ERROR(["Unsupport gauge group choice Nc = ${ac_Nc}"]);;
 esac
 
+############### Symplectic group
+AC_ARG_ENABLE([Sp],
+    [AC_HELP_STRING([--enable-Sp=yes|no], [enable gauge group Sp2n])],
+    [ac_ENABLE_SP=${enable_Sp}], [ac_ENABLE_SP=no])
+
+AM_CONDITIONAL(BUILD_SP, [ test "${ac_ENABLE_SP}X" == "yesX" ])
+
+case ${ac_ENABLE_SP} in
+   yes)
+        AC_DEFINE([Sp2n_config],[1],[gauge group Sp2n], [have_sp2n=true]);;
+   no)
+        AC_DEFINE([Sp2n_config],[0],[gauge group SUn], [have_sp2n=false]);;
+    *)
+        AC_MSG_ERROR(["--enable-Sp is either yes or no"]);;
+esac
+
 ############### FP16 conversions
 AC_ARG_ENABLE([sfw-fp16],
     [AS_HELP_STRING([--enable-sfw-fp16=yes|no],[enable software fp16 comms])],
@@ -737,7 +755,7 @@ case ${ac_TIMERS} in
 esac
 
 ############### Chroma regression test
-AC_ARG_ENABLE([chroma],[AS_HELP_STRING([--enable-chroma],[Expect chroma compiled under c++11 ])],ac_CHROMA=yes,ac_CHROMA=no)
+AC_ARG_ENABLE([chroma],[AS_HELP_STRING([--enable-chroma],[Expect chroma compiled under c++14 ])],ac_CHROMA=yes,ac_CHROMA=no)
 
 case ${ac_CHROMA} in
      yes|no)
@@ -819,6 +837,7 @@ FFTW                        : `if test "x$have_fftw" = xtrue; then echo yes; els
 LIME (ILDG support)         : `if test "x$have_lime" = xtrue; then echo yes; else echo no; fi`
 HDF5                        : `if test "x$have_hdf5" = xtrue; then echo yes; else echo no; fi`
 build DOXYGEN documentation : `if test "$DX_FLAG_doc" = '1'; then echo yes; else echo no; fi`
+Sp2n                        : ${ac_ENABLE_SP}
 ----- BUILD FLAGS -------------------------------------
 CXXFLAGS:
 `echo ${AM_CXXFLAGS} ${CXXFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
@@ -847,6 +866,7 @@ AC_CONFIG_FILES(tests/lanczos/Makefile)
 AC_CONFIG_FILES(tests/smearing/Makefile)
 AC_CONFIG_FILES(tests/qdpxx/Makefile)
 AC_CONFIG_FILES(tests/testu01/Makefile)
+AC_CONFIG_FILES(tests/sp2n/Makefile)
 AC_CONFIG_FILES(benchmarks/Makefile)
 AC_CONFIG_FILES(examples/Makefile)
 AC_OUTPUT

documentation/GridXcode/readme.md

@@ -10,9 +10,8 @@ For first time setup of the Xcode and Grid build environment on Mac OS, you will
 
 1. Install Xcode and the Xcode command-line utilities
 2. Set Grid environment variables
-3. Install and build Open MPI ***optional***
-4. Install and build Grid pre-requisites
-5. Install, Configure and Build Grid
+3. Install and build Grid pre-requisites
+4. Install, Configure and Build Grid
 
 Apple's [Xcode website][Xcode] is the go-to reference for 1, and the definitive reference for 4 and 5 is the [Grid Documentation][GridDoc].
 
@@ -92,60 +91,33 @@ launchctl setenv GridPkg /opt/local</string>
 </plist>
 ```
 
-## 3. Install and build Open MPI -- ***optional***
-
-Download the latest version of [Open MPI][OMPI] version 3.1 (I used 3.1.5) and build it like so:
-
-[OMPI]: https://www.open-mpi.org/software/ompi/v3.1/
-
-    ../configure CC=clang CXX=clang++ CXXFLAGS=-g --prefix=$GridPre/bin
-    make -j 4 all install
-
-***Note the `/bin` at the end of the prefix - this is required. As a quirk of the OpenMPI installer, `--prefix` must point to the `bin` subdirectory, with other files installed in `$GridPre/include`, `$GridPre/lib`, `$GridPre/share`, etc.***
-
-Grid does not have any dependencies on fortran, however many standard scientific packages do, so you may wish to download GNU fortran (e.g. MacPorts ``gfortran`` package) and add the following to your configure invocation:
-
-    F77=gfortran FC=gfortran
-
-## 4. Install and build Grid pre-requisites
+## 3. Install and build Grid pre-requisites
 
 To simplify the installation of **Grid pre-requisites**, you can use your favourite package manager, e.g.:
 
-### 1. [MacPorts][MacPorts]
+### 3.1. [MacPorts][MacPorts]
 
 [MacPorts]: https://www.macports.org "MacPorts package manager"
 
 Install [MacPorts][MacPorts] if you haven't done so already, and then install packages with:
 
-    sudo port install <portname>
+    sudo port install openmpi git-flow-avh gmp hdf5 mpfr fftw-3-single lapack wget autoconf automake bison cmake gawk libomp
 
-These are the `portname`s for mandatory Grid libraries:
+On a Mac without GPUs:
 
-* git-flow-avh
-* gmp
-* hdf5
-* mpfr
+    sudo port install OpenBLAS +native
 
-and these are the `portname`s for optional Grid libraries:
+To use `Gnu sha256sum`:
 
-* fftw-3-single
-* lapack
-* doxygen
-* OpenBLAS
+    pushd /opt/local/bin; sudo ln -s gsha256sum sha256sum; popd 
 
-***Please update this list with any packages I've missed! ... and double-check whether OpenBLAS is really for Grid. NB: lapack doesn't seem to work. Should it be scalapack?***
+These `port`s are not strictly necessary, but they are helpful:
 
-### 2. [Homebrew][Homebrew]
+    sudo port install gnuplot gsl h5utils nasm rclone texinfo tree xorg-server
 
-[Homebrew]: https://brew.sh "Homebrew package manager"
+***Please update this list with any packages I've missed!***
 
-Install [Homebrew][Homebrew] if you haven't done so already, and then install packages with:
-
-    sudo brew install <packagename>
-
-The same packages are available as from MacPorts.
-
-### Install LIME ***optional***
+#### Install LIME
 
 There isn't currently a port for [C-LIME][C-LIME], so download the source and then build it:
 
@@ -154,9 +126,19 @@ There isn't currently a port for [C-LIME][C-LIME], so download the source and th
     ../configure CC=clang --prefix=$GridPre
     make -j 4 all install
 
-## 5. Install, Configure and Build Grid
+### 3.2. [Homebrew][Homebrew]
 
-### 5.1 Install Grid
+[Homebrew]: https://brew.sh "Homebrew package manager"
+
+Install [Homebrew][Homebrew] if you haven't done so already, and then install packages with:
+
+    sudo brew install <packagename>
+
+I don't use Homebrew, so I'm not sure what the Brew package name equivalents are. ** Please update if you know **
+
+## 4. Install, Configure and Build Grid
+
+### 4.1 Install Grid
 
 [Grid]: https://github.com/paboyle/Grid
 
@@ -174,7 +156,7 @@ or
 
 depending on how many times you like to enter your password.
 
-### 5.2 Configure Grid
+### 4.2 Configure Grid
 
 The Xcode build system supports multiple configurations for each project, by default: `Debug` and `Release`, but more configurations can be defined. We will create separate Grid build directories for each configuration, using the Grid **Autoconf** build system to make each configuration. NB: it is **not** necessary to run `make install` on them once they are built (IDE features such as *jump to definition* will work better of you don't).
 
@@ -198,7 +180,7 @@ Debug configuration with MPI:
 
     ../configure CXX=clang++ CXXFLAGS="-I$GridPkg/include/libomp -Xpreprocessor -fopenmp -std=c++11" LDFLAGS="-L$GridPkg/lib/libomp" LIBS="-lomp" --with-hdf5=$GridPkg --with-gmp=$GridPkg --with-mpfr=$GridPkg --with-fftw=$GridPkg --with-lime=$GridPre --enable-simd=GEN --enable-comms=mpi-auto MPICXX=$GridPre/bin/mpicxx --prefix=$GridPre/MPIDebug
 
-### 5.3 Build Grid
+### 4.3 Build Grid
 
 Each configuration must be built before they can be used. You can either:
 

documentation/manual.rst

@@ -2778,47 +2778,81 @@ and there are associated reconstruction routines for assembling four spinors fro
 
 These ca
 
-
-SU(N)
+Gauge Group
 --------
+A generic Nc qcd/utils/GaugeGroup.h is provided. This defines a template class that can be specialised to different gauge groups::
 
-A generic Nc qcd/utils/SUn.h is provided. This defines a template class::
+  template <int ncolour, class group_name>
+  class GaugeGroup {...}
 
-  template <int ncolour> class SU ;
+Supported groups are SU(N) and Sp(2N). The group can be specified through the GroupName namespace::
 
-The most important external methods are::
+  namespace GroupName {
+  class SU {};
+  class Sp {};
+  }
+
+A simpler interface is achieved by aliasing the GaugeGroup class with a specific group::
+
+  template <int ncolour>
+  using SU = GaugeGroup<ncolour, GroupName::SU>;
+
+  template <int ncolour>
+  using Sp = GaugeGroup<ncolour, GroupName::Sp>;
+  
+Specific aliases are then defined::
+
+  typedef SU<2> SU2;
+  typedef SU<3> SU3;
+  typedef SU<4> SU4;
+  typedef SU<5> SU5;
+  typedef Sp<2> Sp2;
+  typedef Sp<4> Sp4;
+  typedef Sp<6> Sp6;
+  typedef Sp<8> Sp8;
+
+Some methods are common to both gauge groups. Common external methods are::
 
-  static void printGenerators(void) ;
   template <class cplx>  static void generator(int lieIndex, iSUnMatrix<cplx> &ta) ;
+  static void GaussianFundamentalLieAlgebraMatrix(GridParallelRNG &pRNG, LatticeMatrix &out, Real scale = 1.0) ;
+  static void HotConfiguration(GridParallelRNG &pRNG, GaugeField &out) ;
+  static void TepidConfiguration(GridParallelRNG &pRNG,GaugeField &out);
+  static void ColdConfiguration(GaugeField &out);
+  static void taProj( const LatticeMatrixType &in,  LatticeMatrixType &out);
+  static void taExp(const LatticeMatrixType &x, LatticeMatrixType &ex) ;
+  static void printGenerators(void) ;
+   
+Whenever needed, a different implementation of these methods for the gauge groups is achieved by overloading. For example,::
+
+  template <typename LatticeMatrixType> //  shared interface for the traceless-antihermitian projection
+  static void taProj(const LatticeMatrixType &in, LatticeMatrixType &out) {
+    taProj(in, out, group_name());
+  }
+  
+  template <typename LatticeMatrixType> //  overloaded function to SU(N) simply perform Ta
+  static void taProj(const LatticeMatrixType &in, LatticeMatrixType &out, GroupName::SU) {
+    out = Ta(in);
+  }
+  
+  template <typename LatticeMatrixType> //  overloaded function to Sp(2N) must use a modified Ta function
+  static void taProj(const LatticeMatrixType &in, LatticeMatrixType &out, GroupName::Sp) {
+    out = SpTa(in);
+  }
+
+Gauge Group: SU(N)
+--------
+The specialisation of GaugeGroup to SU(N), formally part of qcd/utils/GaugeGroup.h, is found in the file qcd/utils/SUn.impl
+It contains methods that are only implemented for SU(N), and specialisations of shared methods to the special unitary group
+
+Public methods are::
 
   static void SubGroupHeatBath(GridSerialRNG &sRNG, GridParallelRNG &pRNG, RealD beta,  // coeff multiplying staple in action (with no 1/Nc)
                                LatticeMatrix &link,
 			       const LatticeMatrix &barestaple,  // multiplied by action coeffs so th
 			       int su2_subgroup, int nheatbath, LatticeInteger &wheremask);
-
-  static void GaussianFundamentalLieAlgebraMatrix(GridParallelRNG &pRNG,
-                                                  LatticeMatrix &out,
-                                                  Real scale = 1.0) ;
   static void GaugeTransform( GaugeField &Umu, GaugeMat &g)
   static void RandomGaugeTransform(GridParallelRNG &pRNG, GaugeField &Umu, GaugeMat &g);
 
-  static void HotConfiguration(GridParallelRNG &pRNG, GaugeField &out) ;
-  static void TepidConfiguration(GridParallelRNG &pRNG,GaugeField &out);
-  static void ColdConfiguration(GaugeField &out);
-
-  static void taProj( const LatticeMatrixType &in,  LatticeMatrixType &out);
-  static void taExp(const LatticeMatrixType &x, LatticeMatrixType &ex) ;
-
-  static int su2subgroups(void) ; // returns how many subgroups
-
-
-Specific instantiations are defined::
-
-	 typedef SU<2> SU2;
-	 typedef SU<3> SU3;
-	 typedef SU<4> SU4;
-	 typedef SU<5> SU5;
-
 For example, Quenched QCD updating may be run as (tests/core/Test_quenched_update.cc)::
 
   for(int sweep=0;sweep<1000;sweep++){
@@ -2857,6 +2891,16 @@ For example, Quenched QCD updating may be run as (tests/core/Test_quenched_updat
     }
   }
 
+Gauge Group: Sp(2N)
+--------
+The specialisation of GaugeGroup to Sp(2N), formally part of qcd/utils/GaugeGroup.h, is found in the file qcd/utils/Sp(2N).impl
+It contains methods that are only implemented for Sp(2N), and specialisations of shared methods to the special unitary group
+
+External methods are::
+
+  static void Omega(LatticeColourMatrixD &in) // Symplectic matrix left invariant by Sp(2N)
+
+Generation of Sp(2N) gauge fields is only supported via HMC.
 
 Space time grids
 ----------------

scripts/filelist

@@ -15,6 +15,8 @@ STAG_FERMION_FILES=`  find . -name '*.cc' -path '*/instantiation/*' -path '*/ins
 GP_FERMION_FILES=`    find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/Gparity*' `
 ADJ_FERMION_FILES=`   find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/WilsonAdj*' `
 TWOIND_FERMION_FILES=`find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/WilsonTwoIndex*'`
+SP_FERMION_FILES=`find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/SpWilsonImpl*'`
+SP_TWOIND_FERMION_FILES=`find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/SpWilsonTwo*'`
 
 HPPFILES=`find . -type f -name '*.hpp'`
 echo HFILES=$HFILES $HPPFILES > Make.inc
@@ -27,13 +29,14 @@ echo STAG_FERMION_FILES=$STAG_FERMION_FILES   >> Make.inc
 echo GP_FERMION_FILES=$GP_FERMION_FILES   >> Make.inc
 echo ADJ_FERMION_FILES=$ADJ_FERMION_FILES   >> Make.inc
 echo TWOIND_FERMION_FILES=$TWOIND_FERMION_FILES   >> Make.inc
+echo SP_FERMION_FILES=$SP_FERMION_FILES >> Make.inc
+echo SP_TWOIND_FERMION_FILES=$SP_TWOIND_FERMION_FILES >> Make.inc
 
 # tests Make.inc
 cd $home/tests
 dirs=`find . -type d -not -path '*/\.*'`
 for subdir in $dirs; do
     cd $home/tests/$subdir
-    pwd
     TESTS=`ls T*.cc`
     TESTLIST=`echo ${TESTS} | sed s/.cc//g `
     PREF=`[ $subdir = '.' ] && echo noinst || echo EXTRA`

systems/Frontier/benchmarks/bench2.slurm

@@ -0,0 +1,43 @@
+#!/bin/bash -l
+#SBATCH --job-name=bench
+##SBATCH --partition=small-g
+#SBATCH --nodes=2
+#SBATCH --ntasks-per-node=8
+#SBATCH --cpus-per-task=7
+#SBATCH --gpus-per-node=8
+#SBATCH --time=00:10:00
+#SBATCH --account=phy157_dwf
+#SBATCH --gpu-bind=none
+#SBATCH --exclusive
+#SBATCH --mem=0
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3 7 6 5 4)
+export NUMA_MAP=(3 3 1 1 2 2 0 0)
+export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
+export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
+export HIP_VISIBLE_DEVICES=\$GPU
+unset ROCR_VISIBLE_DEVICES
+echo RANK \$SLURM_LOCALID using GPU \$GPU    
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./select_gpu
+
+root=$HOME/Frontier/Grid/systems/Frontier/
+source ${root}/sourceme.sh
+
+export OMP_NUM_THREADS=7
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+
+for vol in 32.32.32.64
+do
+srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-overlap --shm 2048 --shm-mpi 0 --grid $vol  > log.shm0.ov.$vol
+srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-overlap --shm 2048 --shm-mpi 1 --grid $vol  > log.shm1.ov.$vol
+
+srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-sequential --shm 2048 --shm-mpi 0 --grid $vol  > log.shm0.seq.$vol
+srun ./select_gpu ./Benchmark_dwf_fp32 --mpi 2.2.2.2 --accelerator-threads 8 --comms-sequential --shm 2048 --shm-mpi 1 --grid $vol > log.shm1.seq.$vol
+done
+

systems/Frontier/config-command

@@ -0,0 +1,23 @@
+CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
+../../configure --enable-comms=mpi-auto \
+--with-lime=$CLIME \
+--enable-unified=no \
+--enable-shm=nvlink \
+--enable-tracing=timer \
+--enable-accelerator=hip \
+--enable-gen-simd-width=64 \
+--disable-gparity \
+--disable-fermion-reps \
+--enable-simd=GPU \
+--enable-accelerator-cshift \
+--with-gmp=$OLCF_GMP_ROOT \
+--with-fftw=$FFTW_DIR/.. \
+--with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
+--disable-fermion-reps \
+CXX=hipcc MPICXX=mpicxx \
+CXXFLAGS="-fPIC -I{$ROCM_PATH}/include/ -std=c++14 -I${MPICH_DIR}/include -L/lib64 " \
+ LDFLAGS="-L/lib64 -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64 "
+
+
+
+

systems/Frontier/mpiwrapper.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+lrank=$SLURM_LOCALID
+lgpu=(0 1 2 3 7 6 5 4)
+
+export ROCR_VISIBLE_DEVICES=${lgpu[$lrank]}
+
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES "
+
+$*
+
+
+

systems/Frontier/sourceme.sh

@@ -0,0 +1,13 @@
+. /autofs/nccs-svm1_home1/paboyle/Crusher/Grid/spack/share/spack/setup-env.sh
+spack load c-lime
+#export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/sw/crusher/spack-envs/base/opt/cray-sles15-zen3/gcc-11.2.0/gperftools-2.9.1-72ubwtuc5wcz2meqltbfdb76epufgzo2/lib
+module load emacs 
+module load PrgEnv-gnu
+module load rocm
+module load cray-mpich/8.1.23
+module load gmp
+module load cray-fftw
+module load craype-accel-amd-gfx90a
+export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
+#Hack for lib
+#export LD_LIBRARY_PATH=`pwd`:$LD_LIBRARY_PATH

systems/Frontier/wrap.sh

@@ -0,0 +1,9 @@
+#!/bin/sh
+
+export HIP_VISIBLE_DEVICES=$ROCR_VISIBLE_DEVICES
+unset ROCR_VISIBLE_DEVICES
+
+#rank=$SLURM_PROCID
+#rocprof -d rocprof.$rank -o rocprof.$rank/results.rank$SLURM_PROCID.csv --sys-trace $@
+
+$@

systems/OEM/README

@@ -0,0 +1,53 @@
+1. Prerequisites:
+===================
+Make sure you have the latest Intel ipcx release loaded (via modules or similar)
+Make sure you have SYCL aware MPICH or Intel MPI loaded (assumed as mpicxx)
+
+2. Obtain Grid:
+===================
+
+bash$
+git clone https://github.com/paboyle/Grid
+cd Grid
+./bootstrap.sh
+cd systems/PVC
+
+3. Build Grid:
+===================
+
+Here, configure command is stored in file config-command:
+
+bash$
+../../configure \
+	--enable-simd=GPU \
+	--enable-gen-simd-width=64 \
+	--enable-comms=mpi-auto \
+	--enable-accelerator-cshift \
+	--disable-gparity \
+	--disable-fermion-reps \
+	--enable-shm=nvlink \
+	--enable-accelerator=sycl \
+	--enable-unified=no \
+	MPICXX=mpicxx \
+	CXX=icpx \
+	LDFLAGS="-fiopenmp  -fsycl -fsycl-device-code-split=per_kernel -fsycl-device-lib=all -lze_loader " \
+	CXXFLAGS="-fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare "
+
+make all
+
+4. Run a benchmark:
+===================
+
+*** Assumes interactive access to node. ***
+
+run Benchmark_dwf_fp32 using benchmarks/bench.sh
+
+bash$
+cd benchmarks
+./bench.sh
+
+
+
+
+
+

systems/OEM/benchmarks/bench.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+export EnableImplicitScaling=0
+export ZE_ENABLE_PCI_ID_DEVICE_ORDER=1
+export ZE_AFFINITY_MASK=$gpu_id.$tile_id
+export ONEAPI_DEVICE_FILTER=gpu,level_zero
+export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=0
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:2
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
+
+mpiexec -launcher ssh -n 1 -host localhost ./select_gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads 16 --shm-mpi 1 --shm 2048 --device-mem 32768 | tee 1tile.log
+mpiexec -launcher ssh -n 2 -host localhost ./select_gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads 16 --shm-mpi 1 --shm 2048 --device-mem 32768 | tee 2tile.log
+
+#mpiexec -launcher ssh -n 4 -host localhost ./select_gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.2.2 --grid 16.16.64.64 --accelerator-threads 16 --shm-mpi 0 --shm 2048 --device-mem 32768 | tee 4tile.log
+#mpiexec -launcher ssh -n 8 -host localhost ./select_gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.2.4 --grid 16.16.64.128 --accelerator-threads 16 --shm-mpi 0 --shm 2048 --device-mem 32768 | tee 8tile.log
+
+

systems/OEM/benchmarks/select_gpu.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+num_tile=2
+
+gpu_id=$(( (MPI_LOCAL_RANKID % num_tile ) ))
+tile_id=$((MPI_LOCAL_RANKID / num_tile))
+
+export ZE_AFFINITY_MASK=$gpu_id.$tile_id
+
+echo "local rank $MPI_LOCALRANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK"
+
+"$@"
+