Better messaging

code

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)
-  { 
-    verbose=0;
+      _FineLinop(FineLinop),
+      _Smoother(Smoother)
+  {
+    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;
+
+    std::cout<<GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" target rsq "<<rsq<<" ssq "<<ssq<<std::endl;
     
-      int peri_k  = k % mmax;
-      int peri_kp = (k+1) % mmax;
+    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[peri_k],x);
-      RealD rn = axpy_norm(r,-a,mmp[peri_k],r);
+      axpy(x,a,p,x);
+      RealD rn = axpy_norm(r,-a,mmp,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]);
-	
-	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;
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: Pcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+
+	_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);
-    
-    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;
+    Field r(this->grid);
+    Field mmp(this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
+
+    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 +1,449 @@
+/*************************************************************************************
+
+    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 iVector<CComplex,nbasis >  Cvec;
+  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;
+  std::vector<CoarseVector> MultTemporaries;
+
+  ///////////////////////
+  // 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; RealD ttemps=0; RealD tcopy=0;
+
+    ttot=-usecond();
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+    CoarseVector tin=in;
+
+    texch-=usecond();
+    CoarseVector pin = Cell.ExchangePeriodic(tin);
+    texch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+    
+    int64_t osites=pin.Grid()->oSites();
+
+    RealD flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    RealD bytes = 1.0*osites*sizeof(siteMatrix)*npoint
+                + 2.0*osites*sizeof(siteVector)*npoint;
+      
+    {
+      tviews-=usecond();
+      autoView( in_v , pin, AcceleratorRead);
+      autoView( out_v , pout, AcceleratorWriteDiscard);
+      autoView( Stencil_v  , Stencil, AcceleratorRead);
+      tviews+=usecond();
+
+      // Static and prereserve to keep UVM region live and not resized across multiple calls
+      ttemps-=usecond();
+      MultTemporaries.resize(npoint,pin.Grid());       
+      ttemps+=usecond();
+      std::vector<Aview> AcceleratorViewContainer_h;
+      std::vector<Vview> AcceleratorVecViewContainer_h; 
+
+      tviews-=usecond();
+      for(int p=0;p<npoint;p++) {
+	AcceleratorViewContainer_h.push_back(      A[p].View(AcceleratorRead));
+	AcceleratorVecViewContainer_h.push_back(MultTemporaries[p].View(AcceleratorWrite));
+      }
+      tviews+=usecond();
+
+      static deviceVector<Aview> AcceleratorViewContainer; AcceleratorViewContainer.resize(npoint);
+      static deviceVector<Vview> AcceleratorVecViewContainer; AcceleratorVecViewContainer.resize(npoint); 
+      
+      auto Aview_p = &AcceleratorViewContainer[0];
+      auto Vview_p = &AcceleratorVecViewContainer[0];
+      tcopy-=usecond();
+      acceleratorCopyToDevice(&AcceleratorViewContainer_h[0],&AcceleratorViewContainer[0],npoint *sizeof(Aview));
+      acceleratorCopyToDevice(&AcceleratorVecViewContainer_h[0],&AcceleratorVecViewContainer[0],npoint *sizeof(Vview));
+      tcopy+=usecond();
+
+      tmult-=usecond();
+      accelerator_for(spb, osites*nbasis*npoint, Nsimd, {
+	  typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+	  int32_t ss   = spb/(nbasis*npoint);
+	  int32_t bp   = spb%(nbasis*npoint);
+	  int32_t b    = bp/npoint;
+	  int32_t point= bp%npoint;
+	  auto SE  = Stencil_v.GetEntry(point,ss);
+	  auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
+	  auto res = coalescedRead(Aview_p[point][ss](b,0))*nbr(0);
+	  for(int bb=1;bb<nbasis;bb++) {
+	    res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
+	  }
+	  coalescedWrite(Vview_p[point][ss](b),res);
+      });
+      accelerator_for(sb, osites*nbasis, Nsimd, {
+	  int ss = sb/nbasis;
+	  int b  = sb%nbasis;
+	  auto res = coalescedRead(Vview_p[0][ss](b));
+	  for(int point=1;point<npoint;point++){
+	    res = res + coalescedRead(Vview_p[point][ss](b));
+	  }
+	  coalescedWrite(out_v[ss](b),res);
+      });
+      tmult+=usecond();
+      for(int p=0;p<npoint;p++) {
+	AcceleratorViewContainer_h[p].ViewClose();
+	AcceleratorVecViewContainer_h[p].ViewClose();
+      }
+    }
+
+    text-=usecond();
+    out = Cell.Extract(pout);
+    text+=usecond();
+    ttot+=usecond();
+    
+    std::cout << GridLogPerformance<<"Coarse Mult Aviews "<<tviews<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult exch "<<texch<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult ext  "<<text<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult temps "<<ttemps<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult copy  "<<tcopy<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult tot  "<<ttot<<" us"<<std::endl;
+    //    std::cout << GridLogPerformance<<std::endl;
+    //    std::cout << GridLogPerformance<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogPerformance<<"Coarse Kernel bytes/s"<< bytes/tmult<<" MB/s"<<std::endl;
+    //    std::cout << GridLogPerformance<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogPerformance<<"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/allocator/MemoryManager.h

@@ -209,9 +209,9 @@ private:
   static void     CpuViewClose(uint64_t Ptr);
   static uint64_t CpuViewOpen(uint64_t  CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
 #endif
-  static void NotifyDeletion(void * CpuPtr);
 
  public:
+  static void NotifyDeletion(void * CpuPtr);
   static void Print(void);
   static void PrintAll(void);
   static void PrintState( void* CpuPtr);

Grid/allocator/MemoryManagerCache.cc

@@ -8,7 +8,7 @@ NAMESPACE_BEGIN(Grid);
 static char print_buffer [ MAXLINE ];
 
 #define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug << print_buffer;
 //#define dprintf(...) 
 
 
@@ -111,7 +111,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -141,7 +141,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld\n",
 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
   if (AccCache.accLock!=0) return;
@@ -155,7 +155,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)NULL;
     AccCache.state=CpuDirty; // CPU primary now
     DeviceBytes   -=AccCache.bytes;
-    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   //  uint64_t CpuPtr = AccCache.CpuPtr;
   DeviceEvictions++;
@@ -169,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -184,7 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone AccPtr %lx <- CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;

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/Communicator_base.h

@@ -138,6 +138,14 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
+  void CommsComplete(std::vector<CommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+			   void *xmit,
+			   int dest,
+			   void *recv,
+			   int from,
+			   int bytes,int dir);
+  
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,

Grid/communicator/Communicator_mpi3.cc

@@ -306,6 +306,44 @@ void CartesianCommunicator::GlobalSumVector(double *d,int N)
   int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
   assert(ierr==0);
 }
+
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes,int dir)
+{
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  assert(dest != _processor);
+  assert(from != _processor);
+
+  int tag;
+
+  tag= dir+from*32;
+  int ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator,&rrq);
+  assert(ierr==0);
+  list.push_back(rrq);
+  
+  tag= dir+_processor*32;
+  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator,&xrq);
+  assert(ierr==0);
+  list.push_back(xrq);
+}
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
+{
+  int nreq=list.size();
+
+  if (nreq==0) return;
+
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+  list.resize(0);
+}
+
 // Basic Halo comms primitive
 void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,

Grid/communicator/Communicator_none.cc

@@ -91,6 +91,17 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes,int dir)
+{
+  assert(0);
+}
+
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   bcopy(in,out,bytes*words);

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_crc.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
+template<class vobj> void DumpSliceNorm(std::string s,const Lattice<vobj> &f,int mu=-1)
 {
   auto ff = localNorm2(f);
   if ( mu==-1 ) mu = f.Grid()->Nd()-1;

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_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;
@@ -1054,7 +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

@@ -45,6 +45,170 @@ struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::ve
   typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
 };  
 
+
+/*
+ *
+ * TODO: 
+ *  -- address elementsof vobj via thread block in Scatter/Gather
+ *  -- overlap comms with motion in Face_exchange
+ *
+ */
+
+template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
+					      Lattice<vobj> &lat,
+					      int x,
+					      int dim,
+					      int offset=0)
+{
+  const int Nsimd=vobj::Nsimd();
+  typedef typename vobj::scalar_object sobj;
+
+  GridBase *grid = lat.Grid();
+  Coordinate simd = grid->_simd_layout;
+  int Nd          = grid->Nd();
+  int block       = grid->_slice_block[dim];
+  int stride      = grid->_slice_stride[dim];
+  int nblock      = grid->_slice_nblock[dim];
+  int rd          = grid->_rdimensions[dim];
+
+  int ox = x%rd;
+  int ix = x/rd;
+
+  int isites = 1; for(int d=0;d<Nd;d++) if( d!=dim) isites*=simd[d];
+
+  Coordinate rsimd= simd;  rsimd[dim]=1; // maybe reduce Nsimd
+
+  int rNsimd = 1; for(int d=0;d<Nd;d++) rNsimd*=rsimd[d];
+  int rNsimda= Nsimd/simd[dim]; // should be equal
+  assert(rNsimda==rNsimd);
+  int face_ovol=block*nblock;
+
+  //  assert(buf.size()==face_ovol*rNsimd);
+
+  /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
+  //Let's make it work on GPU and then make a special accelerator_for that
+  //doesn't hide the SIMD direction and keeps explicit in the threadIdx
+  //for cross platform
+  // FIXME -- can put internal indices into thread loop
+  auto buf_p = & buf[0];
+  autoView(lat_v, lat, AcceleratorRead);
+  accelerator_for(ss, face_ovol/simd[dim],Nsimd,{
+
+    // scalar layout won't coalesce
+    int blane=acceleratorSIMTlane(Nsimd); // buffer lane
+    int olane=blane%rNsimd;               // reduced lattice lane
+    int obit =blane/rNsimd;
+
+    ///////////////////////////////////////////////////////////////
+    // osite -- potentially one bit from simd in the buffer: (ss<<1)|obit
+    ///////////////////////////////////////////////////////////////
+    int ssp = ss*simd[dim]+obit;
+    int b    = ssp%block;
+    int n    = ssp/block;
+    int osite= b+n*stride + ox*block;
+
+    ////////////////////////////////////////////
+    // isite -- map lane within buffer to lane within lattice
+    ////////////////////////////////////////////
+    Coordinate icoor;
+    int lane;
+    Lexicographic::CoorFromIndex(icoor,olane,rsimd);
+    icoor[dim]=ix;
+    Lexicographic::IndexFromCoor(icoor,lane,simd);
+
+    ///////////////////////////////////////////
+    // Transfer into lattice - will coalesce
+    ///////////////////////////////////////////
+    sobj obj = extractLane(blane,buf_p[ss+offset]);
+    insertLane(lane,lat_v[osite],obj);
+  });
+}
+
+template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
+					     const Lattice<vobj> &lat,
+					     int x,
+					     int dim,
+					     int offset=0)
+{
+  const int Nsimd=vobj::Nsimd();
+  typedef typename vobj::scalar_object sobj;
+
+  autoView(lat_v, lat, AcceleratorRead);
+
+  GridBase *grid = lat.Grid();
+  Coordinate simd = grid->_simd_layout;
+  int Nd          = grid->Nd();
+  int block       = grid->_slice_block[dim];
+  int stride      = grid->_slice_stride[dim];
+  int nblock      = grid->_slice_nblock[dim];
+  int rd          = grid->_rdimensions[dim];
+
+  int ox = x%rd;
+  int ix = x/rd;
+
+  int isites = 1; for(int d=0;d<Nd;d++) if( d!=dim) isites*=simd[d];
+
+  Coordinate rsimd= simd;  rsimd[dim]=1; // maybe reduce Nsimd
+
+  int rNsimd = 1; for(int d=0;d<Nd;d++) rNsimd*=rsimd[d];
+  
+  int face_ovol=block*nblock;
+
+  //  assert(buf.size()==face_ovol*rNsimd);
+
+  /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
+  //Let's make it work on GPU and then make a special accelerator_for that
+  //doesn't hide the SIMD direction and keeps explicit in the threadIdx
+  //for cross platform
+  //For CPU perhaps just run a loop over Nsimd
+  auto buf_p = & buf[0];
+  accelerator_for(ss, face_ovol/simd[dim],Nsimd,{
+
+    // scalar layout won't coalesce
+    int blane=acceleratorSIMTlane(Nsimd); // buffer lane
+    int olane=blane%rNsimd;               // reduced lattice lane
+    int obit =blane/rNsimd;
+
+    ////////////////////////////////////////////
+    // osite
+    ////////////////////////////////////////////
+    int ssp = ss*simd[dim]+obit;
+    int b    = ssp%block;
+    int n    = ssp/block;
+    int osite= b+n*stride + ox*block;
+
+    ////////////////////////////////////////////
+    // isite -- map lane within buffer to lane within lattice
+    ////////////////////////////////////////////
+    Coordinate icoor;
+    int lane;
+    Lexicographic::CoorFromIndex(icoor,olane,rsimd);
+    icoor[dim]=ix;
+    Lexicographic::IndexFromCoor(icoor,lane,simd);
+
+    ///////////////////////////////////////////
+    // Take out of lattice
+    ///////////////////////////////////////////
+
+    sobj obj = extractLane(lane,lat_v[osite]);
+    insertLane(blane,buf_p[ss+offset],obj);
+
+  });
+  /*
+  int words =block*nblock/simd[dim];
+  std::vector<vobj> tbuf(words);
+  acceleratorCopyFromDevice((void *)&buf[offset],(void *)&tbuf[0],words*sizeof(vobj));
+  typedef typename vobj::scalar_type scalar;
+  scalar *sbuf = (scalar *)&tbuf[0];
+  scalar tmp=0.0;
+  for(int w=0;w<words*sizeof(vobj)/sizeof(scalar);w++){
+    tmp=tmp+conjugate(sbuf[w])*sbuf[w];
+  }
+  std::cout << " Gathered buffer norm "<<tmp<<std::endl;
+  */
+}
+
+
 class PaddedCell {
 public:
   GridCartesian * unpadded_grid;
@@ -63,8 +227,9 @@ 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)
@@ -81,27 +246,36 @@ public:
     Coordinate processors=unpadded_grid->_processors;
     Coordinate plocal    =unpadded_grid->LocalDimensions();
     Coordinate global(dims);
-
+    GridCartesian *old_grid = unpadded_grid;
     // 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];
+	}
 
-      for(int d=0;d<dims;d++){
-	global[d] = plocal[d]*processors[d];
+	old_grid = new GridCartesian(global,simd,processors);
       }
-
-      grids.push_back(new GridCartesian(global,simd,processors));
+      grids.push_back(old_grid);
     }
   };
   template<class vobj>
   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;
   }
@@ -116,10 +290,22 @@ public:
     }
     return tmp;
   }
+  template<class vobj>
+  inline Lattice<vobj> ExchangePeriodic(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 = ExpandPeriodic(d,tmp,cshift); // rvalue && assignment
+    }
+    return tmp;
+  }
   // expand up one dim at a time
   template<class vobj>
   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);
@@ -129,46 +315,251 @@ 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;
+    //    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
 
     double tins=0, tshift=0;
-    
-    // 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;
+    int islocal = 0 ;
+    if ( processors[dim] == 1 ) islocal = 1;
 
+    if ( islocal ) {
+
+      // replace with a copy and maybe grid swizzle
+      double t = usecond();
+      padded = in;
+      tins += usecond() - t;
+      
+    } else {
+
+      //////////////////////////////////////////////
+      // Replace sequence with
+      // ---------------------
+      // (i) Gather high face(s); start comms
+      // (ii) Gather low  face(s); start comms
+      // (iii) Copy middle bit with localCopyRegion
+      // (iv) Complete high face(s), insert slice(s)
+      // (iv) Complete low  face(s), insert slice(s)
+      //////////////////////////////////////////////
+      // 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 << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
     
     return padded;
   }
 
+  template<class vobj>
+  inline Lattice<vobj> ExpandPeriodic(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);
+    Lattice<vobj> shifted(old_grid);    
+    Coordinate local     =old_grid->LocalDimensions();
+    Coordinate plocal    =new_grid->LocalDimensions();
+    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;
+    double tins=0, tshift=0;
+
+    int islocal = 0 ;
+    if ( processors[dim] == 1 ) islocal = 1;
+
+    if ( islocal ) {
+
+      // replace with a copy and maybe grid swizzle
+      double t = usecond();
+      padded = in;
+      tins += usecond() - t;
+      
+    } else {
+
+      //////////////////////////////////////////////
+      // Replace sequence with
+      // ---------------------
+      // (i) Gather high face(s); start comms
+      // (ii) Gather low  face(s); start comms
+      // (iii) Copy middle bit with localCopyRegion
+      // (iv) Complete high face(s), insert slice(s)
+      // (iv) Complete low  face(s), insert slice(s)
+      //////////////////////////////////////////////
+      Face_exchange(in,padded,dim,depth);
+    }
+    return padded;
+  }
+  template<class vobj>
+  void Face_exchange(const Lattice<vobj> &from,
+		     Lattice<vobj> &to,
+		     int dimension,int depth) const
+  {
+    typedef typename vobj::vector_type vector_type;
+    typedef typename vobj::scalar_type scalar_type;
+    typedef typename vobj::scalar_object sobj;
+
+    RealD t_gather=0.0;
+    RealD t_scatter=0.0;
+    RealD t_comms=0.0;
+    RealD t_copy=0.0;
+    
+    //    std::cout << GridLogMessage << "dimension " <<dimension<<std::endl;
+    //    DumpSliceNorm(std::string("Face_exchange from"),from,dimension);
+    GridBase *grid=from.Grid();
+    GridBase *new_grid=to.Grid();
+
+    Coordinate lds = from.Grid()->_ldimensions;
+    Coordinate nlds=   to.Grid()->_ldimensions;
+    Coordinate simd= from.Grid()->_simd_layout;
+    int ld    = lds[dimension];
+    int nld   = to.Grid()->_ldimensions[dimension];
+    const int Nsimd = vobj::Nsimd();
+
+    assert(depth<=lds[dimension]); // A must be on neighbouring node
+    assert(depth>0);   // A caller bug if zero
+    assert(ld+2*depth==nld);
+    ////////////////////////////////////////////////////////////////////////////
+    // Face size and byte calculations
+    ////////////////////////////////////////////////////////////////////////////
+    int buffer_size = 1;
+    for(int d=0;d<lds.size();d++){
+      if ( d!= dimension) buffer_size=buffer_size*lds[d];
+    }
+    buffer_size = buffer_size  / Nsimd;
+    int rNsimd = Nsimd / simd[dimension];
+    assert( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
+
+    static cshiftVector<vobj> send_buf; 
+    static cshiftVector<vobj> recv_buf;
+    send_buf.resize(buffer_size*2*depth);    
+    recv_buf.resize(buffer_size*2*depth);
+
+    std::vector<CommsRequest_t> fwd_req;   
+    std::vector<CommsRequest_t> bwd_req;   
+
+    int words = buffer_size;
+    int bytes = words * sizeof(vobj);
+
+    ////////////////////////////////////////////////////////////////////////////
+    // Communication coords
+    ////////////////////////////////////////////////////////////////////////////
+    int comm_proc = 1;
+    int xmit_to_rank;
+    int recv_from_rank;
+    grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+    ////////////////////////////////////////////////////////////////////////////
+    // Gather all surface terms up to depth "d"
+    ////////////////////////////////////////////////////////////////////////////
+    RealD t;
+    int plane=0;
+    for ( int d=0;d < depth ; d ++ ) {
+      int tag = d*1024 + dimension*2+0;
+
+      t=usecond();
+      GatherSlice(send_buf,from,d,dimension,plane*buffer_size); plane++;
+      t_gather+=usecond()-t;
+
+      t=usecond();
+      grid->SendToRecvFromBegin(fwd_req,
+				(void *)&send_buf[d*buffer_size], xmit_to_rank,
+				(void *)&recv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+      t_comms+=usecond()-t;
+     }
+    for ( int d=0;d < depth ; d ++ ) {
+      int tag = d*1024 + dimension*2+1;
+
+      t=usecond();
+      GatherSlice(send_buf,from,ld-depth+d,dimension,plane*buffer_size); plane++;
+      t_gather+= usecond() - t;
+
+      t=usecond();
+      grid->SendToRecvFromBegin(bwd_req,
+				(void *)&send_buf[(d+depth)*buffer_size], recv_from_rank,
+				(void *)&recv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+      t_comms+=usecond()-t;
+    }
+
+    ////////////////////////////////////////////////////////////////////////////
+    // Copy interior -- overlap this with comms
+    ////////////////////////////////////////////////////////////////////////////
+    int Nd = new_grid->Nd();
+    Coordinate LL(Nd,0);
+    Coordinate sz = grid->_ldimensions;
+    Coordinate toLL(Nd,0);
+    toLL[dimension]=depth;
+    t=usecond();
+    localCopyRegion(from,to,LL,toLL,sz);
+    t_copy= usecond() - t;
+    
+    ////////////////////////////////////////////////////////////////////////////
+    // Scatter all faces
+    ////////////////////////////////////////////////////////////////////////////
+    //    DumpSliceNorm(std::string("Face_exchange to before scatter"),to,dimension);
+
+    plane=0;
+
+    t=usecond();
+    grid->CommsComplete(fwd_req);
+    t_comms+= usecond() - t;
+
+    t=usecond();
+    for ( int d=0;d < depth ; d ++ ) {
+      ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++;
+    }
+    t_scatter= usecond() - t;
+
+    t=usecond();
+    grid->CommsComplete(bwd_req);
+    t_comms+= usecond() - t;
+    
+    t=usecond();
+    for ( int d=0;d < depth ; d ++ ) {
+      ScatterSlice(recv_buf,to,d,dimension,plane*buffer_size); plane++;
+    }
+    t_scatter+= usecond() - t;
+    //    DumpSliceNorm(std::string("Face_exchange to scatter 1st "),to,dimension);
+
+    //DumpSliceNorm(std::string("Face_exchange to done"),to,dimension);
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << t_gather/1000  << "ms"<<std::endl;
+    //    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << 2.0*bytes/t_gather << "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << t_scatter/1000   << "ms"<<std::endl;
+    //    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << 2.0*bytes/t_scatter<< "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: copy   :" << t_copy/1000      << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms  :" << t_comms/1000     << "ms"<<std::endl;
+    //    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms  :" << (RealD)4.0*bytes/t_comms   << "MB/s"<<std::endl;
+  }
+  
 };
  
 
 NAMESPACE_END(Grid);
 
+

Grid/parallelIO/BinaryIO.h

@@ -165,7 +165,7 @@ class BinaryIO {
 	 * FIXME -- 128^3 x 256 x 16 will overflow.
 	 */
 	
-	int global_site;
+	int64_t global_site;
 
 	Lexicographic::CoorFromIndex(coor,local_site,local_vol);
 
@@ -175,8 +175,8 @@ class BinaryIO {
 
 	Lexicographic::IndexFromCoor(coor,global_site,global_vol);
 
-	uint32_t gsite29   = global_site%29;
-	uint32_t gsite31   = global_site%31;
+	uint64_t gsite29   = global_site%29;
+	uint64_t gsite31   = global_site%31;
 	
 	site_crc = crc32(0,(unsigned char *)site_buf,sizeof(fobj));
 	//	std::cout << "Site "<<local_site << " crc "<<std::hex<<site_crc<<std::dec<<std::endl;
@@ -545,7 +545,9 @@ class BinaryIO {
 				       const std::string &format,
 				       uint32_t &nersc_csum,
 				       uint32_t &scidac_csuma,
-				       uint32_t &scidac_csumb)
+				       uint32_t &scidac_csumb,
+				       int control=BINARYIO_LEXICOGRAPHIC
+				       )
   {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
@@ -556,7 +558,7 @@ class BinaryIO {
     std::vector<sobj> scalardata(lsites); 
     std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
     
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|control,
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     GridStopWatch timer; 
@@ -582,7 +584,8 @@ class BinaryIO {
 					  const std::string &format,
 					  uint32_t &nersc_csum,
 					  uint32_t &scidac_csuma,
-					  uint32_t &scidac_csumb)
+					  uint32_t &scidac_csumb,
+					  int control=BINARYIO_LEXICOGRAPHIC)
   {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
@@ -607,7 +610,7 @@ class BinaryIO {
     while (attemptsLeft >= 0)
     {
       grid->Barrier();
-      IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
+      IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|control,
 	             nersc_csum,scidac_csuma,scidac_csumb);
       if (checkWrite)
       {
@@ -617,7 +620,7 @@ class BinaryIO {
 
         std::cout << GridLogMessage << "writeLatticeObject: read back object" << std::endl;
         grid->Barrier();
-        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|control,
 	               cknersc_csum,ckscidac_csuma,ckscidac_csumb);
         if ((cknersc_csum != nersc_csum) or (ckscidac_csuma != scidac_csuma) or (ckscidac_csumb != scidac_csumb))
         {

Grid/parallelIO/IldgIO.h

@@ -206,7 +206,7 @@ class GridLimeReader : public BinaryIO {
   // Read a generic lattice field and verify checksum
   ////////////////////////////////////////////
   template<class vobj>
-  void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
+  void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name,int control=BINARYIO_LEXICOGRAPHIC)
   {
     typedef typename vobj::scalar_object sobj;
     scidacChecksum scidacChecksum_;
@@ -238,7 +238,7 @@ class GridLimeReader : public BinaryIO {
 	uint64_t offset= ftello(File);
 	//	std::cout << " ReadLatticeObject from offset "<<offset << std::endl;
 	BinarySimpleMunger<sobj,sobj> munge;
-	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb,control);
 	std::cout << GridLogMessage << "SciDAC checksum A " << std::hex << scidac_csuma << std::dec << std::endl;
 	std::cout << GridLogMessage << "SciDAC checksum B " << std::hex << scidac_csumb << std::dec << std::endl;
 	/////////////////////////////////////////////
@@ -408,7 +408,7 @@ class GridLimeWriter : public BinaryIO
   // in communicator used by the field.Grid()
   ////////////////////////////////////////////////////
   template<class vobj>
-  void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
+  void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name,int control=BINARYIO_LEXICOGRAPHIC)
   {
     ////////////////////////////////////////////////////////////////////
     // NB: FILE and iostream are jointly writing disjoint sequences in the
@@ -459,7 +459,7 @@ class GridLimeWriter : public BinaryIO
     ///////////////////////////////////////////
     std::string format = getFormatString<vobj>();
     BinarySimpleMunger<sobj,sobj> munge;
-    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb);
+    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb,control);
 
     ///////////////////////////////////////////
     // Wind forward and close the record
@@ -512,7 +512,8 @@ class ScidacWriter : public GridLimeWriter {
   ////////////////////////////////////////////////
   template <class vobj, class userRecord>
   void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
-                              const unsigned int recordScientificPrec = 0) 
+                              const unsigned int recordScientificPrec = 0,
+			      int control=BINARYIO_LEXICOGRAPHIC)
   {
     GridBase * grid = field.Grid();
 
@@ -534,7 +535,7 @@ class ScidacWriter : public GridLimeWriter {
       writeLimeObject(0,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
     }
     // Collective call
-    writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
+    writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA),control);      // Closes message with checksum
   }
 };
 
@@ -553,7 +554,8 @@ class ScidacReader : public GridLimeReader {
   // Write generic lattice field in scidac format
   ////////////////////////////////////////////////
   template <class vobj, class userRecord>
-  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord) 
+  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord,
+			     int control=BINARYIO_LEXICOGRAPHIC) 
   {
     typedef typename vobj::scalar_object sobj;
     GridBase * grid = field.Grid();
@@ -571,7 +573,7 @@ class ScidacReader : public GridLimeReader {
     readLimeObject(header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
     readLimeObject(_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
     readLimeObject(_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
-    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));
+    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA),control);
   }
   void skipPastBinaryRecord(void) {
     std::string rec_name(ILDG_BINARY_DATA);

Grid/qcd/utils/WilsonLoops.h

@@ -487,7 +487,7 @@ public:
     for(int mu=0;mu<Nd;mu++){
       { //view scope
 	autoView( gStaple_v , gStaple, AcceleratorWrite);
-	auto gStencil_v = gStencil.View();
+	auto gStencil_v = gStencil.View(AcceleratorRead);
 	
 	accelerator_for(ss, ggrid->oSites(), ggrid->Nsimd(), {
 	    decltype(coalescedRead(Ug_dirs_v[0][0])) stencil_ss;
@@ -1199,7 +1199,7 @@ public:
 
       { //view scope
 	autoView( gStaple_v , gStaple, AcceleratorWrite);
-	auto gStencil_v = gStencil.View();
+	auto gStencil_v = gStencil.View(AcceleratorRead);
 
 	accelerator_for(ss, ggrid->oSites(), ggrid->Nsimd(), {
 	    decltype(coalescedRead(Ug_dirs_v[0][0])) stencil_ss;

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;
   }

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/Init.cc

@@ -283,6 +283,7 @@ void GridBanner(void)
     std::cout << "Build " << GRID_BUILD_STR(GRID_BUILD_REF) << std::endl;
 #endif
     std::cout << std::endl;
+    std::cout << std::setprecision(9);
 }
 
 void Grid_init(int *argc,char ***argv)

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--){

scripts/prequisites.sh

@@ -0,0 +1,44 @@
+#!/bin/bash
+
+if [ $1 = "install" ]
+then
+    dir=`pwd`
+    cd $HOME
+    git clone -c feature.manyFiles=true https://github.com/spack/spack.git
+    source $HOME/spack/share/spack/setup-env.sh
+
+    spack install autoconf
+    spack install automake
+    spack install c-lime cppflags=-fPIE
+    spack install fftw
+    spack install llvm
+    spack install gmp
+    spack install mpfr
+    spack install cuda@11.8
+    spack install openmpi
+    spack install openssl
+    spack install hdf5
+else
+    source $HOME/spack/share/spack/setup-env.sh
+fi
+
+spack load autoconf
+spack load automake
+spack load c-lime
+spack load fftw
+spack load llvm
+spack load gmp
+spack load mpfr
+spack load cuda@11.8
+spack load openmpi
+spack load openssl
+spack load hdf5
+
+export FFTW=`spack find --paths fftw    | grep ^fftw   | awk '{print $2}' `
+export HDF5=`spack find --paths hdf5    | grep ^hdf5   | awk '{print $2}' `
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+export NVIDIA=$CUDA_HOME
+export NVIDIALIB=$NVIDIA/targets/x86_64-linux/lib/
+export LD_LIBRARY_PATH=$NVIDIALIB:$FFTW/lib/:$MPFR/lib:$LD_LIBRARY_PATH

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/mac-arm/config-command-mpi

@@ -1,4 +1,3 @@
 BREW=/opt/local/
 MPICXX=mpicxx ../../configure --enable-simd=GEN --enable-comms=mpi-auto --enable-unified=yes --prefix $HOME/QCD/GridInstall --with-lime=/Users/peterboyle/QCD/SciDAC/install/ --with-openssl=$BREW --disable-fermion-reps --disable-gparity --disable-debug
 
-

tests/debug/Test_general_coarse.cc

@@ -0,0 +1,235 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+using namespace std;
+using namespace Grid;
+
+///////////////////////
+// Tells little dirac op to use MdagM as the .Op()
+///////////////////////
+template<class Field>
+class HermOpAdaptor : public LinearOperatorBase<Field>
+{
+  LinearOperatorBase<Field> & wrapped;
+public:
+  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    wrapped.HermOp(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    wrapped.HermOp(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    wrapped.HermOp(in,out);
+  }
+};
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=4;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplex::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  // Construct a coarsened grid
+  Coordinate clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/2;
+  }
+
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt,
+							    GridDefaultSimd(Nd,vComplex::Nsimd()),
+							    GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  LatticeFermion    src(FGrid); random(RNG5,src);
+  LatticeFermion result(FGrid); result=Zero();
+  LatticeFermion    ref(FGrid); ref=Zero();
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+  LatticeGaugeField Umu(UGrid);
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+  //  Umu=Zero();
+  
+  RealD mass=0.1;
+  RealD M5=1.8;
+
+  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+
+  const int nbasis = 16;
+  const int cb = 0 ;
+  LatticeFermion prom(FGrid);
+
+  std::vector<LatticeFermion> subspace(nbasis,FGrid);
+
+  std::cout<<GridLogMessage<<"Calling Aggregation class" <<std::endl;
+
+  ///////////////////////////////////////////////////////////
+  // Squared operator is in HermOp
+  ///////////////////////////////////////////////////////////
+  MdagMLinearOperator<DomainWallFermionD,LatticeFermion> HermDefOp(Ddwf);
+
+  ///////////////////////////////////////////////////
+  // Random aggregation space
+  ///////////////////////////////////////////////////
+  std::cout<<GridLogMessage << "Building random aggregation class"<< std::endl;
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
+  Subspace Aggregates(Coarse5d,FGrid,cb);
+  Aggregates.CreateSubspaceRandom(RNG5);
+
+  ///////////////////////////////////////////////////
+  // Build little dirac op
+  ///////////////////////////////////////////////////
+  std::cout<<GridLogMessage << "Building little Dirac operator"<< std::endl;
+
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNearestStencilGeometry5D geom(Coarse5d);
+  LittleDiracOperator LittleDiracOp(geom,FGrid,Coarse5d);
+  LittleDiracOperator LittleDiracOpCol(geom,FGrid,Coarse5d);
+
+  HermOpAdaptor<LatticeFermionD> HOA(HermDefOp);
+
+  int pp=16;
+  LittleDiracOp.CoarsenOperator(HOA,Aggregates);
+  
+  ///////////////////////////////////////////////////
+  // Test the operator
+  ///////////////////////////////////////////////////
+  CoarseVector c_src (Coarse5d);
+  CoarseVector c_res (Coarse5d);
+  CoarseVector c_res_dag(Coarse5d);
+  CoarseVector c_proj(Coarse5d);
+
+  subspace=Aggregates.subspace;
+
+  //  random(CRNG,c_src);
+  c_src = 1.0;
+
+  blockPromote(c_src,err,subspace);
+
+  prom=Zero();
+  for(int b=0;b<nbasis;b++){
+    prom=prom+subspace[b];
+  }
+  err=err-prom; 
+  std::cout<<GridLogMessage<<"Promoted back from subspace: err "<<norm2(err)<<std::endl;
+  std::cout<<GridLogMessage<<"c_src "<<norm2(c_src)<<std::endl;
+  std::cout<<GridLogMessage<<"prom  "<<norm2(prom)<<std::endl;
+
+  HermDefOp.HermOp(prom,tmp);
+
+  blockProject(c_proj,tmp,subspace);
+  std::cout<<GridLogMessage<<" Called Big Dirac Op "<<norm2(tmp)<<std::endl;
+
+  std::cout<<GridLogMessage<<" Calling little Dirac Op "<<std::endl;
+  LittleDiracOp.M(c_src,c_res);
+  LittleDiracOp.Mdag(c_src,c_res_dag);
+
+  std::cout<<GridLogMessage<<"Little dop : "<<norm2(c_res)<<std::endl;
+  std::cout<<GridLogMessage<<"Little dop dag : "<<norm2(c_res_dag)<<std::endl;
+  std::cout<<GridLogMessage<<"Big dop in subspace : "<<norm2(c_proj)<<std::endl;
+
+  c_proj = c_proj - c_res;
+  std::cout<<GridLogMessage<<" ldop error: "<<norm2(c_proj)<<std::endl;
+
+  c_res_dag = c_res_dag - c_res;
+  std::cout<<GridLogMessage<<"Little dopDag - dop: "<<norm2(c_res_dag)<<std::endl;
+
+  std::cout<<GridLogMessage << "Testing Hermiticity stochastically "<< std::endl;
+  CoarseVector phi(Coarse5d);
+  CoarseVector chi(Coarse5d);
+  CoarseVector Aphi(Coarse5d);
+  CoarseVector Achi(Coarse5d);
+
+  random(CRNG,phi);
+  random(CRNG,chi);
+
+  std::cout<<GridLogMessage<<"Made randoms "<<norm2(phi)<<" " << norm2(chi)<<std::endl;
+
+  LittleDiracOp.M(phi,Aphi);
+
+  LittleDiracOp.Mdag(chi,Achi);
+
+  std::cout<<GridLogMessage<<"Aphi "<<norm2(Aphi)<<" A chi" << norm2(Achi)<<std::endl;
+
+  ComplexD pAc = innerProduct(chi,Aphi);
+  ComplexD cAp = innerProduct(phi,Achi);
+  ComplexD cAc = innerProduct(chi,Achi);
+  ComplexD pAp = innerProduct(phi,Aphi);
+
+  std::cout<<GridLogMessage<< "pAc "<<pAc<<" cAp "<< cAp<< " diff "<<pAc-adj(cAp)<<std::endl;
+  std::cout<<GridLogMessage<< "pAp "<<pAp<<" cAc "<< cAc<<"Should be real"<< std::endl;
+
+  std::cout<<GridLogMessage<<"Testing linearity"<<std::endl;
+  CoarseVector PhiPlusChi(Coarse5d);
+  CoarseVector APhiPlusChi(Coarse5d);
+  CoarseVector linerr(Coarse5d);
+  PhiPlusChi = phi+chi;
+  LittleDiracOp.M(PhiPlusChi,APhiPlusChi);
+
+  linerr= APhiPlusChi-Aphi;
+  linerr= linerr-Achi;
+  std::cout<<GridLogMessage<<"**Diff "<<norm2(linerr)<<std::endl;
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  
+  Grid_finalize();
+  return 0;
+}

tests/debug/Test_general_coarse_hdcg.cc

@@ -0,0 +1,426 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_general_coarse_hdcg.cc
+
+    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 */
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+//#include <Grid/algorithms/GeneralCoarsenedMatrix.h>
+#include <Grid/algorithms/iterative/AdefGeneric.h>
+
+using namespace std;
+using namespace Grid;
+
+template<class Coarsened>
+void SaveOperator(Coarsened &Operator,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Operator.Grid()->IsBoss());
+  assert(Operator._A.size()==Operator.geom.npoint);
+  WR.open(file);
+  for(int p=0;p<Operator._A.size();p++){
+    auto tmp = Operator.Cell.Extract(Operator._A[p]);
+    WR.writeScidacFieldRecord(tmp,record);
+  }
+  WR.close();
+#endif
+}
+template<class Coarsened>
+void LoadOperator(Coarsened &Operator,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  Grid::ScidacReader RD ;
+  RD.open(file);
+  assert(Operator._A.size()==Operator.geom.npoint);
+  for(int p=0;p<Operator.geom.npoint;p++){
+    conformable(Operator._A[p].Grid(),Operator.CoarseGrid());
+    RD.readScidacFieldRecord(Operator._A[p],record);
+  }    
+  RD.close();
+  Operator.ExchangeCoarseLinks();
+#endif
+}
+template<class aggregation>
+void SaveBasis(aggregation &Agg,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Agg.FineGrid->IsBoss());
+  WR.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    WR.writeScidacFieldRecord(Agg.subspace[b],record);
+  }
+  WR.close();
+#endif
+}
+template<class aggregation>
+void LoadBasis(aggregation &Agg, std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  RD.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    RD.readScidacFieldRecord(Agg.subspace[b],record);
+  }    
+  RD.close();
+#endif
+}
+
+
+template<class Field> class TestSolver : public LinearFunction<Field> {
+public:
+  TestSolver() {};
+  void operator() (const Field &in, Field &out){    out = Zero();  }     
+};
+
+
+RealD InverseApproximation(RealD x){
+  return 1.0/x;
+}
+
+// Want Op in CoarsenOp to call MatPcDagMatPc
+template<class Field>
+class HermOpAdaptor : public LinearOperatorBase<Field>
+{
+  LinearOperatorBase<Field> & wrapped;
+public:
+  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
+  void Op     (const Field &in, Field &out)   { wrapped.HermOp(in,out);  }
+  void HermOp(const Field &in, Field &out)    { wrapped.HermOp(in,out); }
+  void AdjOp     (const Field &in, Field &out){ wrapped.HermOp(in,out);  }
+  void OpDiag (const Field &in, Field &out)                  {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out)  {    assert(0);  };
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+};
+template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _SmootherOperator;
+  Chebyshev<Field> Cheby;
+  ChebyshevSmoother(RealD _lo,RealD _hi,int _ord, FineOperator &SmootherOperator) :
+    _SmootherOperator(SmootherOperator),
+    Cheby(_lo,_hi,_ord,InverseApproximation)
+  {
+    std::cout << GridLogMessage<<" Chebyshev smoother order "<<_ord<<" ["<<_lo<<","<<_hi<<"]"<<std::endl;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+    Field tmp(in.Grid());
+    tmp = in;
+    Cheby(_SmootherOperator,tmp,out);
+  }
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=24;
+  const int nbasis = 40;
+  const int cb = 0 ;
+  RealD mass=0.00078;
+  RealD M5=1.8;
+  RealD b=1.5;
+  RealD c=0.5;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplex::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  // Construct a coarsened grid with 4^4 cell
+  Coordinate clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/4;
+  }
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt,
+							    GridDefaultSimd(Nd,vComplex::Nsimd()),
+							    GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  ///////////////////////// RNGs /////////////////////////////////
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  ///////////////////////// Configuration /////////////////////////////////
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  //////////////////////// Fermion action //////////////////////////////////
+  MobiusFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c);
+
+  SchurDiagMooeeOperator<MobiusFermionD, LatticeFermion> HermOpEO(Ddwf);
+
+  typedef HermOpAdaptor<LatticeFermionD> HermFineMatrix;
+  HermFineMatrix FineHermOp(HermOpEO);
+  
+  LatticeFermion result(FrbGrid); result=Zero();
+
+  LatticeFermion    src(FrbGrid); random(RNG5,src);
+
+  // Run power method on FineHermOp
+  PowerMethod<LatticeFermion>       PM;   PM(HermOpEO,src);
+
+ 
+  ////////////////////////////////////////////////////////////
+  ///////////// Coarse basis and Little Dirac Operator ///////
+  ////////////////////////////////////////////////////////////
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
+  NearestStencilGeometry5D geom_nn(Coarse5d);
+  
+  // Warning: This routine calls PVdagM.Op, not PVdagM.HermOp
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
+  Subspace Aggregates(Coarse5d,FrbGrid,cb);
+
+  ////////////////////////////////////////////////////////////
+  // Need to check about red-black grid coarsening
+  ////////////////////////////////////////////////////////////
+  LittleDiracOperator LittleDiracOp(geom,FrbGrid,Coarse5d);
+
+  bool load=false;
+  if ( load ) {
+    LoadBasis(Aggregates,"/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.scidac");
+    LoadOperator(LittleDiracOp,"/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/LittleDiracOp.scidac");
+  } else {
+    Aggregates.CreateSubspaceChebyshev(RNG5,HermOpEO,nbasis,
+				       95.0,0.1,
+				       //				     400,200,200 -- 48 iters
+				       //				     600,200,200 -- 38 iters, 162s
+				       //				     600,200,100 -- 38 iters, 169s
+				       //				     600,200,50  -- 88 iters. 370s 
+				       800,
+				       200,
+				       100,
+				       0.0);
+    LittleDiracOp.CoarsenOperator(FineHermOp,Aggregates);
+    SaveBasis(Aggregates,"/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.scidac");
+    SaveOperator(LittleDiracOp,"/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/LittleDiracOp.scidac");
+  }
+  
+  // Try projecting to one hop only
+  LittleDiracOperator LittleDiracOpProj(geom_nn,FrbGrid,Coarse5d);
+  LittleDiracOpProj.ProjectNearestNeighbour(0.01,LittleDiracOp); // smaller shift 0.02? n
+
+  typedef HermitianLinearOperator<LittleDiracOperator,CoarseVector> HermMatrix;
+  HermMatrix CoarseOp     (LittleDiracOp);
+  HermMatrix CoarseOpProj (LittleDiracOpProj);
+  
+  //////////////////////////////////////////
+  // Build a coarse lanczos
+  //////////////////////////////////////////
+  Chebyshev<CoarseVector>      IRLCheby(0.2,40.0,71);  // 1 iter
+  FunctionHermOp<CoarseVector> IRLOpCheby(IRLCheby,CoarseOp);
+  PlainHermOp<CoarseVector>    IRLOp    (CoarseOp);
+  int Nk=48;
+  int Nm=64;
+  int Nstop=Nk;
+  ImplicitlyRestartedLanczos<CoarseVector> IRL(IRLOpCheby,IRLOp,Nstop,Nk,Nm,1.0e-5,20);
+
+  int Nconv;
+  std::vector<RealD>            eval(Nm);
+  std::vector<CoarseVector>     evec(Nm,Coarse5d);
+  CoarseVector c_src(Coarse5d);
+  //c_src=1.0;
+  random(CRNG,c_src);
+
+  CoarseVector c_res(Coarse5d); 
+  CoarseVector c_ref(Coarse5d); 
+
+  PowerMethod<CoarseVector>       cPM;   cPM(CoarseOp,c_src);
+
+  IRL.calc(eval,evec,c_src,Nconv);
+  DeflatedGuesser<CoarseVector> DeflCoarseGuesser(evec,eval);
+
+  //////////////////////////////////////////
+  // Build a coarse space solver
+  //////////////////////////////////////////
+  int maxit=20000;
+  ConjugateGradient<CoarseVector>  CG(1.0e-8,maxit,false);
+  ConjugateGradient<LatticeFermionD>  CGfine(1.0e-8,10000,false);
+  ZeroGuesser<CoarseVector> CoarseZeroGuesser;
+
+  //  HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,CoarseZeroGuesser);
+  HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,DeflCoarseGuesser);
+  c_res=Zero();
+  HPDSolve(c_src,c_res); c_ref = c_res;
+  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  std::cout << GridLogMessage<<"ref norm "<<norm2(c_ref)<<std::endl;
+  //////////////////////////////////////////////////////////////////////////
+  // Deflated (with real op EV's) solve for the projected coarse op
+  // Work towards ADEF1 in the coarse space
+  //////////////////////////////////////////////////////////////////////////
+  HPDSolver<CoarseVector> HPDSolveProj(CoarseOpProj,CG,DeflCoarseGuesser);
+  c_res=Zero();
+  HPDSolveProj(c_src,c_res);
+  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  std::cout << GridLogMessage<<"res norm "<<norm2(c_res)<<std::endl;
+  c_res = c_res - c_ref;
+  std::cout << "Projected solver error "<<norm2(c_res)<<std::endl;
+
+  //////////////////////////////////////////////////////////////////////
+  // Coarse ADEF1 with deflation space
+  //////////////////////////////////////////////////////////////////////
+  ChebyshevSmoother<CoarseVector,HermMatrix >
+    CoarseSmoother(1.0,37.,8,CoarseOpProj);  // just go to sloppy 0.1 convergence
+    //  CoarseSmoother(0.1,37.,8,CoarseOpProj);  //
+  //  CoarseSmoother(0.5,37.,6,CoarseOpProj);  //  8 iter 0.36s
+  //    CoarseSmoother(0.5,37.,12,CoarseOpProj);  // 8 iter, 0.55s
+  //    CoarseSmoother(0.5,37.,8,CoarseOpProj);// 7-9 iter
+  //  CoarseSmoother(1.0,37.,8,CoarseOpProj); // 0.4 - 0.5s solve to 0.04, 7-9 iter
+  //  ChebyshevSmoother<CoarseVector,HermMatrix > CoarseSmoother(0.5,36.,10,CoarseOpProj);  // 311
+
+  ////////////////////////////////////////////////////////
+  // CG, Cheby mode spacing 200,200
+  // Unprojected Coarse CG solve to 1e-8 : 190 iters, 4.9s
+  // Unprojected Coarse CG solve to 4e-2 :  33 iters, 0.8s
+  // Projected Coarse CG solve to 1e-8 : 100 iters, 0.36s
+  ////////////////////////////////////////////////////////
+  // CoarseSmoother(1.0,48.,8,CoarseOpProj); 48 evecs 
+  ////////////////////////////////////////////////////////
+  // ADEF1 Coarse solve to 1e-8 : 44 iters, 2.34s  2.1x gain
+  // ADEF1 Coarse solve to 4e-2 : 7 iters, 0.4s
+  // HDCG 38 iters 162s
+  //
+  // CoarseSmoother(1.0,40.,8,CoarseOpProj); 48 evecs 
+  // ADEF1 Coarse solve to 1e-8 : 37 iters, 2.0s  2.1x gain
+  // ADEF1 Coarse solve to 4e-2 : 6 iters, 0.36s
+  // HDCG 38 iters 169s
+
+  TwoLevelADEF1defl<CoarseVector>
+    cADEF1(1.0e-8, 500,
+	   CoarseOp,
+	   CoarseSmoother,
+	   evec,eval);
+
+  c_res=Zero();
+  cADEF1(c_src,c_res);
+  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  std::cout << GridLogMessage<<"cADEF1 res norm "<<norm2(c_res)<<std::endl;
+  c_res = c_res - c_ref;
+  std::cout << "cADEF1 solver error "<<norm2(c_res)<<std::endl;
+  
+  //  cADEF1.Tolerance = 4.0e-2;
+  //  cADEF1.Tolerance = 1.0e-1;
+  cADEF1.Tolerance = 5.0e-2;
+  c_res=Zero();
+  cADEF1(c_src,c_res);
+  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  std::cout << GridLogMessage<<"cADEF1 res norm "<<norm2(c_res)<<std::endl;
+  c_res = c_res - c_ref;
+  std::cout << "cADEF1 solver error "<<norm2(c_res)<<std::endl;
+  
+  //////////////////////////////////////////
+  // Build a smoother
+  //////////////////////////////////////////
+  //  ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(10.0,100.0,10,FineHermOp); //499
+  //  ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(3.0,100.0,10,FineHermOp);  //383
+  //  ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(1.0,100.0,10,FineHermOp);  //328
+  //  std::vector<RealD> los({0.5,1.0,3.0}); // 147/142/146 nbasis 1
+  //  std::vector<RealD> los({1.0,2.0}); // Nbasis 24: 88,86 iterations
+  //  std::vector<RealD> los({2.0,4.0}); // Nbasis 32 == 52, iters
+  //  std::vector<RealD> los({2.0,4.0}); // Nbasis 40 == 36,36 iters
+
+  //
+  // Turns approx 2700 iterations into 340 fine multiplies with Nbasis 40
+  // Need to measure cost of coarse space.
+  //
+  // -- i) Reduce coarse residual   -- 0.04
+  // -- ii) Lanczos on coarse space -- done
+  // -- iii) Possible 1 hop project and/or preconditioning it - easy - PrecCG it and
+  //         use a limited stencil. Reread BFM code to check on evecs / deflation strategy with prec
+  //
+  std::vector<RealD> los({3.0}); // Nbasis 40 == 36,36 iters
+
+  //  std::vector<int> ords({7,8,10}); // Nbasis 40 == 40,38,36 iters (320,342,396 mults)
+  std::vector<int> ords({7}); // Nbasis 40 == 40 iters (320 mults)  
+
+  for(int l=0;l<los.size();l++){
+
+    RealD lo = los[l];
+
+    for(int o=0;o<ords.size();o++){
+
+      ConjugateGradient<CoarseVector>  CGsloppy(4.0e-2,maxit,false);
+      HPDSolver<CoarseVector> HPDSolveSloppy(CoarseOp,CGsloppy,DeflCoarseGuesser);
+      
+      //    ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(lo,92,10,FineHermOp); // 36 best case
+      ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(lo,92,ords[o],FineHermOp);  // 311
+
+      //////////////////////////////////////////
+      // Build a HDCG solver
+      //////////////////////////////////////////
+      TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
+	HDCG(1.0e-8, 100,
+	     FineHermOp,
+	     Smoother,
+	     HPDSolveSloppy,
+	     HPDSolve,
+	     Aggregates);
+
+      TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
+	HDCGdefl(1.0e-8, 100,
+		 FineHermOp,
+		 Smoother,
+		 cADEF1,
+		 HPDSolve,
+		 Aggregates);
+      
+      result=Zero();
+      HDCGdefl(src,result);
+
+      result=Zero();
+      HDCG(src,result);
+
+      
+    }
+  }
+
+  // Standard CG
+  result=Zero();
+  CGfine(HermOpEO, src, result);
+  
+  Grid_finalize();
+  return 0;
+}

tests/debug/Test_general_coarse_hdcg_phys.cc

@@ -0,0 +1,429 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_general_coarse_hdcg.cc
+
+    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 */
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+//#include <Grid/algorithms/GeneralCoarsenedMatrix.h>
+#include <Grid/algorithms/iterative/AdefGeneric.h>
+
+using namespace std;
+using namespace Grid;
+
+template<class Coarsened>
+void SaveOperator(Coarsened &Operator,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Operator.Grid()->IsBoss());
+  assert(Operator._A.size()==Operator.geom.npoint);
+  WR.open(file);
+  for(int p=0;p<Operator._A.size();p++){
+    auto tmp = Operator.Cell.Extract(Operator._A[p]);
+    WR.writeScidacFieldRecord(tmp,record,0,0);
+    //    WR.writeScidacFieldRecord(tmp,record,0,BINARYIO_LEXICOGRAPHIC);
+  }
+  WR.close();
+#endif
+}
+template<class Coarsened>
+void LoadOperator(Coarsened &Operator,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  Grid::ScidacReader RD ;
+  RD.open(file);
+  assert(Operator._A.size()==Operator.geom.npoint);
+  for(int p=0;p<Operator.geom.npoint;p++){
+    conformable(Operator._A[p].Grid(),Operator.CoarseGrid());
+    //    RD.readScidacFieldRecord(Operator._A[p],record,BINARYIO_LEXICOGRAPHIC);
+    RD.readScidacFieldRecord(Operator._A[p],record,0);
+  }    
+  RD.close();
+  Operator.ExchangeCoarseLinks();
+#endif
+}
+template<class aggregation>
+void SaveBasis(aggregation &Agg,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Agg.FineGrid->IsBoss());
+  WR.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    //WR.writeScidacFieldRecord(Agg.subspace[b],record,0,BINARYIO_LEXICOGRAPHIC);
+    WR.writeScidacFieldRecord(Agg.subspace[b],record,0,0);
+  }
+  WR.close();
+#endif
+}
+template<class aggregation>
+void LoadBasis(aggregation &Agg, std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  RD.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    //    RD.readScidacFieldRecord(Agg.subspace[b],record,BINARYIO_LEXICOGRAPHIC);
+    RD.readScidacFieldRecord(Agg.subspace[b],record,0);
+  }    
+  RD.close();
+#endif
+}
+
+
+template<class Field> class TestSolver : public LinearFunction<Field> {
+public:
+  TestSolver() {};
+  void operator() (const Field &in, Field &out){    out = Zero();  }     
+};
+
+
+RealD InverseApproximation(RealD x){
+  return 1.0/x;
+}
+
+// Want Op in CoarsenOp to call MatPcDagMatPc
+template<class Field>
+class HermOpAdaptor : public LinearOperatorBase<Field>
+{
+  LinearOperatorBase<Field> & wrapped;
+public:
+  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
+  void Op     (const Field &in, Field &out)   { wrapped.HermOp(in,out);  }
+  void HermOp(const Field &in, Field &out)    { wrapped.HermOp(in,out); }
+  void AdjOp     (const Field &in, Field &out){ wrapped.HermOp(in,out);  }
+  void OpDiag (const Field &in, Field &out)                  {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out)  {    assert(0);  };
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+};
+template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _SmootherOperator;
+  Chebyshev<Field> Cheby;
+  ChebyshevSmoother(RealD _lo,RealD _hi,int _ord, FineOperator &SmootherOperator) :
+    _SmootherOperator(SmootherOperator),
+    Cheby(_lo,_hi,_ord,InverseApproximation)
+  {
+    std::cout << GridLogMessage<<" Chebyshev smoother order "<<_ord<<" ["<<_lo<<","<<_hi<<"]"<<std::endl;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+    Field tmp(in.Grid());
+    tmp = in;
+    Cheby(_SmootherOperator,tmp,out);
+  }
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=24;
+  const int nbasis = 40;
+  const int cb = 0 ;
+  RealD mass=0.00078;
+  RealD M5=1.8;
+  RealD b=1.5;
+  RealD c=0.5;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplex::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  // Construct a coarsened grid with 4^4 cell
+  Coordinate clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/4;
+  }
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt,
+							    GridDefaultSimd(Nd,vComplex::Nsimd()),
+							    GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  ///////////////////////// RNGs /////////////////////////////////
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  ///////////////////////// Configuration /////////////////////////////////
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("ckpoint_lat.2250");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  //////////////////////// Fermion action //////////////////////////////////
+  MobiusFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c);
+
+  SchurDiagMooeeOperator<MobiusFermionD, LatticeFermion> HermOpEO(Ddwf);
+
+  typedef HermOpAdaptor<LatticeFermionD> HermFineMatrix;
+  HermFineMatrix FineHermOp(HermOpEO);
+  
+  LatticeFermion result(FrbGrid); result=Zero();
+
+  LatticeFermion    src(FrbGrid); random(RNG5,src);
+
+  // Run power method on FineHermOp
+  PowerMethod<LatticeFermion>       PM;   PM(HermOpEO,src);
+
+ 
+  ////////////////////////////////////////////////////////////
+  ///////////// Coarse basis and Little Dirac Operator ///////
+  ////////////////////////////////////////////////////////////
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
+  NearestStencilGeometry5D geom_nn(Coarse5d);
+  
+  // Warning: This routine calls PVdagM.Op, not PVdagM.HermOp
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
+  Subspace Aggregates(Coarse5d,FrbGrid,cb);
+
+  ////////////////////////////////////////////////////////////
+  // Need to check about red-black grid coarsening
+  ////////////////////////////////////////////////////////////
+  LittleDiracOperator LittleDiracOp(geom,FrbGrid,Coarse5d);
+
+  std::string subspace_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys.nolex.scidac");
+  std::string ldop_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/LittleDiracOp.phys.nolex.scidac");
+  bool load_agg=true;
+  bool load_mat=true;
+  if ( load_agg ) {
+    LoadBasis(Aggregates,subspace_file);
+  } else {
+    Aggregates.CreateSubspaceChebyshev(RNG5,HermOpEO,nbasis,
+				       95.0,0.05,
+				       1000,
+				       200,
+				       200,
+				       0.0);
+    SaveBasis(Aggregates,subspace_file);
+  }
+  if ( load_mat ) {
+    LoadOperator(LittleDiracOp,ldop_file);
+  } else {
+    LittleDiracOp.CoarsenOperator(FineHermOp,Aggregates);
+    SaveOperator(LittleDiracOp,ldop_file);
+  }
+  
+  // Try projecting to one hop only
+  LittleDiracOperator LittleDiracOpProj(geom_nn,FrbGrid,Coarse5d);
+  LittleDiracOpProj.ProjectNearestNeighbour(0.01,LittleDiracOp); // smaller shift 0.02? n
+
+  typedef HermitianLinearOperator<LittleDiracOperator,CoarseVector> HermMatrix;
+  HermMatrix CoarseOp     (LittleDiracOp);
+  HermMatrix CoarseOpProj (LittleDiracOpProj);
+  
+  //////////////////////////////////////////
+  // Build a coarse lanczos
+  //////////////////////////////////////////
+  //  Chebyshev<CoarseVector>      IRLCheby(0.01,44.0,201);  // 1 iter
+  Chebyshev<CoarseVector>      IRLCheby(0.005,44.0,401);  // 1 iter
+  FunctionHermOp<CoarseVector> IRLOpCheby(IRLCheby,CoarseOp);
+  PlainHermOp<CoarseVector>    IRLOp    (CoarseOp);
+  int Nk=160;
+  int Nm=240;
+  int Nstop=Nk;
+  ImplicitlyRestartedLanczos<CoarseVector> IRL(IRLOpCheby,IRLOp,Nstop,Nk,Nm,1.0e-3,20);
+
+  int Nconv;
+  std::vector<RealD>            eval(Nm);
+  std::vector<CoarseVector>     evec(Nm,Coarse5d);
+  CoarseVector c_src(Coarse5d); c_src=1.0;
+  CoarseVector c_res(Coarse5d); 
+  CoarseVector c_ref(Coarse5d); 
+
+  PowerMethod<CoarseVector>       cPM;   cPM(CoarseOp,c_src);
+
+  IRL.calc(eval,evec,c_src,Nconv);
+  DeflatedGuesser<CoarseVector> DeflCoarseGuesser(evec,eval);
+
+  //////////////////////////////////////////
+  // Build a coarse space solver
+  //////////////////////////////////////////
+  int maxit=20000;
+  ConjugateGradient<CoarseVector>  CG(1.0e-8,maxit,false);
+  ConjugateGradient<LatticeFermionD>  CGfine(1.0e-8,10000,false);
+  ZeroGuesser<CoarseVector> CoarseZeroGuesser;
+
+  //  HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,CoarseZeroGuesser);
+  HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,DeflCoarseGuesser);
+  c_res=Zero();
+  //  HPDSolve(c_src,c_res); c_ref = c_res;
+  //  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  //  std::cout << GridLogMessage<<"ref norm "<<norm2(c_ref)<<std::endl;
+  //////////////////////////////////////////////////////////////////////////
+  // Deflated (with real op EV's) solve for the projected coarse op
+  // Work towards ADEF1 in the coarse space
+  //////////////////////////////////////////////////////////////////////////
+  HPDSolver<CoarseVector> HPDSolveProj(CoarseOpProj,CG,DeflCoarseGuesser);
+  c_res=Zero();
+  //  HPDSolveProj(c_src,c_res);
+  //  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  //  std::cout << GridLogMessage<<"res norm "<<norm2(c_res)<<std::endl;
+  //  c_res = c_res - c_ref;
+  //  std::cout << "Projected solver error "<<norm2(c_res)<<std::endl;
+
+  //////////////////////////////////////////////////////////////////////
+  // Coarse ADEF1 with deflation space
+  //////////////////////////////////////////////////////////////////////
+  ChebyshevSmoother<CoarseVector,HermMatrix >
+    CoarseSmoother(1.0,37.,8,CoarseOpProj);  // just go to sloppy 0.1 convergence
+    //  CoarseSmoother(0.1,37.,8,CoarseOpProj);  //
+  //  CoarseSmoother(0.5,37.,6,CoarseOpProj);  //  8 iter 0.36s
+  //    CoarseSmoother(0.5,37.,12,CoarseOpProj);  // 8 iter, 0.55s
+  //    CoarseSmoother(0.5,37.,8,CoarseOpProj);// 7-9 iter
+  //  CoarseSmoother(1.0,37.,8,CoarseOpProj); // 0.4 - 0.5s solve to 0.04, 7-9 iter
+  //  ChebyshevSmoother<CoarseVector,HermMatrix > CoarseSmoother(0.5,36.,10,CoarseOpProj);  // 311
+
+  ////////////////////////////////////////////////////////
+  // CG, Cheby mode spacing 200,200
+  // Unprojected Coarse CG solve to 1e-8 : 190 iters, 4.9s
+  // Unprojected Coarse CG solve to 4e-2 :  33 iters, 0.8s
+  // Projected Coarse CG solve to 1e-8 : 100 iters, 0.36s
+  ////////////////////////////////////////////////////////
+  // CoarseSmoother(1.0,48.,8,CoarseOpProj); 48 evecs 
+  ////////////////////////////////////////////////////////
+  // ADEF1 Coarse solve to 1e-8 : 44 iters, 2.34s  2.1x gain
+  // ADEF1 Coarse solve to 4e-2 : 7 iters, 0.4s
+  // HDCG 38 iters 162s
+  //
+  // CoarseSmoother(1.0,40.,8,CoarseOpProj); 48 evecs 
+  // ADEF1 Coarse solve to 1e-8 : 37 iters, 2.0s  2.1x gain
+  // ADEF1 Coarse solve to 4e-2 : 6 iters, 0.36s
+  // HDCG 38 iters 169s
+
+  TwoLevelADEF1defl<CoarseVector>
+    cADEF1(1.0e-8, 500,
+	   CoarseOp,
+	   CoarseSmoother,
+	   evec,eval);
+
+  //  c_res=Zero();
+  //  cADEF1(c_src,c_res);
+  //  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  //  std::cout << GridLogMessage<<"cADEF1 res norm "<<norm2(c_res)<<std::endl;
+  //  c_res = c_res - c_ref;
+  //  std::cout << "cADEF1 solver error "<<norm2(c_res)<<std::endl;
+  
+  //  cADEF1.Tolerance = 4.0e-2;
+  //  cADEF1.Tolerance = 1.0e-1;
+  //  cADEF1.Tolerance = 5.0e-2;
+  //  c_res=Zero();
+  //  cADEF1(c_src,c_res);
+  //  std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
+  //  std::cout << GridLogMessage<<"cADEF1 res norm "<<norm2(c_res)<<std::endl;
+  //  c_res = c_res - c_ref;
+  //  std::cout << "cADEF1 solver error "<<norm2(c_res)<<std::endl;
+  
+  //////////////////////////////////////////
+  // Build a smoother
+  //////////////////////////////////////////
+  //  ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(10.0,100.0,10,FineHermOp); //499
+  //  ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(3.0,100.0,10,FineHermOp);  //383
+  //  ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(1.0,100.0,10,FineHermOp);  //328
+  //  std::vector<RealD> los({0.5,1.0,3.0}); // 147/142/146 nbasis 1
+  //  std::vector<RealD> los({1.0,2.0}); // Nbasis 24: 88,86 iterations
+  //  std::vector<RealD> los({2.0,4.0}); // Nbasis 32 == 52, iters
+  //  std::vector<RealD> los({2.0,4.0}); // Nbasis 40 == 36,36 iters
+
+  //
+  // Turns approx 2700 iterations into 340 fine multiplies with Nbasis 40
+  // Need to measure cost of coarse space.
+  //
+  // -- i) Reduce coarse residual   -- 0.04
+  // -- ii) Lanczos on coarse space -- done
+  // -- iii) Possible 1 hop project and/or preconditioning it - easy - PrecCG it and
+  //         use a limited stencil. Reread BFM code to check on evecs / deflation strategy with prec
+  //
+  std::vector<RealD> los({3.0}); // Nbasis 40 == 36,36 iters
+
+  //  std::vector<int> ords({7,8,10}); // Nbasis 40 == 40,38,36 iters (320,342,396 mults)
+  std::vector<int> ords({7}); // Nbasis 40 == 40 iters (320 mults)  
+
+  for(int l=0;l<los.size();l++){
+
+    RealD lo = los[l];
+
+    for(int o=0;o<ords.size();o++){
+
+      ConjugateGradient<CoarseVector>  CGsloppy(5.0e-2,maxit,false);
+      HPDSolver<CoarseVector> HPDSolveSloppy(CoarseOp,CGsloppy,DeflCoarseGuesser);
+      
+      //    ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(lo,92,10,FineHermOp); // 36 best case
+      ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(lo,92,ords[o],FineHermOp);  // 311
+
+      //////////////////////////////////////////
+      // Build a HDCG solver
+      //////////////////////////////////////////
+      TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
+	HDCG(1.0e-8, 100,
+	     FineHermOp,
+	     Smoother,
+	     HPDSolveSloppy,
+	     HPDSolve,
+	     Aggregates);
+
+      TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
+	HDCGdefl(1.0e-8, 100,
+		 FineHermOp,
+		 Smoother,
+		 cADEF1,
+		 HPDSolve,
+		 Aggregates);
+      
+      //      result=Zero();
+      //      HDCGdefl(src,result);
+
+      result=Zero();
+      HDCG(src,result);
+      
+    }
+  }
+
+  // Standard CG
+  result=Zero();
+  CGfine(HermOpEO, src, result);
+  
+  Grid_finalize();
+  return 0;
+}

tests/debug/Test_general_coarse_pvdagm.cc

@@ -0,0 +1,267 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_padded_cell.cc
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h>
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+#include <Grid/algorithms/iterative/BiCGSTAB.h>
+
+using namespace std;
+using namespace Grid;
+
+template<class Field>
+class HermOpAdaptor : public LinearOperatorBase<Field>
+{
+  LinearOperatorBase<Field> & wrapped;
+public:
+  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    wrapped.HermOp(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    wrapped.HermOp(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    wrapped.HermOp(in,out);
+  }
+  
+};
+
+template<class Matrix,class Field>
+class PVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+public:
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+  void HermOp(const Field &in, Field &out){
+    std::cout << "HermOp"<<std::endl;
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    _PV.M(out,tmp);
+    _Mat.Mdag(tmp,out);
+    std::cout << "HermOp done "<<norm2(out)<<std::endl;
+    
+  }
+};
+
+template<class Field> class DumbOperator  : public LinearOperatorBase<Field> {
+public:
+  LatticeComplex scale;
+  DumbOperator(GridBase *grid) : scale(grid)
+  {
+    scale = 0.0;
+    LatticeComplex scalesft(grid);
+    LatticeComplex scaletmp(grid);
+    for(int d=0;d<4;d++){
+      Lattice<iScalar<vInteger> > x(grid); LatticeCoordinate(x,d+1);
+      LatticeCoordinate(scaletmp,d+1);
+      scalesft = Cshift(scaletmp,d+1,1);
+      scale = 100.0*scale + where( mod(x    ,2)==(Integer)0, scalesft,scaletmp);
+    }
+    std::cout << " scale\n" << scale << std::endl;
+  }
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {};
+  void OpDir  (const Field &in, Field &out,int dir,int disp){};
+  void OpDirAll  (const Field &in, std::vector<Field> &out) {};
+
+  void Op     (const Field &in, Field &out){
+    out = scale * in;
+  }
+  void AdjOp  (const Field &in, Field &out){
+    out = scale * in;
+  }
+  void HermOp(const Field &in, Field &out){
+    double n1, n2;
+    HermOpAndNorm(in,out,n1,n2);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,double &n1,double &n2){
+    ComplexD dot;
+
+    out = scale * in;
+
+    dot= innerProduct(in,out);
+    n1=real(dot);
+
+    dot = innerProduct(out,out);
+    n2=real(dot);
+  }
+};
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=2;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  // Construct a coarsened grid
+  Coordinate clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/4;
+  }
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  LatticeFermion    src(FGrid); random(RNG5,src);
+  LatticeFermion result(FGrid); result=Zero();
+  LatticeFermion    ref(FGrid); ref=Zero();
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("ckpoint_lat.4000");
+  NerscIO::readConfiguration(Umu,header,file);
+  //Umu = 1.0;
+  
+  RealD mass=0.5;
+  RealD M5=1.8;
+
+  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionD Dpv(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
+
+  const int nbasis = 1;
+  const int cb = 0 ;
+  LatticeFermion prom(FGrid);
+
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNearestStencilGeometry5D geom(Coarse5d);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  
+  PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM(Ddwf,Dpv);
+  HermOpAdaptor<LatticeFermionD> HOA(PVdagM);
+
+  // Run power method on HOA??
+  PowerMethod<LatticeFermion>       PM;   PM(HOA,src);
+ 
+  // Warning: This routine calls PVdagM.Op, not PVdagM.HermOp
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
+  Subspace AggregatesPD(Coarse5d,FGrid,cb);
+  AggregatesPD.CreateSubspaceChebyshev(RNG5,
+				       HOA,
+				       nbasis,
+				       5000.0,
+				       0.02,
+				       100,
+				       50,
+				       50,
+				       0.0);
+  
+  LittleDiracOperator LittleDiracOpPV(geom,FGrid,Coarse5d);
+  LittleDiracOpPV.CoarsenOperator(PVdagM,AggregatesPD);
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"Testing coarsened operator "<<std::endl;
+  
+  CoarseVector c_src (Coarse5d);
+  CoarseVector c_res (Coarse5d);
+  CoarseVector c_proj(Coarse5d);
+
+  std::vector<LatticeFermion> subspace(nbasis,FGrid);
+  subspace=AggregatesPD.subspace;
+
+  Complex one(1.0);
+  c_src = one;  // 1 in every element for vector 1.
+  blockPromote(c_src,err,subspace);
+
+  prom=Zero();
+  for(int b=0;b<nbasis;b++){
+    prom=prom+subspace[b];
+  }
+  err=err-prom; 
+  std::cout<<GridLogMessage<<"Promoted back from subspace: err "<<norm2(err)<<std::endl;
+  std::cout<<GridLogMessage<<"c_src "<<norm2(c_src)<<std::endl;
+  std::cout<<GridLogMessage<<"prom  "<<norm2(prom)<<std::endl;
+
+  PVdagM.Op(prom,tmp);
+  blockProject(c_proj,tmp,subspace);
+  std::cout<<GridLogMessage<<" Called Big Dirac Op "<<norm2(tmp)<<std::endl;
+
+  LittleDiracOpPV.M(c_src,c_res);
+  std::cout<<GridLogMessage<<" Called Little Dirac Op c_src "<< norm2(c_src) << "  c_res "<< norm2(c_res) <<std::endl;
+
+  std::cout<<GridLogMessage<<"Little dop : "<<norm2(c_res)<<std::endl;
+  //  std::cout<<GridLogMessage<<" Little "<< c_res<<std::endl;
+
+  std::cout<<GridLogMessage<<"Big dop in subspace : "<<norm2(c_proj)<<std::endl;
+  //  std::cout<<GridLogMessage<<" Big "<< c_proj<<std::endl;
+  c_proj = c_proj - c_res;
+  std::cout<<GridLogMessage<<" ldop error: "<<norm2(c_proj)<<std::endl;
+  //  std::cout<<GridLogMessage<<" error "<< c_proj<<std::endl;
+
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
+  std::cout<<GridLogMessage<<std::endl;
+  std::cout<<GridLogMessage << "Done "<< std::endl;
+
+  Grid_finalize();
+  return 0;
+}