MultiRHS coarse

It is block project limited.
Could be sped up with calls to Batched GEMM and a data layout change.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,54 @@
+name: Bug report
+description: Report a bug.
+title: "<insert title>"
+labels: [bug]
+
+body:
+  - type: markdown
+    attributes:
+      value: >
+        Thank you for taking the time to file a bug report.
+        Please check that the code is pointing to the HEAD of develop
+        or any commit in master which is tagged with a version number.
+
+  - type: textarea
+    attributes:
+      label: "Describe the issue:"
+      description: >
+        Describe the issue and any previous attempt to solve it.
+    validations:
+      required: true
+
+  - type: textarea
+    attributes:
+      label: "Code example:"
+      description: >
+        If relevant, show how to reproduce the issue using a minimal working
+        example.
+      placeholder: |
+        << your code here >>
+      render: shell
+    validations:
+      required: false
+
+  - type: textarea
+    attributes:
+      label: "Target platform:"
+      description: >
+        Give a description of the target platform (CPU, network, compiler).
+        Please give the full CPU part description, using for example
+        `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux)
+        or `sysctl machdep.cpu.brand_string` (macOS) and the full output
+        the `--version` option of your compiler.
+    validations:
+      required: true
+
+  - type: textarea
+    attributes:
+      label: "Configure options:"
+      description: >
+        Please give the exact configure command used and attach
+        `config.log`, `grid.config.summary` and the output of `make V=1`.
+      render: shell
+    validations:
+      required: true

Grid/Makefile.am

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

Grid/algorithms/Algorithms.h

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

Grid/algorithms/LinearOperator.h

@@ -145,6 +145,44 @@ public:
   }
 };
 
+////////////////////////////////////////////////////////////////////
+// Create a shifted HermOp
+////////////////////////////////////////////////////////////////////
+template<class Field>
+class ShiftedHermOpLinearOperator : public LinearOperatorBase<Field> {
+  LinearOperatorBase<Field> &_Mat;
+  RealD _shift;
+public:
+  ShiftedHermOpLinearOperator(LinearOperatorBase<Field> &Mat,RealD shift): _Mat(Mat), _shift(shift){};
+  // Support for coarsening to a multigrid
+  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){
+    assert(0);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    assert(0);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+    _Mat.HermOp(in,out);
+    out = out + _shift*in;
+  }
+};
+
+
 ////////////////////////////////////////////////////////////////////
 // Wrap an already herm matrix
 ////////////////////////////////////////////////////////////////////
@@ -542,6 +580,7 @@ public:
       (*this)(in[i], out[i]);
     }
   }
+  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {

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/approx/MultiShiftFunction.h

@@ -40,7 +40,7 @@ public:
   RealD norm;
   RealD lo,hi;
 
-  MultiShiftFunction(int n,RealD _lo,RealD _hi): poles(n), residues(n), lo(_lo), hi(_hi) {;};
+  MultiShiftFunction(int n,RealD _lo,RealD _hi): poles(n), residues(n), tolerances(n), lo(_lo), hi(_hi) {;};
   RealD approx(RealD x);
   void csv(std::ostream &out);
   void gnuplot(std::ostream &out);

Grid/algorithms/iterative/AdefGeneric.h

@@ -33,109 +33,110 @@ 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)
+  {
+    std::cout << GridLogMessage<<"HDCG: fPcg starting"<<std::endl;
     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);
+    int mmax = 5;
+    std::cout << GridLogMessage<<"HDCG: fPcg allocating"<<std::endl;
+    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 z(grid);
     Field tmp(grid);
-    Field r  (grid);
-    Field mu (grid);
-  
+    Field  mp (grid);
+    Field  r  (grid);
+    Field  mu (grid);
+    
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated"<<std::endl;
+    //Initial residual computation & set up
+    RealD guess   = norm2(x);
+    std::cout << GridLogMessage<<"HDCG: fPcg guess nrm "<<guess<<std::endl;
+    RealD src_nrm = norm2(src);
+    std::cout << GridLogMessage<<"HDCG: fPcg src nrm "<<src_nrm<<std::endl;
+    
+    if ( src_nrm == 0.0 ) {
+      std::cout << GridLogMessage<<"HDCG: fPcg given trivial source norm "<<src_nrm<<std::endl;
+      x=Zero();
+    }
+    RealD tn;
+    
+    GridStopWatch HDCGTimer;
+    HDCGTimer.Start();
     //////////////////////////
     // x0 = Vstart -- possibly modify guess
     //////////////////////////
-    x=src;
     Vstart(x,src);
-
+    
     // r0 = b -A x0
-    HermOp(x,mmp); // Shouldn't this be something else?
+    _FineLinop.HermOp(x,mmp[0]);
     axpy (r, -1.0,mmp[0], src);    // Recomputes r=src-Ax0
+    {
+      double n1 = norm2(x);
+      double n2 = norm2(mmp[0]);
+      double n3 = norm2(r);
+      std::cout<<GridLogMessage<<"x,vstart,r = "<<n1<<" "<<n2<<" "<<n3<<std::endl;
+    }
 
     //////////////////////////////////
     // 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 
+    // Def2 p = 1 - Q Az = Pright z
     // Other algos M2 is trivial
     ///////////////////////////////////////
-    M2(z,p[0]);
+    PcgM2(z,p[0]);
+
+    RealD ssq =  norm2(src);
+    RealD rsq =  ssq*Tolerance*Tolerance;
+
+    std::cout << GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" rsq "<<rsq<<"\n";
+
+    Field pp(grid);
 
     for (int k=0;k<=MaxIterations;k++){
     
@@ -143,31 +144,46 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       int peri_kp = (k+1) % mmax;
 
       rtz=rtzp;
-      d= M3(p[peri_k],mp,mmp[peri_k],tmp);
+      d= PcgM3(p[peri_k],mmp[peri_k]);
       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);
 
       // Compute z = M x
-      M1(r,z,tmp,mp);
-
+      PcgM1(r,z);
+      
+      {
+	RealD n1,n2;
+	n1=norm2(r);
+	n2=norm2(z);
+	std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : vector r,z "<<n1<<" "<<n2<<"\n";
+      }
       rtzp =real(innerProduct(r,z));
+      std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : inner rtzp "<<rtzp<<"\n";
 
-      M2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+      //    PcgM2(z,p[0]);
+      PcgM2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+      
+      p[peri_kp]=mu;
 
-      p[peri_kp]=p[peri_k];
-
-      // Standard search direction  p -> z + b p    ; b = 
+      // Standard search direction  p -> z + b p    
       b = (rtzp)/rtz;
-
+      
       int northog;
+      // k=zero  <=> peri_kp=1;        northog = 1
+      // k=1     <=> peri_kp=2;        northog = 2
+      // ...               ...                  ...
+      // k=mmax-2<=> peri_kp=mmax-1;   northog = mmax-1
+      // k=mmax-1<=> peri_kp=0;        northog = 1
+
       //    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
     
+      std::cout<<GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : orthogonalising to last "<<northog<<" vectors\n";
       for(int back=0; back < northog; back++){
 	int peri_back = (k-back)%mmax;
 	RealD pbApk= real(innerProduct(mmp[peri_back],p[peri_kp]));
@@ -176,75 +192,315 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       }
 
       RealD rrn=sqrt(rn/ssq);
-      std::cout<<GridLogMessage<<"TwoLevelfPcg: k= "<<k<<" residual = "<<rrn<<std::endl;
+      RealD rtn=sqrt(rtz/ssq);
+      RealD rtnp=sqrt(rtzp/ssq);
+
+      std::cout<<GridLogMessage<<"HDCG: fPcg k= "<<k<<" residual = "<<rrn<<"\n";
 
       // Stopping condition
       if ( rn <= rsq ) { 
 
-	HermOp(x,mmp); // Shouldn't this be something else?
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	
+	_FineLinop.HermOp(x,mmp[0]);			  
 	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;
+	RealD  mmpnorm = sqrt(norm2(mmp[0]));
+	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);
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    RealD  xnorm   = sqrt(norm2(x));
+    RealD  srcnorm = sqrt(norm2(src));
+    std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
   }
 
+
+
+  virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    std::vector<RealD> f(nrhs);
+    std::vector<RealD> rtzp(nrhs);
+    std::vector<RealD> rtz(nrhs);
+    std::vector<RealD> a(nrhs);
+    std::vector<RealD> d(nrhs);
+    std::vector<RealD> b(nrhs);
+    std::vector<RealD> rptzp(nrhs);
+    /////////////////////////////
+    // Set up history vectors
+    /////////////////////////////
+    int mmax = 2;
+    std::cout << GridLogMessage<<"HDCG: fPcg allocating"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<Field> > p(nrhs);   for(int r=0;r<nrhs;r++)  p[r].resize(mmax,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated p"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<Field> > mmp(nrhs); for(int r=0;r<nrhs;r++) mmp[r].resize(mmax,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated mmp"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<RealD> > pAp(nrhs); for(int r=0;r<nrhs;r++) pAp[r].resize(mmax);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated pAp"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<Field> z(nrhs,grid);
+    std::vector<Field>  mp (nrhs,grid);
+    std::vector<Field>  r  (nrhs,grid);
+    std::vector<Field>  mu (nrhs,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated z,mp,r,mu"<<std::endl;
+    src[0].Grid()->Barrier();
+
+    //Initial residual computation & set up
+    std::vector<RealD> src_nrm(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      src_nrm[rhs]=norm2(src[rhs]);
+      assert(src_nrm[rhs]!=0.0);
+    }
+    std::vector<RealD> tn(nrhs);
+
+    GridStopWatch HDCGTimer;
+    HDCGTimer.Start();
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    for(int rhs=0;rhs<nrhs;rhs++){
+      Vstart(x[rhs],src[rhs]);
+
+      // r0 = b -A x0
+      _FineLinop.HermOp(x[rhs],mmp[rhs][0]);
+      axpy (r[rhs], -1.0,mmp[rhs][0], src[rhs]);    // Recomputes r=src-Ax0
+    }
+
+    //////////////////////////////////
+    // Compute z = M1 x
+    //////////////////////////////////
+    // This needs a multiRHS version for acceleration
+    PcgM1(r,z);
+
+    std::vector<RealD> ssq(nrhs);
+    std::vector<RealD> rsq(nrhs);
+    std::vector<Field> pp(nrhs,grid);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+      p[rhs][0]=z[rhs];
+      ssq[rhs]=norm2(src[rhs]);
+      rsq[rhs]=  ssq[rhs]*Tolerance*Tolerance;
+      std::cout << GridLogMessage<<"mrhs HDCG: "<<rhs<<" k=0 residual "<<rtzp[rhs]<<" rsq "<<rsq[rhs]<<"\n";
+    }
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+    
+      int peri_k  = k % mmax;
+      int peri_kp = (k+1) % mmax;
+
+      for(int rhs=0;rhs<nrhs;rhs++){
+	rtz[rhs]=rtzp[rhs];
+	d[rhs]= PcgM3(p[rhs][peri_k],mmp[rhs][peri_k]);
+	a[rhs] = rtz[rhs]/d[rhs];
+    
+	// Memorise this
+	pAp[rhs][peri_k] = d[rhs];
+
+	axpy(x[rhs],a[rhs],p[rhs][peri_k],x[rhs]);
+	rn[rhs] = axpy_norm(r[rhs],-a[rhs],mmp[rhs][peri_k],r[rhs]);
+      }
+
+      // Compute z = M x (for *all* RHS)
+      PcgM1(r,z);
+
+      RealD max_rn=0.0;
+      for(int rhs=0;rhs<nrhs;rhs++){
+
+	rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+
+	std::cout << GridLogMessage<<"HDCG::fPcg rhs"<<rhs<<" iteration "<<k<<" : inner rtzp "<<rtzp[rhs]<<"\n";
+	
+	mu[rhs]=z[rhs];
+
+	p[rhs][peri_kp]=mu[rhs];
+
+	// Standard search direction p == z + b p 
+	b[rhs] = (rtzp[rhs])/rtz[rhs];
+
+	int northog = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
+	std::cout<<GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : orthogonalising to last "<<northog<<" vectors\n";
+	for(int back=0; back < northog; back++){
+	  int peri_back = (k-back)%mmax;
+	  RealD pbApk= real(innerProduct(mmp[rhs][peri_back],p[rhs][peri_kp]));
+	  RealD beta = -pbApk/pAp[rhs][peri_back];
+	  axpy(p[rhs][peri_kp],beta,p[rhs][peri_back],p[rhs][peri_kp]);
+	}
+
+	RealD rrn=sqrt(rn[rhs]/ssq[rhs]);
+	RealD rtn=sqrt(rtz[rhs]/ssq[rhs]);
+	RealD rtnp=sqrt(rtzp[rhs]/ssq[rhs]);
+	
+	std::cout<<GridLogMessage<<"HDCG: rhs "<<rhs<<"fPcg k= "<<k<<" residual = "<<rrn<<"\n";
+	if ( rrn > max_rn ) max_rn = rrn;
+      }
+
+      // Stopping condition based on worst case
+      if ( max_rn <= Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+	  _FineLinop.HermOp(x[rhs],mmp[rhs][0]);			  
+	  Field tmp(grid);
+	  axpy(tmp,-1.0,src[rhs],mmp[rhs][0]);
+      
+	  RealD  mmpnorm = sqrt(norm2(mmp[rhs][0]));
+	  RealD  xnorm   = sqrt(norm2(x[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(tmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<" mmp "<<mmpnorm	  
+		   <<std::endl;
+	}
+	return;
+      }
+      
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    for(int rhs=0;rhs<nrhs;rhs++){
+      RealD  xnorm   = sqrt(norm2(x[rhs]));
+      RealD  srcnorm = sqrt(norm2(src[rhs]));
+      std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+    }
+  }
+  
+
  public:
 
-  virtual void M(Field & in,Field & out,Field & tmp) {
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out)
+  {
+    std::cout << "PcgM1 default (cheat) mrhs versoin"<<std::endl;
+    for(int rhs=0;rhs<in.size();rhs++){
+      this->PcgM1(in[rhs],out[rhs]);
+    }
+  }
+  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)
+  {
+    GRID_TRACE("MultiGridPreconditioner ");
     // [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)
+  {
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart "<<std::endl;
     ///////////////////////////////////
     // Choose x_0 such that 
     // x_0 = guess +  (A_ss^inv) r_s = guess + Ass_inv [src -Aguess]
@@ -256,142 +512,157 @@ 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);
+
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart projecting "<<std::endl;
+    this->_Aggregates.ProjectToSubspace(PleftProj,src);     
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart coarse solve "<<std::endl;
+    this->_CoarseSolverPrecise(PleftProj,PleftMss_proj); // Ass^{-1} r_s
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart promote "<<std::endl;
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,x);  
 
   }
 
+};
+
+template<class Field, class CoarseField, class Aggregation>
+class TwoLevelADEF2mrhs : public TwoLevelADEF2<Field,CoarseField,Aggregation>
+{
+public:
+  GridBase *coarsegridmrhs;
+  LinearFunction<CoarseField> &_CoarseSolverMrhs;
+  LinearFunction<CoarseField> &_CoarseGuesser;
+  TwoLevelADEF2mrhs(RealD tol,
+		    Integer maxit,
+		    LinearOperatorBase<Field>    &FineLinop,
+		    LinearFunction<Field>        &Smoother,
+		    LinearFunction<CoarseField>  &CoarseSolver,
+		    LinearFunction<CoarseField>  &CoarseSolverPrecise,
+		    LinearFunction<CoarseField>  &CoarseSolverMrhs,
+		    LinearFunction<CoarseField>  &CoarseGuesser,
+		    GridBase *rhsgrid,
+		    Aggregation &Aggregates) :
+    TwoLevelADEF2<Field,CoarseField,Aggregation>(tol, maxit,FineLinop,Smoother,CoarseSolver,CoarseSolverPrecise,Aggregates),
+    _CoarseSolverMrhs(CoarseSolverMrhs),
+    _CoarseGuesser(CoarseGuesser)
+  {
+    coarsegridmrhs = rhsgrid;
+  };
+  
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out){
+
+    int nrhs=in.size();
+    std::cout << " mrhs PcgM1 for "<<nrhs<<" right hand sides"<<std::endl;
+    // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
+    Field tmp(this->grid);
+    std::vector<Field> Min(nrhs,this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
+
+    CoarseField PleftProjMrhs(this->coarsegridmrhs);
+    CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
+
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      this->grid->Barrier();
+      std::cout << " Calling smoother for "<<rhs<<std::endl;
+      this->grid->Barrier();
+      this->_Smoother(in[rhs],Min[rhs]);
+      this->grid->Barrier();
+      std::cout << " smoother done "<<rhs<<std::endl;
+      this->grid->Barrier();
+      this->_FineLinop.HermOp(Min[rhs],out[rhs]);
+      this->grid->Barrier();
+      std::cout << " Hermop for "<<rhs<<std::endl;
+      this->grid->Barrier();
+      axpy(tmp,-1.0,out[rhs],in[rhs]);          // tmp  = in - A Min
+      this->grid->Barrier();
+      std::cout << " axpy "<<rhs<<std::endl;
+      this->grid->Barrier();
+      this->_Aggregates.ProjectToSubspace(PleftProj,tmp);     // can optimise later
+      this->grid->Barrier();
+      std::cout << " project "<<rhs<<std::endl;
+      this->grid->Barrier();
+      InsertSlice(PleftProj,PleftProjMrhs,rhs,0);
+      this->grid->Barrier();
+      std::cout << " insert rhs "<<rhs<<std::endl;
+      this->grid->Barrier();
+      this->_CoarseGuesser(PleftProj,PleftMss_proj);
+      this->grid->Barrier();
+      std::cout << " insert guess "<<rhs<<std::endl;
+      this->grid->Barrier();
+      InsertSlice(PleftMss_proj,PleftMss_projMrhs,rhs,0);
+    }
+
+    std::cout << " Coarse solve "<<std::endl;
+    this->_CoarseSolverMrhs(PleftProjMrhs,PleftMss_projMrhs); // Ass^{-1} [in - A Min]_s
+
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      ExtractSlice(PleftMss_proj,PleftMss_projMrhs,rhs,0);
+      this->_Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+      axpy(out[rhs],1.0,Min[rhs],tmp); // Min+tmp
+    }
+    std::cout << " Extracted "<<std::endl;
+  }
+};
+  
+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)
+  {
+    GRID_TRACE("EvecPreconditioner ");
+    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;
 
@@ -207,7 +207,8 @@ public:
 
     TrueResidual = sqrt(norm2(p)/ssq);
 
-    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
+    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations
+	      <<" residual "<< TrueResidual<< std::endl;
 
     if (ErrorOnNoConverge) assert(0);
     IterationsToComplete = k;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -144,7 +144,7 @@ public:
     for(int s=0;s<nshift;s++){
       rsq[s] = cp * mresidual[s] * mresidual[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShift: shift "<<s
-	       <<" target resid "<<rsq[s]<<std::endl;
+	       <<" target resid^2 "<<rsq[s]<<std::endl;
       ps[s] = src;
     }
     // r and p for primary

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -166,16 +166,16 @@ public:
       rsqf[s] =rsq[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
       //      ps_d[s] = src_d;
-      precisionChangeFast(ps_f[s],src_d);
+      precisionChange(ps_f[s],src_d);
     }
     // r and p for primary
     p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
     r_d = p_d;
     
     //MdagM+m[0]
-    precisionChangeFast(p_f,p_d);
+    precisionChange(p_f,p_d);
     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    precisionChangeFast(tmp_d,mmp_f);
+    precisionChange(tmp_d,mmp_f);
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
@@ -204,7 +204,7 @@ public:
   
     for(int s=0;s<nshift;s++) {
       axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
-      precisionChangeFast(psi_f[s],psi_d[s]);
+      precisionChange(psi_f[s],psi_d[s]);
     }
   
     ///////////////////////////////////////
@@ -225,7 +225,7 @@ public:
       AXPYTimer.Stop();
 
       PrecChangeTimer.Start();
-      precisionChangeFast(r_f, r_d);
+      precisionChange(r_f, r_d);
       PrecChangeTimer.Stop();
 
       AXPYTimer.Start();
@@ -243,13 +243,13 @@ public:
 
       cp=c;
       PrecChangeTimer.Start();
-      precisionChangeFast(p_f, p_d); //get back single prec search direction for linop
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
       PrecChangeTimer.Stop();
       MatrixTimer.Start();  
       Linop_f.HermOp(p_f,mmp_f);
       MatrixTimer.Stop();  
       PrecChangeTimer.Start();
-      precisionChangeFast(mmp_d, mmp_f); // From Float to Double
+      precisionChange(mmp_d, mmp_f); // From Float to Double
       PrecChangeTimer.Stop();
 
       d=real(innerProduct(p_d,mmp_d));    
@@ -311,7 +311,7 @@ public:
 	SolverTimer.Stop();
 
 	for(int s=0;s<nshift;s++){
-	  precisionChangeFast(psi_d[s],psi_f[s]);
+	  precisionChange(psi_d[s],psi_f[s]);
 	}
 
 	

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -211,7 +211,7 @@ public:
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
-    //    assert(norm2(tmp_d)< 1.0e-4);
+    assert(norm2(tmp_d)< 1.0);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -79,14 +79,16 @@ template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public Imp
     RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
 
     std::cout.precision(13);
-    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
-	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
-	     <<std::endl;
 
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
 
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<" target " << eresid*eresid << " conv " <<conv
+	     <<std::endl;
+
     return conv;
   }
 };
@@ -419,14 +421,15 @@ until convergence
 	}
       }
 
-      if ( Nconv < Nstop )
+      if ( Nconv < Nstop ) {
 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-
+	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
+      }
       eval=eval2;
       
       //Keep only converged
-      eval.resize(Nconv);// Nstop?
-      evec.resize(Nconv,grid);// Nstop?
+      eval.resize(Nstop);// was Nconv
+      evec.resize(Nstop,grid);// was Nconv
       basisSortInPlace(evec,eval,reverse);
       
     }
@@ -456,7 +459,7 @@ 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 );
 
@@ -464,7 +467,7 @@ 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];
 
@@ -479,18 +482,18 @@ 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 +1,381 @@
+/*************************************************************************************
+
+    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);
+
+inline RealD AggregatePowerLaw(RealD x)
+{
+  //  return std::pow(x,-4);
+  //  return std::pow(x,-3);
+  return std::pow(x,-5);
+}
+
+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);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+
+      // Refine
+      Chebyshev<FineField> PowerLaw(lo,hi,1000,AggregatePowerLaw);
+      noise = Mn;
+      PowerLaw(hermop,noise,Mn);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+
+      // normalise
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+
+  }
+
+  virtual void CreateSubspaceChebyshevPowerLaw(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					       int nn,
+					       double hi,
+					       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<<" [0,"<<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(0.0,hi,orderfilter,AggregatePowerLaw);
+      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;
+    }
+
+  }
+
+  virtual void CreateSubspaceMultishift(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					double Lo,double tol,int maxit)
+  {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Multishift subspace : Lo "<<Lo<<std::endl;
+
+    // Filter
+    // [ 1/6(x+Lo)  - 1/2(x+2Lo) + 1/2(x+3Lo)  -1/6(x+4Lo) = Lo^3 /[ (x+1Lo)(x+2Lo)(x+3Lo)(x+4Lo) ]
+    //
+    // 1/(x+Lo)  - 1/(x+2 Lo)
+    double epsilon      = Lo/3;
+    std::vector<RealD> alpha({1.0/6.0,-1.0/2.0,1.0/2.0,-1.0/6.0});
+    std::vector<RealD> shifts({Lo,Lo+epsilon,Lo+2*epsilon,Lo+3*epsilon});
+    std::vector<RealD> tols({tol,tol,tol,tol});
+    std::cout << "sizes "<<alpha.size()<<" "<<shifts.size()<<" "<<tols.size()<<std::endl;
+
+    MultiShiftFunction msf(4,0.0,95.0);
+    std::cout << "msf constructed "<<std::endl;
+    msf.poles=shifts;
+    msf.residues=alpha;
+    msf.tolerances=tols;
+    msf.norm=0.0;
+    msf.order=alpha.size();
+    ConjugateGradientMultiShift<FineField> MSCG(maxit,msf);
+    
+    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;
+
+      MSCG(hermop,noise,Mn);
+      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;
+
+    }
+
+  }
+  virtual void RefineSubspace(LinearOperatorBase<FineField> &hermop,
+			      double Lo,double tol,int maxit)
+  {
+    FineField tmp(FineGrid);
+    for(int b =0;b<nbasis;b++)
+    {
+      RealD MirsShift = Lo;
+      ConjugateGradient<FineField>  CGsloppy(tol,maxit,false);
+      ShiftedHermOpLinearOperator<FineField> ShiftedFineHermOp(hermop,MirsShift);
+      CGsloppy(hermop,subspace[b],tmp);
+      subspace[b]=tmp;
+    }
+  }
+
+  
+  
+};
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/BatchedBlas.h

@@ -0,0 +1,537 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.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
+
+#ifdef GRID_HIP
+#include <hipblas/hipblas.h>
+#endif
+#ifdef GRID_CUDA
+#include <hipblas/hipblas.h>
+#endif
+#ifdef GRID_SYCL
+#error // need oneMKL version
+#endif
+
+///////////////////////////////////////////////////////////////////////	  
+// Need to rearrange lattice data to be in the right format for a
+// batched multiply. Might as well make these static, dense packed
+///////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+#ifdef GRID_HIP
+  typedef hipblasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_CUDA
+  typedef cudablasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_SYCL
+  typedef int32_t gridblasHandle_t;
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+  typedef int32_t gridblasHandle_t;
+#endif
+
+class GridBLAS {
+public:
+
+  static gridblasHandle_t gridblasHandle;
+  static int            gridblasInit;
+  
+  static void Init(void)
+  {
+    if ( ! gridblasInit ) {
+#ifdef GRID_CUDA
+      std::cout << "cublasCreate"<<std::endl;
+      cublasCreate(&gridblasHandle);
+#endif
+#ifdef GRID_HIP
+      std::cout << "hipblasCreate"<<std::endl;
+      hipblasCreate(&gridblasHandle);
+#endif
+#ifdef GRID_SYCL
+#endif
+    }
+  }
+  
+  // Force construct once
+  GridBLAS() { Init(); };
+  ~GridBLAS() { };
+  
+  /////////////////////////////////////////////////////////////////////////////////////
+  // BLAS GEMM conventions:
+  /////////////////////////////////////////////////////////////////////////////////////
+  // - C = alpha A * B + beta C
+  // Dimensions:
+  // - C_m.n
+  // - A_m.k
+  // - B_k.n
+  // - Flops = 8 M N K
+  // - Bytes = 2*sizeof(word) * (MN+MK+KN)
+  // M=60, N=12
+  // Flop/Byte = 8 . 60.60.12 / (60.12+60.60+60.12)/16 = 4 so expect about 4 TF/s on a GCD
+  /////////////////////////////////////////////////////////////////////////////////////
+  void synchronise(void)
+  {
+#ifdef GRID_HIP
+    auto err = hipDeviceSynchronize();
+    assert(err==hipSuccess);
+#endif
+#ifdef GRID_CUDA
+    auto err = cudaDeviceSynchronize();
+    assert(err==cudaSuccess);
+#endif
+#ifdef GRID_SYCL
+    accelerator_barrier();
+#endif
+  }
+  void benchmark(int nbasis, int nrhs, int coarseVol, int nstencil)
+  {
+    int32_t N_A = nbasis*nbasis*coarseVol*nstencil;
+    int32_t N_B = nbasis*nrhs*coarseVol*nstencil; // One leg of stencil at a time
+    int32_t N_C = nbasis*nrhs*coarseVol*nstencil; 
+    deviceVector<ComplexD> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(ComplexD));
+    deviceVector<ComplexD> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(ComplexD));
+    deviceVector<ComplexD> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(ComplexD));
+    ComplexD alpha(1.0);
+    ComplexD beta (1.0);
+    for(int i=0;i<10;i++){
+      RealD t0 = usecond();
+      for(int s=0;s<nstencil;s++){
+	gemmStridedBatched(nbasis,nrhs,nbasis,
+			   alpha,
+			   &A[0], // m x k 
+			   &B[0], // k x n
+			   beta, 
+			   &C[0], // m x n
+			   coarseVol);
+      }
+      synchronise();
+      RealD t1 = usecond();
+      RealD flops = 8.0*nbasis*nbasis*nrhs*coarseVol*nstencil;
+      RealD bytes = 1.0*sizeof(ComplexD)*(nbasis*nbasis+nbasis*nrhs*3)*coarseVol*nstencil;
+      std::cout << " batched Blas call "<<i<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+      std::cout << " batched Blas call "<<i<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+    }
+  }
+
+  void gemmBatched(int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    static deviceVector<ComplexD> alpha_p(1);
+    static deviceVector<ComplexD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+    RealD t0=usecond();
+    //       std::cout << "hipblasZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    auto err = hipblasZgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (hipblasDoubleComplex *) &alpha_p[0],
+				   (hipblasDoubleComplex **)&Amk[0], lda,
+				   (hipblasDoubleComplex **)&Bkn[0], ldb,
+				   (hipblasDoubleComplex *) &beta_p[0],
+				   (hipblasDoubleComplex **)&Cmn[0], ldc,
+				   batchCount);
+    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasZgemmBatched(gridblasHandle,
+				  CUBLAS_OP_N,
+				  CUBLAS_OP_N,
+				  m,n,k,
+				  (cuDoubleComplex *) &alpha_p[0],
+				  (cuDoubleComplex **)&Amk[0], lda,
+				  (cuDoubleComplex **)&Bkn[0], ldb,
+				  (cuDoubleComplex *) &beta_p[0],
+				  (cuDoubleComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexD c_mn(0.0);
+	  for (int kk = 0; kk < k, ++kk)
+	    c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	  Cmn[mm + nn*ldc + p*sdc] =  (*alpha_p)*c_mn + (*beta_p)*Cmn[mm + nn*ldc + p*sdc];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+
+  void gemmBatched(int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    static deviceVector<ComplexF> alpha_p(1);
+    static deviceVector<ComplexF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
+    RealD t0=usecond();
+    //       std::cout << "hipblasZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    auto err = hipblasCgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (hipblasComplex *) &alpha_p[0],
+				   (hipblasComplex **)&Amk[0], lda,
+				   (hipblasComplex **)&Bkn[0], ldb,
+				   (hipblasComplex *) &beta_p[0],
+				   (hipblasComplex **)&Cmn[0], ldc,
+				   batchCount);
+    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasCgemmBatched(gridblasHandle,
+				  CUBLAS_OP_N,
+				  CUBLAS_OP_N,
+				  m,n,k,
+				  (cuComplex *) &alpha_p[0],
+				  (cuComplex **)&Amk[0], lda,
+				  (cuComplex **)&Bkn[0], ldb,
+				  (cuComplex *) &beta_p[0],
+				  (cuComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexD c_mn(0.0);
+	  for (int kk = 0; kk < k, ++kk)
+	    c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	  Cmn[mm + nn*ldc + p*sdc] =  (*alpha_p)*c_mn + (*beta_p)*Cmn[mm + nn*ldc + p*sdc];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Single precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    static deviceVector<RealF> alpha_p(1);
+    static deviceVector<RealF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
+    RealD t0=usecond();
+    //       std::cout << "hipblasZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    auto err = hipblasSgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (float *) &alpha_p[0],
+				   (float **)&Amk[0], lda,
+				   (float **)&Bkn[0], ldb,
+				   (float *) &beta_p[0],
+				   (float **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasSgemmBatched(gridblasHandle,
+				  CUBLAS_OP_N,
+				  CUBLAS_OP_N,
+				  m,n,k,
+				  (float *) &alpha_p[0],
+				  (float **)&Amk[0], lda,
+				  (float **)&Bkn[0], ldb,
+				  (float *) &beta_p[0],
+				  (float **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k, ++kk)
+	    c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	  Cmn[mm + nn*ldc + p*sdc] =  (*alpha_p)*c_mn + (*beta_p)*Cmn[mm + nn*ldc + p*sdc];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+  
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Double precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    static deviceVector<RealD> alpha_p(1);
+    static deviceVector<RealD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
+    RealD t0=usecond();
+    //       std::cout << "hipblasZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    auto err = hipblasDgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (double *) &alpha_p[0],
+				   (double **)&Amk[0], lda,
+				   (double **)&Bkn[0], ldb,
+				   (double *) &beta_p[0],
+				   (double **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasDgemmBatched(gridblasHandle,
+				  CUBLAS_OP_N,
+				  CUBLAS_OP_N,
+				  m,n,k,
+				  (double *) &alpha_p[0],
+				  (double **)&Amk[0], lda,
+				  (double **)&Bkn[0], ldb,
+				  (double *) &beta_p[0],
+				  (double **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    /*
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t batchCount64=batchCount;
+      oneapi::mkl::blas::column_major::gemm_batch(*theGridAccelerator,
+      onemkl::transpose::N,
+      onemkl::transpose::N,
+      &m64,&n64,&k64,
+      (double *) &alpha_p[0],
+      (double **)&Amk[0], lda,
+      (double **)&Bkn[0], ldb,
+      (double *) &beta_p[0],
+      (double **)&Cmn[0], ldc,
+      1,&batchCount64);
+     */
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k, ++kk)
+	    c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	  Cmn[mm + nn*ldc + p*sdc] =  (*alpha_p)*c_mn + (*beta_p)*Cmn[mm + nn*ldc + p*sdc];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+  
+
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // Strided case used by benchmark, but generally unused in Grid
+  // Keep a code example in double complex, but don't generate the single and real variants for now
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  
+  void gemmStridedBatched(int m,int n, int k,
+			  ComplexD alpha,
+			  ComplexD* Amk,  // pointer list to matrices
+			  ComplexD* Bkn,
+			  ComplexD beta,
+			  ComplexD* Cmn,
+			  int batchCount)
+  {
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    int sda = m*k;
+    int sdb = k*n;
+    int sdc = m*n;
+    deviceVector<ComplexD> alpha_p(1);
+    deviceVector<ComplexD> beta_p(1);
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+    std::cout << "blasZgemmStridedBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    std::cout << "blasZgemmStridedBatched ld   "<<lda<<","<<ldb<<","<<ldc<<std::endl;
+    std::cout << "blasZgemmStridedBatched sd   "<<sda<<","<<sdb<<","<<sdc<<std::endl;
+#ifdef GRID_HIP
+    auto err = hipblasZgemmStridedBatched(gridblasHandle,
+					  HIPBLAS_OP_N,
+					  HIPBLAS_OP_N,
+					  m,n,k,
+					  (hipblasDoubleComplex *) &alpha_p[0],
+					  (hipblasDoubleComplex *) Amk, lda, sda,
+					  (hipblasDoubleComplex *) Bkn, ldb, sdb,
+					  (hipblasDoubleComplex *) &beta_p[0],
+					  (hipblasDoubleComplex *) Cmn, ldc, sdc,
+					  batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasZgemmStridedBatched(gridblasHandle,
+			      CUBLAS_OP_N,
+			      CUBLAS_OP_N,
+			      m,n,k,
+			      (cuDoubleComplex *) &alpha_p[0],
+			      (cuDoubleComplex *) Amk, lda, sda,
+			      (cuDoubleComplex *) Bkn, ldb, sdb,
+			      (cuDoubleComplex *) &beta_p[0],
+			      (cuDoubleComplex *) Cmn, ldc, sdc,
+			      batchCount);
+#endif
+#ifdef GRID_SYCL
+     #warning "oneMKL implementation not made "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+     // Need a default/reference implementation
+     for (int p = 0; p < batchCount; ++p) {
+       for (int mm = 0; mm < m; ++mm) {
+	 for (int nn = 0; nn < n; ++nn) {
+	   ComplexD c_mn(0.0);
+	   for (int kk = 0; kk < k, ++kk)
+	     c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	   Cmn[mm + nn*ldc + p*sdc] =  (*alpha_p)*c_mn + (*beta_p)*Cmn[mm + nn*ldc + p*sdc];
+	 }
+       }
+     }
+#endif
+  }
+
+
+
+
+};
+
+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 +1,467 @@
+/*************************************************************************************
+
+    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 Lattice<CComplex >    FineComplexField;
+  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 _CoarseGrid; };   // 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 ShiftMatrix(RealD shift)
+  {
+    int Nd=_FineGrid->Nd(); 
+    Coordinate zero_shift(Nd,0);
+    for(int p=0;p<geom.npoint;p++){
+      if ( zero_shift==geom.shifts[p] ) {
+	_A[p] = _A[p]+shift;
+	_Adag[p] = _Adag[p]+shift;
+      }
+    }    
+  }
+  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;
+    }
+    _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;
+    RealD tmult2=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 point= bp/nbasis;
+	  int32_t b    = bp%nbasis;
+	  auto SE  = Stencil_v.GetEntry(point,ss);
+	  auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
+	  auto res = coalescedRead(Aview_p[point][ss](0,b))*nbr(0);
+	  for(int bb=1;bb<nbasis;bb++) {
+	    res = res + coalescedRead(Aview_p[point][ss](bb,b))*nbr(bb);
+	  }
+	  coalescedWrite(Vview_p[point][ss](b),res);
+      });
+      tmult2-=usecond();
+      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);
+      });
+      tmult2+=usecond();
+      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 1rhs 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<<" of which mult2  "<<tmult2<<" 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 flops "<< flops<<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 tphaseBZ=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
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid());
+    
+    CoarseVector coarseInner(CoarseGrid());
+    
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    tphase=-usecond();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid());
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+      
+    }
+    tphase+=usecond();
+    
+    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
+	tphaseBZ-=usecond();
+	phaV = phaF[p]*Subspace.subspace[i];
+	tphaseBZ+=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();
+	blockProjectFast(coarseInner,MphaV,Subspace.subspace);
+	coarseInner = conjugate(pha[p]) * 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](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    // 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 phaseBZ "<<tphaseBZ<<" 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.ExchangePeriodic(_A[p]);
+      _Adag[p]= Cell.ExchangePeriodic(_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/GeneralCoarsenedMatrixMultiRHS.h

@@ -0,0 +1,402 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrixMultiRHS.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/algorithms/multigrid/BatchedBlas.h>
+
+NAMESPACE_BEGIN(Grid);
+
+
+// Move this to accelerator.h
+// Also give a copy device.
+// Rename acceleratorPut
+// Rename acceleratorGet
+template<class T> void deviceSet(T& dev,T&host)
+{
+  acceleratorCopyToDevice(&host,&dev,sizeof(T));
+}
+template<class T> T deviceGet(T& dev)
+{
+  T host;
+  acceleratorCopyFromDevice(&dev,&host,sizeof(T));
+  return host;
+}
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class MultiGeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+  typedef typename CComplex::scalar_object SComplex;
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef MultiGeneralCoarsenedMatrix<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
+
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  typedef iVector<SComplex,nbasis >           calcVector;
+  typedef iMatrix<SComplex,nbasis >           calcMatrix;
+  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
+  ////////////////////
+  GridCartesian *       _CoarseGridMulti; 
+  GridCartesian *       _CoarseGrid;
+  GeneralCoarseOp &     _Op;
+  NonLocalStencilGeometry geom;
+  PaddedCell Cell;
+  GeneralLocalStencil Stencil;
+
+  deviceVector<calcVector> BLAS_B;
+  deviceVector<calcVector> BLAS_C;
+  std::vector<deviceVector<calcMatrix> > BLAS_A;
+
+  std::vector<deviceVector<ComplexD *> > BLAS_AP;
+  std::vector<deviceVector<ComplexD *> > BLAS_BP;
+  deviceVector<ComplexD *>               BLAS_CP;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase      * Grid(void)           { return _CoarseGridMulti; };   // this is all the linalg routines need to know
+  GridCartesian * CoarseGrid(void)     { return _CoarseGridMulti; };   // this is all the linalg routines need to know
+
+  MultiGeneralCoarsenedMatrix(GeneralCoarseOp & Op,GridCartesian *CoarseGridMulti) :
+    _Op(Op),
+    _CoarseGrid(Op.CoarseGrid()),
+    _CoarseGridMulti(CoarseGridMulti),
+    geom(_CoarseGridMulti,Op.geom.hops,Op.geom.skip+1),
+    Cell(Op.geom.Depth(),_CoarseGridMulti),
+    Stencil(Cell.grids.back(),geom.shifts) // padded cell stencil
+  {
+    int32_t padded_sites   = _Op._A[0].Grid()->lSites();
+    int32_t unpadded_sites = _CoarseGrid->lSites();
+    
+    int32_t nrhs  = CoarseGridMulti->FullDimensions()[0];  // # RHS
+    int32_t orhs  = nrhs/CComplex::Nsimd();
+
+    /////////////////////////////////////////////////
+    // Device data vector storage
+    /////////////////////////////////////////////////
+    BLAS_A.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_A[p].resize (unpadded_sites); // no ghost zone, npoint elements
+    }
+    BLAS_B.resize(nrhs *padded_sites);   // includes ghost zone
+    BLAS_C.resize(nrhs *unpadded_sites); // no ghost zone
+
+    BLAS_AP.resize(geom.npoint);
+    BLAS_BP.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_AP[p].resize(unpadded_sites);
+      BLAS_BP[p].resize(unpadded_sites);
+    }
+    BLAS_CP.resize(unpadded_sites);
+
+    /////////////////////////////////////////////////
+    // Pointers to data
+    /////////////////////////////////////////////////
+
+    // Site identity mapping for A, C
+    for(int p=0;p<geom.npoint;p++){
+      for(int ss=0;ss<unpadded_sites;ss++){
+	ComplexD *ptr = (ComplexD *)&BLAS_A[p][ss];
+	//ComplexD *ptr = (ComplexD *)&BLAS_A[p][0]; std::cout << " A ptr "<<std::hex<<ptr<<std::dec<<" "<<ss<<"/"<<BLAS_A[p].size()<<std::endl;
+	deviceSet(BLAS_AP[p][ss],ptr);
+      }
+    }
+    for(int ss=0;ss<unpadded_sites;ss++){
+      ComplexD *ptr = (ComplexD *)&BLAS_C[ss*nrhs];
+      //ComplexD *ptr = (ComplexD *)&BLAS_C[0];  std::cout << " C ptr "<<std::hex<<ptr<<std::dec<<" "<<ss<<"/"<<BLAS_C.size()<<std::endl;
+      deviceSet(BLAS_CP[ss],ptr);
+    }
+
+    /////////////////////////////////////////////////
+    // Neighbour table is more complicated
+    /////////////////////////////////////////////////
+    int32_t j=0; // Interior point counter (unpadded)
+    for(int32_t s=0;s<padded_sites;s++){ // 4 volume, padded
+      int ghost_zone=0;
+      for(int32_t point = 0 ; point < geom.npoint; point++){
+	int i=s*orhs*geom.npoint+point;
+	if( Stencil._entries[i]._wrap ) { // stencil is indexed by the oSite of the CoarseGridMulti, hence orhs factor
+	  ghost_zone=1; // If general stencil wrapped in any direction, wrap=1
+	}
+      }
+      //      GeneralStencilEntryReordered tmp;
+      if( ghost_zone==0) {
+	for(int32_t point = 0 ; point < geom.npoint; point++){
+	  int i=s*orhs*geom.npoint+point;
+ 	  int32_t nbr = Stencil._entries[i]._offset*CComplex::Nsimd(); // oSite -> lSite
+	  //	  std::cout << " B ptr "<< nbr<<"/"<<BLAS_B.size()<<std::endl;
+	  assert(nbr<BLAS_B.size());
+	  ComplexD * ptr = (ComplexD *)&BLAS_B[nbr];
+	  //	  ComplexD * ptr = (ComplexD *)&BLAS_B[0];
+	  //	  std::cout << " B ptr unpadded "<<std::hex<<ptr<<std::dec<<" "<<s<<"/"<<padded_sites<<std::endl;
+	  //	  std::cout << " B ptr   padded "<<std::hex<<ptr<<std::dec<<" "<<j<<"/"<<unpadded_sites<<std::endl;
+	  deviceSet(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
+	  //	  auto tmp = deviceGet(*BLAS_BP[point][j]);  // debug trigger SEGV if bad ptr
+	}
+	j++;
+      }
+    }
+    assert(j==unpadded_sites);
+    CopyMatrix();
+  }
+  template<class vobj> void GridtoBLAS(const Lattice<vobj> &from,deviceVector<typename vobj::scalar_object> &to)
+  {
+#if 0
+    std::vector<typename vobj::scalar_object> tmp;
+    unvectorizeToLexOrdArray(tmp,from);
+    assert(tmp.size()==from.Grid()->lSites());
+    assert(tmp.size()==to.size());
+    to.resize(tmp.size());
+    acceleratorCopyToDevice(&tmp[0],&to[0],sizeof(typename vobj::scalar_object)*tmp.size());
+#else
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Fg = from.Grid();
+  assert(!Fg->_isCheckerBoarded);
+  int nd = Fg->_ndimension;
+
+  to.resize(Fg->lSites());
+
+  Coordinate LocalLatt = Fg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+
+  autoView(from_v,from,AcceleratorRead);
+  auto to_v = &to[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate from_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	from_coor[i] = base[i];
+      }
+      int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      scalar_type* to = (scalar_type *)&to_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	to[w] = stmp;
+      }
+    });
+#endif
+  }    
+  template<class vobj> void BLAStoGrid(Lattice<vobj> &grid,deviceVector<typename vobj::scalar_object> &in)
+  {
+#if 0
+    std::vector<typename vobj::scalar_object> tmp;
+    tmp.resize(in.size());
+    //    std::cout << "BLAStoGrid volume " <<tmp.size()<<" "<< grid.Grid()->lSites()<<std::endl;
+    assert(in.size()==grid.Grid()->lSites());
+    acceleratorCopyFromDevice(&in[0],&tmp[0],sizeof(typename vobj::scalar_object)*in.size());
+    vectorizeFromLexOrdArray(tmp,grid);
+#else
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Tg = grid.Grid();
+  assert(!Tg->_isCheckerBoarded);
+  int nd = Tg->_ndimension;
+  
+  assert(in.size()==Tg->lSites());
+
+  Coordinate LocalLatt = Tg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
+
+  autoView(to_v,grid,AcceleratorWrite);
+  auto from_v = &in[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate to_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	to_coor[i] = base[i];
+      }
+      int to_oidx = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      scalar_type* from = (scalar_type *)&from_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp=from[w];
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+#endif
+  }
+  void CopyMatrix (void)
+  {
+    // Clone "A" to be lexicographic in the physics coords
+    // Use unvectorisetolexordarray
+    // Copy to device
+    for(int p=0;p<geom.npoint;p++){
+      //Unpadded
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      //      Coordinate coor({0,0,0,0,0});
+      //      auto sval = peekSite(Aup,coor);
+      //      std::cout << "CopyMatrix: p "<<p<<" Aup[0] :"<<sval<<std::endl;
+      //      sval = peekSite(_Op._A[p],coor);
+      //      std::cout << "CopyMatrix: p "<<p<<" _Op._Ap[0] :"<<sval<<std::endl;
+      GridtoBLAS(Aup,BLAS_A[p]);
+      //      std::cout << "Copy Matrix p "<<p<<" "<< deviceGet(BLAS_A[p][0])<<std::endl;
+    }
+  }
+  void Mdag(const CoarseVector &in, CoarseVector &out)
+  {
+    this->M(in,out);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    std::cout << GridLogMessage << "New Mrhs coarse"<<std::endl;
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    RealD t_tot;
+    RealD t_exch;
+    RealD t_GtoB;
+    RealD t_BtoG;
+    RealD t_mult;
+
+    t_tot=-usecond();
+    CoarseVector tin=in;
+    t_exch=-usecond();
+    CoarseVector pin = Cell.ExchangePeriodic(tin); //padded input
+    t_exch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef calcMatrix* Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+
+    RealD flops,bytes;
+    int64_t osites=in.Grid()->oSites(); // unpadded
+    int64_t unpadded_vol = _CoarseGrid->lSites();
+    
+    flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    bytes = 1.0*osites*sizeof(siteMatrix)*npoint/pin.Grid()->GlobalDimensions()[0]
+          + 2.0*osites*sizeof(siteVector)*npoint;
+    
+    int64_t nrhs  =pin.Grid()->GlobalDimensions()[0];
+    assert(nrhs>=1);
+
+    std::cout << GridLogMessage << "New Mrhs GridtoBLAS in sizes "<<in.Grid()->lSites()<<" "<<pin.Grid()->lSites()<<std::endl;
+    t_GtoB=-usecond();
+    GridtoBLAS(pin,BLAS_B);
+    //    out = Zero();
+    //    GridtoBLAS(out,BLAS_C);
+    t_GtoB+=usecond();
+
+    GridBLAS BLAS;
+
+    t_mult=-usecond();
+    for(int p=0;p<geom.npoint;p++){
+      RealD c = 1.0;
+      if (p==0) c = 0.0;
+      ComplexD beta(c);
+      //      std::cout << GridLogMessage << "New Mrhs coarse gemmBatched "<<p<<std::endl;
+      BLAS.gemmBatched(nbasis,nrhs,nbasis,
+		       ComplexD(1.0),
+		       BLAS_AP[p], 
+		       BLAS_BP[p], 
+		       ComplexD(c), 
+		       BLAS_CP);
+    }
+    BLAS.synchronise();
+    t_mult+=usecond();
+    //    std::cout << GridLogMessage << "New Mrhs coarse BLAStoGrid "<<std::endl;
+    t_BtoG=-usecond();
+    BLAStoGrid(out,BLAS_C);
+    t_BtoG+=usecond();
+    t_tot+=usecond();
+    //    auto check =deviceGet(BLAS_C[0]);
+    //      std::cout << "C[0] "<<check<<std::endl;
+    //    Coordinate coor({0,0,0,0,0,0});
+    //    peekLocalSite(check,out,coor);
+    //    std::cout << "C[0] "<< check<<std::endl;
+    std::cout << GridLogMessage << "New Mrhs coarse DONE "<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult exch "<<t_exch<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult mult "<<t_mult<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult GtoB  "<<t_GtoB<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult BtoG  "<<t_BtoG<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult tot  "<<t_tot<<" us"<<std::endl;
+    std::cout << GridLogMessage<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Kernel flops "<< flops<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Kernel flop/s "<< flops/t_mult<<" mflop/s"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Kernel bytes/s "<< bytes/t_mult/1000<<" GB/s"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse overall flops/s "<< flops/t_tot<<" mflop/s"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+  };
+  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 +1,238 @@
+/*************************************************************************************
+
+    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 skip;
+  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
+  int DimSkip(void){return skip;};
+
+  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,int _skip) : grid(_coarse_grid), hops(_hops), skip(_skip)
+  {
+    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;
+     }
+    }
+    this->BuildShifts();
+  };
+
+};
+
+// Need to worry about red-black now
+class NonLocalStencilGeometry4D : public NonLocalStencilGeometry {
+public:
+  virtual int DerivedDimSkip(void) { return 0;};
+  NonLocalStencilGeometry4D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,0) { };
+  virtual ~NonLocalStencilGeometry4D() {};
+};
+class NonLocalStencilGeometry5D : public NonLocalStencilGeometry {
+public:
+  virtual int DerivedDimSkip(void) { return 1; }; 
+  NonLocalStencilGeometry5D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,1)  { };
+  virtual ~NonLocalStencilGeometry5D() {};
+};
+/*
+ * Bunch of different options classes
+ */
+class NextToNextToNextToNearestStencilGeometry4D : public NonLocalStencilGeometry4D {
+public:
+  NextToNextToNextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,4)
+  {
+  };
+};
+class NextToNextToNextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNextToNextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,4)
+  {
+  };
+};
+class NextToNearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,2)
+  {
+  };
+};
+class NextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,2)
+  {
+  };
+};
+class NearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,1)
+  {
+  };
+};
+class NearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,1)
+  {
+  };
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/MultiGrid.h

@@ -0,0 +1,35 @@
+    /*************************************************************************************
+
+    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/BatchedBlas.h>
+#include <Grid/algorithms/multigrid/CoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h>

Grid/allocator/AlignedAllocator.h

@@ -175,8 +175,56 @@ template<class T> using cshiftAllocator = std::allocator<T>;
 
 template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
 template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,devAllocator<T> >;
-template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
+template<class T> using commVector    = std::vector<T,devAllocator<T> >;
+template<class T> using deviceVector  = std::vector<T,devAllocator<T> >;
+template<class T> using cshiftVector  = std::vector<T,cshiftAllocator<T> >;
+
+/*
+template<class T> class vecView
+{
+ protected:
+  T * data;
+  uint64_t size;
+  ViewMode mode;
+  void * cpu_ptr;
+ public:
+  accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
+  vecView(std::vector<T> &refer_to_me,ViewMode _mode)
+  {
+    cpu_ptr = &refer_to_me[0];
+    size = refer_to_me.size();
+    mode = _mode;
+    data =(T *) MemoryManager::ViewOpen(cpu_ptr,
+					size*sizeof(T),
+					mode,
+					AdviseDefault);
+  }
+  void ViewClose(void)
+  { // Inform the manager
+    MemoryManager::ViewClose(this->cpu_ptr,this->mode);    
+  }
+};
+
+template<class T> vecView<T> VectorView(std::vector<T> &vec,ViewMode _mode)
+{
+  vecView<T> ret(vec,_mode); // does the open
+  return ret;                // must be closed
+}
+
+// Little autoscope assister
+template<class View> 
+class VectorViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+ public:
+  VectorViewCloser(View &_v) : v(_v) {};
+  ~VectorViewCloser() { auto ptr = v.cpu_ptr; v.ViewClose();  MemoryManager::NotifyDeletion(ptr);}
+};
+
+#define autoVecView(v_v,v,mode)					\
+  auto v_v = VectorView(v,mode);				\
+  ViewCloser<decltype(v_v)> _autoView##v_v(v_v);
+*/
 
 NAMESPACE_END(Grid);
 

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/communicator/SharedMemoryMPI.cc

@@ -27,9 +27,10 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
+#define Mheader "SharedMemoryMpi: "
+
 #include <Grid/GridCore.h>
 #include <pwd.h>
-#include <syscall.h>
 
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
@@ -39,11 +40,118 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #endif
 #ifdef GRID_SYCL
 #define GRID_SYCL_LEVEL_ZERO_IPC
+#include <syscall.h>
+#define SHM_SOCKETS 
+#endif
+
+#include <sys/socket.h>
+#include <sys/un.h>
+
+NAMESPACE_BEGIN(Grid); 
+
+#ifdef SHM_SOCKETS
+
+/*
+ * Barbaric extra intranode communication route in case we need sockets to pass FDs
+ * Forced by level_zero not being nicely designed
+ */
+static int sock;
+static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
+static char sock_path[256];
+class UnixSockets {
+public:
+  static void Open(int rank)
+  {
+    int errnum;
+
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
+    unlink(sa_un.sun_path);
+    if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
+      perror("bind failure");
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  static int RecvFileDescriptor(void)
+  {
+    int n;
+    int fd;
+    char buf[1];
+    struct iovec iov;
+    struct msghdr msg;
+    struct cmsghdr *cmsg;
+    char cms[CMSG_SPACE(sizeof(int))];
+
+    iov.iov_base = buf;
+    iov.iov_len = 1;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_name = 0;
+    msg.msg_namelen = 0;
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+
+    msg.msg_control = (caddr_t)cms;
+    msg.msg_controllen = sizeof cms;
+
+    if((n=recvmsg(sock, &msg, 0)) < 0) {
+      perror("recvmsg failed");
+      return -1;
+    }
+    if(n == 0){
+      perror("recvmsg returned 0");
+      return -1;
+    }
+    cmsg = CMSG_FIRSTHDR(&msg);
+
+    memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
+
+    return fd;
+  }
+
+  static void SendFileDescriptor(int fildes,int xmit_to_rank)
+  {
+    struct msghdr msg;
+    struct iovec iov;
+    struct cmsghdr *cmsg = NULL;
+    char ctrl[CMSG_SPACE(sizeof(int))];
+    char data = ' ';
+
+    memset(&msg, 0, sizeof(struct msghdr));
+    memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
+    iov.iov_base = &data;
+    iov.iov_len = sizeof(data);
+    
+    sprintf(sock_path,sock_path_fmt,xmit_to_rank);
+    
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
+
+    msg.msg_name = (void *)&sa_un;
+    msg.msg_namelen = sizeof(sa_un);
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+    msg.msg_controllen =  CMSG_SPACE(sizeof(int));
+    msg.msg_control = ctrl;
+
+    cmsg = CMSG_FIRSTHDR(&msg);
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int));
+
+    *((int *) CMSG_DATA(cmsg)) = fildes;
+
+    sendmsg(sock, &msg, 0);
+  };
+};
 #endif
 
 
-NAMESPACE_BEGIN(Grid); 
-#define header "SharedMemoryMpi: "
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -66,8 +174,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
   if ( WorldRank == 0) {
-    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
-    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+    std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl;
   }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
@@ -344,7 +452,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -429,7 +537,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
 
-  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+  std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
 
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -472,7 +580,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -480,8 +588,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef SHM_SOCKETS
+  UnixSockets::Open(WorldShmRank);
+#endif
   for(int r=0;r<WorldShmSize;r++){
 
+    MPI_Barrier(WorldShmComm);
+
 #ifndef GRID_MPI3_SHM_NONE
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
@@ -489,24 +602,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void * thisBuf = ShmCommBuf;
     if(!Stencil_force_mpi) {
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+    typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
-
+    
     if ( r==WorldShmRank ) { 
       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
+      memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
+#ifdef SHM_SOCKETS
+      for(int rr=0;rr<WorldShmSize;rr++){
+	if(rr!=r){
+	  UnixSockets::SendFileDescriptor(handle.fd,rr);
+	}
+      }
+#endif
     }
 #endif
 #ifdef GRID_CUDA
@@ -534,6 +655,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
     { 
+      MPI_Barrier(WorldShmComm);
       int ierr=MPI_Bcast(&handle,
 			 sizeof(handle),
 			 MPI_BYTE,
@@ -549,6 +671,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     if ( r!=WorldShmRank ) {
       thisBuf = nullptr;
+      int myfd;
+#ifdef SHM_SOCKETS
+      myfd=UnixSockets::RecvFileDescriptor();
+#else
       std::cout<<"mapping seeking remote pid/fd "
 	       <<handle.pid<<"/"
 	       <<handle.fd<<std::endl;
@@ -556,16 +682,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
       std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
-      int myfd  = syscall(438,pidfd,handle.fd,0);
-
-      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-      
+      myfd  = syscall(438,pidfd,handle.fd,0);
+      int err_t = errno;
+      if (myfd < 0) {
+        fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr);
+	perror("pidfd_getfd failed ");
+	assert(0);
+      }
+#endif
+      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
       auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
@@ -600,6 +732,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
+    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
@@ -611,7 +744,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -648,7 +781,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -658,7 +791,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -705,7 +838,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_ET.h

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

Grid/lattice/Lattice_base.h

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

Grid/lattice/Lattice_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_reduction.h

@@ -203,6 +203,27 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
   return real(nrm); 
 }
 
+
+template<class Op,class T1>
+inline auto norm2(const LatticeUnaryExpression<Op,T1> & expr)  ->RealD
+{
+  return norm2(closure(expr));
+}
+
+template<class Op,class T1,class T2>
+inline auto norm2(const LatticeBinaryExpression<Op,T1,T2> & expr)      ->RealD
+{
+  return norm2(closure(expr));
+}
+
+
+template<class Op,class T1,class T2,class T3>
+inline auto norm2(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)      ->RealD
+{
+  return norm2(closure(expr));
+}
+
+
 //The global maximum of the site norm2
 template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
 {

Grid/lattice/Lattice_reduction_gpu.h

@@ -30,7 +30,7 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   cudaGetDevice(&device);
 #endif
 #ifdef GRID_HIP
-  hipGetDevice(&device);
+  auto discard=hipGetDevice(&device);
 #endif
   
   Iterator warpSize            = gpu_props[device].warpSize;

Grid/lattice/Lattice_rng.h

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

Grid/lattice/Lattice_trace.h

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

Grid/lattice/Lattice_transfer.h

@@ -265,8 +265,8 @@ inline auto localInnerProductD(const Lattice<vobj> &lhs,const Lattice<vobj> &rhs
 ////////////////////////////////////////////////////////////////////////////////////////////
 template<class vobj,class CComplex,int nbasis,class VLattice>
 inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
-			   const             Lattice<vobj>   &fineData,
-			   const VLattice &Basis)
+			 const             Lattice<vobj>   &fineData,
+			 const VLattice &Basis)
 {
   GridBase * fine  = fineData.Grid();
   GridBase * coarse= coarseData.Grid();
@@ -276,18 +276,65 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
 
   autoView( coarseData_ , coarseData, AcceleratorWrite);
   autoView( ip_         , ip,         AcceleratorWrite);
+  RealD t_IP=0;
+  RealD t_co=0;
+  RealD t_za=0;
   for(int v=0;v<nbasis;v++) {
+    t_IP-=usecond();
     blockInnerProductD(ip,Basis[v],fineDataRed); // ip = <basis|fine>
+    t_IP+=usecond();
+    t_co-=usecond();
     accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
 	convertType(coarseData_[sc](v),ip_[sc]);
     });
+    t_co+=usecond();
 
     // improve numerical stability of projection
     // |fine> = |fine> - <basis|fine> |basis>
     ip=-ip;
+    t_za-=usecond();
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
+    t_za+=usecond();
   }
+  //  std::cout << GridLogPerformance << " blockProject : blockInnerProduct :  "<<t_IP<<" us"<<std::endl;
+  //  std::cout << GridLogPerformance << " blockProject : conv              :  "<<t_co<<" us"<<std::endl;
+  //  std::cout << GridLogPerformance << " blockProject : blockZaxpy        :  "<<t_za<<" us"<<std::endl;
 }
+
+
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void blockProjectFast(Lattice<iVector<CComplex,nbasis > > &coarseData,
+			     const             Lattice<vobj>   &fineData,
+			     const VLattice &Basis)
+{
+  GridBase * fine  = fineData.Grid();
+  GridBase * coarse= coarseData.Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  Lattice<vobj>     fineDataRed = fineData;
+
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( ip_         , ip,         AcceleratorWrite);
+  RealD t_IP=0;
+  RealD t_co=0;
+  for(int v=0;v<nbasis;v++) {
+    t_IP-=usecond();
+    blockInnerProductD(ip,Basis[v],fineData); // ip = <basis|fine>
+    t_IP+=usecond();
+    t_co-=usecond();
+    accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+	convertType(coarseData_[sc](v),ip_[sc]);
+    });
+    t_co+=usecond();
+  }
+  //  std::cout << GridLogPerformance << " blockProjectFast : blockInnerProduct :  "<<t_IP<<" us"<<std::endl;
+  //  std::cout << GridLogPerformance << " blockProjectFast : conv              :  "<<t_co<<" us"<<std::endl;
+}
+
+
+// This only minimises data motion from CPU to GPU
+// there is chance of better implementation that does a vxk loop of inner products to data share
+// at the GPU thread level
 template<class vobj,class CComplex,int nbasis,class VLattice>
 inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
                                const std::vector<Lattice<vobj>> &fineData,
@@ -393,8 +440,15 @@ template<class vobj,class CComplex>
   Lattice<dotp> coarse_inner(coarse);
 
   // Precision promotion
+  RealD t;
+  t=-usecond();
   fine_inner = localInnerProductD<vobj>(fineX,fineY);
+  //  t+=usecond(); std::cout << GridLogPerformance << " blockInnerProduct : localInnerProductD "<<t<<" us"<<std::endl;
+  
+  t=-usecond();
   blockSum(coarse_inner,fine_inner);
+  //  t+=usecond(); std::cout << GridLogPerformance << " blockInnerProduct : blockSum "<<t<<" us"<<std::endl;
+  t=-usecond();
   {
     autoView( CoarseInner_  , CoarseInner,AcceleratorWrite);
     autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
@@ -402,6 +456,7 @@ template<class vobj,class CComplex>
       convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
     });
   }
+  //  t+=usecond(); std::cout << GridLogPerformance << " blockInnerProduct : convertType "<<t<<" us"<<std::endl;
  
 }
 
@@ -444,6 +499,9 @@ inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 template<class vobj>
 inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData) 
 {
+  const int maxsubsec=256;
+  typedef iVector<vobj,maxsubsec> vSubsec;
+
   GridBase * fine  = fineData.Grid();
   GridBase * coarse= coarseData.Grid();
 
@@ -463,37 +521,62 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   autoView( coarseData_ , coarseData, AcceleratorWrite);
   autoView( fineData_   , fineData, AcceleratorRead);
 
-  auto coarseData_p = &coarseData_[0];
-  auto fineData_p = &fineData_[0];
+  auto coarseData_p  = &coarseData_[0];
+  auto fineData_p    = &fineData_[0];
   
   Coordinate fine_rdimensions = fine->_rdimensions;
   Coordinate coarse_rdimensions = coarse->_rdimensions;
 
   vobj zz = Zero();
-  
-  accelerator_for(sc,coarse->oSites(),1,{
 
+  // Somewhat lazy calculation
+  // Find the biggest power of two subsection divisor less than or equal to maxsubsec
+  int subsec=maxsubsec;
+  int subvol;
+  subvol=blockVol/subsec;
+  while(subvol*subsec!=blockVol){
+    subsec = subsec/2;
+    subvol=blockVol/subsec;
+  };
+
+  Lattice<vSubsec> coarseTmp(coarse);
+  autoView( coarseTmp_, coarseTmp, AcceleratorWriteDiscard);
+  auto coarseTmp_p= &coarseTmp_[0];
+  
+  // Sum within subsecs in a first kernel
+  accelerator_for(sce,subsec*coarse->oSites(),vobj::Nsimd(),{
+
+      int sc=sce/subsec;
+      int e=sce%subsec;
+      
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
 
-      vobj cd = zz;
-      
-      for(int sb=0;sb<blockVol;sb++){
-
+      auto cd = coalescedRead(zz);
+      for(int sb=e*subvol;sb<MIN((e+1)*subvol,blockVol);sb++){
 	int sf;
 	Coordinate coor_b(_ndimension);
 	Coordinate coor_f(_ndimension);
 	Lexicographic::CoorFromIndex(coor_b,sb,block_r);               // Block sub coordinate
 	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(coarseTmp_[sc](e),cd);
 
     });
+   // Sum across subsecs in a second kernel
+   accelerator_for(sc,coarse->oSites(),vobj::Nsimd(),{
+      auto cd = coalescedRead(coarseTmp_p[sc](0));
+      for(int e=1;e<subsec;e++){
+	cd=cd+coalescedRead(coarseTmp_p[sc](e));
+      }
+      coalescedWrite(coarseData_p[sc],cd);
+   });
+
   return;
 }
 
@@ -550,7 +633,7 @@ inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> >
   blockOrthonormalize(ip,Basis);
 }
 
-#if 0
+#ifdef GRID_ACCELERATED
 // TODO: CPU optimized version here
 template<class vobj,class CComplex,int nbasis>
 inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
@@ -576,26 +659,37 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   autoView( fineData_   , fineData, AcceleratorWrite);
   autoView( coarseData_ , coarseData, AcceleratorRead);
 
+  typedef LatticeView<vobj> Vview;
+  std::vector<Vview> AcceleratorVecViewContainer_h; 
+  for(int v=0;v<nbasis;v++) {
+    AcceleratorVecViewContainer_h.push_back(Basis[v].View(AcceleratorRead));
+  }
+  static deviceVector<Vview> AcceleratorVecViewContainer; AcceleratorVecViewContainer.resize(nbasis); 
+  acceleratorCopyToDevice(&AcceleratorVecViewContainer_h[0],&AcceleratorVecViewContainer[0],nbasis *sizeof(Vview));
+  auto Basis_p = &AcceleratorVecViewContainer[0];
   // Loop with a cache friendly loop ordering
-  accelerator_for(sf,fine->oSites(),1,{
+  Coordinate frdimensions=fine->_rdimensions;
+  Coordinate crdimensions=coarse->_rdimensions;
+  accelerator_for(sf,fine->oSites(),vobj::Nsimd(),{
     int sc;
     Coordinate coor_c(_ndimension);
     Coordinate coor_f(_ndimension);
 
-    Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
+    Lexicographic::CoorFromIndex(coor_f,sf,frdimensions);
     for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-    Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+    Lexicographic::IndexFromCoor(coor_c,sc,crdimensions);
 
-    for(int i=0;i<nbasis;i++) {
-      /*      auto basis_ = Basis[i],  );*/
-      if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
-      else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
-    }
+    auto sum= coarseData_(sc)(0) *Basis_p[0](sf);
+    for(int i=1;i<nbasis;i++) sum = sum + coarseData_(sc)(i)*Basis_p[i](sf);
+    coalescedWrite(fineData_[sf],sum);
   });
+  for(int v=0;v<nbasis;v++) {
+    AcceleratorVecViewContainer_h[v].ViewClose();
+  }
   return;
-  
 }
 #else
+// CPU version
 template<class vobj,class CComplex,int nbasis,class VLattice>
 inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 			 Lattice<vobj>   &fineData,
@@ -682,8 +776,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
-  static const int words=sizeof(vobj)/sizeof(vector_type);
-
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // the checks should guarantee that the operations are local
+  ////////////////////////////////////////////////////////////////////////////////////////////////
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
   assert(!Fg->_isCheckerBoarded);
@@ -697,42 +792,48 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
 
-  // the above should guarantee that the operations are local
-  Coordinate ldf = Fg->_ldimensions;
-  Coordinate rdf = Fg->_rdimensions;
-  Coordinate isf = Fg->_istride;
-  Coordinate osf = Fg->_ostride;
-  Coordinate rdt = Tg->_rdimensions;
-  Coordinate ist = Tg->_istride;
-  Coordinate ost = Tg->_ostride;
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
 
-  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, AcceleratorRead);
-  accelerator_for(idx,Fg->lSites(),1,{
-    sobj s;
-    Coordinate Fcoor(nd);
-    Coordinate Tcoor(nd);
-    Lexicographic::CoorFromIndex(Fcoor,idx,ldf);
-    int in_region=1;
-    for(int d=0;d<nd;d++){
-      if ( (Fcoor[d] < FromLowerLeft[d]) || (Fcoor[d]>=FromLowerLeft[d]+RegionSize[d]) ){ 
-	in_region=0;
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate from_coor, to_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,RegionSize);
+      for(int i=0;i<nd;i++){
+	from_coor[i] = base[i] + FromLowerLeft[i];
+	to_coor[i] = base[i] + ToLowerLeft[i];
       }
-      Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
-    }
-    if (in_region) {
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
-      vector_type * fp = (vector_type *)&f_v[odx_f];
-      vector_type * tp = (vector_type *)&t_v[odx_t];
+      int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+      int to_oidx   = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane   = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
       for(int w=0;w<words;w++){
-	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
       }
-    }
-  });
+    });
 }
 
 
@@ -824,7 +925,9 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 
 }
 
-
+//FIXME: make this run entirely on GPU
+//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
+//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -841,11 +944,70 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-  }
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
+#if 1
+  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+  
+  thread_for(idx,nsite,{
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
+    lcoor[orthog] = slice_lo;
+    hcoor[orthog] = slice_hi;
+    size_t rem = idx;
+    for(int mu=0;mu<nl;mu++){
+      if(mu != orthog){
+	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
+	lcoor[mu] = hcoor[mu] = xmu;
+      }
+    }
+    int loidx = lg->oIndex(lcoor);
+    int liidx = lg->iIndex(lcoor);
+    int hoidx = hg->oIndex(hcoor);
+    int hiidx = hg->iIndex(hcoor);
+    int* tt = table + 4*idx;
+    tt[0] = loidx;
+    tt[1] = liidx;
+    tt[2] = hoidx;
+    tt[3] = hiidx;
+    });
+   
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(lowDim_v,lowDim,AcceleratorRead);
+  autoView(higherDim_v,higherDim,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
+      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+#else
   // the above should guarantee that the operations are local
   autoView(lowDimv,lowDim,CpuRead);
   autoView(higherDimv,higherDim,CpuWrite);
@@ -861,6 +1023,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       pokeLocalSite(s,higherDimv,hcoor);
     }
   });
+#endif
 }
 
 
@@ -925,7 +1088,7 @@ 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++){
@@ -1611,5 +1774,32 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   }
 }
 
+//////////////////////////////////////////////////////
+// MultiRHS interface support for coarse space
+// -- Simplest possible implementation to begin with
+//////////////////////////////////////////////////////
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void blockProjectMany(Lattice<iVector<CComplex,nbasis > > &coarseIP,
+			     Lattice<iVector<CComplex,nbasis > > &coarseTMP,
+			     const VLattice &fineData, // Basis and fineData necessarily same type
+			     const VLattice &Basis)
+{
+  for(int r=0;r<fineData.size();r++){
+    blockProject(coarseTMP,fineData[r],Basis);
+    InsertSliceLocal(coarseTMP, coarseIP,r,r,0);
+  }
+}
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void blockPromoteMany(Lattice<iVector<CComplex,nbasis > > &coarseIP,
+			     Lattice<iVector<CComplex,nbasis > > &coarseTMP,
+			     const VLattice &fineData, // Basis and fineData necessarily same type
+			     const VLattice &Basis)
+{
+  for(int r=0;r<fineData.size();r++){
+    ExtractSliceLocal(coarseTMP, coarseIP,r,r,0);
+    blockPromote(coarseTMP,fineData[r],Basis);
+  }
+}
+
 NAMESPACE_END(Grid);
 

Grid/lattice/PaddedCell.h

@@ -0,0 +1,571 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/PaddedCell.h
+
+    Copyright (C) 2019
+
+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/cshift/Cshift.h>
+
+NAMESPACE_BEGIN(Grid);
+
+//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
+template<typename vobj>
+struct CshiftImplBase{
+  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
+  virtual ~CshiftImplBase(){}
+};
+template<typename vobj>
+struct CshiftImplDefault: public CshiftImplBase<vobj>{
+  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
+};
+template<typename Gimpl>
+struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
+  typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
+};  
+
+
+/*
+ *
+ * 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;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  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, AcceleratorWrite);
+  accelerator_for(ss, face_ovol/simd[dim],Nsimd,{
+
+    // scalar layout won't coalesce
+#ifdef GRID_SIMT
+      {
+	int blane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int blane=0;blane<Nsimd;blane++) {
+#endif
+	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);
+	const int words=sizeof(vobj)/sizeof(vector_type);
+	vector_type * from = (vector_type *)&buf_p[ss+offset];
+	vector_type * to   = (vector_type *)&lat_v[osite];
+	scalar_type stmp;
+	for(int w=0;w<words;w++){
+	  stmp = getlane(from[w], blane);
+	  putlane(to[w], stmp, lane);
+	}
+      }
+  });
+}
+
+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;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  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
+#ifdef GRID_SIMT
+      {
+	int blane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int blane=0;blane<Nsimd;blane++) {
+#endif
+	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);
+	const int words=sizeof(vobj)/sizeof(vector_type);
+	vector_type * to    = (vector_type *)&buf_p[ss+offset];
+	vector_type * from  = (vector_type *)&lat_v[osite];
+	scalar_type stmp;
+	for(int w=0;w<words;w++){
+	  stmp = getlane(from[w], lane);
+	  putlane(to[w], stmp, blane);
+	}
+      }
+  });
+}
+
+
+class PaddedCell {
+public:
+  GridCartesian * unpadded_grid;
+  int dims;
+  int depth;
+  std::vector<GridCartesian *> grids;
+
+  ~PaddedCell()
+  {
+    DeleteGrids();
+  }
+  PaddedCell(int _depth,GridCartesian *_grid)
+  {
+    unpadded_grid = _grid;
+    depth=_depth;
+    dims=_grid->Nd();
+    AllocateGrids();
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate procs     =unpadded_grid->ProcessorGrid();
+    for(int d=0;d<dims;d++){
+      if ( procs[d] > 1 ) assert(local[d]>=depth);
+    }
+  }
+  void DeleteGrids(void)
+  {
+    Coordinate processors=unpadded_grid->_processors;
+    for(int d=0;d<grids.size();d++){
+      if ( processors[d] > 1 ) { 
+	delete grids[d];
+      }
+    }
+    grids.resize(0);
+  };
+  void AllocateGrids(void)
+  {
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate simd      =unpadded_grid->_simd_layout;
+    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++){
+
+      if ( processors[d] > 1 ) { 
+	plocal[d] += 2*depth; 
+      
+	for(int d=0;d<dims;d++){
+	  global[d] = plocal[d]*processors[d];
+	}
+
+	old_grid = 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();
+    // depends on the MPI spread      
+    Coordinate fll(dims,depth);
+    Coordinate tll(dims,0); // depends on the MPI spread
+    for(int d=0;d<dims;d++){
+      if( processors[d]==1 ) fll[d]=0;
+    }
+    localCopyRegion(in,out,fll,tll,local);
+    return out;
+  }
+  template<class vobj>
+  inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = Expand(d,tmp,cshift); // rvalue && assignment
+    }
+    return tmp;
+  }
+  template<class vobj>
+  inline Lattice<vobj> ExchangePeriodic(const Lattice<vobj> &in) 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); // 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);
+    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]);
+
+    double tins=0, tshift=0;
+
+    int islocal = 0 ;
+    if ( processors[dim] == 1 ) islocal = 1;
+
+    if ( islocal ) {
+
+      // replace with a copy and maybe grid swizzle
+      // return in;??
+      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
+  {
+    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 ) {
+      padded=in; // slightly different interface could avoid a copy operation
+    } else {
+      Face_exchange(in,padded,dim,depth);
+      return padded;
+    }
+    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;
+    RealD t_tot=-usecond();
+    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
+    ////////////////////////////////////////////////////////////////////////////
+    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;
+    t_tot+=usecond();
+
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << t_gather/1000  << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << t_scatter/1000   << "ms"<<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: total  :" << t_tot/1000     << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << depth*4.0*bytes/t_gather << "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << depth*4.0*bytes/t_scatter<< "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms  :" << (RealD)4.0*bytes/t_comms   << "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: face bytes  :" << depth*bytes/1e6 << "MB"<<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/QCD.h

@@ -104,6 +104,7 @@ template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iSca
 template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
 template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
 template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
+template<typename vtype> using iLorentzComplex            = iVector<iScalar<iScalar<vtype> >, Nd > ;
 template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
 template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >;
 template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >;
@@ -178,6 +179,15 @@ typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF;
 typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD;
 typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2;
 
+// LorentzComplex
+typedef iLorentzComplex<Complex  > LorentzComplex;
+typedef iLorentzComplex<ComplexF > LorentzComplexF;
+typedef iLorentzComplex<ComplexD > LorentzComplexD;
+
+typedef iLorentzComplex<vComplex > vLorentzComplex;
+typedef iLorentzComplex<vComplexF> vLorentzComplexF;
+typedef iLorentzComplex<vComplexD> vLorentzComplexD;
+
 // DoubleStored gauge field
 typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
@@ -307,6 +317,10 @@ typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF;
 typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD;
 typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2;
 
+typedef Lattice<vLorentzComplex>  LatticeLorentzComplex;
+typedef Lattice<vLorentzComplexF> LatticeLorentzComplexF;
+typedef Lattice<vLorentzComplexD> LatticeLorentzComplexD;
+
 // DoubleStored gauge field
 typedef Lattice<vDoubleStoredColourMatrix>   LatticeDoubleStoredColourMatrix;
 typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF;

Grid/qcd/action/ActionBase.h

@@ -34,10 +34,24 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+///////////////////////////////////
+// Smart configuration base class
+///////////////////////////////////
+template< class Field >
+class ConfigurationBase
+{
+public:
+  ConfigurationBase() {}
+  virtual ~ConfigurationBase() {}
+  virtual void set_Field(Field& U) =0;
+  virtual void smeared_force(Field&) = 0;
+  virtual Field& get_SmearedU() =0;
+  virtual Field &get_U(bool smeared = false) = 0;
+};
+
 template <class GaugeField >
 class Action 
 {
-
 public:
   bool is_smeared = false;
   RealD deriv_norm_sum;
@@ -77,11 +91,39 @@ public:
   void refresh_timer_stop(void)  { refresh_us+=usecond(); }
   void S_timer_start(void)       { S_us-=usecond(); }
   void S_timer_stop(void)        { S_us+=usecond(); }
+  /////////////////////////////
   // Heatbath?
+  /////////////////////////////
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
   virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
+
+  /////////////////////////////////////////////////////////////
+  // virtual smeared interface through configuration container
+  /////////////////////////////////////////////////////////////
+  virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
+  {
+    refresh(U.get_U(is_smeared),sRNG,pRNG);
+  }
+  virtual RealD S(ConfigurationBase<GaugeField>& U)
+  {
+    return S(U.get_U(is_smeared));
+  }
+  virtual RealD Sinitial(ConfigurationBase<GaugeField>& U) 
+  {
+    return Sinitial(U.get_U(is_smeared));
+  }
+  virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU)
+  {
+    deriv(U.get_U(is_smeared),dSdU); 
+    if ( is_smeared ) {
+      U.smeared_force(dSdU);
+    }
+  }
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
   virtual std::string action_name()    = 0;                             // return the action name
   virtual std::string LogParameters()  = 0;                             // prints action parameters
   virtual ~Action(){}

Grid/qcd/action/ActionCore.h

@@ -30,6 +30,8 @@ directory
 #ifndef QCD_ACTION_CORE
 #define QCD_ACTION_CORE
 
+#include <Grid/qcd/action/gauge/GaugeImplementations.h>
+
 #include <Grid/qcd/action/ActionBase.h>
 NAMESPACE_CHECK(ActionBase);
 #include <Grid/qcd/action/ActionSet.h>

Grid/qcd/action/fermion/Fermion.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -507,6 +507,7 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
+    accelerator_barrier();
   }
 
 };

Grid/qcd/action/fermion/WilsonImpl.h

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

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

@@ -332,8 +332,7 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   /////////////////////////////
   {
     GRID_TRACE("Gather");
-    st.HaloExchangeOptGather(in,compressor);
-    accelerator_barrier();
+    st.HaloExchangeOptGather(in,compressor); // Put the barrier in the routine
   }
   
   std::vector<std::vector<CommsRequest_t> > requests;

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

@@ -423,14 +423,14 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
 #define KERNEL_CALL_EXT(A)						\
-  const uint64_t    NN = Nsite*Ls;					\
   const uint64_t    sz = st.surface_list.size();			\
   auto ptr = &st.surface_list[0];					\
   accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
       int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
+      int sU = sF/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
+    });									\
+  accelerator_barrier();
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
@@ -474,9 +474,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
+     // dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
@@ -506,9 +507,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     // Dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteDagExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Grid/qcd/action/gauge/Gauge.h

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

Grid/qcd/action/gauge/GaugeImplTypes.h

@@ -61,7 +61,7 @@ NAMESPACE_BEGIN(Grid);
   typedef typename Impl::Field Field;
 
 // hardcodes the exponential approximation in the template
-template <class S, int Nrepresentation = Nc, int Nexp = 12 > class GaugeImplTypes {
+template <class S, int Nrepresentation = Nc, int Nexp = 12, class Group = SU<Nc> > class GaugeImplTypes {
 public:
   typedef S Simd;
   typedef typename Simd::scalar_type scalar_type;
@@ -78,8 +78,6 @@ public:
   typedef Lattice<SiteLink>    LinkField; 
   typedef Lattice<SiteField>   Field;
 
-  typedef SU<Nrepresentation> Group;
-
   // Guido: we can probably separate the types from the HMC functions
   // this will create 2 kind of implementations
   // probably confusing the users
@@ -119,6 +117,7 @@ public:
     //
     LinkField Pmu(P.Grid());
     Pmu = Zero();
+
     for (int mu = 0; mu < Nd; mu++) {
       Group::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
       RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR) ;
@@ -126,8 +125,12 @@ public:
       PokeIndex<LorentzIndex>(P, Pmu, mu);
     }
   }
-
-  static inline Field projectForce(Field &P) { return Ta(P); }
+    
+  static inline Field projectForce(Field &P) {
+      Field ret(P.Grid());
+      Group::taProj(P, ret);
+      return ret;
+    }
 
   static inline void update_field(Field& P, Field& U, double ep){
     //static std::chrono::duration<double> diff;
@@ -137,14 +140,15 @@ public:
     autoView(P_v,P,AcceleratorRead);
     accelerator_for(ss, P.Grid()->oSites(),1,{
       for (int mu = 0; mu < Nd; mu++) {
-        U_v[ss](mu) = ProjectOnGroup(Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu));
+          U_v[ss](mu) = Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu);
+          U_v[ss](mu) = Group::ProjectOnGeneralGroup(U_v[ss](mu));
       }
     });
    //auto end = std::chrono::high_resolution_clock::now();
    // diff += end - start;
    // std::cout << "Time to exponentiate matrix " << diff.count() << " s\n";
   }
-
+    
   static inline RealD FieldSquareNorm(Field& U){
     LatticeComplex Hloc(U.Grid());
     Hloc = Zero();
@@ -157,7 +161,7 @@ public:
   }
 
   static inline void Project(Field &U) {
-    ProjectSUn(U);
+    Group::ProjectOnSpecialGroup(U);
   }
 
   static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
@@ -171,6 +175,7 @@ public:
   static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
     Group::ColdConfiguration(pRNG, U);
   }
+
 };
 
 
@@ -178,10 +183,17 @@ typedef GaugeImplTypes<vComplex, Nc> GimplTypesR;
 typedef GaugeImplTypes<vComplexF, Nc> GimplTypesF;
 typedef GaugeImplTypes<vComplexD, Nc> GimplTypesD;
 
+typedef GaugeImplTypes<vComplex, Nc, 12, Sp<Nc> > SpGimplTypesR;
+typedef GaugeImplTypes<vComplexF, Nc, 12, Sp<Nc> > SpGimplTypesF;
+typedef GaugeImplTypes<vComplexD, Nc, 12, Sp<Nc> > SpGimplTypesD;
+
 typedef GaugeImplTypes<vComplex, SU<Nc>::AdjointDimension> GimplAdjointTypesR;
 typedef GaugeImplTypes<vComplexF, SU<Nc>::AdjointDimension> GimplAdjointTypesF;
 typedef GaugeImplTypes<vComplexD, SU<Nc>::AdjointDimension> GimplAdjointTypesD;
 
+
+
+
 NAMESPACE_END(Grid);
 
 #endif // GRID_GAUGE_IMPL_TYPES_H

Grid/qcd/action/gauge/GaugeImplementations.h

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

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

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

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h

@@ -53,9 +53,10 @@ NAMESPACE_BEGIN(Grid);
       Integer ReliableUpdateFreq;
     protected:
 
+      //Action evaluation
       //Allow derived classes to override the multishift CG
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
-#if 0
+#if 1
 	SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOpD : DenOpD);
 	ConjugateGradientMultiShift<FermionFieldD> msCG(MaxIter, approx);
 	msCG(schurOp,in, out);
@@ -70,9 +71,10 @@ NAMESPACE_BEGIN(Grid);
 	msCG(schurOpD, in, out);
 #endif
       }
+      //Force evaluation
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
 	SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
-	SchurDifferentiableOperator<ImplF>  schurOpF (numerator ? NumOpF  : DenOpF);
+	SchurDifferentiableOperator<ImplF>  schurOpF(numerator ? NumOpF  : DenOpF);
 
 	FermionFieldD inD(NumOpD.FermionRedBlackGrid());
 	FermionFieldD outD(NumOpD.FermionRedBlackGrid());
@@ -84,20 +86,15 @@ NAMESPACE_BEGIN(Grid);
       virtual void ImportGauge(const typename ImplD::GaugeField &Ud){
 
 	typename ImplF::GaugeField Uf(NumOpF.GaugeGrid());
-	typename ImplD::GaugeField Ud2(NumOpD.GaugeGrid());
 	precisionChange(Uf, Ud);
-	precisionChange(Ud2, Ud);
 
-	std::cout << "Importing "<<norm2(Ud)<<" "<< norm2(Uf)<<" " << norm2(Ud2)<<std::endl;
+	std::cout << "Importing "<<norm2(Ud)<<" "<< norm2(Uf)<<" " <<std::endl;
 	
 	NumOpD.ImportGauge(Ud);
 	DenOpD.ImportGauge(Ud);
 
 	NumOpF.ImportGauge(Uf);
 	DenOpF.ImportGauge(Uf);
-
-	NumOpD.ImportGauge(Ud2);
-	DenOpD.ImportGauge(Ud2);
       }
       
     public:

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -207,20 +207,27 @@ NAMESPACE_BEGIN(Grid);
         //X = (Mdag M)^-1 V^dag phi
         //Y = (Mdag)^-1 V^dag  phi
         Vpc.MpcDag(PhiOdd,Y);          // Y= Vdag phi
+	std::cout << GridLogMessage <<" Y "<<norm2(Y)<<std::endl;
         X=Zero();
         DerivativeSolver(Mpc,Y,X);     // X= (MdagM)^-1 Vdag phi
+	std::cout << GridLogMessage <<" X "<<norm2(X)<<std::endl;
         Mpc.Mpc(X,Y);                  // Y=  Mdag^-1 Vdag phi
+	std::cout << GridLogMessage <<" Y "<<norm2(Y)<<std::endl;
 
         // phi^dag V (Mdag M)^-1 dV^dag  phi
         Vpc.MpcDagDeriv(force , X, PhiOdd );   dSdU = force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
   
         // phi^dag dV (Mdag M)^-1 V^dag  phi
         Vpc.MpcDeriv(force , PhiOdd, X );      dSdU = dSdU+force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
 
         //    -    phi^dag V (Mdag M)^-1 Mdag dM   (Mdag M)^-1 V^dag  phi
         //    -    phi^dag V (Mdag M)^-1 dMdag M   (Mdag M)^-1 V^dag  phi
         Mpc.MpcDeriv(force,Y,X);              dSdU = dSdU-force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
         Mpc.MpcDagDeriv(force,X,Y);           dSdU = dSdU-force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
 
         // FIXME No force contribution from EvenEven assumed here
         // Needs a fix for clover.

Grid/qcd/hmc/GenericHMCrunner.h

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

Grid/qcd/hmc/HMC.h

@@ -283,12 +283,13 @@ public:
       std::cout << GridLogHMC << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
 
       TheIntegrator.print_timer();
-
+      
+      TheIntegrator.Smearer.set_Field(Ucur);
       for (int obs = 0; obs < Observables.size(); obs++) {
       	std::cout << GridLogDebug << "Observables # " << obs << std::endl;
       	std::cout << GridLogDebug << "Observables total " << Observables.size() << std::endl;
       	std::cout << GridLogDebug << "Observables pointer " << Observables[obs] << std::endl;
-        Observables[obs]->TrajectoryComplete(traj + 1, Ucur, sRNG, pRNG);
+        Observables[obs]->TrajectoryComplete(traj + 1, TheIntegrator.Smearer, sRNG, pRNG);
       }
       std::cout << GridLogHMC << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
     }

Grid/qcd/hmc/checkpointers/BaseCheckpointer.h

@@ -35,13 +35,16 @@ class CheckpointerParameters : Serializable {
 public:
   GRID_SERIALIZABLE_CLASS_MEMBERS(CheckpointerParameters, 
 				  std::string, config_prefix, 
+				  std::string, smeared_prefix, 
 				  std::string, rng_prefix, 
 				  int, saveInterval, 
+				  bool, saveSmeared, 
 				  std::string, format, );
 
-  CheckpointerParameters(std::string cf = "cfg", std::string rn = "rng",
+  CheckpointerParameters(std::string cf = "cfg", std::string sf="cfg_smr" , std::string rn = "rng",
 			 int savemodulo = 1, const std::string &f = "IEEE64BIG")
     : config_prefix(cf),
+      smeared_prefix(sf),
       rng_prefix(rn),
       saveInterval(savemodulo),
       format(f){};
@@ -61,13 +64,21 @@ template <class Impl>
 class BaseHmcCheckpointer : public HmcObservable<typename Impl::Field> {
 public:
   void build_filenames(int traj, CheckpointerParameters &Params,
-                       std::string &conf_file, std::string &rng_file) {
+                       std::string &conf_file,
+                       std::string &smear_file,
+		       std::string &rng_file) {
     {
       std::ostringstream os;
       os << Params.rng_prefix << "." << traj;
       rng_file = os.str();
     }
 
+    {
+      std::ostringstream os;
+      os << Params.smeared_prefix << "." << traj;
+      smear_file = os.str();
+    }
+
     {
       std::ostringstream os;
       os << Params.config_prefix << "." << traj;
@@ -84,6 +95,11 @@ public:
   }
   virtual void initialize(const CheckpointerParameters &Params) = 0;
 
+  virtual void TrajectoryComplete(int traj,
+                                  typename Impl::Field &U,
+                                  GridSerialRNG &sRNG,
+                                  GridParallelRNG &pRNG) { assert(0); } ; // HMC should pass the smart config with smeared and unsmeared
+  
   virtual void CheckpointRestore(int traj, typename Impl::Field &U,
                                  GridSerialRNG &sRNG,
                                  GridParallelRNG &pRNG) = 0;

Grid/qcd/hmc/checkpointers/BinaryCheckpointer.h

@@ -61,11 +61,14 @@ public:
     fout.close();
   }
 
-  void TrajectoryComplete(int traj, Field &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
+  void TrajectoryComplete(int traj,
+			  ConfigurationBase<Field> &SmartConfig,
+			  GridSerialRNG &sRNG, GridParallelRNG &pRNG)
+  {
 
     if ((traj % Params.saveInterval) == 0) {
-      std::string config, rng;
-      this->build_filenames(traj, Params, config, rng);
+      std::string config, rng, smr;
+      this->build_filenames(traj, Params, config, smr, rng);
 
       uint32_t nersc_csum;
       uint32_t scidac_csuma;
@@ -74,9 +77,15 @@ public:
       BinarySimpleUnmunger<sobj_double, sobj> munge;
       truncate(rng);
       BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
-      truncate(config);
+      std::cout << GridLogMessage << "Written Binary RNG " << rng
+                << " checksum " << std::hex 
+		<< nersc_csum   <<"/"
+		<< scidac_csuma   <<"/"
+		<< scidac_csumb 
+		<< std::dec << std::endl;
 
-      BinaryIO::writeLatticeObject<vobj, sobj_double>(U, config, munge, 0, Params.format,
+      truncate(config);
+      BinaryIO::writeLatticeObject<vobj, sobj_double>(SmartConfig.get_U(false), config, munge, 0, Params.format,
 						      nersc_csum,scidac_csuma,scidac_csumb);
 
       std::cout << GridLogMessage << "Written Binary Configuration " << config
@@ -85,6 +94,18 @@ public:
 		<< scidac_csuma   <<"/"
 		<< scidac_csumb 
 		<< std::dec << std::endl;
+
+      if ( Params.saveSmeared ) {
+	truncate(smr);
+	BinaryIO::writeLatticeObject<vobj, sobj_double>(SmartConfig.get_U(true), smr, munge, 0, Params.format,
+							nersc_csum,scidac_csuma,scidac_csumb);
+	std::cout << GridLogMessage << "Written Binary Smeared Configuration " << smr
+                << " checksum " << std::hex 
+		<< nersc_csum   <<"/"
+		<< scidac_csuma   <<"/"
+		<< scidac_csumb 
+		<< std::dec << std::endl;
+      }
     }
 
   };

Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h

@@ -69,17 +69,27 @@ public:
     }
   }
 
-  void TrajectoryComplete(int traj, GaugeField &U, GridSerialRNG &sRNG,
+  void TrajectoryComplete(int traj,
+			  ConfigurationBase<GaugeField> &SmartConfig,
+			  GridSerialRNG &sRNG,
                           GridParallelRNG &pRNG) {
     if ((traj % Params.saveInterval) == 0) {
-      std::string config, rng;
+      std::string config, rng, smr;
       this->build_filenames(traj, Params, config, rng);
-      GridBase *grid = U.Grid();
+      GridBase *grid = SmartConfig.get_U(false).Grid();
       uint32_t nersc_csum,scidac_csuma,scidac_csumb;
       BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
+      std::cout << GridLogMessage << "Written BINARY RNG " << rng
+                << " checksum " << std::hex 
+		<< nersc_csum<<"/"
+		<< scidac_csuma<<"/"
+		<< scidac_csumb
+		<< std::dec << std::endl;
+
+      
       IldgWriter _IldgWriter(grid->IsBoss());
       _IldgWriter.open(config);
-      _IldgWriter.writeConfiguration<GaugeStats>(U, traj, config, config);
+      _IldgWriter.writeConfiguration<GaugeStats>(SmartConfig.get_U(false), traj, config, config);
       _IldgWriter.close();
 
       std::cout << GridLogMessage << "Written ILDG Configuration on " << config
@@ -88,6 +98,21 @@ public:
 		<< scidac_csuma<<"/"
 		<< scidac_csumb
 		<< std::dec << std::endl;
+
+      if ( Params.saveSmeared ) { 
+	IldgWriter _IldgWriter(grid->IsBoss());
+	_IldgWriter.open(smr);
+	_IldgWriter.writeConfiguration<GaugeStats>(SmartConfig.get_U(true), traj, config, config);
+	_IldgWriter.close();
+
+	std::cout << GridLogMessage << "Written ILDG Configuration on " << smr
+                << " checksum " << std::hex 
+		<< nersc_csum<<"/"
+		<< scidac_csuma<<"/"
+		<< scidac_csumb
+		<< std::dec << std::endl;
+      }
+
     }
   };
 

Grid/qcd/hmc/checkpointers/NerscCheckpointer.h

@@ -52,23 +52,29 @@ public:
     Params.format = "IEEE64BIG";  // fixed, overwrite any other choice
   }
 
-  void TrajectoryComplete(int traj, GaugeField &U, GridSerialRNG &sRNG,
-                          GridParallelRNG &pRNG) {
+  virtual void TrajectoryComplete(int traj,
+                                  ConfigurationBase<GaugeField> &SmartConfig,
+                                  GridSerialRNG &sRNG,
+                                  GridParallelRNG &pRNG)
+  {
     if ((traj % Params.saveInterval) == 0) {
-      std::string config, rng;
-      this->build_filenames(traj, Params, config, rng);
-
+      std::string config, rng, smr;
+      this->build_filenames(traj, Params, config, smr, rng);
+      
       int precision32 = 1;
       int tworow = 0;
       NerscIO::writeRNGState(sRNG, pRNG, rng);
-      NerscIO::writeConfiguration<GaugeStats>(U, config, tworow, precision32);
+      NerscIO::writeConfiguration<GaugeStats>(SmartConfig.get_U(false), config, tworow, precision32);
+      if ( Params.saveSmeared ) {
+	NerscIO::writeConfiguration<GaugeStats>(SmartConfig.get_U(true), smr, tworow, precision32);
+      }
     }
   };
 
   void CheckpointRestore(int traj, GaugeField &U, GridSerialRNG &sRNG,
                          GridParallelRNG &pRNG) {
-    std::string config, rng;
-    this->build_filenames(traj, Params, config, rng);
+    std::string config, rng, smr;
+    this->build_filenames(traj, Params, config, smr, rng );
     this->check_filename(rng);
     this->check_filename(config);
 

Grid/qcd/hmc/checkpointers/ScidacCheckpointer.h

@@ -70,19 +70,37 @@ class ScidacHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
     }
   }
 
-  void TrajectoryComplete(int traj, Field &U, GridSerialRNG &sRNG,
+  void TrajectoryComplete(int traj, 
+			  ConfigurationBase<Field> &SmartConfig,
+			  GridSerialRNG &sRNG,
                           GridParallelRNG &pRNG) {
     if ((traj % Params.saveInterval) == 0) {
-      std::string config, rng;
-      this->build_filenames(traj, Params, config, rng);
-      GridBase *grid = U.Grid();
+      std::string config, rng,smr;
+      this->build_filenames(traj, Params, config, smr, rng);
+      GridBase *grid = SmartConfig.get_U(false).Grid();
       uint32_t nersc_csum,scidac_csuma,scidac_csumb;
       BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
-      ScidacWriter _ScidacWriter(grid->IsBoss());
-      _ScidacWriter.open(config);
-      _ScidacWriter.writeScidacFieldRecord(U, MData);
-      _ScidacWriter.close();
+      std::cout << GridLogMessage << "Written Binary RNG " << rng
+                << " checksum " << std::hex 
+		<< nersc_csum   <<"/"
+		<< scidac_csuma   <<"/"
+		<< scidac_csumb 
+		<< std::dec << std::endl;
 
+
+      {
+	ScidacWriter _ScidacWriter(grid->IsBoss());
+	_ScidacWriter.open(config);
+	_ScidacWriter.writeScidacFieldRecord(SmartConfig.get_U(false), MData);
+	_ScidacWriter.close();
+      }
+      
+      if ( Params.saveSmeared ) {
+	ScidacWriter _ScidacWriter(grid->IsBoss());
+	_ScidacWriter.open(smr);
+	_ScidacWriter.writeScidacFieldRecord(SmartConfig.get_U(true), MData);
+	_ScidacWriter.close();
+      }
       std::cout << GridLogMessage << "Written Scidac Configuration on " << config << std::endl;
     }
   };

Grid/qcd/hmc/integrators/Integrator.h

@@ -66,6 +66,7 @@ public:
 template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
+public:
   typedef FieldImplementation_ FieldImplementation;
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
@@ -96,7 +97,6 @@ protected:
   {
     t_P[level] += ep;
     update_P(P, U, level, ep);
-
     std::cout << GridLogIntegrator << "[" << level << "] P " << " dt " << ep << " : t_P " << t_P[level] << std::endl;
   }
 
@@ -130,28 +130,20 @@ protected:
       Field force(U.Grid());
       conformable(U.Grid(), Mom.Grid());
 
-      Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
 
-      std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] before"<<std::endl;
-      
       as[level].actions.at(a)->deriv_timer_start();
-      as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
+      as[level].actions.at(a)->deriv(Smearer, force);  // deriv should NOT include Ta
       as[level].actions.at(a)->deriv_timer_stop();
 
-      std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] after"<<std::endl;
-
-      std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
       auto name = as[level].actions.at(a)->action_name();
-      if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
 
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
-
-      //      DumpSliceNorm("force ",force,Nd-1);
+      
       MomFilter->applyFilter(force);
+
       std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<<" dt "<<ep<<  std::endl;
-      DumpSliceNorm("force filtered ",force,Nd-1);
       
       Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
       Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
@@ -377,14 +369,9 @@ public:
 	auto name = as[level].actions.at(actionID)->action_name();
         std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
 
-        Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
-
-	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] before"<<std::endl;
-
 	as[level].actions.at(actionID)->refresh_timer_start();
-        as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
+        as[level].actions.at(actionID)->refresh(Smearer, sRNG, pRNG);
 	as[level].actions.at(actionID)->refresh_timer_stop();
-	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] after"<<std::endl;
 
       }
 
@@ -425,10 +412,9 @@ public:
 
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
-        Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
 	        as[level].actions.at(actionID)->S_timer_start();
-        Hterm = as[level].actions.at(actionID)->S(Us);
+        Hterm = as[level].actions.at(actionID)->S(Smearer);
    	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
@@ -469,12 +455,11 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
-        Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
-	        as[level].actions.at(actionID)->S_timer_start();
 
-        Hterm = as[level].actions.at(actionID)->Sinitial(Us);
-   	        as[level].actions.at(actionID)->S_timer_stop();
+	as[level].actions.at(actionID)->S_timer_start();
+        Hterm = as[level].actions.at(actionID)->S(Smearer);
+	as[level].actions.at(actionID)->S_timer_stop();
 
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;

Grid/qcd/observables/hmc_observable.h

@@ -34,6 +34,13 @@ NAMESPACE_BEGIN(Grid);
 template <class Field>
 class HmcObservable {
  public:
+  virtual void TrajectoryComplete(int traj,
+                                  ConfigurationBase<Field> &SmartConfig,
+                                  GridSerialRNG &sRNG,
+                                  GridParallelRNG &pRNG)
+  {
+    TrajectoryComplete(traj,SmartConfig.get_U(false),sRNG,pRNG); // Unsmeared observable
+  };
   virtual void TrajectoryComplete(int traj,
                                   Field &U,
                                   GridSerialRNG &sRNG,

Grid/qcd/observables/plaquette.h

@@ -42,6 +42,18 @@ public:
   // necessary for HmcObservable compatibility
   typedef typename Impl::Field Field;
 
+  virtual void TrajectoryComplete(int traj,
+                                  ConfigurationBase<Field> &SmartConfig,
+                                  GridSerialRNG &sRNG,
+                                  GridParallelRNG &pRNG)
+  {
+    std::cout << GridLogMessage << "+++++++++++++++++++"<<std::endl;
+    std::cout << GridLogMessage << "Unsmeared plaquette"<<std::endl;
+    TrajectoryComplete(traj,SmartConfig.get_U(false),sRNG,pRNG); // Unsmeared observable
+    std::cout << GridLogMessage << "Smeared plaquette"<<std::endl;
+    TrajectoryComplete(traj,SmartConfig.get_U(true),sRNG,pRNG); // Unsmeared observable
+    std::cout << GridLogMessage << "+++++++++++++++++++"<<std::endl;
+  };
   void TrajectoryComplete(int traj,
                           Field &U,
                           GridSerialRNG &sRNG,

Grid/qcd/representations/fundamental.h

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

Grid/qcd/representations/two_index.h

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

Grid/qcd/smearing/GaugeConfiguration.h

@@ -7,26 +7,27 @@
 
 NAMESPACE_BEGIN(Grid);
 
+
 //trivial class for no smearing
 template< class Impl >
-class NoSmearing
+class NoSmearing : public ConfigurationBase<typename Impl::Field>
 {
 public:
   INHERIT_FIELD_TYPES(Impl);
 
-  Field* ThinField;
+  Field* ThinLinks;
 
-  NoSmearing(): ThinField(NULL) {}
+  NoSmearing(): ThinLinks(NULL) {}
 
-  void set_Field(Field& U) { ThinField = &U; }
+  virtual void set_Field(Field& U) { ThinLinks = &U; }
 
-  void smeared_force(Field&) const {}
+  virtual void smeared_force(Field&) {}
 
-  Field& get_SmearedU() { return *ThinField; }
+  virtual Field& get_SmearedU() { return *ThinLinks; }
 
-  Field &get_U(bool smeared = false)
+  virtual Field &get_U(bool smeared = false)
   {
-    return *ThinField;
+    return *ThinLinks;
   }
 };
 
@@ -42,19 +43,24 @@ public:
   It stores a list of smeared configurations.
 */
 template <class Gimpl>
-class SmearedConfiguration
+class SmearedConfiguration : public ConfigurationBase<typename Gimpl::Field>
 {
 public:
   INHERIT_GIMPL_TYPES(Gimpl);
 
-private:
+protected:
   const unsigned int smearingLevels;
   Smear_Stout<Gimpl> *StoutSmearing;
   std::vector<GaugeField> SmearedSet;
-
+public:
+  GaugeField*  ThinLinks; /* Pointer to the thin links configuration */ // move to base???
+protected:
+  
   // Member functions
   //====================================================================
-  void fill_smearedSet(GaugeField &U)
+
+  // Overridden in masked version
+  virtual void fill_smearedSet(GaugeField &U)
   {
     ThinLinks = &U;  // attach the smearing routine to the field U
 
@@ -82,9 +88,10 @@ private:
       }
     }
   }
-  //====================================================================
-  GaugeField AnalyticSmearedForce(const GaugeField& SigmaKPrime,
-                                  const GaugeField& GaugeK) const 
+
+  //overridden in masked verson
+  virtual GaugeField AnalyticSmearedForce(const GaugeField& SigmaKPrime,
+					  const GaugeField& GaugeK) const 
   {
     GridBase* grid = GaugeK.Grid();
     GaugeField C(grid), SigmaK(grid), iLambda(grid);
@@ -213,8 +220,6 @@ private:
 
   //====================================================================
 public:
-  GaugeField*
-      ThinLinks; /* Pointer to the thin links configuration */
 
   /* Standard constructor */
   SmearedConfiguration(GridCartesian* UGrid, unsigned int Nsmear,
@@ -230,7 +235,7 @@ public:
     : smearingLevels(0), StoutSmearing(nullptr), SmearedSet(), ThinLinks(NULL) {}
 
   // attach the smeared routines to the thin links U and fill the smeared set
-  void set_Field(GaugeField &U)
+  virtual void set_Field(GaugeField &U)
   {
     double start = usecond();
     fill_smearedSet(U);
@@ -240,7 +245,7 @@ public:
   }
 
   //====================================================================
-  void smeared_force(GaugeField &SigmaTilde) const
+  virtual void smeared_force(GaugeField &SigmaTilde) 
   {
     if (smearingLevels > 0)
     {
@@ -267,14 +272,16 @@ public:
       }
       double end = usecond();
       double time = (end - start)/ 1e3;
-      std::cout << GridLogMessage << "Smearing force in " << time << " ms" << std::endl;  
+      std::cout << GridLogMessage << " GaugeConfiguration: Smeared Force chain rule took " << time << " ms" << std::endl;
     }  // if smearingLevels = 0 do nothing
+    SigmaTilde=Gimpl::projectForce(SigmaTilde); // Ta
+      
   }
   //====================================================================
 
-  GaugeField& get_SmearedU() { return SmearedSet[smearingLevels - 1]; }
+  virtual GaugeField& get_SmearedU() { return SmearedSet[smearingLevels - 1]; }
 
-  GaugeField &get_U(bool smeared = false)
+  virtual GaugeField &get_U(bool smeared = false)
   {
     // get the config, thin links by default
     if (smeared)

Grid/qcd/smearing/GaugeConfigurationMasked.h

@@ -0,0 +1,813 @@
+/*!
+  @file GaugeConfiguration.h
+  @brief Declares the GaugeConfiguration class
+*/
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+/*!
+  @brief Smeared configuration masked container
+  Modified for a multi-subset smearing (aka Luscher Flowed HMC)
+*/
+template <class Gimpl>
+class SmearedConfigurationMasked : public SmearedConfiguration<Gimpl>
+{
+public:
+  INHERIT_GIMPL_TYPES(Gimpl);
+
+private:
+  // These live in base class
+  //  const unsigned int smearingLevels;
+  //  Smear_Stout<Gimpl> *StoutSmearing;
+  //  std::vector<GaugeField> SmearedSet;
+  
+  std::vector<LatticeLorentzComplex> masks;
+
+  typedef typename SU3Adjoint::AMatrix AdjMatrix;
+  typedef typename SU3Adjoint::LatticeAdjMatrix  AdjMatrixField;
+  typedef typename SU3Adjoint::LatticeAdjVector  AdjVectorField;
+
+  // Adjoint vector to GaugeField force
+  void InsertForce(GaugeField &Fdet,AdjVectorField &Fdet_nu,int nu)
+  {
+    Complex ci(0,1);
+    GaugeLinkField Fdet_pol(Fdet.Grid());
+    Fdet_pol=Zero();
+    for(int e=0;e<8;e++){
+      ColourMatrix te;
+      SU3::generator(e, te);
+      auto tmp=peekColour(Fdet_nu,e);
+      Fdet_pol=Fdet_pol + ci*tmp*te; // but norm of te is different.. why?
+    }
+    pokeLorentz(Fdet, Fdet_pol, nu);
+  }
+  void Compute_MpInvJx_dNxxdSy(const GaugeLinkField &PlaqL,const GaugeLinkField &PlaqR, AdjMatrixField MpInvJx,AdjVectorField &Fdet2 )
+  {
+    GaugeLinkField UtaU(PlaqL.Grid());
+    GaugeLinkField D(PlaqL.Grid());
+    AdjMatrixField Dbc(PlaqL.Grid());
+    LatticeComplex tmp(PlaqL.Grid());
+    const int Ngen = SU3Adjoint::Dimension;
+    Complex ci(0,1);
+    ColourMatrix   ta,tb,tc;
+    
+    for(int a=0;a<Ngen;a++) {
+      SU3::generator(a, ta);
+      // Qlat Tb = 2i Tb^Grid
+      UtaU= 2.0*ci*adj(PlaqL)*ta*PlaqR;
+      for(int c=0;c<Ngen;c++) {
+	SU3::generator(c, tc);
+	D = Ta( (2.0)*ci*tc *UtaU);
+	for(int b=0;b<Ngen;b++){
+	  SU3::generator(b, tb);
+	  tmp =-trace(ci*tb*D); 
+	  PokeIndex<ColourIndex>(Dbc,tmp,b,c);  // Adjoint rep
+	}
+      }
+      tmp = trace(MpInvJx * Dbc);
+      PokeIndex<ColourIndex>(Fdet2,tmp,a);
+    }
+  }
+  
+  void ComputeNxy(const GaugeLinkField &PlaqL,const GaugeLinkField &PlaqR,AdjMatrixField &NxAd)
+  {
+    GaugeLinkField Nx(PlaqL.Grid());
+    const int Ngen = SU3Adjoint::Dimension;
+    Complex ci(0,1);
+    ColourMatrix   tb;
+    ColourMatrix   tc;
+    for(int b=0;b<Ngen;b++) {
+      SU3::generator(b, tb);
+      Nx = (2.0)*Ta( adj(PlaqL)*ci*tb * PlaqR );
+      for(int c=0;c<Ngen;c++) {
+	SU3::generator(c, tc);
+	auto tmp =closure( -trace(ci*tc*Nx)); 
+	PokeIndex<ColourIndex>(NxAd,tmp,c,b); 
+      }
+    }
+  }
+  void ApplyMask(GaugeField &U,int smr)
+  {
+    LatticeComplex tmp(U.Grid());
+    GaugeLinkField Umu(U.Grid());
+    for(int mu=0;mu<Nd;mu++){
+      Umu=PeekIndex<LorentzIndex>(U,mu);
+      tmp=PeekIndex<LorentzIndex>(masks[smr],mu);
+      Umu=Umu*tmp;
+      PokeIndex<LorentzIndex>(U, Umu, mu);
+    }
+  }
+public:
+
+  void logDetJacobianForceLevel(const GaugeField &U, GaugeField &force ,int smr)
+  {
+    GridBase* grid = U.Grid();
+    ColourMatrix   tb;
+    ColourMatrix   tc;
+    ColourMatrix   ta;
+    GaugeField C(grid);
+    GaugeField Umsk(grid);
+    std::vector<GaugeLinkField> Umu(Nd,grid);
+    GaugeLinkField Cmu(grid); // U and staple; C contains factor of epsilon
+    GaugeLinkField Zx(grid);  // U times Staple, contains factor of epsilon
+    GaugeLinkField Nxx(grid);  // Nxx fundamental space
+    GaugeLinkField Utmp(grid);
+    GaugeLinkField PlaqL(grid);
+    GaugeLinkField PlaqR(grid);
+    const int Ngen = SU3Adjoint::Dimension;
+    AdjMatrix TRb;
+    ColourMatrix Ident;
+    LatticeComplex  cplx(grid);
+    
+    AdjVectorField  dJdXe_nMpInv(grid); 
+    AdjVectorField  dJdXe_nMpInv_y(grid); 
+    AdjMatrixField  MpAd(grid);    // Mprime luchang's notes
+    AdjMatrixField  MpAdInv(grid); // Mprime inverse
+    AdjMatrixField  NxxAd(grid);    // Nxx in adjoint space
+    AdjMatrixField  JxAd(grid);     
+    AdjMatrixField  ZxAd(grid);
+    AdjMatrixField  mZxAd(grid);
+    AdjMatrixField  X(grid);
+    Complex ci(0,1);
+
+    RealD t0 = usecond();
+    Ident = ComplexD(1.0);
+    for(int d=0;d<Nd;d++){
+      Umu[d] = peekLorentz(U, d);
+    }
+    int mu= (smr/2) %Nd;
+
+    ////////////////////////////////////////////////////////////////////////////////
+    // Mask the gauge field
+    ////////////////////////////////////////////////////////////////////////////////
+    auto mask=PeekIndex<LorentzIndex>(masks[smr],mu); // the cb mask
+
+    Umsk = U;
+    ApplyMask(Umsk,smr);
+    Utmp = peekLorentz(Umsk,mu);
+
+    ////////////////////////////////////////////////////////////////////////////////
+    // Retrieve the eps/rho parameter(s) -- could allow all different but not so far
+    ////////////////////////////////////////////////////////////////////////////////
+    double rho=this->StoutSmearing->SmearRho[1];
+    int idx=0;
+    for(int mu=0;mu<4;mu++){
+    for(int nu=0;nu<4;nu++){
+      if ( mu!=nu) assert(this->StoutSmearing->SmearRho[idx]==rho);
+      else         assert(this->StoutSmearing->SmearRho[idx]==0.0);
+      idx++;
+    }}
+    //////////////////////////////////////////////////////////////////
+    // Assemble the N matrix
+    //////////////////////////////////////////////////////////////////
+    // Computes ALL the staples -- could compute one only and do it here
+    RealD time;
+    time=-usecond();
+    this->StoutSmearing->BaseSmear(C, U);
+    Cmu = peekLorentz(C, mu);
+
+    //////////////////////////////////////////////////////////////////
+    // Assemble Luscher exp diff map J matrix 
+    //////////////////////////////////////////////////////////////////
+    // Ta so Z lives in Lie algabra
+    Zx  = Ta(Cmu * adj(Umu[mu]));
+    time+=usecond();
+    std::cout << GridLogMessage << "Z took "<<time<< " us"<<std::endl;
+
+    time=-usecond();
+    // Move Z to the Adjoint Rep == make_adjoint_representation
+    ZxAd = Zero();
+    for(int b=0;b<8;b++) {
+      // Adj group sets traceless antihermitian T's -- Guido, really????
+      SU3::generator(b, tb);         // Fund group sets traceless hermitian T's
+      SU3Adjoint::generator(b,TRb);
+      TRb=-TRb;
+      cplx = 2.0*trace(ci*tb*Zx); // my convention 1/2 delta ba
+      ZxAd = ZxAd + cplx * TRb; // is this right? YES - Guido used Anti herm Ta's and with bloody wrong sign.
+    }
+    time+=usecond();
+    std::cout << GridLogMessage << "ZxAd took "<<time<< " us"<<std::endl;
+
+    //////////////////////////////////////
+    // J(x) = 1 + Sum_k=1..N (-Zac)^k/(k+1)!
+    //////////////////////////////////////
+    time=-usecond();
+    X=1.0; 
+    JxAd = X;
+    mZxAd = (-1.0)*ZxAd; 
+    RealD kpfac = 1;
+    for(int k=1;k<12;k++){
+      X=X*mZxAd;
+      kpfac = kpfac /(k+1);
+      JxAd = JxAd + X * kpfac;
+    }
+    time+=usecond();
+    std::cout << GridLogMessage << "Jx took "<<time<< " us"<<std::endl;
+
+    //////////////////////////////////////
+    // dJ(x)/dxe
+    //////////////////////////////////////
+    time=-usecond();
+    std::vector<AdjMatrixField>  dJdX;    dJdX.resize(8,grid);
+    AdjMatrixField tbXn(grid);
+    AdjMatrixField sumXtbX(grid);
+    AdjMatrixField t2(grid);
+    AdjMatrixField dt2(grid);
+    AdjMatrixField t3(grid);
+    AdjMatrixField dt3(grid);
+    AdjMatrixField aunit(grid);
+    for(int b=0;b<8;b++){
+      aunit = ComplexD(1.0);
+      SU3Adjoint::generator(b, TRb); //dt2
+
+      X  = (-1.0)*ZxAd; 
+      t2 = X;
+      dt2 = TRb;
+      for (int j = 20; j > 1; --j) {
+	t3 = t2*(1.0 / (j + 1))  + aunit;
+	dt3 = dt2*(1.0 / (j + 1));
+	t2 = X * t3;
+	dt2 = TRb * t3 + X * dt3;
+      }
+      dJdX[b] = -dt2; 
+    }
+    time+=usecond();
+    std::cout << GridLogMessage << "dJx took "<<time<< " us"<<std::endl;
+    /////////////////////////////////////////////////////////////////
+    // Mask Umu for this link
+    /////////////////////////////////////////////////////////////////
+    time=-usecond();
+    PlaqL = Ident;
+    PlaqR = Utmp*adj(Cmu);
+    ComputeNxy(PlaqL,PlaqR,NxxAd);
+    time+=usecond();
+    std::cout << GridLogMessage << "ComputeNxy took "<<time<< " us"<<std::endl;
+    
+    ////////////////////////////
+    // Mab
+    ////////////////////////////
+    MpAd = Complex(1.0,0.0);
+    MpAd = MpAd - JxAd * NxxAd;
+
+    /////////////////////////
+    // invert the 8x8
+    /////////////////////////
+    time=-usecond();
+    MpAdInv = Inverse(MpAd);
+    time+=usecond();
+    std::cout << GridLogMessage << "MpAdInv took "<<time<< " us"<<std::endl;
+    
+    RealD t3a = usecond();
+    /////////////////////////////////////////////////////////////////
+    // Nxx Mp^-1
+    /////////////////////////////////////////////////////////////////
+    AdjVectorField  FdetV(grid);
+    AdjVectorField  Fdet1_nu(grid);
+    AdjVectorField  Fdet2_nu(grid);
+    AdjVectorField  Fdet2_mu(grid);
+    AdjVectorField  Fdet1_mu(grid);
+
+    AdjMatrixField nMpInv(grid);
+    nMpInv= NxxAd *MpAdInv;
+
+    AdjMatrixField MpInvJx(grid);
+    AdjMatrixField MpInvJx_nu(grid);
+    MpInvJx = (-1.0)*MpAdInv * JxAd;// rho is on the plaq factor
+
+    Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx,FdetV);
+    Fdet2_mu=FdetV;
+    Fdet1_mu=Zero();
+    
+    for(int e =0 ; e<8 ; e++){
+      LatticeComplexD tr(grid);
+      ColourMatrix te;
+      SU3::generator(e, te);
+      tr = trace(dJdX[e] * nMpInv);
+      pokeColour(dJdXe_nMpInv,tr,e);
+    }
+    ///////////////////////////////
+    // Mask it off
+    ///////////////////////////////
+    auto tmp=PeekIndex<LorentzIndex>(masks[smr],mu);
+    dJdXe_nMpInv = dJdXe_nMpInv*tmp;
+    
+    //    dJdXe_nMpInv needs to multiply:
+    //       Nxx_mu (site local)                           (1)
+    //       Nxy_mu one site forward  in each nu direction (3)
+    //       Nxy_mu one site backward in each nu direction (3)
+    //       Nxy_nu 0,0  ; +mu,0; 0,-nu; +mu-nu   [ 3x4 = 12]
+    // 19 terms.
+    AdjMatrixField Nxy(grid);
+
+    GaugeField Fdet1(grid);
+    GaugeField Fdet2(grid);
+    GaugeLinkField Fdet_pol(grid); // one polarisation
+
+    RealD t4 = usecond();
+    for(int nu=0;nu<Nd;nu++){
+
+      if (nu!=mu) {
+	///////////////// +ve nu /////////////////
+	//     __
+	//    |  |
+	//    x==    // nu polarisation -- clockwise
+
+	time=-usecond();
+	PlaqL=Ident;
+
+	PlaqR=(-rho)*Gimpl::CovShiftForward(Umu[nu], nu,
+ 	       Gimpl::CovShiftForward(Umu[mu], mu,
+	         Gimpl::CovShiftBackward(Umu[nu], nu,
+		   Gimpl::CovShiftIdentityBackward(Utmp, mu))));
+	time+=usecond();
+	std::cout << GridLogMessage << "PlaqLR took "<<time<< " us"<<std::endl;
+
+	time=-usecond();
+	dJdXe_nMpInv_y =   dJdXe_nMpInv;
+	ComputeNxy(PlaqL,PlaqR,Nxy);
+	Fdet1_nu = transpose(Nxy)*dJdXe_nMpInv_y;
+	time+=usecond();
+	std::cout << GridLogMessage << "ComputeNxy (occurs 6x) took "<<time<< " us"<<std::endl;
+
+	time=-usecond();
+	PlaqR=(-1.0)*PlaqR;
+	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx,FdetV);
+	Fdet2_nu = FdetV;
+	time+=usecond();
+	std::cout << GridLogMessage << "Compute_MpInvJx_dNxxSy (occurs 6x) took "<<time<< " us"<<std::endl;
+	
+	//    x==
+	//    |  |
+	//    .__|    // nu polarisation -- anticlockwise
+
+	PlaqR=(rho)*Gimpl::CovShiftForward(Umu[nu], nu,
+		      Gimpl::CovShiftBackward(Umu[mu], mu,
+    	 	        Gimpl::CovShiftIdentityBackward(Umu[nu], nu)));
+
+	PlaqL=Gimpl::CovShiftIdentityBackward(Utmp, mu);
+
+	dJdXe_nMpInv_y = Cshift(dJdXe_nMpInv,mu,-1);
+	ComputeNxy(PlaqL, PlaqR,Nxy);
+	Fdet1_nu = Fdet1_nu+transpose(Nxy)*dJdXe_nMpInv_y;
+	
+
+	MpInvJx_nu = Cshift(MpInvJx,mu,-1);
+	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Fdet2_nu = Fdet2_nu+FdetV;
+	
+	///////////////// -ve nu /////////////////
+	//  __
+	// |  |
+	// x==          // nu polarisation -- clockwise
+
+	PlaqL=(rho)* Gimpl::CovShiftForward(Umu[mu], mu,
+		       Gimpl::CovShiftForward(Umu[nu], nu,
+			 Gimpl::CovShiftIdentityBackward(Utmp, mu)));
+
+        PlaqR = Gimpl::CovShiftIdentityForward(Umu[nu], nu);
+
+	dJdXe_nMpInv_y = Cshift(dJdXe_nMpInv,nu,1);
+	ComputeNxy(PlaqL,PlaqR,Nxy);
+	Fdet1_nu = Fdet1_nu + transpose(Nxy)*dJdXe_nMpInv_y;
+
+	MpInvJx_nu = Cshift(MpInvJx,nu,1);
+	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Fdet2_nu = Fdet2_nu+FdetV;
+	
+	// x==
+	// |  |
+	// |__|         // nu polarisation
+
+	PlaqL=(-rho)*Gimpl::CovShiftForward(Umu[nu], nu,
+ 	        Gimpl::CovShiftIdentityBackward(Utmp, mu));
+
+	PlaqR=Gimpl::CovShiftBackward(Umu[mu], mu,
+	        Gimpl::CovShiftIdentityForward(Umu[nu], nu));
+
+	dJdXe_nMpInv_y = Cshift(dJdXe_nMpInv,mu,-1);
+	dJdXe_nMpInv_y = Cshift(dJdXe_nMpInv_y,nu,1);
+
+	ComputeNxy(PlaqL,PlaqR,Nxy);
+	Fdet1_nu = Fdet1_nu + transpose(Nxy)*dJdXe_nMpInv_y;
+
+	MpInvJx_nu = Cshift(MpInvJx,mu,-1);
+	MpInvJx_nu = Cshift(MpInvJx_nu,nu,1);
+	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Fdet2_nu = Fdet2_nu+FdetV;
+
+	/////////////////////////////////////////////////////////////////////
+	// Set up the determinant force contribution in 3x3 algebra basis
+	/////////////////////////////////////////////////////////////////////
+	InsertForce(Fdet1,Fdet1_nu,nu);
+	InsertForce(Fdet2,Fdet2_nu,nu);
+	
+	//////////////////////////////////////////////////
+	// Parallel direction terms
+	//////////////////////////////////////////////////
+
+        //     __
+	//    |  "
+	//    |__"x    // mu polarisation
+	PlaqL=(-rho)*Gimpl::CovShiftForward(Umu[mu], mu,
+		      Gimpl::CovShiftBackward(Umu[nu], nu,
+   		        Gimpl::CovShiftIdentityBackward(Utmp, mu)));
+
+	PlaqR=Gimpl::CovShiftIdentityBackward(Umu[nu], nu);
+	
+	dJdXe_nMpInv_y = Cshift(dJdXe_nMpInv,nu,-1);
+
+	ComputeNxy(PlaqL,PlaqR,Nxy);
+	Fdet1_mu = Fdet1_mu + transpose(Nxy)*dJdXe_nMpInv_y;
+
+	MpInvJx_nu = Cshift(MpInvJx,nu,-1);
+
+	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Fdet2_mu = Fdet2_mu+FdetV;
+
+	//  __
+	// "  |
+	// x__|          // mu polarisation
+
+	PlaqL=(-rho)*Gimpl::CovShiftForward(Umu[mu], mu,
+		       Gimpl::CovShiftForward(Umu[nu], nu,
+		 	 Gimpl::CovShiftIdentityBackward(Utmp, mu)));
+
+        PlaqR=Gimpl::CovShiftIdentityForward(Umu[nu], nu);
+
+	dJdXe_nMpInv_y = Cshift(dJdXe_nMpInv,nu,1);
+
+	ComputeNxy(PlaqL,PlaqR,Nxy);
+	Fdet1_mu = Fdet1_mu + transpose(Nxy)*dJdXe_nMpInv_y;
+
+	MpInvJx_nu = Cshift(MpInvJx,nu,1);
+
+	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Fdet2_mu = Fdet2_mu+FdetV;
+	
+      }
+    }
+    RealD t5 = usecond();
+
+    Fdet1_mu = Fdet1_mu + transpose(NxxAd)*dJdXe_nMpInv;
+
+    InsertForce(Fdet1,Fdet1_mu,mu);
+    InsertForce(Fdet2,Fdet2_mu,mu);
+
+    force= (-0.5)*( Fdet1 + Fdet2);
+    RealD t1 = usecond();
+    std::cout << GridLogMessage << " logDetJacobianForce level took "<<t1-t0<<" us "<<std::endl;
+    std::cout << GridLogMessage << " logDetJacobianForce t3-t0 "<<t3a-t0<<" us "<<std::endl;
+    std::cout << GridLogMessage << " logDetJacobianForce t4-t3 dJdXe_nMpInv "<<t4-t3a<<" us "<<std::endl;
+    std::cout << GridLogMessage << " logDetJacobianForce t5-t4 mu nu loop "<<t5-t4<<" us "<<std::endl;
+    std::cout << GridLogMessage << " logDetJacobianForce t1-t5 "<<t1-t5<<" us "<<std::endl;
+    std::cout << GridLogMessage << " logDetJacobianForce level took "<<t1-t0<<" us "<<std::endl;
+  }
+  RealD logDetJacobianLevel(const GaugeField &U,int smr)
+  {
+    GridBase* grid = U.Grid();
+    GaugeField C(grid);
+    GaugeLinkField Nb(grid);
+    GaugeLinkField Z(grid);
+    GaugeLinkField Umu(grid), Cmu(grid);
+    ColourMatrix   Tb;
+    ColourMatrix   Tc;
+    typedef typename SU3Adjoint::AMatrix AdjMatrix;
+    typedef typename SU3Adjoint::LatticeAdjMatrix  AdjMatrixField;
+    typedef typename SU3Adjoint::LatticeAdjVector  AdjVectorField;
+    const int Ngen = SU3Adjoint::Dimension;
+    AdjMatrix TRb;
+    LatticeComplex       cplx(grid); 
+    AdjVectorField  AlgV(grid); 
+    AdjMatrixField  Mab(grid);
+    AdjMatrixField  Ncb(grid);
+    AdjMatrixField  Jac(grid);
+    AdjMatrixField  Zac(grid);
+    AdjMatrixField  mZac(grid);
+    AdjMatrixField  X(grid);
+
+    int mu= (smr/2) %Nd;
+
+    auto mask=PeekIndex<LorentzIndex>(masks[smr],mu); // the cb mask
+
+    //////////////////////////////////////////////////////////////////
+    // Assemble the N matrix
+    //////////////////////////////////////////////////////////////////
+    // Computes ALL the staples -- could compute one only here
+    this->StoutSmearing->BaseSmear(C, U);
+    Cmu = peekLorentz(C, mu);
+    Umu = peekLorentz(U, mu);
+    Complex ci(0,1);
+    for(int b=0;b<Ngen;b++) {
+      SU3::generator(b, Tb);
+      // Qlat Tb = 2i Tb^Grid
+      Nb = (2.0)*Ta( ci*Tb * Umu * adj(Cmu));
+      for(int c=0;c<Ngen;c++) {
+	SU3::generator(c, Tc);
+	auto tmp = -trace(ci*Tc*Nb); // Luchang's norm: (2Tc) (2Td) N^db = -2 delta cd N^db // - was important
+	PokeIndex<ColourIndex>(Ncb,tmp,c,b); 
+      }
+    }      
+
+    //////////////////////////////////////////////////////////////////
+    // Assemble Luscher exp diff map J matrix 
+    //////////////////////////////////////////////////////////////////
+    // Ta so Z lives in Lie algabra
+    Z  = Ta(Cmu * adj(Umu));
+
+    // Move Z to the Adjoint Rep == make_adjoint_representation
+    Zac = Zero();
+    for(int b=0;b<8;b++) {
+      // Adj group sets traceless antihermitian T's -- Guido, really????
+      // Is the mapping of these the same? Same structure constants
+      // Might never have been checked.
+      SU3::generator(b, Tb);         // Fund group sets traceless hermitian T's
+      SU3Adjoint::generator(b,TRb);
+      TRb=-TRb;
+      cplx = 2.0*trace(ci*Tb*Z); // my convention 1/2 delta ba
+      Zac = Zac + cplx * TRb; // is this right? YES - Guido used Anti herm Ta's and with bloody wrong sign.
+    }
+
+    //////////////////////////////////////
+    // J(x) = 1 + Sum_k=1..N (-Zac)^k/(k+1)!
+    //////////////////////////////////////
+    X=1.0; 
+    Jac = X;
+    mZac = (-1.0)*Zac; 
+    RealD kpfac = 1;
+    for(int k=1;k<12;k++){
+      X=X*mZac;
+      kpfac = kpfac /(k+1);
+      Jac = Jac + X * kpfac;
+    }
+
+    ////////////////////////////
+    // Mab
+    ////////////////////////////
+    Mab = Complex(1.0,0.0);
+    Mab = Mab - Jac * Ncb;
+
+    ////////////////////////////
+    // det
+    ////////////////////////////
+    LatticeComplex       det(grid); 
+    det = Determinant(Mab);
+
+    ////////////////////////////
+    // ln det
+    ////////////////////////////
+    LatticeComplex       ln_det(grid); 
+    ln_det = log(det);
+
+    ////////////////////////////
+    // Masked sum
+    ////////////////////////////
+    ln_det = ln_det * mask;
+    Complex result = sum(ln_det);
+    return result.real();
+  }
+public:
+  RealD logDetJacobian(void)
+  {
+    RealD ln_det = 0;
+    if (this->smearingLevels > 0)
+    {
+      double start = usecond();
+      for (int ismr = this->smearingLevels - 1; ismr > 0; --ismr) {
+	ln_det+= logDetJacobianLevel(this->get_smeared_conf(ismr-1),ismr);
+      }
+      ln_det +=logDetJacobianLevel(*(this->ThinLinks),0);
+
+      double end = usecond();
+      double time = (end - start)/ 1e3;
+      std::cout << GridLogMessage << "GaugeConfigurationMasked: logDetJacobian took " << time << " ms" << std::endl;  
+    }
+    return ln_det;
+  }
+  void logDetJacobianForce(GaugeField &force)
+  {
+    force =Zero();
+    GaugeField force_det(force.Grid());
+
+    if (this->smearingLevels > 0)
+    {
+      double start = usecond();
+
+      GaugeLinkField tmp_mu(force.Grid());
+
+      for (int ismr = this->smearingLevels - 1; ismr > 0; --ismr) {
+
+	// remove U in UdSdU...
+	for (int mu = 0; mu < Nd; mu++) {
+	  tmp_mu = adj(peekLorentz(this->get_smeared_conf(ismr), mu)) * peekLorentz(force, mu);
+	  pokeLorentz(force, tmp_mu, mu);
+	}
+	
+      	// Propagate existing force
+        force = this->AnalyticSmearedForce(force, this->get_smeared_conf(ismr - 1), ismr);
+
+	// Add back U in UdSdU...
+	for (int mu = 0; mu < Nd; mu++) {
+	  tmp_mu = peekLorentz(this->get_smeared_conf(ismr - 1), mu) * peekLorentz(force, mu);
+	  pokeLorentz(force, tmp_mu, mu);
+	}
+    	
+	// Get this levels determinant force
+	force_det = Zero();
+	logDetJacobianForceLevel(this->get_smeared_conf(ismr-1),force_det,ismr);
+
+	// Sum the contributions
+	force = force + force_det;
+      }
+    
+      // remove U in UdSdU...
+      for (int mu = 0; mu < Nd; mu++) {
+	tmp_mu = adj(peekLorentz(this->get_smeared_conf(0), mu)) * peekLorentz(force, mu);
+	pokeLorentz(force, tmp_mu, mu);
+      }
+
+      force = this->AnalyticSmearedForce(force, *this->ThinLinks,0);
+
+      for (int mu = 0; mu < Nd; mu++) {
+	tmp_mu = peekLorentz(*this->ThinLinks, mu) * peekLorentz(force, mu);
+	pokeLorentz(force, tmp_mu, mu);
+      }
+
+      force_det = Zero();
+
+      logDetJacobianForceLevel(*this->ThinLinks,force_det,0);
+
+      force = force + force_det;
+
+      force=Ta(force); // Ta
+      
+      double end = usecond();
+      double time = (end - start)/ 1e3;
+      std::cout << GridLogMessage << "GaugeConfigurationMasked: lnDetJacobianForce took " << time << " ms" << std::endl;  
+    }  // if smearingLevels = 0 do nothing
+  }
+
+private:
+  //====================================================================
+  // Override base clas here to mask it
+  virtual void fill_smearedSet(GaugeField &U)
+  {
+    this->ThinLinks = &U;  // attach the smearing routine to the field U
+
+    // check the pointer is not null
+    if (this->ThinLinks == NULL)
+      std::cout << GridLogError << "[SmearedConfigurationMasked] Error in ThinLinks pointer\n";
+
+    if (this->smearingLevels > 0)
+    {
+      std::cout << GridLogMessage << "[SmearedConfigurationMasked] Filling SmearedSet\n";
+      GaugeField previous_u(this->ThinLinks->Grid());
+
+      GaugeField smeared_A(this->ThinLinks->Grid());
+      GaugeField smeared_B(this->ThinLinks->Grid());
+
+      previous_u = *this->ThinLinks;
+      double start = usecond();
+      for (int smearLvl = 0; smearLvl < this->smearingLevels; ++smearLvl)
+      {
+        this->StoutSmearing->smear(smeared_A, previous_u);
+	ApplyMask(smeared_A,smearLvl);
+	smeared_B = previous_u;
+	ApplyMask(smeared_B,smearLvl);
+	// Replace only the masked portion
+	this->SmearedSet[smearLvl] = previous_u-smeared_B + smeared_A;
+        previous_u = this->SmearedSet[smearLvl];
+
+        // For debug purposes
+        RealD impl_plaq = WilsonLoops<Gimpl>::avgPlaquette(previous_u);
+        std::cout << GridLogMessage << "[SmearedConfigurationMasked] smeared Plaq: " << impl_plaq << std::endl;
+      }
+      double end = usecond();
+      double time = (end - start)/ 1e3;
+      std::cout << GridLogMessage << "GaugeConfigurationMasked: Link smearing took " << time << " ms" << std::endl;  
+    }
+  }
+  //====================================================================
+  // Override base to add masking
+  virtual GaugeField AnalyticSmearedForce(const GaugeField& SigmaKPrime,
+					  const GaugeField& GaugeK,int level) 
+  {
+    GridBase* grid = GaugeK.Grid();
+    GaugeField C(grid), SigmaK(grid), iLambda(grid);
+    GaugeField SigmaKPrimeA(grid);
+    GaugeField SigmaKPrimeB(grid);
+    GaugeLinkField iLambda_mu(grid);
+    GaugeLinkField iQ(grid), e_iQ(grid);
+    GaugeLinkField SigmaKPrime_mu(grid);
+    GaugeLinkField GaugeKmu(grid), Cmu(grid);
+    
+    this->StoutSmearing->BaseSmear(C, GaugeK);
+    SigmaK = Zero();
+    iLambda = Zero();
+
+    SigmaKPrimeA = SigmaKPrime;
+    ApplyMask(SigmaKPrimeA,level);
+    SigmaKPrimeB = SigmaKPrime - SigmaKPrimeA;
+    
+    // Could get away with computing only one polarisation here
+    // int mu= (smr/2) %Nd;
+    // SigmaKprime_A has only one component
+    for (int mu = 0; mu < Nd; mu++)
+    {
+      Cmu = peekLorentz(C, mu);
+      GaugeKmu = peekLorentz(GaugeK, mu);
+      SigmaKPrime_mu = peekLorentz(SigmaKPrimeA, mu);
+      iQ = Ta(Cmu * adj(GaugeKmu));
+      this->set_iLambda(iLambda_mu, e_iQ, iQ, SigmaKPrime_mu, GaugeKmu);
+      pokeLorentz(SigmaK, SigmaKPrime_mu * e_iQ + adj(Cmu) * iLambda_mu, mu);
+      pokeLorentz(iLambda, iLambda_mu, mu);
+    }
+    this->StoutSmearing->derivative(SigmaK, iLambda,GaugeK);  // derivative of SmearBase
+
+    ////////////////////////////////////////////////////////////////////////////////////
+    // propagate the rest of the force as identity map, just add back
+    ////////////////////////////////////////////////////////////////////////////////////
+    SigmaK = SigmaK+SigmaKPrimeB;
+
+    return SigmaK;
+  }
+
+public:
+
+  /* Standard constructor */
+  SmearedConfigurationMasked(GridCartesian* _UGrid, unsigned int Nsmear, Smear_Stout<Gimpl>& Stout)
+    : SmearedConfiguration<Gimpl>(_UGrid, Nsmear,Stout)
+  {
+    assert(Nsmear%(2*Nd)==0); // Or multiply by 8??
+
+    // was resized in base class
+    assert(this->SmearedSet.size()==Nsmear);
+    
+    GridRedBlackCartesian * UrbGrid;
+    UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(_UGrid);
+    LatticeComplex one(_UGrid); one = ComplexD(1.0,0.0);
+    LatticeComplex tmp(_UGrid);
+
+    for (unsigned int i = 0; i < this->smearingLevels; ++i) {
+
+      masks.push_back(*(new LatticeLorentzComplex(_UGrid)));
+
+      int mu= (i/2) %Nd;
+      int cb= (i%2);
+      LatticeComplex tmpcb(UrbGrid);
+	
+      masks[i]=Zero();
+      ////////////////////
+      // Setup the mask
+      ////////////////////
+      tmp = Zero();
+      pickCheckerboard(cb,tmpcb,one);
+      setCheckerboard(tmp,tmpcb);
+      PokeIndex<LorentzIndex>(masks[i],tmp, mu);
+	
+    }
+    delete UrbGrid;
+  }
+  
+  virtual void smeared_force(GaugeField &SigmaTilde) 
+  {
+    if (this->smearingLevels > 0)
+    {
+      double start = usecond();
+      GaugeField force = SigmaTilde; // actually = U*SigmaTilde
+      GaugeLinkField tmp_mu(SigmaTilde.Grid());
+
+      // Remove U from UdSdU
+      for (int mu = 0; mu < Nd; mu++)
+      {
+        // to get just SigmaTilde
+        tmp_mu = adj(peekLorentz(this->SmearedSet[this->smearingLevels - 1], mu)) * peekLorentz(force, mu);
+        pokeLorentz(force, tmp_mu, mu);
+      }
+
+      for (int ismr = this->smearingLevels - 1; ismr > 0; --ismr) {
+        force = this->AnalyticSmearedForce(force, this->get_smeared_conf(ismr - 1),ismr);
+      }
+      
+      force = this->AnalyticSmearedForce(force, *this->ThinLinks,0);
+
+      // Add U to UdSdU
+      for (int mu = 0; mu < Nd; mu++)
+      {
+        tmp_mu = peekLorentz(*this->ThinLinks, mu) * peekLorentz(force, mu);
+        pokeLorentz(SigmaTilde, tmp_mu, mu);
+      }
+
+
+      double end = usecond();
+      double time = (end - start)/ 1e3;
+      std::cout << GridLogMessage << " GaugeConfigurationMasked: Smeared Force chain rule took " << time << " ms" << std::endl;
+
+    }  // if smearingLevels = 0 do nothing
+    SigmaTilde=Gimpl::projectForce(SigmaTilde); // Ta
+  }
+
+};
+
+NAMESPACE_END(Grid);
+