Updated workaround

broad unicode elimination effort

CLAUDE.md

@@ -106,7 +106,7 @@ Tests and benchmarks that need optional fermion representations are guarded by `
 ### GPU acceleration and the view/memory-manager discipline
 ### Multigrid (`Grid/algorithms/multigrid/`)
 
-Aggregation-based algebraic multigrid for Wilson-type fermions. Key files: `GeneralCoarsenedMatrixMultiRHSV2.h` (the multi-RHS coarse operator, "V2": coarsening by Fourier probing, batched-GEMM apply on a D+1 grid with rhs innermost and unvectorised), `MultiRHSBlockProject.h` (in `deflation/`; the batched-GEMM transfer operators), `PVdagMMultiGrid.h` (three-level non-Hermitian PVdagM chain with dense bottom) and `HDCGMultiGrid.h` (two-level Hermitian HDCG chain) with their `*Params.h` (XML-serialisable parameters), `MrhsMultiGrid.h` (mrhs V-cycle, preconditioner interface, fp64/fp32 seam), `Smoothers.h`, `MultiGridIO.h`, `Aggregates.h` (near-null vector construction), `Geometry.h` (coarse stencils: max-norm-1 boxes of Manhattan radius 1, 2 or 4), `CoarsenedMatrix.h` (the original nearest-neighbour Wilson multigrid). `deprecated/` holds the V1 coarse operators and the Aggregation-based single-RHS ADEF-2, kept only for the pre-2026 drivers. `MultiGrid.h` is the top-level include.
+Aggregation-based algebraic multigrid for Wilson-type fermions. Key files: `GeneralCoarsenedMatrixMultiRHS.h` (`MultiGeneralCoarsenedOperator`, the coarse operator for any number of right-hand sides: coarsening by Fourier probing, batched-GEMM apply on a D+1 grid with rhs innermost and unvectorised), `MultiRHSBlockProject.h` (in `deflation/`; the batched-GEMM transfer operators), `PVdagMMultiGrid.h` (three-level non-Hermitian PVdagM chain with dense bottom) and `HDCGMultiGrid.h` (two-level Hermitian HDCG chain) with their `*Params.h` (XML-serialisable parameters), `MrhsMultiGrid.h` (mrhs V-cycle, preconditioner interface, fp64/fp32 seam), `Smoothers.h`, `MultiGridIO.h`, `Aggregates.h` (near-null vector construction), `Geometry.h` (coarse stencils: max-norm-1 boxes of Manhattan radius 1, 2 or 4), `CoarsenedMatrix.h` (the original nearest-neighbour Wilson multigrid). `deprecated/` holds the superseded coarse operators and the Aggregation-based single-RHS ADEF-2, with class names prefixed `Deprecated` (`DeprecatedGeneralCoarsenedMatrix`, `DeprecatedMultiGeneralCoarsenedMatrix`, `DeprecatedTwoLevelADEF2`); nothing in the library uses them, only older drivers in `examples/` and `tests/debug/`. `MultiGrid.h` is the top-level include.
 
 ### GPU acceleration
 
@@ -174,7 +174,7 @@ Every programme is wrapped in `Grid_init(&argc, &argv)` / `Grid_finalize()` (`Gr
 - Template structure: most classes are templated on `<_FImpl>` (fermion impl) or `<Gimpl>` (gauge impl), which encode the representation and precision. Instantiation is controlled by `--enable-fermion-instantiations`.
 - **Tensor indices are positional, not labelled.** The `Grid/tensors/` arithmetic recurses structurally over the `iScalar`/`iVector`/`iMatrix` nest: each level defines only the {scalar,vector,matrix}² products at its own level, with element types resolved by automatic type deduction, so every colour/spin/lorentz combination composes from ~200 lines (versus the pre-C++11 QDP++/PETE approach of machine-generating every case). An index's meaning derives entirely from its nesting depth counted from the outside; `iScalar` is the identity/broadcast case at every level. Never insert or remove a nesting level casually — the multiplication tables contract by position.
 - **Multigrid coarsening deepens the tensor nest by one level.** A coarse site vector is `iVector<CComplex,nbasis>`, and `innerProduct` on it returns `iScalar<CComplex>` — one level deeper than the fine block scalar. So the block-inner-product scalar type gains one `iScalar` wrapper per MG level (fine: `vTComplex`; level 2: `iScalar<vTComplex>`; see `examples/Example_pvdagm_3level.cc`). When calling `blockInnerProduct`/`blockZAXPY`/`blockOrthogonalise` on coarse fields, the coarse scalar type must match `decltype(innerProduct(siteVector(),siteVector()))` exactly; a wrong depth fails to compile (no viable `operator=` deep in the instantiation chain) rather than mis-contracting.
-- **Grids are borrowed, never owned.** `conformable` is pointer identity, so every object that interoperates must hold the *same* `GridCartesian *`; a class that minted its own grid internally could never conform with anything else. Ownership is therefore not available, and lifetime is managed by scope discipline instead of reference counting: whoever creates a grid retains it beyond every object it handed a reference to. Anything *derived* from a grid inherits this — `~PaddedCell` dereferences its `unpadded_grid`, so a `PaddedCell` cannot even be **destroyed** after its parent grid, only used. Where a consumer must let go early, it offers an explicit hand-back (`MultiGeneralCoarsenedOperatorV2::ReleaseGrid()`) to be called *before* the grid is destroyed.
+- **Grids are borrowed, never owned.** `conformable` is pointer identity, so every object that interoperates must hold the *same* `GridCartesian *`; a class that minted its own grid internally could never conform with anything else. Ownership is therefore not available, and lifetime is managed by scope discipline instead of reference counting: whoever creates a grid retains it beyond every object it handed a reference to. Anything *derived* from a grid inherits this — `~PaddedCell` dereferences its `unpadded_grid`, so a `PaddedCell` cannot even be **destroyed** after its parent grid, only used. Where a consumer must let go early, it offers an explicit hand-back (`MultiGeneralCoarsenedOperator::ReleaseGrid()`) to be called *before* the grid is destroyed.
 
 - The `RealD`/`RealF`/`ComplexD`/`ComplexF` typedefs are used everywhere; avoid raw `double`/`float`.
 - Use `GRID_ASSERT(cond)` (defined in `Grid/GridStd.h`), not bare `assert` — it prints a Grid-formatted message and aborts cleanly under MPI.

Grid/algorithms/LinearOperator.h

@@ -667,6 +667,9 @@ public:
       (*this)(in[i], out[i]);
     }
   }
+  // The identity, if a derived class says so.  A caller can then use its input
+  // where it would otherwise have materialised the output, and skip the copy.
+  virtual int isTrivial(void) { return 0; };
   virtual ~LinearFunction(){};
 };
 

Grid/algorithms/Preconditioner.h

@@ -45,6 +45,7 @@ public:
   virtual void operator()(const Field &src, Field & psi){
     psi = src;
   }
+  virtual int isTrivial(void) { return 1; };
   TrivialPrecon(void){};
 };
 

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/algorithms/iterative/PrecGeneralisedConjugateResidual.h
 
@@ -28,213 +28,48 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #ifndef GRID_PREC_GCR_H
 #define GRID_PREC_GCR_H
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
 
-///////////////////////////////////////////////////////////////////////////////////////////////////////
-//VPGCR Abe and Zhang, 2005.
-//INTERNATIONAL JOURNAL OF NUMERICAL ANALYSIS AND MODELING
-//Computing and Information Volume 2, Number 2, Pages 147-161
-//NB. Likely not original reference since they are focussing on a preconditioner variant.
-//    but VPGCR was nicely written up in their paper
-///////////////////////////////////////////////////////////////////////////////////////////////////////
 NAMESPACE_BEGIN(Grid);
 
-#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" " 
-
+///////////////////////////////////////////////////////////////////////////////////////////////////////
+// Hermitian spelling of PrecGeneralisedConjugateResidualNonHermitian.
+//
+// The algorithm is the same; only the operator entry point differs.  The solver
+// drives Linop.Op(), while a caller of this class means "invert HermOp", so the
+// operator is wrapped in HermOpAdaptor.  The wrap is what makes the two classes
+// one: for an operator whose Op() already IS its HermOp() (HermitianLinearOperator,
+// Gamma5R5HermitianLinearOperator, HermOpAdaptor) it changes nothing, and for one
+// where they differ (MdagMLinearOperator: Op()=M, HermOp()=MdagM) it preserves the
+// Hermitian meaning the caller asked for.
+//
+// A non-zero initial guess is respected.  Callers that guarantee x0 = 0 can call
+// SetZeroGuess(1) to skip the first r0 = src - A x0 apply.
+///////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class Field>
 class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
-public:                                                
+  // Declaration order is initialisation order: the adaptor is complete before
+  // the solver binds its reference to it.
+  HermOpAdaptor<Field>                                HermLinop;
+  PrecGeneralisedConjugateResidualNonHermitian<Field> GCR;
+public:
   using LinearFunction<Field>::operator();
-  RealD   Tolerance;
-  Integer MaxIterations;
-  int verbose;
-  int mmax;
-  int nstep;
-  int steps;
-  int level;
-  GridStopWatch PrecTimer;
-  GridStopWatch MatTimer;
-  GridStopWatch LinalgTimer;
 
-  LinearFunction<Field>     &Preconditioner;
-  LinearOperatorBase<Field> &Linop;
+  PrecGeneralisedConjugateResidual(RealD tol,Integer maxit,LinearOperatorBase<Field> &Linop,
+                                   LinearFunction<Field> &Prec,int mmax,int nstep)
+    : HermLinop(Linop),
+      GCR(tol,maxit,HermLinop,Prec,mmax,nstep) {};
 
-  void Level(int lv) { level=lv; };
+  void operator() (const Field &src, Field &psi) { GCR(src,psi); };
 
-  PrecGeneralisedConjugateResidual(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
-    Tolerance(tol), 
-    MaxIterations(maxit),
-    Linop(_Linop),
-    Preconditioner(Prec),
-    mmax(_mmax),
-    nstep(_nstep)
-  { 
-    level=1;
-    verbose=1;
-  };
-
-  void operator() (const Field &src, Field &psi){
-
-    psi=Zero();
-    RealD cp, ssq,rsq;
-    ssq=norm2(src);
-    rsq=Tolerance*Tolerance*ssq;
-      
-    Field r(src.Grid());
-
-    PrecTimer.Reset();
-    MatTimer.Reset();
-    LinalgTimer.Reset();
-
-    GridStopWatch SolverTimer;
-    SolverTimer.Start();
-
-    steps=0;
-    for(int k=0;k<MaxIterations;k++){
-
-      cp=GCRnStep(src,psi,rsq);
-
-      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
-
-      if(cp<rsq) {
-
-	SolverTimer.Stop();
-
-	Linop.HermOp(psi,r);
-	axpy(r,-1.0,src,r);
-	RealD tr = norm2(r);
-	GCRLogLevel<<"PGCR: Converged on iteration " <<steps
-		 << " computed residual "<<sqrt(cp/ssq)
-		 << " true residual "    <<sqrt(tr/ssq)
-		 << " target "           <<Tolerance <<std::endl;
-
-	GCRLogLevel<<"PGCR Time elapsed: Total  "<< SolverTimer.Elapsed() <<std::endl;
-	/*
-	  GCRLogLevel<<"PGCR Time elapsed: Precon "<<   PrecTimer.Elapsed() <<std::endl;
-	  GCRLogLevel<<"PGCR Time elapsed: Matrix "<<    MatTimer.Elapsed() <<std::endl;
-	  GCRLogLevel<<"PGCR Time elapsed: Linalg "<< LinalgTimer.Elapsed() <<std::endl;
-	*/
-	return;
-      }
-
-    }
-    GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    GRID_ASSERT(0);
-  }
-
-  RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
-
-    RealD cp;
-    RealD a, b;
-    RealD zAz, zAAz;
-    RealD rq;
-
-    GridBase *grid = src.Grid();
-
-    Field r(grid);
-    Field z(grid);
-    Field tmp(grid);
-    Field ttmp(grid);
-    Field Az(grid);
-
-    ////////////////////////////////
-    // history for flexible orthog
-    ////////////////////////////////
-    std::vector<Field> q(mmax,grid);
-    std::vector<Field> p(mmax,grid);
-    std::vector<RealD> qq(mmax);
-      
-    GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
-
-    //////////////////////////////////
-    // initial guess x0 is taken as nonzero.
-    // r0=src-A x0 = src
-    //////////////////////////////////
-    MatTimer.Start();
-    Linop.HermOpAndNorm(psi,Az,zAz,zAAz); 
-    MatTimer.Stop();
-    
-
-    LinalgTimer.Start();
-    r=src-Az;
-    LinalgTimer.Stop();
-    GCRLogLevel<< "PGCR true residual r = src - A psi   "<<norm2(r) <<std::endl;
-    
-    /////////////////////
-    // p = Prec(r)
-    /////////////////////
-
-    PrecTimer.Start();
-    Preconditioner(r,z);
-    PrecTimer.Stop();
-
-    MatTimer.Start();
-    Linop.HermOpAndNorm(z,Az,zAz,zAAz); 
-    MatTimer.Stop();
-
-    LinalgTimer.Start();
-
-    //p[0],q[0],qq[0] 
-    p[0]= z;
-    q[0]= Az;
-    qq[0]= zAAz;
-    
-    cp =norm2(r);
-    LinalgTimer.Stop();
-
-    for(int k=0;k<nstep;k++){
-
-      steps++;
-
-      int kp     = k+1;
-      int peri_k = k %mmax;
-      int peri_kp= kp%mmax;
-
-      LinalgTimer.Start();
-      rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
-      a = rq/qq[peri_k];
-
-      axpy(psi,a,p[peri_k],psi);         
-
-      cp = axpy_norm(r,-a,q[peri_k],r);
-      LinalgTimer.Stop();
-
-      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
-
-      if((k==nstep-1)||(cp<rsq)){
-	return cp;
-      }
-
-
-      PrecTimer.Start();
-      Preconditioner(r,z);// solve Az = r
-      PrecTimer.Stop();
-
-      MatTimer.Start();
-      Linop.HermOpAndNorm(z,Az,zAz,zAAz);
-      MatTimer.Stop();
-
-      LinalgTimer.Start();
-
-      q[peri_kp]=Az;
-      p[peri_kp]=z;
-
-      int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
-      for(int back=0;back<northog;back++){
-
-	int peri_back=(k-back)%mmax;   	  GRID_ASSERT((k-back)>=0);
-
-	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
-	p[peri_kp]=p[peri_kp]+b*p[peri_back];
-	q[peri_kp]=q[peri_kp]+b*q[peri_back];
-
-      }
-      qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
-      LinalgTimer.Stop();
-    }
-    GRID_ASSERT(0); // never reached
-    return cp;
-  }
+  void Name(std::string n)                       { GCR.Name(n); };
+  void Level(int n)                              { GCR.Level(n); };
+  void SetZeroGuess(int z)                       { GCR.SetZeroGuess(z); };
+  int  Steps(void) const                         { return GCR.Steps(); };
+  void LogCoefficients(int l)                    { GCR.LogCoefficients(l); };
+  void SetCoefficientRecorder(GCRCoefficients *r){ GCR.SetCoefficientRecorder(r); };
+  void ReleaseHistory(void)                      { GCR.ReleaseHistory(); };
 };
+
 NAMESPACE_END(Grid);
-#undef GCRLogLevel
 #endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -43,8 +43,51 @@ NAMESPACE_BEGIN(Grid);
 
 template<class Field>
 class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
-public:                                                
+public:
   using LinearFunction<Field>::operator();
+
+  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) :
+    Tolerance(tol),
+    MaxIterations(maxit),
+    mmax(_mmax),
+    nstep(_nstep),
+    Preconditioner(Prec),
+    Linop(_Linop)
+  {
+    Level(1);
+    verbose=1;
+    trivial_prec = Preconditioner.isTrivial();
+  };
+
+  // The name is also the trace range's, so a profile separates the four
+  // instances (Fouter, Fsmoother, Couter, Csmoother) that otherwise nest
+  // indistinguishably as one shared range.
+  void Name(std::string _name) { name = _name; trace_step = name + " PGCR_step"; };
+
+  void Level(int n) { Name("Level " + std::to_string(n)); level = n; }
+
+  void SetZeroGuess(int z) { ZeroGuess = z; };
+
+  // Steps taken by the last solve.
+  int  Steps(void) const { return steps; };
+
+  // Coefficient logging: one line per step with the step length a_k and the
+  // orthogonalisation coefficients b_{k,j}.  These are the data from which a
+  // FIXED polynomial smoother can be harvested: if they are stable from call
+  // to call, the adaptive GCR can be replaced by a stationary p(A) with the
+  // same applies and no reductions.  Off by default; boss rank prints.
+  void LogCoefficients(int l) { LogCoeffs = l; };
+
+  // Optional recorder of the per-step coefficients (means over calls), for
+  // replay by GCRReplaySmoother (Smoothers.h).  Records only; no effect on
+  // the iteration.
+  void SetCoefficientRecorder(GCRCoefficients *r) { Recorder = r; if(r) r->mmax = mmax; };
+
+  // Free the persistent history (e.g. when this solver is replaced by a
+  // replayed polynomial): re-made on the next call if ever needed again.
+  void ReleaseHistory(void) { q.clear(); p.clear(); qq.clear(); hist_grid = nullptr; };
+
+private:
   RealD   Tolerance;
   RealD   SSQ;
   Integer MaxIterations;
@@ -65,45 +108,14 @@ public:
   std::vector<Field>  p;
   std::vector<RealD>  qq;
   int FirstCycle = 0;
+  int  LogCoeffs = 0;
+  int  trivial_prec;      // Preconditioner is the identity: p == r, copies fold away
+  GCRCoefficients *Recorder = nullptr;
 
   LinearFunction<Field>     &Preconditioner;
   LinearOperatorBase<Field> &Linop;
 
-  // The name is also the trace range's, so a profile separates the four
-  // instances (Fouter, Fsmoother, Couter, Csmoother) that otherwise nest
-  // indistinguishably as one shared range.
-  void Name(std::string _name) { name = _name; trace_step = name + " PGCR_step"; };
-
-  void Level(int n) { Name("Level " + std::to_string(n)); level = n; }
-
-  void SetZeroGuess(int z) { ZeroGuess = z; };
-  // Coefficient logging: one line per step with the step length a_k and the
-  // orthogonalisation coefficients b_{k,j}.  These are the data from which a
-  // FIXED polynomial smoother can be harvested: if they are stable from call
-  // to call, the adaptive GCR can be replaced by a stationary p(A) with the
-  // same applies and no reductions.  Off by default; boss rank prints.
-  int  LogCoeffs = 0;
-  void LogCoefficients(int l) { LogCoeffs = l; };
-  // Optional recorder of the per-step coefficients (means over calls), for
-  // replay by GCRReplaySmoother (Smoothers.h).  Records only; no effect on
-  // the iteration.
-  GCRCoefficients *Recorder = nullptr;
-  void SetCoefficientRecorder(GCRCoefficients *r) { Recorder = r; if(r) r->mmax = mmax; };
-  // Free the persistent history (e.g. when this solver is replaced by a
-  // replayed polynomial): re-made on the next call if ever needed again.
-  void ReleaseHistory(void) { q.clear(); p.clear(); qq.clear(); hist_grid = nullptr; };
-
-  PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
-    Tolerance(tol), 
-    MaxIterations(maxit),
-    Linop(_Linop),
-    Preconditioner(Prec),
-    mmax(_mmax),
-    nstep(_nstep)
-  {
-    Level(1);
-    verbose=1;
-  };
+public:
 
   void operator() (const Field &src, Field &psi){
 
@@ -151,6 +163,7 @@ public:
     //    GRID_ASSERT(0);
   }
 
+private:
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
 
     RealD cp;
@@ -189,18 +202,20 @@ public:
     // contract (enforced here), so r0 = src exactly; skip the apply.
     // Restart cycles (psi!=0) always do the full computation.
     //////////////////////////////////
-    if (ZeroGuess && FirstCycle) {
-      psi = Zero();
-      LinalgTimer.Start();
-      r = src;
-      LinalgTimer.Stop();
-    } else {
+    // The residual is READ through rp and only materialised in r when it is
+    // first written, at the k=0 update.  On a zero-guess first cycle r0 IS src,
+    // so no copy is made; a restart cycle computes r and rp points at it.
+    const Field *rp = &src;
+    GRID_ASSERT(nstep>=1);   // psi is first written at k=0; see zero_start below
+    int zero_start = (ZeroGuess && FirstCycle);
+    if ( !zero_start ) {
       MatTimer.Start();
       Linop.Op(psi,Az);
       MatTimer.Stop();
       LinalgTimer.Start();
       r=src-Az;
       LinalgTimer.Stop();
+      rp = &r;
     }
     FirstCycle=0;
 
@@ -208,9 +223,11 @@ public:
     // p = Prec(r)
     /////////////////////
 
-    // p[0] = Prec(r), q[0] = A p[0], written straight into the history slots
+    // p[0] = Prec(r), q[0] = A p[0], written straight into the history slots.
+    // This copy is forced even for a trivial preconditioner: p[0] is history and
+    // must survive while the residual evolves.
     PrecTimer.Start();
-    Preconditioner(r,p[0]);
+    Preconditioner(*rp,p[0]);
     PrecTimer.Stop();
 
     MatTimer.Start();
@@ -221,7 +238,7 @@ public:
 
     qq[0]= norm2(q[0]);
 
-    cp =norm2(r);
+    cp =norm2(*rp);
     LinalgTimer.Stop();
     GCRLogLevel<< "PGCR true residual "<< sqrt(cp/SSQ)     <<std::endl;
 
@@ -234,13 +251,32 @@ public:
       int peri_kp= kp%mmax;
 
       LinalgTimer.Start();
-      rq= innerProduct(q[peri_k],r); // what if rAr not real?
+      rq= innerProduct(q[peri_k],*rp); // what if rAr not real?
       a = rq/qq[peri_k];
       if ( Recorder ) Recorder->RecordA(k,a);
 
-      axpy(psi,a,p[peri_k],psi);         
+      // On a zero-guess first step psi is still unwritten, and the update would
+      // be a*p added to zero: assign it instead, so psi is never zeroed.
+      if ( zero_start && (k==0) ) psi = a*p[peri_k];
+      else                        axpy(psi,a,p[peri_k],psi);
 
-      cp = axpy_norm(r,-a,q[peri_k],r);
+#ifdef GRID_GCR_RESIDUAL_RECURRENCE
+      // |r - a q|^2 = |r|^2 - |<q,r>|^2/|q|^2 with a = <q,r>/|q|^2, so the new
+      // residual norm follows from scalars already in hand and the reduction in
+      // axpy_norm is not needed.  cp then DRIFTS rather than being measured; the
+      // true residual is recomputed at restart and at convergence, which bounds
+      // the exposure.  The drift only matters for an instance driven to a tight
+      // tolerance -- the outer solver -- and there it is the same computed-vs-true
+      // residual gap that CG lives with; the smoother and coarse instances run at
+      // fixed work or loose tolerance and are indifferent to it.
+      // Explicit re/im: ComplexD is thrust::complex under HIP.
+      axpy(r,-a,q[peri_k],*rp);
+      cp = cp - (real(rq)*real(rq)+imag(rq)*imag(rq))/qq[peri_k];
+      if ( cp < 0.0 ) cp = 0.0;          // drift can undershoot; keeps the log finite
+#else
+      cp = axpy_norm(r,-a,q[peri_k],*rp);
+#endif
+      rp = &r;   // the residual now lives in r; src is untouched from here on
       LinalgTimer.Stop();
       if ( LogCoeffs ) {
         GCRLogLevel<<"coeff["<<k<<"] a=("<<real(a)<<","<<imag(a)<<")"
@@ -254,13 +290,21 @@ public:
       }
 
       // New direction written straight into its history slot: p = Prec(r), q = A p.
-      PrecTimer.Start();
-      Preconditioner(r,p[peri_kp]);
-      PrecTimer.Stop();
-
-      MatTimer.Start();
-      Linop.Op(p[peri_kp],q[peri_kp]);
-      MatTimer.Stop();
+      // For a trivial preconditioner p[kp] would only be a copy of the residual,
+      // so the operator is applied to the residual directly and the copy folds
+      // into the orthogonalisation pass below.
+      if ( trivial_prec ) {
+        MatTimer.Start();
+        Linop.Op(*rp,q[peri_kp]);
+        MatTimer.Stop();
+      } else {
+        PrecTimer.Start();
+        Preconditioner(*rp,p[peri_kp]);
+        PrecTimer.Stop();
+        MatTimer.Start();
+        Linop.Op(p[peri_kp],q[peri_kp]);
+        MatTimer.Stop();
+      }
 
       LinalgTimer.Start();
 
@@ -289,16 +333,19 @@ public:
         if ( LogCoeffs ) bs<<" b["<<back<<"]="<<bcoef[back];
       }
       if ( Recorder && northog ) Recorder->RecordB(k,bcoef);
+      // axpyMultiNorm returns |q|^2 from the update pass, so qq is complete here:
+      // a separate norm2 would be a second reduction and a second GlobalSum.
       if ( northog ) {
-	axpyMulti(p[peri_kp],bcoef,pwin);
-	qq[peri_kp]=axpyMultiNorm(q[peri_kp],bcoef,qwin);
+        if ( trivial_prec ) axpyMultiFrom(p[peri_kp],*rp,bcoef,pwin);
+        else                axpyMulti(p[peri_kp],bcoef,pwin);
+        qq[peri_kp]=axpyMultiNorm(q[peri_kp],bcoef,qwin);
       } else {
-	qq[peri_kp]=norm2(q[peri_kp]);
+        if ( trivial_prec ) p[peri_kp] = *rp;   // no pass to fold the copy into
+        qq[peri_kp]=norm2(q[peri_kp]);
       }
       if ( LogCoeffs && northog ) {
         GCRLogLevel<<"coeff["<<k<<"]"<<bs.str()<<std::endl;
       }
-      qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
     }
     GRID_ASSERT(0); // never reached

Grid/algorithms/multigrid/BlockCyclic.h

@@ -39,8 +39,7 @@ typedef ComplexD DenseInverseScalar;
 // BlockCyclicLayout: the index arithmetic of a 2D block-cyclic distribution
 // of an N x N matrix over a Pr x Pc logical process grid with block size nb.
 //
-// This is stage 1 of the 2D distributed dense inverse
-// (documentation/DistributedDenseInverse2D.tex).  It is deliberately
+// This is stage 1 of the 2D distributed dense inverse.  It is deliberately
 // COMMUNICATOR-FREE: every mapping is a static pure function of
 // (N, nb, Pr, Pc), so the whole layout is exhaustively unit-testable on one
 // rank with no MPI in the loop (Test_blockcyclic).  A thin instance layer

Grid/algorithms/multigrid/BlockCyclicRedistribute.h

@@ -273,11 +273,15 @@ public:
 
   static void RowsToCyclic(GridBase *grid, const std::vector<int64_t> &rowStart,
                            DenseInverseScalar *rows1d, int64_t myrows, BlockCyclicMatrix &A)
-  { Redistribute(+1, grid, rowStart, rows1d, myrows, A); }
+  {
+    Redistribute(+1, grid, rowStart, rows1d, myrows, A);
+  }
 
   static void CyclicToRows(GridBase *grid, const std::vector<int64_t> &rowStart,
                            BlockCyclicMatrix &A, DenseInverseScalar *rows1d, int64_t myrows)
-  { Redistribute(-1, grid, rowStart, rows1d, myrows, A); }
+  {
+    Redistribute(-1, grid, rowStart, rows1d, myrows, A);
+  }
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/multigrid/BlockCyclicSchurInverse.h

@@ -215,7 +215,11 @@ public:
   // SendToRecvFrom with root (symmetric byte count: the reverse direction
   // carries a same-size dummy -- a leaf-local cost, accepted for simplicity).
   ///////////////////////////////////////////////////////////////////////////
-  static int64_t FirstBlock(int64_t b0, int p, int Pg){ int64_t r = ((b0 % Pg) <= p) ? b0 - (b0 % Pg) + p : b0 - (b0 % Pg) + Pg + p; return r; }
+  static int64_t FirstBlock(int64_t b0, int p, int Pg)
+  {
+    int64_t r = ((b0 % Pg) <= p) ? b0 - (b0 % Pg) + p : b0 - (b0 % Pg) + Pg + p;
+    return r;
+  }
   void BigLeaf(BlockCyclicMatrix &A, int64_t b0, int64_t b1)
   {
     GRID_TRACE("SchurBigLeaf");

Grid/algorithms/multigrid/DenseCoarseMatrix.h

@@ -38,7 +38,7 @@ NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////////////////////
 // DenseCoarseMatrix: a coarsened operator treated as a DENSE matrix -- explicit,
-// row-distributed A^{-1} of a GeneralCoarsenedMatrix.
+// row-distributed A^{-1} of a coarsened operator.
 //
 //  - Stencil -> dense DIRECT IMPORT.  The coarse operator IS the dense matrix
 //    unrolled: Dense[(s,a),(s+shift_p,b)] += A[p][s]_{a,b}.  Rows of my sites are
@@ -176,6 +176,7 @@ public:
   {
     double t0 = usecond();
     {
+      CheckMrhsSlices(Op);    // the batched apply keeps its slices apart
       ImportDense(Op);        // slab <- my rows of A   (LOCAL, no comms)
       ImportCertificate(Op);  // dense apply == Op.M, before inversion
       InvertDense(Op);        // slab <- my rows of A^{-1}
@@ -274,32 +275,67 @@ public:
     GRID_ASSERT(mgrid->_ndimension == nd+1);
     int nr = mgrid->_fdimensions[0];
 
+    // Every slice carries the same vector; slice 0 is the answer.  The
+    // slices are not a check here -- see CheckMrhsSlices, which does that
+    // once on a field chosen for the purpose.
+    Field min(mgrid), mout(mgrid);
+    for(int r=0;r<nr;r++) InsertSliceFast(in,min,r,0);
+    Op.M(min,mout);
+    ExtractSliceFast(out,mout,0,0);
+  }
+
+  ////////////////////////////////////////////////////////////////////
+  // Do the multiRHS slices stay independent?  Slice r carries (r+1) times
+  // one RANDOM field, so linearity says the results scale likewise, and a
+  // slot reading another slot's data is an O(1) error.
+  //
+  // The field is random on purpose.  Run on the output of the inverse, as
+  // this check once was, it measures the cancellation in A applied to
+  // A^-1 x instead: at the production point that is 1.3e-4 in an fp32
+  // sector with an fp32 inversion, so the check then fires on any change
+  // of summation order while saying nothing about slice independence.
+  // Once per import, not per apply.
+  ////////////////////////////////////////////////////////////////////
+  template<class CoarseOp>
+  void CheckMrhsSlices(CoarseOp &Op)
+  {
+    if ( Op.Grid() == grid ) return;               // single rhs: nothing to mix
+
+    GridBase *mgrid = Op.Grid();
+    GRID_ASSERT(mgrid->_ndimension == nd+1);
+    int nr = mgrid->_fdimensions[0];
+    if ( nr < 2 ) return;
+
+    Field in(grid);
+    GridParallelRNG rng(grid); rng.SeedFixedIntegers(std::vector<int>({7,8,9,10}));
+    random(rng,in);
+
     Field min(mgrid), mout(mgrid);
     for(int r=0;r<nr;r++){
       Field scaled(grid);
       scaled = CoarseScalar(r+1.0,0.0)*in;
       InsertSliceFast(scaled,min,r,0);
     }
-
     Op.M(min,mout);
 
-    ExtractSliceFast(out,mout,0,0);
+    Field s0(grid);
+    ExtractSliceFast(s0,mout,0,0);
+    RealD worst=0.0;
     for(int r=1;r<nr;r++){
       Field sr(grid),d(grid);
       ExtractSliceFast(sr,mout,r,0);
-      d = sr - CoarseScalar(r+1.0,0.0)*out;
+      d = sr - CoarseScalar(r+1.0,0.0)*s0;
       RealD rel = std::sqrt(norm2(d)/norm2(sr));
-      // Slices differ by rounding amplified by the operator's conditioning
-      // when the product cancels (A applied to A^-1 x): the tolerance follows
-      // the coarse precision (fp32 measured ~5e-6 here).  A genuine rhs
-      // mix-up is O(1).
+      worst = std::max(worst,rel);
       const RealD otol = (sizeof(CoarseScalar)==sizeof(ComplexF)) ? 1.0e-4 : 1.0e-6;
       if ( rel >= otol ) {
-        std::cout << GridLogMessage << "DenseCoarseMatrix: oracle rhs "<<r
-                  <<" inconsistent with rhs 0, rel "<<rel<<std::endl;
+        std::cout << GridLogMessage << "DenseCoarseMatrix: mrhs slice "<<r
+                  <<" inconsistent with slice 0, rel "<<rel<<std::endl;
       }
       GRID_ASSERT( rel < otol );
     }
+    std::cout << GridLogMessage << "DenseCoarseMatrix: mrhs slices independent to "
+              << worst << " over " << nr << " right hand sides" << std::endl;
   }
 
   ////////////////////////////////////////////////////////////////////
@@ -317,9 +353,8 @@ public:
 
     for(int p=0; p<Op.Geometry().npoint; p++){
       Coordinate shift = Op.Geometry().shifts[p];
-      // _A[p] is PADDED after ExchangeCoarseLinks (end of CoarsenOperator):
-      // extract the unpadded field before peeking with unpadded coordinates
-      // (exactly as MultiGeneralCoarsenedMatrix::CopyMatrix does).
+      // ExtractMatrix hands back the unpadded matrix field, which is what the
+      // unpadded coordinates below index.
       CoarseMatrix Aun(grid);  Op.ExtractMatrix(p,Aun);
       autoView(Av, Aun, CpuRead);
       thread_for(ss, lsites, {

Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h

@@ -34,13 +34,13 @@ NAMESPACE_BEGIN(Grid);
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>
-class MultiGeneralCoarsenedOperatorV2 : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+class MultiGeneralCoarsenedOperator : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
 public:
   typedef typename CComplex::scalar_object SComplex;
   // The BLAS scalar of this level follows the coefficient precision:
   // ComplexF for an fp32 coarse space, ComplexD for fp64.
   typedef typename GridTypeMapper<CComplex>::scalar_type CoarseBLASScalar;
-  typedef MultiGeneralCoarsenedOperatorV2<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
+  typedef MultiGeneralCoarsenedOperator<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
 
   typedef iVector<CComplex,nbasis >           siteVector;
   typedef iMatrix<CComplex,nbasis >           siteMatrix;
@@ -82,7 +82,13 @@ public:
 
   deviceVector<calcVector> BLAS_B;
   deviceVector<calcVector> BLAS_C;
-  std::vector<deviceVector<calcMatrix> > BLAS_A;
+  // The matrix, ONE buffer, site major: a site's npoint blocks are contiguous.
+  // Both organisations read it through their pointer tables, the folded one a
+  // site at a time and the unfolded one a point at a time with a stride of the
+  // whole stencil, so nothing is duplicated.  Writing or reading one point
+  // goes through a single-point scratch.
+  deviceVector<calcMatrix>               BLAS_Amat;   // [site][point]
+  deviceVector<calcMatrix>               BLAS_Ascr;   // one point, for I/O
 
   std::vector<deviceVector<CoarseBLASScalar *> > BLAS_AP;
   std::vector<deviceVector<CoarseBLASScalar *> > BLAS_BP;
@@ -143,20 +149,106 @@ public:
       for(int l=0;l<legs;l++){
 	int p = g*LegGroup+l;
 	for(int32_t ss=0;ss<sites;ss++){
-	  CoarseBLASScalar *ptr = (CoarseBLASScalar *)&BLAS_A[p][ss];
+	  CoarseBLASScalar *ptr = (CoarseBLASScalar *)&BLAS_Amat[(uint64_t)ss*geom.npoint+p];
 	  acceleratorPut(BLAS_APg[g][l*sites+ss],ptr);
 	}
       }
     }
   }
 
-  void SetLegGroup(int G)
+  ///////////////////////////////////////////////////////////////////////////
+  // Stencil legs folded along the CONTRACTION instead of the batch.
+  //
+  // One GEMM per site of [nbasis x (KFold*nbasis)] . [(KFold*nbasis) x nrhs],
+  // accumulating straight into C, so there are no partial results to combine.
+  // It needs the site's matrices contiguous, which is a repack done once, and
+  // the neighbour vectors gathered into one block per site, which replaces the
+  // copy the ungrouped path already does.
+  //
+  // Measured on MI250X at nbasis 60, batch 1024, single precision: the short
+  // form (K=60) reaches 487 GB/s at one right-hand side and 625 at twelve,
+  // the folded form (K=1980) 912 and 823.  Beyond about K=2000 it collapses:
+  // K=4860, the full 81 point box, gives 304 GB/s.  So the fold is chunked to
+  // keep K near 2000, and the chunks accumulate.
+  //
+  // The gather costs one extra write of nbasis*nrhs per leg per site, which is
+  // 3% of the traffic at one right-hand side and 20% at twelve; the fold wins
+  // at both, and by most at one, which is the HMC case.
+  ///////////////////////////////////////////////////////////////////////////
+  static const int TargetK = 2048;
+
+  int KFold = 0;                                     // legs per GEMM; 0 = off
+  deviceVector<CoarseBLASScalar>                 BLAS_BK;   // [site][rhs][leg][nbasis]
+  std::vector<deviceVector<CoarseBLASScalar *> > BLAS_AKP;  // per chunk, per site
+  std::vector<deviceVector<CoarseBLASScalar *> > BLAS_BKP;  // per chunk, per site
+  deviceVector<CoarseBLASScalar *>               BLAS_BPflat; // [point][site] neighbours
+
+  int KChunks(void) const { return (geom.npoint + KFold - 1)/KFold; }
+  int KLegs(int g)  const { return std::min(KFold, geom.npoint - g*KFold); }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Repack the matrix so a site's stencil points are contiguous: the folded
+  // GEMM reads each site's block as one column-major nbasis x (legs*nbasis)
+  // matrix, which is also a single sequential stream rather than npoint
+  // interleaved ones.  Costs one pass over the matrix, once per solve setup.
+  ///////////////////////////////////////////////////////////////////////////
+  void PackFoldedMatrix(void)
   {
-    GRID_ASSERT( G>=1 );
-    GRID_ASSERT( G<=geom.npoint );   // every partial must get an initialising call
-    LegGroup = G;
-    BuildGroupedA();
-    if ( _CoarseGridMulti ) SetGrid(_CoarseGridMulti);   // rebuild the rest
+    // Nothing to pack: the matrix is already site major.  Only the pointer
+    // table is built, one entry per site per chunk, at the chunk's first leg.
+    int32_t sites = _CoarseGrid->lSites();
+    const int np = geom.npoint;
+    BLAS_AKP.resize(KChunks());
+    for(int g=0;g<KChunks();g++){
+      BLAS_AKP[g].resize(sites);
+      for(int32_t ss=0;ss<sites;ss++){
+	CoarseBLASScalar *ptr = (CoarseBLASScalar *)&BLAS_Amat[(uint64_t)ss*np + (uint64_t)g*KFold];
+	acceleratorPut(BLAS_AKP[g][ss],ptr);
+      }
+    }
+  }
+
+  // Legs per GEMM that keeps K near where the kernel runs best.
+  int AutoKFold(void) const
+  {
+    int K = TargetK/nbasis;
+    if ( K < 1 )             K = 1;
+    if ( K > geom.npoint )   K = geom.npoint;
+    return K;
+  }
+  // Legs per call that brings the batch up to where the kernel saturates.
+  int AutoLegGroup(void) const
+  {
+    int32_t sites = _CoarseGrid->lSites();
+    int G = (TargetBatch + sites - 1)/sites;
+    if ( G < 1 )             G = 1;
+    if ( G > geom.npoint )   G = geom.npoint;
+    return G;
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // The fold along K gathers the neighbours, which costs nbasis*nrhs per leg
+  // per site: 16 MB of traffic per apply at one right hand side and 195 at
+  // twelve, against a matrix of 973.  Measured at the production point it is
+  // worth 7.6% at one right hand side and costs 7.6% at twelve, so the fold
+  // is used only for a small batch, and a large one carries its legs in the
+  // batch dimension instead.
+  ///////////////////////////////////////////////////////////////////////////
+  static const int FoldNrhsMax = 4;
+
+  void ChooseOrganisation(int nrhs)
+  {
+    if ( nrhs <= FoldNrhsMax ) {
+      KFold = AutoKFold();
+    } else {
+      KFold = 0;
+    }
+    LegGroup = AutoLegGroup();
+    if ( KFold ) {
+      PackFoldedMatrix();
+    } else {
+      BuildGroupedA();
+    }
   }
 
   ///////////////////////
@@ -171,7 +263,7 @@ public:
   {
     if ( _CoarseGridMulti == nullptr ) {
       std::cout << GridLogError
-		<< "MultiGeneralCoarsenedOperatorV2: the multiRHS grid has not been set."
+		<< "MultiGeneralCoarsenedOperator: the multiRHS grid has not been set."
 		<< std::endl;
       std::cout << GridLogError
 		<< "  Call SetGrid(CoarseGridMulti) with the D+1 dimensional grid your"
@@ -184,33 +276,58 @@ public:
   }
 
   //////////////////////////////////////////////////////////////////////////
-  // Bilingual accessors, matching GeneralCoarsenedMatrix. Note Grid() is the
-  // D+1 multiRHS grid here, so a consumer wanting the space the elements live
-  // on must ask for CoarseGridD().
+  // Accessors.  Note Grid() is the D+1 multiRHS grid, so a consumer wanting
+  // the space the matrix elements live on must ask for CoarseGridD().
   //////////////////////////////////////////////////////////////////////////
   // Resident device memory this operator holds (matrix elements and the
   // Nrhs-dependent BLAS buffers).  Not evictable.
   uint64_t DeviceBytes(void)
   {
-    uint64_t b = (uint64_t)(BLAS_B.capacity()+BLAS_C.capacity()+BLAS_Cg.capacity())*sizeof(calcVector);
-    for(int p=0;p<(int)BLAS_A.size();p++) b += (uint64_t)BLAS_A[p].capacity()*sizeof(calcMatrix);
+    uint64_t b = (uint64_t)(BLAS_B.capacity()+BLAS_C.capacity()+BLAS_Cg.capacity())*sizeof(calcVector)
+               + (uint64_t)BLAS_BK.capacity()*sizeof(CoarseBLASScalar);
+    b += (uint64_t)(BLAS_Amat.capacity()+BLAS_Ascr.capacity())*sizeof(calcMatrix);
     return b;
   }
 
   NonLocalStencilGeometry & Geometry(void)      { return geom_srhs; };
-  void ExtractMatrix(int p,CoarseMatrix &A)     { BLAStoGrid(A,BLAS_A[p]); };
+
+  // One stencil point in or out of the site-major buffer.
+  void MatrixPointIn(int p,const deviceVector<calcMatrix> &src)
+  {
+    int32_t sites = _CoarseGrid->lSites();
+    const int np = geom.npoint;
+    calcMatrix       *dst = &BLAS_Amat[0];
+    const calcMatrix *sp  = &src[0];
+    accelerator_for(ss,sites,1,{ dst[(uint64_t)ss*np+p] = sp[ss]; });
+  }
+  void MatrixPointOut(int p,deviceVector<calcMatrix> &dst)
+  {
+    int32_t sites = _CoarseGrid->lSites();
+    const int np = geom.npoint;
+    const calcMatrix *sp = &BLAS_Amat[0];
+    dst.resize(sites);
+    calcMatrix *dp = &dst[0];
+    accelerator_for(ss,sites,1,{ dp[ss] = sp[(uint64_t)ss*np+p]; });
+  }
+  void ExtractMatrix(int p,CoarseMatrix &A)
+  {
+    MatrixPointOut(p,BLAS_Ascr);
+    BLAStoGrid(A,BLAS_Ascr);
+  }
 
   // I/O on the operator matrices, via the BLAS layout array. The parameter is
   // a vector over the geometry points; the body indexes A[p].
   void SetMatrix (int p,std::vector<CoarseMatrix> & A)
   {
     GRID_ASSERT(A.size()==geom_srhs.npoint);
-    GridtoBLAS(A[p],BLAS_A[p]);
+    GridtoBLAS(A[p],BLAS_Ascr);
+    MatrixPointIn(p,BLAS_Ascr);
   }
   void GetMatrix (int p,std::vector<CoarseMatrix> & A)
   {
     GRID_ASSERT(A.size()==geom_srhs.npoint);
-    BLAStoGrid(A[p],BLAS_A[p]);
+    MatrixPointOut(p,BLAS_Ascr);
+    BLAStoGrid(A[p],BLAS_Ascr);
   }
 
   ///////////////////////////////////////////////////////////////////////////
@@ -218,7 +335,7 @@ public:
   // is independent of Nrhs, in particular the matrix elements, which must
   // survive a change of Nrhs untouched.
   ///////////////////////////////////////////////////////////////////////////
-  MultiGeneralCoarsenedOperatorV2(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGrid) :
+  MultiGeneralCoarsenedOperator(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGrid) :
     _CoarseGrid(CoarseGrid),
     geom_srhs(_geom),
     geom(CoarseGrid,_geom.hops,_geom.skip),
@@ -233,28 +350,26 @@ public:
     /////////////////////////////////////////////////
     // Matrix elements and their pointer table
     /////////////////////////////////////////////////
-    BLAS_A.resize(geom.npoint);
+    BLAS_Amat.resize((uint64_t)unpadded_sites*geom.npoint);
+    BLAS_Ascr.resize(unpadded_sites);
     BLAS_AP.resize(geom.npoint);
     for(int p=0;p<geom.npoint;p++){
-      BLAS_A[p].resize (unpadded_sites); // no ghost zone, npoint elements
       BLAS_AP[p].resize(unpadded_sites);
     }
 
-    // Site identity mapping for A
+    // Point p at site ss: the unfolded path strides by the whole stencil.
     for(int p=0;p<geom.npoint;p++){
       for(int ss=0;ss<unpadded_sites;ss++){
-	CoarseBLASScalar *ptr = (CoarseBLASScalar *)&BLAS_A[p][ss];
+	CoarseBLASScalar *ptr = (CoarseBLASScalar *)&BLAS_Amat[(uint64_t)ss*geom.npoint+p];
 	acceleratorPut(BLAS_AP[p][ss],ptr);
       }
     }
 
-    // Enough legs to bring the batch up to where the kernel saturates, and
-    // never more legs than the stencil has.
-    LegGroup = (TargetBatch + unpadded_sites - 1)/unpadded_sites;
-    if ( LegGroup < 1 )            LegGroup = 1;
-    if ( LegGroup > geom.npoint )  LegGroup = geom.npoint;
+    // The organisation is chosen in SetGrid, where Nrhs is known.
+    KFold    = 0;
+    LegGroup = AutoLegGroup();
     BuildGroupedA();
-    std::cout << GridLogMessage << "MultiGeneralCoarsenedOperatorV2: stencil "
+    std::cout << GridLogMessage << "MultiGeneralCoarsenedOperator: stencil "
 	      << geom.npoint << " points, local coarse volume " << unpadded_sites
 	      << ", legs per GEMM " << LegGroup
 	      << " (batch " << LegGroup*unpadded_sites << ", "
@@ -263,7 +378,7 @@ public:
 	      << ")" << std::endl;
   }
 
-  virtual ~MultiGeneralCoarsenedOperatorV2()
+  virtual ~MultiGeneralCoarsenedOperator()
   {
     ReleaseGrid();
   }
@@ -285,6 +400,10 @@ public:
     BLAS_Cg.resize(0);
     BLAS_CPg.resize(0);
     BLAS_CPr.resize(0);
+    BLAS_BK.resize(0);
+    BLAS_BPflat.resize(0);
+    for(int g=0;g<(int)BLAS_BKP.size();g++){ BLAS_BKP[g].resize(0); }
+    BLAS_BKP.resize(0);
     for(int g=0;g<(int)BLAS_BPg.size();g++){ BLAS_BPg[g].resize(0); }
     BLAS_BPg.resize(0);
     for(int p=0;p<BLAS_BP.size();p++){
@@ -342,6 +461,19 @@ public:
     GRID_ASSERT(CellMulti->grids.back()->lSites() == nrhs*padded_sites);
     GRID_ASSERT(_CoarseGridMulti->lSites()        == nrhs*unpadded_sites);
 
+    /////////////////////////////////////////////////
+    // Choose the organisation FIRST: every table below is sized from
+    // LegGroup or KFold, and is indexed with them in the apply, so the two
+    // must be decided before anything is allocated.
+    /////////////////////////////////////////////////
+    ChooseOrganisation(nrhs);
+    std::cout << GridLogMessage << "MultiGeneralCoarsenedOperator: Nrhs " << nrhs
+	      << ", coarse apply "
+	      << ( KFold ? "folds "+std::to_string(KFold)+" legs along K (K = "
+		           +std::to_string(KFold*nbasis)+", "+std::to_string(KChunks())+" chunk(s))"
+		         : "carries "+std::to_string(LegGroup)+" legs in the batch" )
+	      << std::endl;
+
     /////////////////////////////////////////////////
     // Device data vector storage
     /////////////////////////////////////////////////
@@ -382,6 +514,24 @@ public:
       BLAS_BPg[g].resize(legs*unpadded_sites);
     }
 
+    // Folded-along-K working set: one chunk of gathered neighbour vectors,
+    // laid out per site as a column-major (legs*nbasis) x nrhs block, plus a
+    // pointer table per chunk because a short last chunk has a shorter site
+    // stride.  BLAS_BPflat is the neighbour table the gather kernel reads.
+    if ( KFold ) {
+      BLAS_BK.resize((uint64_t)unpadded_sites*nrhs*KFold*nbasis);
+      BLAS_BKP.resize(KChunks());
+      for(int g=0;g<KChunks();g++){
+	int legs = KLegs(g);
+	BLAS_BKP[g].resize(unpadded_sites);
+	for(int32_t ss=0;ss<unpadded_sites;ss++){
+	  CoarseBLASScalar *ptr = &BLAS_BK[(uint64_t)ss*nrhs*legs*nbasis];
+	  acceleratorPut(BLAS_BKP[g][ss],ptr);
+	}
+      }
+      BLAS_BPflat.resize((uint64_t)geom.npoint*unpadded_sites);
+    }
+
     // Neighbour table is more complicated
     int32_t j=0; // Interior point counter (unpadded)
     for(int32_t s=0;s<padded_sites;s++){ // D volume, padded
@@ -402,6 +552,7 @@ public:
 	  CoarseBLASScalar * ptr = (CoarseBLASScalar *)&BLAS_B[nbr];
 	  acceleratorPut(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
 	  acceleratorPut(BLAS_BPg[point/LegGroup][(point%LegGroup)*unpadded_sites+j],ptr);
+	  if ( KFold ) acceleratorPut(BLAS_BPflat[(uint64_t)point*unpadded_sites+j],ptr);
 	}
 	j++;
       }
@@ -707,7 +858,7 @@ public:
   // orthonormalised basis into it, and the caller keeps it for the solve.
   // The overload below makes its own and is kept for the pre-2026 drivers;
   // it costs a second full basis store, which is why nothing new should use
-  // it (documentation/MultiGridMemoryAudit.md).
+  // it.
   void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
 		       GridCartesian *FineGridMulti,
 		       std::vector<FineField> &Subspace,
@@ -815,8 +966,10 @@ public:
     }
 
     for(int p=0;p<npoint;p++){
-      GridtoBLAS(_A[p],BLAS_A[p]);
+      GridtoBLAS(_A[p],BLAS_Ascr);
+      MatrixPointIn(p,BLAS_Ascr);
     }
+    if ( KFold ) PackFoldedMatrix();   // the matrix is only valid from here
     CoarsenReport(tphase,tphaseBZ,tslice,tmat,tproj,tinv);
   }
 
@@ -920,8 +1073,10 @@ public:
     }
 
     for(int p=0;p<npoint;p++){
-      GridtoBLAS(_A[p],BLAS_A[p]);
+      GridtoBLAS(_A[p],BLAS_Ascr);
+      MatrixPointIn(p,BLAS_Ascr);
     }
+    if ( KFold ) PackFoldedMatrix();   // the matrix is only valid from here
     CoarsenReport(tphase,tphaseBZ,tslice,tmat,tproj,tinv);
   }
   void Mdag(const CoarseVector &in, CoarseVector &out)
@@ -944,7 +1099,7 @@ public:
     CheckGridSet();
     if ( in.Grid() != _CoarseGridMulti ) {
       std::cout << GridLogError
-		<< "MultiGeneralCoarsenedOperatorV2::M called with a field on a"
+		<< "MultiGeneralCoarsenedOperator::M called with a field on a"
 		<< std::endl;
       std::cout << GridLogError
 		<< "  different grid object from the one given to SetGrid(). Two"
@@ -955,14 +1110,14 @@ public:
       GRID_ASSERT(in.Grid() == _CoarseGridMulti);
     }
 
-    GRID_TRACE("CoarseV2Mult");
+    GRID_TRACE("CoarseMult");
     t_tot=-usecond();
     t_exch=-usecond();
     // The exchange takes its input by const reference, so it reads the
     // caller's field directly.  lambda scope so the roctx range covers
     // exactly the exchange; the PaddedCellFwd/BwdMPI markers inside it then
     // nest properly.
-    CoarseVector pin = [&](){ GRID_TRACE("CoarseV2Exchange");
+    CoarseVector pin = [&](){ GRID_TRACE("CoarseExchange");
                               return CellMulti->ExchangePeriodic(in); }(); //padded input
     t_exch+=usecond();
 
@@ -985,7 +1140,7 @@ public:
     
 
     t_GtoB=-usecond();
-    { GRID_TRACE("CoarseV2GridToBLAS");
+    { GRID_TRACE("CoarseGridToBLAS");
       GridtoBLAS(pin,BLAS_B);
     }
     t_GtoB+=usecond();
@@ -993,10 +1148,46 @@ public:
     GridBLAS BLAS;
 
     t_mult=-usecond();
-    { GRID_TRACE("CoarseV2StencilGEMM");
+    { GRID_TRACE("CoarseStencilGEMM");
     // The scalar type selects the GEMM: Cgemm for an fp32 coarse space,
     // Zgemm for fp64.
-    if ( LegGroup == 1 ) {
+    if ( KFold ) {
+      // Gather one chunk of neighbour vectors per site, then one GEMM whose
+      // contraction runs over the chunk's legs, accumulating into C.  No
+      // partial results, so nothing to combine afterwards.
+      const int    nb  = nbasis;
+      const int64_t ns = (int64_t)_CoarseGrid->lSites();  // D dimensional: CoarseGrid() is D+1
+      GRID_ASSERT( (int64_t)BLAS_BPflat.size() == (int64_t)geom.npoint*ns );
+      CoarseBLASScalar * const bk = &BLAS_BK[0];
+      CoarseBLASScalar * const * const bp = &BLAS_BPflat[0];
+      for(int g=0;g<KChunks();g++){
+	const int64_t legs = KLegs(g);
+	const int64_t leg0 = (int64_t)g*KFold;
+	{ GRID_TRACE("CoarseLegGather");
+	  accelerator_for(idx, ns*legs*nrhs, 1, {
+	      int64_t j  = idx % nrhs;          // right hand side  (column)
+	      int64_t t  = idx / nrhs;
+	      int64_t l  = t % legs;            // leg within the chunk
+	      int64_t ss = t / legs;            // site
+	      const CoarseBLASScalar *src = bp[(leg0+l)*ns + ss] + j*nb;
+	      CoarseBLASScalar       *dst = bk + ((ss*nrhs + j)*legs + l)*nb;
+	      for(int i=0;i<nb;i++) dst[i] = src[i];
+	    });
+	}
+	RealD c = (g==0) ? 0.0 : 1.0;
+	BLAS.gemmBatched(nbasis,nrhs,legs*nbasis,
+			 CoarseBLASScalar(1.0),
+			 BLAS_AKP[g],
+			 BLAS_BKP[g],
+			 CoarseBLASScalar(c),
+			 BLAS_CP);
+	// One gather buffer serves every chunk and the multiply is issued
+	// asynchronously, so the next chunk's gather must not start until this
+	// multiply has read it.  One synchronise per chunk, none when the whole
+	// stencil folds into a single call.
+	if ( g+1 < KChunks() ) BLAS.synchronise();
+      }
+    } else if ( LegGroup == 1 ) {
       for(int p=0;p<geom.npoint;p++){
 	RealD c = (p==0) ? 0.0 : 1.0;
 	BLAS.gemmBatched(nbasis,nrhs,nbasis,
@@ -1037,7 +1228,7 @@ public:
     }
     t_mult+=usecond();
 
-    if ( LegGroup > 1 ) { GRID_TRACE("CoarseV2LegSum");
+    if ( !KFold && LegGroup > 1 ) { GRID_TRACE("CoarseLegSum");
       // Sum the LegGroup partial results into the single output buffer.
       // Flat over scalars: a calcVector is nbasis of them and the partials
       // are contiguous, one whole C volume after another.
@@ -1053,7 +1244,7 @@ public:
     }
 
     t_BtoG=-usecond();
-    { GRID_TRACE("CoarseV2BLASToGrid");
+    { GRID_TRACE("CoarseBLASToGrid");
       BLAStoGrid(out,BLAS_C);
     }
     t_BtoG+=usecond();

Grid/algorithms/multigrid/HDCGMultiGrid.h

@@ -34,7 +34,7 @@ NAMESPACE_BEGIN(Grid);
 //   HDCGCoarsening                 everything that is a function of the gauge
 //                                  field: the raw near-null basis (and its
 //                                  refinement), the transfer operators, the
-//                                  Galerkin coarse operator (V2, Hermitian),
+//                                  Galerkin coarse operator (Hermitian),
 //                                  the coarse eigenvectors for deflation
 //   HDCGSolver                     the solve chain on a borrowed coarsening:
 //                                  smoother, deflated coarse solve, ADEF-2
@@ -51,7 +51,7 @@ template<class Fobj,class CComplex,int nbasis>
 class HDCGCoarsening {
 public:
   typedef Lattice<Fobj>                                          FineField;
-  typedef MultiGeneralCoarsenedOperatorV2<Fobj,CComplex,nbasis>  CoarseOperator;
+  typedef MultiGeneralCoarsenedOperator<Fobj,CComplex,nbasis>  CoarseOperator;
   typedef typename CoarseOperator::CoarseVector                  CoarseVector;
   typedef typename GridTypeMapper<Fobj>::SinglePrecision         FobjF;
   typedef Lattice<FobjF>                                         FineFieldF;

Grid/algorithms/multigrid/MrhsMultiGrid.h

@@ -50,23 +50,74 @@ public:
   LinearOperatorBase<Field> &Linop;
   MrhsLinearFunction<Field> &Preconditioner;
   std::function<void(int)> OnStep;      // called with the outer step count after every step
-  void Level(int lv){ name = "Level " + std::to_string(lv); level=lv; }
-  void Name(std::string n){ name = n; trace_op = name+" MrhsPGCR::vOp"; trace_orthog = name+" MrhsPGCR orthog"; }
-  void SetZeroGuess(int z){ ZeroGuess=z; }
+  
+  void Level(int lv){
+    name = "Level " + std::to_string(lv);
+    level=lv;
+  }
+
+  void Name(std::string n){
+    name = n;
+    trace_op = name+" MrhsPGCR::vOp";
+    trace_orthog = name+" MrhsPGCR orthog";
+  }
+
+  void SetZeroGuess(int z)
+  {
+    ZeroGuess=z;
+  }
+
   MrhsPGCRNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,MrhsLinearFunction<Field> &Prec,int _mmax,int _nstep)
-    : Tolerance(tol),MaxIterations(maxit),Linop(_Linop),Preconditioner(Prec),mmax(_mmax),nstep(_nstep){ level=1; }
-  static RealD vnorm2(std::vector<Field> &x){ RealD s=0; for(auto &f:x) s+=norm2(f); return s; }
-  static ComplexD vinnerProduct(std::vector<Field> &x,std::vector<Field> &y){ ComplexD s(0); for(int r=0;r<(int)x.size();r++) s+=innerProduct(x[r],y[r]); return s; }
-  static void vaxpy(std::vector<Field> &z,ComplexD a,std::vector<Field> &x,std::vector<Field> &y){ for(int r=0;r<(int)z.size();r++) axpy(z[r],a,x[r],y[r]); }
-  void vOp(std::vector<Field> &in,std::vector<Field> &out){ GRID_TRACE(trace_op.c_str()); for(int r=0;r<(int)in.size();r++) Linop.Op(in[r],out[r]); }
+    : Tolerance(tol),MaxIterations(maxit),Linop(_Linop),Preconditioner(Prec),mmax(_mmax),nstep(_nstep)
+  {
+    level=1;
+  }
+
+  static RealD vnorm2(std::vector<Field> &x)
+  {
+    RealD s=0;
+    for(auto &f:x) s+=norm2(f);
+    return s;
+  }
+
+  static ComplexD vinnerProduct(std::vector<Field> &x,std::vector<Field> &y){
+    ComplexD s(0);
+    for(int r=0;r<(int)x.size();r++) {
+      s+=innerProduct(x[r],y[r]);
+    }
+    return s;
+  }
+
+  static void vaxpy(std::vector<Field> &z,ComplexD a,std::vector<Field> &x,std::vector<Field> &y)
+  {
+    for(int r=0;r<(int)z.size();r++) {
+      axpy(z[r],a,x[r],y[r]);
+    }
+  }
+
+  void vOp(std::vector<Field> &in,std::vector<Field> &out)
+  {
+    GRID_TRACE(trace_op.c_str());
+    for(int r=0;r<(int)in.size();r++) {
+      Linop.Op(in[r],out[r]);
+    }
+  }
   void operator()(std::vector<Field> &src,std::vector<Field> &psi){
-    RealD cp,ssq,rsq; int nrhs=src.size(); GridBase *grid=src[0].Grid();
-    ssq=vnorm2(src); rsq=Tolerance*Tolerance*ssq;
+    RealD cp,ssq,rsq;
+    int nrhs=src.size();
+    GridBase *grid=src[0].Grid();
+    ssq=vnorm2(src);
+    rsq=Tolerance*Tolerance*ssq;
+
     std::vector<Field> r(nrhs,grid);
-    GridStopWatch T; T.Start(); steps=0; FirstCycle=1;
+    GridStopWatch T;
+    T.Start();
+    steps=0;
+    FirstCycle=1;
     for(int k=0;k<MaxIterations;k++){
       cp=GCRnStep(src,psi,rsq);
-      std::cout<<GridLogMessage<<std::string(level,'\t')<<" "<<name<<" MrhsPGCR("<<mmax<<","<<nstep<<") "<<steps<<" steps cp = "<<cp<<" target "<<rsq<<std::endl;
+      std::cout<<GridLogMessage<<std::string(level,'\t')<<" "<<name<<" MrhsPGCR("<<mmax<<","<<nstep<<") "<<steps
+	       <<" steps cp = "<<cp<<" target "<<rsq<<std::endl;
       if(cp<rsq){
         T.Stop(); vOp(psi,r); for(int rr=0;rr<nrhs;rr++) axpy(r[rr],-1.0,src[rr],r[rr]);
         RealD tr=vnorm2(r);
@@ -78,28 +129,61 @@ public:
     }
     std::cout<<GridLogMessage<<"MrhsPGCR: did not converge"<<std::endl;
   }
-  RealD GCRnStep(std::vector<Field> &src,std::vector<Field> &psi,RealD rsq){
-    RealD cp; ComplexD a,rq; int nrhs=src.size(); GridBase *grid=src[0].Grid();
+  RealD GCRnStep(std::vector<Field> &src,std::vector<Field> &psi,RealD rsq)
+  {
+    RealD cp;
+    ComplexD a,rq;
+    int nrhs=src.size();
+    GridBase *grid=src[0].Grid();
+
     std::vector<Field> r(nrhs,grid),Az(nrhs,grid);   // Az: restart residual scratch only
     std::vector< std::vector<Field> > q(mmax,std::vector<Field>(nrhs,grid));
     std::vector< std::vector<Field> > p(mmax,std::vector<Field>(nrhs,grid));
     std::vector<RealD> qq(mmax);
-    if (ZeroGuess && FirstCycle) { for(int rr=0;rr<nrhs;rr++){ psi[rr]=Zero(); r[rr]=src[rr]; } }
-    else                         { vOp(psi,Az); for(int rr=0;rr<nrhs;rr++) r[rr]=src[rr]-Az[rr]; }
+
+    if (ZeroGuess && FirstCycle) {
+      for(int rr=0;rr<nrhs;rr++){
+	psi[rr]=Zero();
+	r[rr]=src[rr];
+      }
+    } else {
+      vOp(psi,Az);
+      for(int rr=0;rr<nrhs;rr++) {
+	r[rr]=src[rr]-Az[rr];
+      }
+    }
+
     FirstCycle=0;
     // p[0]=Prec(r), q[0]=A p[0], produced directly in the history slots (no copies)
-    Preconditioner(r,p[0]); vOp(p[0],q[0]); qq[0]=vnorm2(q[0]); cp=vnorm2(r);
+
+    Preconditioner(r,p[0]);
+    vOp(p[0],q[0]);
+    qq[0]=vnorm2(q[0]);
+    cp=vnorm2(r);
+    
     for(int k=0;k<nstep;k++){
-      steps++; int kp=k+1, peri_k=k%mmax, peri_kp=kp%mmax;
+      steps++;
+      int kp=k+1, peri_k=k%mmax, peri_kp=kp%mmax;
+      
       if ( OnStep ) OnStep(steps);
-      rq=vinnerProduct(q[peri_k],r); a=rq/qq[peri_k];
-      vaxpy(psi,a,p[peri_k],psi); vaxpy(r,-a,q[peri_k],r); cp=vnorm2(r);
+      
+      rq=vinnerProduct(q[peri_k],r);
+      a=rq/qq[peri_k];
+      vaxpy(psi,a,p[peri_k],psi);
+      vaxpy(r,-a,q[peri_k],r);
+      cp=vnorm2(r);
+
       std::cout<<GridLogMessage<<std::string(level,'\t')<<" "<<name<<" MrhsPGCR step["<<steps<<"]  resid "<<cp<<" target "<<rsq<<std::endl;
+
       if((k==nstep-1)||(cp<rsq)) return cp;
+
       // New direction straight into its history slot: p=Prec(r), q=A p.
       Preconditioner(r,p[peri_kp]);
+
       vOp(p[peri_kp],q[peri_kp]);
+
       int northog=((kp)>(mmax-1))?(mmax-1):(kp);
+
       {
         GRID_TRACE(trace_orthog.c_str());
         // Classical Gram-Schmidt: all coefficients against the UN-updated new q
@@ -110,15 +194,30 @@ public:
         std::vector<ComplexD> bcoef(northog,ComplexD(0.0)), part;
         for(int rr=0;rr<nrhs;rr++){
           std::vector<const Field*> qwin(northog);
-          for(int back=0;back<northog;back++){ int peri_back=(k-back)%mmax; GRID_ASSERT((k-back)>=0); qwin[back]=&q[peri_back][rr]; }
+          for(int back=0;back<northog;back++){
+	    int peri_back=(k-back)%mmax;
+	    GRID_ASSERT((k-back)>=0);
+	    qwin[back]=&q[peri_back][rr];
+	  }
+
           rankInnerProductMulti(part,qwin,q[peri_kp][rr]);
+	  
           for(int back=0;back<northog;back++) bcoef[back]+=part[back];
         }
+
         if(northog) grid->GlobalSumVector(&bcoef[0],northog);
-        for(int back=0;back<northog;back++){ int peri_back=(k-back)%mmax; bcoef[back]=-bcoef[back]/qq[peri_back]; }
+
+        for(int back=0;back<northog;back++){
+	  int peri_back=(k-back)%mmax;
+	  bcoef[back]=-bcoef[back]/qq[peri_back];
+	}
         for(int rr=0;rr<nrhs;rr++){
           std::vector<const Field*> qwin(northog), pwin(northog);
-          for(int back=0;back<northog;back++){ int peri_back=(k-back)%mmax; qwin[back]=&q[peri_back][rr]; pwin[back]=&p[peri_back][rr]; }
+          for(int back=0;back<northog;back++){
+	    int peri_back=(k-back)%mmax;
+	    qwin[back]=&q[peri_back][rr];
+	    pwin[back]=&p[peri_back][rr];
+	  }
           axpyMulti(p[peri_kp][rr],bcoef,pwin);
           axpyMulti(q[peri_kp][rr],bcoef,qwin);
         }
@@ -168,24 +267,42 @@ public:
       _Coarse5d(Coarse5d), _CoarseCoarse5d(CoarseCoarse5d), _CoarseCoarseMrhs(CoarseCoarseMrhs), _nrhs(nrhs) {}
   using LinearFunction<CoarseField>::operator();
   virtual void operator()(const CoarseField &in, CoarseField &out) {
+    GRID_TRACE("MGCoarseVcycle");
     CoarseField vec1(in.Grid());
     CoarseField vec2(in.Grid());
-    out = in;
-    _CoarseOp.Op(out,vec1);  sub(vec1,in,vec1);
-
-    // restrict, through the mixed blockProject: D+1 coarse in, D+1 cc out
     CoarseCoarseField CCsrc(_CoarseCoarseMrhs);
     CoarseCoarseField CCsol(_CoarseCoarseMrhs);
-    _Projector.blockProject(vec1,CCsrc);
-
-    _CoarseCoarseSolve(CCsrc,CCsol);
-
-    _Projector.blockPromote(vec1,CCsol);
-    add(out,out,vec1);
-
-    _CoarseOp.Op(out,vec1);  sub(vec1,in,vec1);
-    _CoarseSmoother(vec1,vec2);
-    add(out,out,vec2);
+    // The cycle starts from x0 = in, one unit-step Richardson iteration ahead of
+    // a zero start; measured, it pays for its apply by cutting Couter steps 22%.
+    // x0 is never materialised in out: the operator is applied to in directly and
+    // the copy folds into the add that lands the coarse-coarse correction.
+    {
+      GRID_TRACE("MGCoarseResidual");
+      _CoarseOp.Op(in,vec1);  sub(vec1,in,vec1);
+    }
+    // restrict, through the mixed blockProject: D+1 coarse in, D+1 cc out
+    {
+      GRID_TRACE("MGCCProject");
+      _Projector.blockProject(vec1,CCsrc);
+    }
+    {
+      GRID_TRACE("MGCCSolve");
+      _CoarseCoarseSolve(CCsrc,CCsol);
+    }
+    {
+      GRID_TRACE("MGCCPromote");
+      _Projector.blockPromote(vec1,CCsol);
+      add(out,in,vec1);
+    }
+    {
+      GRID_TRACE("MGCoarseResidual2");
+      _CoarseOp.Op(out,vec1);  sub(vec1,in,vec1);
+    }
+    {
+      GRID_TRACE("MGCoarseSmooth");
+      _CoarseSmoother(vec1,vec2);
+      add(out,out,vec2);
+    }
   }
 };
 
@@ -211,37 +328,47 @@ public:
   GridBase *_CoarseGrid, *_CoarseGridMrhs;
   std::function<void(int)> SetSloppy = [](int){};
   int SloppyComms = 0;                 // value passed to SetSloppy on entry
+
   MrhsTwoLevelMG(LinearOperatorBase<FineField> &FineOp, FineSmoother &Post,
                  Projector_t &Projector, LinearFunction<CoarseVector> &CoarseSolve,
                  GridBase *CoarseGrid, GridBase *CoarseGridMrhs)
     : _FineOperator(FineOp),_PostSmoother(Post),_Projector(Projector),_CoarseSolve(CoarseSolve),
       _CoarseGrid(CoarseGrid),_CoarseGridMrhs(CoarseGridMrhs){}
-  virtual void operator()(std::vector<FineField> &in, std::vector<FineField> &out){
+
+  virtual void operator()(std::vector<FineField> &in, std::vector<FineField> &out)
+  {
     GRID_TRACE("MGVcycle");
     SetSloppy(SloppyComms);
     int nrhs=in.size(); GridBase *fgrid=in[0].Grid();
     std::vector<FineField> vec1(nrhs,fgrid),vec2(nrhs,fgrid);
-    for(int r=0;r<nrhs;r++) out[r]=in[r];
-    { GRID_TRACE("MGFineResidual");
-      for(int r=0;r<nrhs;r++){ _FineOperator.Op(out[r],vec1[r]); sub(vec1[r],in[r],vec1[r]); }
+    // x0 = in, never materialised in out: the operator is applied to in directly
+    // and the copy folds into the add that lands the coarse correction.
+    {
+      GRID_TRACE("MGFineResidual");
+      for(int r=0;r<nrhs;r++){ _FineOperator.Op(in[r],vec1[r]); sub(vec1[r],in[r],vec1[r]); }
     }
     // fine vector -> D+1 coarse, via the mixed blockProject
     CoarseVector CsrcMrhs(_CoarseGridMrhs), CsolMrhs(_CoarseGridMrhs);
-    { GRID_TRACE("MGProject");
+    {
+      GRID_TRACE("MGProject");
       _Projector.blockProject(vec1,CsrcMrhs);
     }
     CsolMrhs=Zero();
-    { GRID_TRACE("MGCoarseSolve");
+    {
+      GRID_TRACE("MGCoarseSolve");
       _CoarseSolve(CsrcMrhs,CsolMrhs);
     }
-    { GRID_TRACE("MGPromote");
+    {
+      GRID_TRACE("MGPromote");
       _Projector.blockPromote(vec1,CsolMrhs);
-      for(int r=0;r<nrhs;r++) add(out[r],out[r],vec1[r]);
+      for(int r=0;r<nrhs;r++) add(out[r],in[r],vec1[r]);
     }
-    { GRID_TRACE("MGFineResidual2");
+    {
+      GRID_TRACE("MGFineResidual2");
       for(int r=0;r<nrhs;r++){ _FineOperator.Op(out[r],vec1[r]); sub(vec1[r],in[r],vec1[r]); }
     }
-    { GRID_TRACE("MGPostSmooth");
+    {
+      GRID_TRACE("MGPostSmooth");
       for(int r=0;r<nrhs;r++){
         _PostSmoother(vec1[r],vec2[r]); add(out[r],out[r],vec2[r]);
       }
@@ -271,15 +398,19 @@ public:
   precisionChangeWorkspace    _ws_f2d;   // out fp64, in fp32
   GridBase                   *_gridF;
   std::vector<FieldF>         _in_f, _out_f;
+
   MrhsMixedPrecPreconditioner(MrhsPreconditioner<FieldF> &Inner, GridBase *gridD, GridBase *gridF, int nrhs)
     : _Inner(Inner), _ws_d2f(gridF,gridD), _ws_f2d(gridD,gridF), _gridF(gridF) {}
+
   void Scratch(int nrhs){
     if ( (int)_in_f.size() == nrhs ) return;
     _in_f.clear();  _in_f.reserve(nrhs);
     _out_f.clear(); _out_f.reserve(nrhs);
     for(int r=0;r<nrhs;r++){ _in_f.emplace_back(_gridF); _out_f.emplace_back(_gridF); }
   }
-  virtual void operator()(std::vector<FieldD> &in, std::vector<FieldD> &out){
+
+  virtual void operator()(std::vector<FieldD> &in, std::vector<FieldD> &out)
+  {
     GRID_TRACE("MGPrecisionSeam");
     int nrhs=in.size(); Scratch(nrhs);
     for(int r=0;r<nrhs;r++) precisionChange(_in_f[r],in[r],_ws_d2f);
@@ -287,7 +418,8 @@ public:
     for(int r=0;r<nrhs;r++) precisionChange(out[r],_out_f[r],_ws_f2d);
   }
   // The start and the timers are the inner preconditioner's
-  virtual void Vstart(std::vector<FieldD> &x, std::vector<FieldD> &src){
+  virtual void Vstart(std::vector<FieldD> &x, std::vector<FieldD> &src)
+  {
     GRID_TRACE("MGPrecisionSeamVstart");
     int nrhs=src.size(); Scratch(nrhs);
     for(int r=0;r<nrhs;r++) precisionChange(_in_f[r],src[r],_ws_d2f);

Grid/algorithms/multigrid/MultiGrid.h

@@ -30,12 +30,12 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #include <Grid/algorithms/multigrid/Aggregates.h>
 #include <Grid/algorithms/multigrid/Geometry.h>
 #include <Grid/algorithms/multigrid/CoarsenedMatrix.h>
-#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHSV2.h>
-// DEPRECATED: the V1 coarse operators and the Aggregation-based single-RHS
-// ADEF-2.  Nothing in the library uses them; the PVdagM and HDCG chains are
-// on V2 (PVdagMMultiGrid.h, HDCGMultiGrid.h).  Kept so the pre-2026 drivers
-// in tests/debug and examples still build; removing this block is the
-// deletion gate for them.
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h>
+// DEPRECATED: the superseded coarse operators and the Aggregation-based
+// single-RHS ADEF-2.  Nothing in the library uses them; the PVdagM and HDCG
+// chains are on MultiGeneralCoarsenedOperator (PVdagMMultiGrid.h,
+// HDCGMultiGrid.h).  Kept so the pre-2026 drivers in tests/debug and
+// examples still build; removing this block is the deletion gate for them.
 #include <Grid/algorithms/multigrid/deprecated/GeneralCoarsenedMatrix.h>
 #include <Grid/algorithms/multigrid/deprecated/GeneralCoarsenedMatrixMultiRHS.h>
 #include <Grid/algorithms/multigrid/deprecated/TwoLevelADEF2.h>

Grid/algorithms/multigrid/PVdagMMultiGrid.h

@@ -212,14 +212,14 @@ template<class Fobj,class CComplex,int nbasis>
 class PVdagMMultiGridCoarsening {
 public:
   typedef Lattice<Fobj>                                                 FineField;
-  typedef MultiGeneralCoarsenedOperatorV2<Fobj,CComplex,nbasis>         CoarseOperator;
+  typedef MultiGeneralCoarsenedOperator<Fobj,CComplex,nbasis>         CoarseOperator;
   typedef typename CoarseOperator::CoarseVector                         CoarseVector;
   typedef typename CoarseVector::vector_object                          CoarseSiteObj;
   // Every coarse level carries the same site type iVector<CComplex,nbasis>:
   // the coefficient scalar does not deepen with the level (the operator keeps
   // its own deeper scratch type for the block inner products).  Levels are
   // told apart by their grids, not their C++ types.
-  typedef MultiGeneralCoarsenedOperatorV2<CoarseSiteObj,CComplex,nbasis> CoarseCoarseOperator;
+  typedef MultiGeneralCoarsenedOperator<CoarseSiteObj,CComplex,nbasis> CoarseCoarseOperator;
   typedef typename CoarseCoarseOperator::CoarseVector                   CoarseCoarseVector;
   typedef DenseCoarseMatrix<CComplex,nbasis>                            DenseBottom;
   // The fp32 fine field, derived from the fp64 one: the fp32 fine level of
@@ -359,7 +359,7 @@ public:
     rawPsi.reserve(nbasis);
     for(int k=0;k<nbasis;k++) rawPsi.push_back(psi[k]);
 
-    // L2: the V2 L1 operator is natively multiRHS at the batch grid
+    // L2: the L1 operator is natively multiRHS at the batch grid
     CoarseOpL2.SetGrid(Grids.CoarseCoarseBatch);
     NonHermitianLinearOperator<CoarseOperator,CoarseVector> LinOpCoarse(CoarseOpPV);
     std::cout << GridLogMessage << "PVdagMMultiGridCoarsening: L2 CoarsenOperator, batch "

Grid/algorithms/multigrid/PVdagMOperators.h

@@ -36,17 +36,54 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////////
 template<class Matrix,class Field>
 class PVdagMLinearOperator : public LinearOperatorBase<Field> {
-  Matrix &_Mat; Matrix &_PV;
+  Matrix &_Mat;
+  Matrix &_PV;
 public:
-  PVdagMLinearOperator(Matrix &Mat,Matrix &PV): _Mat(Mat),_PV(PV) {};
-  void SloppyComms(int sloppy) { _Mat.SloppyComms(sloppy); _PV.SloppyComms(sloppy); }
-  void OpDiag (const Field &in, Field &out) { GRID_ASSERT(0); }
-  void OpDir  (const Field &in, Field &out,int dir,int disp) { GRID_ASSERT(0); }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){ GRID_ASSERT(0); };
-  void Op     (const Field &in, Field &out){ Field tmp(in.Grid()); _Mat.M(in,tmp); _PV.Mdag(tmp,out); }
-  void AdjOp  (const Field &in, Field &out){ Field tmp(in.Grid()); _PV.M(in,tmp); _Mat.Mdag(tmp,out); }
-  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ HermOp(in,out); ComplexD d=innerProduct(in,out); n1=real(d); n2=norm2(out); }
-  void HermOp(const Field &in, Field &out){ Field tmp(in.Grid()); Op(in,tmp); AdjOp(tmp,out); }
+  PVdagMLinearOperator(Matrix &Mat,Matrix &PV):
+    _Mat(Mat),_PV(PV)
+  {};
+  void SloppyComms(int sloppy)
+  {
+    _Mat.SloppyComms(sloppy);
+    _PV.SloppyComms(sloppy);
+  }
+  void OpDiag (const Field &in, Field &out)
+  {
+    GRID_ASSERT(0);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp)
+  {
+    GRID_ASSERT(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out)
+  {
+    GRID_ASSERT(0);
+  };
+  void Op     (const Field &in, Field &out)
+  {
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+  }
+  void AdjOp  (const Field &in, Field &out)
+  {
+    Field tmp(in.Grid());
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)
+  {
+    HermOp(in,out);
+    ComplexD d=innerProduct(in,out);
+    n1=real(d);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out)
+  {
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
 };
 
 //////////////////////////////////////////////////////////////////////
@@ -55,17 +92,46 @@ public:
 //////////////////////////////////////////////////////////////////////
 template<class Matrix,class Field>
 class ShiftedPVdagMLinearOperator : public LinearOperatorBase<Field> {
-  Matrix &_Mat; Matrix &_PV;
+  Matrix &_Mat;
+  Matrix &_PV;
 public:
   RealD shift;
   ShiftedPVdagMLinearOperator(RealD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
-  void OpDiag (const Field &in, Field &out) { GRID_ASSERT(0); }
-  void OpDir  (const Field &in, Field &out,int dir,int disp) { GRID_ASSERT(0); }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){ GRID_ASSERT(0); };
-  void Op     (const Field &in, Field &out){ Field tmp(in.Grid()); _Mat.M(in,tmp); _PV.Mdag(tmp,out); out = out + shift*in; }
-  void AdjOp  (const Field &in, Field &out){ Field tmp(in.Grid()); _PV.M(in,tmp); _Mat.Mdag(tmp,out); out = out + shift*in; }
-  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ GRID_ASSERT(0); }
-  void HermOp(const Field &in, Field &out){ Field tmp(in.Grid()); Op(in,tmp); AdjOp(tmp,out); }
+  void OpDiag (const Field &in, Field &out)
+  {
+    GRID_ASSERT(0);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp)
+  {
+    GRID_ASSERT(0);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out)
+  {
+    GRID_ASSERT(0);
+  };
+  void Op     (const Field &in, Field &out)
+  {
+    Field tmp(in.Grid());
+    _Mat.M(in,tmp);
+    _PV.Mdag(tmp,out);
+    out = out + shift*in;
+  }
+  void AdjOp  (const Field &in, Field &out)
+  {
+    Field tmp(in.Grid());
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
+    out = out + shift*in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)
+  {
+    GRID_ASSERT(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
 };
 
 //////////////////////////////////////////////////////////////////////
@@ -73,16 +139,41 @@ public:
 //////////////////////////////////////////////////////////////////////
 template<class Field>
 class ShiftedLinearOperator : public LinearOperatorBase<Field> {
-  LinearOperatorBase<Field> &_Op; RealD shift;
+  LinearOperatorBase<Field> &_Op;
+  RealD shift;
 public:
   ShiftedLinearOperator(RealD _shift, LinearOperatorBase<Field> &Op) : _Op(Op), shift(_shift) {}
-  void OpDiag  (const Field &in, Field &out) { GRID_ASSERT(0); }
-  void OpDir   (const Field &in, Field &out,int dir,int disp) { GRID_ASSERT(0); }
-  void OpDirAll (const Field &in, std::vector<Field> &out) { GRID_ASSERT(0); }
-  void Op      (const Field &in, Field &out) { _Op.Op(in,out);    out = out + shift*in; }
-  void AdjOp   (const Field &in, Field &out) { _Op.AdjOp(in,out); out = out + shift*in; }
-  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ GRID_ASSERT(0); }
-  void HermOp  (const Field &in, Field &out) { Field tmp(in.Grid()); Op(in,tmp); AdjOp(tmp,out); }
+  void OpDiag  (const Field &in, Field &out)
+  {
+    GRID_ASSERT(0);
+  }
+  void OpDir   (const Field &in, Field &out,int dir,int disp)
+  {
+    GRID_ASSERT(0);
+  }
+  void OpDirAll (const Field &in, std::vector<Field> &out)
+  {
+    GRID_ASSERT(0);
+  }
+  void Op      (const Field &in, Field &out)
+  {
+    _Op.Op(in,out);
+    out = out + shift*in;
+  }
+  void AdjOp   (const Field &in, Field &out) {
+    _Op.AdjOp(in,out);
+    out = out + shift*in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)
+  {
+    GRID_ASSERT(0);
+  }
+  void HermOp  (const Field &in, Field &out)
+  {
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
+  }
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/multigrid/deprecated/GeneralCoarsenedMatrix.h

@@ -38,10 +38,10 @@ 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 > > >  {
+class DeprecatedGeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
 public:
 
-  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef DeprecatedGeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
   typedef iVector<CComplex,nbasis >           siteVector;
   typedef iMatrix<CComplex,nbasis >           siteMatrix;
   typedef Lattice<iScalar<CComplex> >         CoarseComplexField;
@@ -103,7 +103,7 @@ public:
   void ProjectNearestNeighbour(RealD shift, GeneralCoarseOp &CopyMe)
   {
     int nfound=0;
-    std::cout << GridLogMessage <<"GeneralCoarsenedMatrix::ProjectNearestNeighbour "<< CopyMe._A[0].Grid()<<std::endl;
+    std::cout << GridLogMessage <<"DeprecatedGeneralCoarsenedMatrix::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
@@ -120,7 +120,7 @@ public:
   }
   */
   
-  GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid,int _herm=1)
+  DeprecatedGeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid,int _herm=1)
     : geom(_geom),
       _FineGrid(FineGrid),
       _CoarseGrid(CoarseGrid),

Grid/algorithms/multigrid/deprecated/GeneralCoarsenedMatrixMultiRHS.h

@@ -34,11 +34,11 @@ NAMESPACE_BEGIN(Grid);
 // 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 > > >  {
+class DeprecatedMultiGeneralCoarsenedMatrix : 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 DeprecatedGeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
 
   typedef iVector<CComplex,nbasis >           siteVector;
   typedef iMatrix<CComplex,nbasis >           siteMatrix;
@@ -78,13 +78,33 @@ public:
   GridCartesian * CoarseGrid(void)     { return _CoarseGridMulti; };   // this is all the linalg routines need to know
 
   //////////////////////////////////////////////////////////////////////////
-  // Bilingual accessors, matching GeneralCoarsenedMatrix. Grid() here is the
+  // Accessors shared with DeprecatedGeneralCoarsenedMatrix. Grid() here is the
   // D+1 multiRHS grid and this class never holds the D dimensional one, so
   // ExtractMatrix writes into whatever grid the caller's lattice is on.
   //////////////////////////////////////////////////////////////////////////
   NonLocalStencilGeometry & Geometry(void)      { return geom_srhs; };
   void ExtractMatrix(int p,CoarseMatrix &A)     { BLAStoGrid(A,BLAS_A[p]); };
 
+  // One stencil point in or out, matching MultiGeneralCoarsenedOperator, so a
+  // caller can read or write a point without knowing which of the two layouts
+  // (per-point here, site-major there) holds it.
+  void MatrixPointIn(int p,const deviceVector<calcMatrix> &src)
+  {
+    int64_t sites = BLAS_A[p].size();
+    GRID_ASSERT((int64_t)src.size() == sites);
+    calcMatrix       *dst = &BLAS_A[p][0];
+    const calcMatrix *sp  = &src[0];
+    accelerator_for(ss,sites,1,{ dst[ss] = sp[ss]; });
+  }
+  void MatrixPointOut(int p,deviceVector<calcMatrix> &dst)
+  {
+    int64_t sites = BLAS_A[p].size();
+    dst.resize(sites);
+    const calcMatrix *sp = &BLAS_A[p][0];
+    calcMatrix       *dp = &dst[0];
+    accelerator_for(ss,sites,1,{ dp[ss] = sp[ss]; });
+  }
+
   // I/O on the operator matrices, via the BLAS layout array. The parameter is
   // a vector over the geometry points; the body indexes A[p].
   void SetMatrix (int p,std::vector<CoarseMatrix> & A)
@@ -121,7 +141,7 @@ public:
   }
   */
   
-  MultiGeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGridMulti) :
+  DeprecatedMultiGeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGridMulti) :
     _CoarseGridMulti(CoarseGridMulti),
     geom_srhs(_geom),
     geom(_CoarseGridMulti,_geom.hops,_geom.skip+1),

Grid/algorithms/multigrid/deprecated/TwoLevelADEF2.h

@@ -27,15 +27,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     /*  END LEGAL */
 #pragma once
 
-// DEPRECATED with the V1 coarse operators: the single-RHS ADEF-2 on an
+// DEPRECATED with the other classes in deprecated/: the single-RHS ADEF-2 on an
 // Aggregation (ProjectToSubspace/PromoteFromSubspace) and a D-dimensional
 // coarse operator.  The mrhs solver in AdefMrhs.h covers one right-hand
-// side through its LinearFunction interface, on the V2 coarse operator.
+// side through its LinearFunction interface, on MultiGeneralCoarsenedOperator.
 
 NAMESPACE_BEGIN(Grid);
 
 template<class Field, class CoarseField, class Aggregation>
-class TwoLevelADEF2 : public TwoLevelCG<Field>
+class DeprecatedTwoLevelADEF2 : public TwoLevelCG<Field>
 {
  public:
   ///////////////////////////////////////////////////////////////////////////////////
@@ -50,7 +50,7 @@ class TwoLevelADEF2 : public TwoLevelCG<Field>
   ///////////////////////////////////////////////////////////////////////////////////
   
   // more most opertor functions
-  TwoLevelADEF2(RealD tol,
+  DeprecatedTwoLevelADEF2(RealD tol,
 		Integer maxit,
 		LinearOperatorBase<Field>    &FineLinop,
 		LinearFunction<Field>        &Smoother,

Grid/communicator/Communicator_base.h

@@ -133,6 +133,7 @@ public:
 
   template<class obj> void GlobalSumP2P(obj &o)
   {
+    GRID_TRACE("GlobalSumP2P");
     std::vector<obj> column;
     obj accum = o;
     int source,dest;

Grid/lattice/Lattice_reduction.h

@@ -898,7 +898,8 @@ template<int B,class vobj>
 void axpyMultiChunk(Lattice<vobj> &z,const ComplexD *b,
 		    const std::vector<const Lattice<vobj>*> &x,int m,
 		    int do_norm,
-		    decltype(innerProduct(vobj(),vobj())) *inner_tmp_v)
+                    decltype(innerProduct(vobj(),vobj())) *inner_tmp_v,
+                    const Lattice<vobj> *from=nullptr)
 {
   typedef decltype(z.View(AcceleratorRead)) View;
   GRID_ASSERT(m>=1 && m<=B);
@@ -917,13 +918,29 @@ void axpyMultiChunk(Lattice<vobj> &z,const ComplexD *b,
   }
   for(int j=m;j<B;j++){ pack.set(j,h_v[0]); pack.b[j] = ComplexD(0.0); }
 
-  autoView(z_v,z,AcceleratorWrite);
-  accelerator_for(ss,sites,nsimd,{
-      auto acc = coalescedRead(z_v[ss]);
-      for(int j=0;j<B;j++) if ( j<m ) acc = acc + pack.b[j]*coalescedRead(pack.v(j)[ss]);
-      coalescedWrite(z_v[ss],acc);
-      if ( do_norm ) coalescedWrite(inner_tmp_v[ss],innerProduct(acc,acc));
-  });
+  if ( from == nullptr ) {
+    autoView(z_v,z,AcceleratorWrite);
+    accelerator_for(ss,sites,nsimd,{
+        auto acc = coalescedRead(z_v[ss]);
+        for(int j=0;j<B;j++) if ( j<m ) acc = acc + pack.b[j]*coalescedRead(pack.v(j)[ss]);
+        coalescedWrite(z_v[ss],acc);
+        if ( do_norm ) coalescedWrite(inner_tmp_v[ss],innerProduct(acc,acc));
+    });
+  } else {
+    // z = from + sum_j b[j] x[j].  The accumulator is seeded from another field,
+    // so z is written without being read and the caller needs no copy into it.
+    conformable(*from,z);
+    GRID_ASSERT(from!=&z);
+    const Lattice<vobj> &y = *from;   // named: autoView is a macro, *p.View() misparses
+    autoView(y_v,y,AcceleratorRead);
+    autoView(z_v,z,AcceleratorWriteDiscard);
+    accelerator_for(ss,sites,nsimd,{
+        auto acc = coalescedRead(y_v[ss]);
+        for(int j=0;j<B;j++) if ( j<m ) acc = acc + pack.b[j]*coalescedRead(pack.v(j)[ss]);
+        coalescedWrite(z_v[ss],acc);
+        if ( do_norm ) coalescedWrite(inner_tmp_v[ss],innerProduct(acc,acc));
+    });
+  }
   for(int j=0;j<m;j++) h_v[j].ViewClose();
 }
 
@@ -931,7 +948,8 @@ void axpyMultiChunk(Lattice<vobj> &z,const ComplexD *b,
 // (last) pass.  Chunks of 16 for windows longer than 16.
 template<class vobj>
 RealD axpyMultiNormImpl(Lattice<vobj> &z,const std::vector<ComplexD> &b,
-			const std::vector<const Lattice<vobj>*> &x,int do_norm)
+                        const std::vector<const Lattice<vobj>*> &x,int do_norm,
+                        const Lattice<vobj> *from=nullptr)
 {
   typedef decltype(innerProduct(vobj(),vobj())) inner_t;
   int m = x.size();
@@ -943,6 +961,7 @@ RealD axpyMultiNormImpl(Lattice<vobj> &z,const std::vector<ComplexD> &b,
   inner_t *inner_tmp_v = &inner_tmp[0];
 
   if ( m==0 ) {
+    if ( from ) z = *from;
     if ( do_norm ) return norm2(z);
     return 0.0;
   }
@@ -951,10 +970,12 @@ RealD axpyMultiNormImpl(Lattice<vobj> &z,const std::vector<ComplexD> &b,
     int last = (j0+mm>=m);
     std::vector<const Lattice<vobj>*> sub(x.begin()+j0,x.begin()+j0+mm);
     int dn = do_norm && last;
-    if      ( mm<=2 ) axpyMultiChunk<2> (z,&b[j0],sub,mm,dn,inner_tmp_v);
-    else if ( mm<=4 ) axpyMultiChunk<4> (z,&b[j0],sub,mm,dn,inner_tmp_v);
-    else if ( mm<=8 ) axpyMultiChunk<8> (z,&b[j0],sub,mm,dn,inner_tmp_v);
-    else              axpyMultiChunk<16>(z,&b[j0],sub,mm,dn,inner_tmp_v);
+    // Only the first chunk is seeded from `from`; later chunks accumulate into z.
+    const Lattice<vobj> *f = (j0==0) ? from : nullptr;
+    if      ( mm<=2 ) axpyMultiChunk<2> (z,&b[j0],sub,mm,dn,inner_tmp_v,f);
+    else if ( mm<=4 ) axpyMultiChunk<4> (z,&b[j0],sub,mm,dn,inner_tmp_v,f);
+    else if ( mm<=8 ) axpyMultiChunk<8> (z,&b[j0],sub,mm,dn,inner_tmp_v,f);
+    else              axpyMultiChunk<16>(z,&b[j0],sub,mm,dn,inner_tmp_v,f);
   }
 
   RealD nrm = 0.0;
@@ -969,6 +990,15 @@ void axpyMulti(Lattice<vobj> &z,const std::vector<ComplexD> &b,const std::vector
 {
   axpyMultiNormImpl(z,b,x,0);
 }
+// z = y + sum_j b[j] x[j].  Same single pass as axpyMulti, but the accumulator
+// starts from y, so a caller whose z would otherwise have to be primed with a
+// copy of y does not make that copy.
+template<class vobj>
+void axpyMultiFrom(Lattice<vobj> &z,const Lattice<vobj> &y,
+                   const std::vector<ComplexD> &b,const std::vector<const Lattice<vobj>*> &x)
+{
+  axpyMultiNormImpl(z,b,x,0,&y);
+}
 template<class vobj>
 RealD axpyMultiNorm(Lattice<vobj> &z,const std::vector<ComplexD> &b,const std::vector<const Lattice<vobj>*> &x)
 {

Grid/lattice/Lattice_transfer.h

@@ -800,6 +800,39 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   autoView(from_v,From,AcceleratorRead);
   autoView(to_v,To,AcceleratorWrite);
 
+  if constexpr ( vobj::Nsimd() == 1 ) {
+
+    // Unvectorised: there is one lane, so the lane indices are identically zero
+    // and the internal word index is free to carry the thread.  Threading over
+    // (site,word) with the word fastest makes consecutive threads read and write
+    // consecutive elements; one thread per site instead leaves them `words`
+    // apart, which is what made this uncoalesced.  The site count and the word
+    // count are folded into one range so the large product lands on the only
+    // block dimension that can hold it.
+    accelerator_for(ss,nsite*words,1,{
+
+        int      w   = ss % words;
+        uint64_t idx = ss / words;
+
+        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];
+        }
+        // rdimensions == ldimensions when unvectorised and the region is local,
+        // so the oSite sums need no modulo and there is no lane term.
+        int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*from_coor[d];
+        int to_oidx   = 0; for(int d=0;d<nd;d++) to_oidx  +=t_ostride[d]*to_coor[d];
+
+        const vector_type* from = (const vector_type *)&from_v[from_oidx];
+        vector_type* to = (vector_type *)&to_v[to_oidx];
+
+        to[w] = from[w];
+    });
+
+  } else {
+
   accelerator_for(idx,nsite,1,{
 
       Coordinate from_coor, to_coor, base;
@@ -815,13 +848,15 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
 
       const vector_type* from = (const vector_type *)&from_v[from_oidx];
       vector_type* to = (vector_type *)&to_v[to_oidx];
-      
+
       scalar_type stmp;
       for(int w=0;w<words;w++){
 	stmp = getlane(from[w], from_lane);
 	putlane(to[w], stmp, to_lane);
       }
   });
+
+  }
 }
 
 template<class vobj>

Grid/lattice/PaddedCell.h

@@ -94,7 +94,7 @@ template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
   // 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,{
+  accelerator_forNB(ss, face_ovol/simd[dim],Nsimd,{
 
     // scalar layout won't coalesce
 #ifdef GRID_SIMT
@@ -180,7 +180,7 @@ template<class vobj> inline void GatherSlice(deviceVector<vobj> &buf,
   //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,{
+  accelerator_forNB(ss, face_ovol/simd[dim],Nsimd,{
 
     // scalar layout won't coalesce
 #ifdef GRID_SIMT
@@ -303,10 +303,17 @@ public:
   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++){
+    Coordinate processors=unpadded_grid->_processors;
+    int dims = in.Grid()->Nd();
+    // An undecomposed dimension is not padded, so Expand on it would only copy
+    // its input onto the same grid: those are skipped.  The first decomposed
+    // dimension expands from `in` itself, so the chain needs no seed copy.
+    int first=-1;
+    for(int d=0;d<dims;d++) if ( processors[d] > 1 ) { first=d; break; }
+    if ( first < 0 ) return in;  // nothing decomposed: the padded grid IS the input grid
+    Lattice<vobj> tmp = Expand(first,in,cshift);
+    for(int d=first+1;d<dims;d++){
+      if ( processors[d] == 1 ) continue;
       tmp = Expand(d,tmp,cshift); // rvalue && assignment
     }
     return tmp;
@@ -314,10 +321,18 @@ public:
   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++){
+    Coordinate processors=unpadded_grid->_processors;
+    int dims = in.Grid()->Nd();
+    // As Exchange: undecomposed dimensions are pure copies and are skipped, and
+    // the first decomposed dimension expands from `in` so there is no seed copy.
+    // On the D+1 mrhs coarse grid (nrhs,1,x,y,z,t) this removes three full
+    // coarse-field copies per operator application.
+    int first=-1;
+    for(int d=0;d<dims;d++) if ( processors[d] > 1 ) { first=d; break; }
+    if ( first < 0 ) return in;  // nothing decomposed: the padded grid IS the input grid
+    Lattice<vobj> tmp = ExpandPeriodic(first,in);
+    for(int d=first+1;d<dims;d++){
+      if ( processors[d] == 1 ) continue;
       tmp = ExpandPeriodic(d,tmp); // rvalue && assignment
     }
     return tmp;
@@ -513,12 +528,18 @@ public:
     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;
-
+    }
+    for ( int d=0;d < depth ; d ++ ) {
+      t=usecond();
+      GatherSlice(send_buf,from,ld-depth+d,dimension,plane*buffer_size); plane++;
+      t_gather+= usecond() - t;
+    }
+    accelerator_barrier();
+    for ( int d=0;d < depth ; d ++ ) {
+      int tag = d*1024 + dimension*2+0;
       t=usecond();
 
       if(d==0) fwd_trace = traceStart("PaddedCellFwdMPI");
@@ -533,14 +554,9 @@ public:
 				(void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag);
 #endif
       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();
       if (d==0) bwd_trace = traceStart("PaddedCellBwdMPI");
 #ifdef ACCELERATOR_AWARE_MPI
@@ -603,9 +619,11 @@ public:
     for ( int d=0;d < depth ; d ++ ) {
       ScatterSlice(recv_buf,to,d,dimension,plane*buffer_size); plane++;
     }
+    accelerator_barrier();
     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;

Grid/qcd/smearing/WilsonFlow.h

@@ -52,7 +52,7 @@ protected:
 public:
 
 //Define the action used to evolve the plaquettes
-//(Lüscher: https://arxiv.org/pdf/1006.4518 eq. 1.4)
+//(Luscher: https://arxiv.org/pdf/1006.4518 eq. 1.4)
 //V'(t) = -g^2 * ( d/dVt S[Vt](g) ) * Vt
 //      = -g^2 * ( d/dVt (1/g^2 * sum_p Re tr{ 1 - Vt(p) } ) ) * Vt
 //      = - d/dVt ( sum_p ( Nc - Re tr Vt(p) ) * Vt
@@ -157,7 +157,7 @@ public:
 ////////////////////////////////////////////////////////////////////////////////
 
 //Compute t^2 <E(t)> for time from the plaquette form
-//(Lüscher: https://arxiv.org/pdf/1006.4518 eq. 3.1)
+//(Luscher: https://arxiv.org/pdf/1006.4518 eq. 3.1)
 //E(t) = 2 * sum_p Retr{ 1 - Vt(p) } =
 //     = 2 * sum_p ( Nc - Retr Vt(p) ) =
 //     = 2 * Nc * sum_p ( 1 - Retr Vt(p)/Nc )

Grid/simd/Grid_vector_types.h

@@ -1002,10 +1002,10 @@ accelerator_inline void precisionChange(vRealD    *out,const vRealF    *in,int n
     Optimization::PrecisionChange::StoD(in[m].v,out[n].v,out[n+1].v);
     // Bug in gcc 10.0.1 and gcc 10.1 using fixed-size SVE ACLE data types  CAS-159553-Y1K4C6
     // function call results in compile-time error:
-    // In function ‘void Grid::precisionChange(Grid::vRealD*, Grid::vRealF*, int)’:
+    // In function 'void Grid::precisionChange(Grid::vRealD*, Grid::vRealF*, int)':
     // .../Grid_vector_types.h:961:56: error:
-    // cannot bind non-const lvalue reference of type ‘vecd&’ {aka ‘svfloat64_t&’}
-    // to an rvalue of type ‘vecd’ {aka ‘svfloat64_t’}
+    // cannot bind non-const lvalue reference of type 'vecd&' {aka 'svfloat64_t&'}
+    // to an rvalue of type 'vecd' {aka 'svfloat64_t'}
     // 961 |     Optimization::PrecisionChange::StoD(in[m].v,out[n].v,out[n+1].v);
     //  |                                                 ~~~~~~~^
   }

MPI_benchmark/halo_mpi.cc

@@ -104,6 +104,119 @@ inline double usecond(void) {
   gettimeofday(&tv,NULL);
   return 1.0e6*tv.tv_sec + 1.0*tv.tv_usec;
 }
+/**************************************************************
+ * Point to point cost, per Cartesian direction.
+ *
+ * PaddedCell::Face_exchange issues one MPI_Sendrecv per dimension and waits,
+ * so the quantity that bounds a halo round is the sendrecv time for that
+ * direction, not a one way ping-pong.  Sweeping the size gives the latency
+ * intercept and the bandwidth slope separately.
+ *
+ * Each direction is labelled on- or off-node, which is what decides whether a
+ * halo round rides the intra-node fabric or the network, and therefore which
+ * dimension ordering costs least.  Times are reduced across ranks: the max is
+ * the one that bounds a collective halo round.
+ **************************************************************
+ */
+void PingPong(std::vector<int> cart_geom,bool use_device,int ncall)
+{
+  int Nd=cart_geom.size();
+  std::vector<int> periodic(Nd,1);
+  std::vector<int> coor(Nd);
+  int rank;
+
+  MPI_Comm communicator;
+  MPI_Cart_create(WorldComm,Nd,&cart_geom[0],&periodic[0],0,&communicator);
+  MPI_Comm_rank(communicator,&rank);
+  MPI_Cart_coords(communicator,rank,Nd,&coor[0]);
+
+  // An identifier every rank on a node agrees on, so a neighbour can be
+  // classified by comparing it.
+  int node_id = WorldRank;
+  MPI_Bcast(&node_id,1,MPI_INT,0,WorldShmComm);
+
+  size_t max_bytes = 2*1024*1024;
+  void *xmit, *recv;
+  if ( use_device ) {
+    xmit = acceleratorAllocDevice(max_bytes);
+    recv = acceleratorAllocDevice(max_bytes);
+  } else {
+    xmit = malloc(max_bytes);
+    recv = malloc(max_bytes);
+  }
+
+  if ( !WorldRank ) {
+    printf("= dim  dir       bytes      us(max)      us(min)        MB/s   off-node ranks\n");
+    fflush(stdout);
+  }
+
+  for(int d=0;d<Nd;d++){
+
+    // An undecomposed dimension is a local wrap in PaddedCell and carries no
+    // message at all; a shift here would only time a rank talking to itself.
+    if ( cart_geom[d] == 1 ) continue;
+
+    for(int sign=-1;sign<=1;sign+=2){
+
+      int from,to;
+      MPI_Cart_shift(communicator,d,sign,&from,&to);
+
+      // Fetch the identifier of the neighbour we SEND to, which means sending
+      // ours the other way round the shift.
+      int peer_node=node_id;
+      MPI_Sendrecv(&node_id, 1,MPI_INT,from,rank,
+		   &peer_node,1,MPI_INT,to,  to,
+		   communicator,MPI_STATUS_IGNORE);
+      int offnode = (peer_node != node_id) ? 1 : 0;
+      int offnode_count=0;
+      MPI_Reduce(&offnode,&offnode_count,1,MPI_INT,MPI_SUM,0,communicator);
+
+      for(size_t bytes=8; bytes<=max_bytes; bytes*=8){
+
+	// Latency needs the repeats; bandwidth does not, and large messages
+	// would otherwise dominate the run time.
+	int nc = (bytes<=32768) ? ncall : (ncall/20 > 10 ? ncall/20 : 10);
+
+	// Untimed warm-up: the first exchange on a fresh buffer pays connection
+	// setup and registration, and would otherwise land in the first row.
+	MPI_Sendrecv(xmit,bytes,MPI_CHAR,to,rank,
+		     recv,bytes,MPI_CHAR,from,from,
+		     communicator,MPI_STATUS_IGNORE);
+
+	MPI_Barrier(communicator);
+	double t0=usecond();
+	for(int i=0;i<nc;i++){
+	  MPI_Sendrecv(xmit,bytes,MPI_CHAR,to,rank,
+		       recv,bytes,MPI_CHAR,from,from,
+		       communicator,MPI_STATUS_IGNORE);
+	}
+	double t1=usecond();
+
+	double us = (t1-t0)/(double)nc;
+	double us_max, us_min;
+	MPI_Reduce(&us,&us_max,1,MPI_DOUBLE,MPI_MAX,0,communicator);
+	MPI_Reduce(&us,&us_min,1,MPI_DOUBLE,MPI_MIN,0,communicator);
+
+	if ( !WorldRank ) {
+	  printf("= %3d  %s  %10zu  %11.2f  %11.2f  %10.1f   %d\n",
+		 d, sign>0?"+":"-", bytes, us_max, us_min,
+		 (double)bytes/us_max, offnode_count);
+	  fflush(stdout);
+	}
+      }
+    }
+  }
+
+  if ( use_device ) {
+    acceleratorFreeDevice(xmit);
+    acceleratorFreeDevice(recv);
+  } else {
+    free(xmit);
+    free(recv);
+  }
+  MPI_Comm_free(&communicator);
+}
+
 /**************************************************************
  * Main benchmark routine
  **************************************************************
@@ -303,6 +416,20 @@ int main(int argc, char **argv)
   }
 
   
+  if( !WorldRank ) {
+    printf("=========================================================\n");
+    printf("= Point to point cost per direction, HOST memory         \n");
+    printf("=========================================================\n");fflush(stdout);
+  }
+  PingPong(mpi,false,1000);
+
+  if( !WorldRank ) {
+    printf("=========================================================\n");
+    printf("= Point to point cost per direction, DEVICE memory       \n");
+    printf("=========================================================\n");fflush(stdout);
+  }
+  PingPong(mpi,true,1000);
+
   if( !WorldRank ) {
     printf("=========================================================\n");
     printf("= Benchmarking HOST memory MPI performance               \n");

MPI_benchmark/simple_reproducer_noncache.cc

@@ -0,0 +1,214 @@
+// Minimal reproducer for the CXI uncached-registration defect.
+//
+// Hypothesis: with the libfabric MR cache disabled, a device pointer at a non-zero offset
+// into a hipMalloc'd allocation is registered as the whole allocation and the offset is
+// discarded, so a send transmits from the base and a receive lands at the base.
+//
+// Each test therefore prints, for both the data and its landing place, what a correct
+// implementation must give, what the hypothesis predicts instead, and what was observed.
+//
+//   hipcc -O2 -o simple_reproducer_noncache simple_reproducer_noncache.cc \
+//         -I$MPICH_DIR/include -L$MPICH_DIR/lib -lmpi -L$MPICH_DIR/gtl/lib -lmpi_gtl_hsa
+//
+//   MPICH_GPU_SUPPORT_ENABLED=1 FI_MR_CACHE_MAX_COUNT=0 \
+//     srun -N2 -n2 --ntasks-per-node=1 ./simple_reproducer_noncache
+
+#include <mpi.h>
+#include <hip/hip_runtime.h>
+#include <cstdio>
+#include <cstdint>
+#include <cstdlib>
+#include <vector>
+
+// The message must be large enough to go rendezvous: an eager message is staged through a host
+// bounce buffer with hipMemcpy, which honours the offset, and never registers the user pointer.
+static const int NWORDS = 262144; // 2 MB buffers
+static int       NMSG   = 16384;  // 128 KB messages          (argv[1], in words)
+static int       OFF    = 16384;  // offending offset in words (argv[2]); clearest at >= NMSG,
+                                  // which keeps the intended and predicted windows disjoint
+
+static int rank, size, peer;
+static uint64_t *A, *B;
+static std::vector<uint64_t> h(NWORDS);
+
+#define HIP(cmd) do { hipError_t e=(cmd); if ( e != hipSuccess ) {			\
+      printf("hip error %s at line %d\n",hipGetErrorString(e),__LINE__);		\
+      MPI_Abort(MPI_COMM_WORLD,1); } } while(0)
+
+static uint64_t word(uint64_t tag,int r,int n)
+{
+  return (tag<<60) | ((uint64_t)r<<32) | (uint64_t)n;
+}
+
+static const char *verdict(int observed,int correct,int predicted)
+{
+  if ( observed == correct   ) return "CORRECT";
+  if ( observed == predicted ) return "WRONG, as predicted";
+  return "WRONG, and not as predicted";
+}
+
+static void refresh(void)
+{
+  for(int n=0;n<NWORDS;n++) h[n] = word(0xa,rank,n);
+  HIP(hipMemcpy(A,h.data(),NWORDS*sizeof(uint64_t),hipMemcpyHostToDevice));
+  for(int n=0;n<NWORDS;n++) h[n] = word(0xb,rank,n);
+  HIP(hipMemcpy(B,h.data(),NWORDS*sizeof(uint64_t),hipMemcpyHostToDevice));
+  MPI_Barrier(MPI_COMM_WORLD);
+}
+
+static void exchange(int sendoff,int recvoff)
+{
+  MPI_Request req[2];
+  MPI_Irecv(B+recvoff,NMSG*sizeof(uint64_t),MPI_BYTE,peer,0,MPI_COMM_WORLD,&req[0]);
+  MPI_Isend(A+sendoff,NMSG*sizeof(uint64_t),MPI_BYTE,peer,0,MPI_COMM_WORLD,&req[1]);
+  MPI_Waitall(2,req,MPI_STATUSES_IGNORE);
+}
+
+// Report the state of one NMSG-word window of B: how much of it was overwritten, and by what.
+static void window(const char *label,int at)
+{
+  int mod=0, m0=-1;
+  for(int i=0;i<NMSG && at+i<NWORDS;i++) {
+    if ( h[at+i] != word(0xb,rank,at+i) ) { if ( m0 < 0 ) m0 = at+i; mod++; }
+  }
+  if ( mod == 0 ) {
+    printf("rank %d    %-9s B[%d..%d] : untouched\n",rank,label,at,at+NMSG-1);
+    return;
+  }
+  uint64_t w = h[m0];
+  printf("rank %d    %-9s B[%d..%d] : %d of %d words overwritten, B[%d] holds A[%d] from rank %d\n",
+	 rank,label,at,at+NMSG-1,mod,NMSG,m0,(int)(w & 0xffffffff),(int)((w>>32) & 0xfffffff));
+}
+
+// Name a word: still the receive pattern, or a word of the neighbour's send buffer.
+static const char *describe(uint64_t w,int idx)
+{
+  static char s[64];
+  if      ( w == word(0xb,rank,idx) ) snprintf(s,sizeof(s),"untouched");
+  else if ( (w>>60) == 0xa )          snprintf(s,sizeof(s),"A[%d] of rank %d",
+					       (int)(w & 0xffffffff),(int)((w>>32) & 0xfffffff));
+  else                                snprintf(s,sizeof(s),"unrecognised");
+  return s;
+}
+
+static void report(const char *name,int sendoff,int recvoff)
+{
+  HIP(hipMemcpy(h.data(),B,NWORDS*sizeof(uint64_t),hipMemcpyDeviceToHost));
+
+  int first=-1, last=-1;
+  for(int n=0;n<NWORDS;n++) {
+    if ( h[n] != word(0xb,rank,n) ) { if ( first < 0 ) first = n; last = n; }
+  }
+
+  printf("rank %d  %s : send &A[%d] -> recv &B[%d]\n",rank,name,sendoff,recvoff);
+  if ( first < 0 ) { printf("rank %d    nothing was written\n",rank); return; }
+
+  uint64_t w   = h[first];
+  int      src = (int)(w & 0xffffffff);
+  int      sr  = (int)((w>>32) & 0xfffffff);
+
+  // The hypothesis predicts both offsets are discarded: the bytes at A[0], landing at B[0].
+  printf("rank %d    data      : want A[%d]  predict A[0]  got A[%d] from rank %d   %s\n",
+	 rank,sendoff,src,sr,verdict(src,sendoff,0));
+  printf("rank %d    place     : want B[%d]  predict B[0]  got B[%d]                %s\n",
+	 rank,recvoff,first,verdict(first,recvoff,0));
+
+  if ( src == sendoff && first == recvoff ) return;
+
+  // Both candidate landing sites in our receive buffer, and both candidate source words in the
+  // neighbour's send buffer, which the fill pattern fixes by construction.
+  printf("rank %d    Boff %d\n",rank,recvoff);
+  printf("rank %d      B[0]      = 0x%016llx  %s\n",
+	 rank,(unsigned long long)h[0],describe(h[0],0));
+  if ( recvoff != 0 )
+    printf("rank %d      B[%d]\t= 0x%016llx  %s\n",
+	   rank,recvoff,(unsigned long long)h[recvoff],describe(h[recvoff],recvoff));
+  printf("rank %d    Neighbour rank %d send buffer holds\n",rank,peer);
+  printf("rank %d      A[0]\t= 0x%016llx\n",
+	 rank,(unsigned long long)word(0xa,peer,0));
+  if ( sendoff != 0 )
+    printf("rank %d      A[%d]\t= 0x%016llx\n",
+	   rank,sendoff,(unsigned long long)word(0xa,peer,sendoff));
+
+  window("intended",recvoff);
+  if ( recvoff != 0 ) window("predicted",0);
+  if ( first < recvoff || last >= recvoff+NMSG ) {
+    if ( first != 0 || last != NMSG-1 ) {
+      printf("rank %d    stray     : modified words span B[%d..%d], outside both windows\n",
+	     rank,first,last);
+    }
+  }
+
+  int bad=0;
+  for(int i=0;i<NMSG && first+i<NWORDS;i++) {
+    if ( h[first+i] != word(0xa,peer,src+i) ) bad++;
+  }
+  if ( bad ) printf("rank %d    body      : %d of %d words are not a contiguous run from A[%d]\n",
+		    rank,bad,NMSG,src);
+}
+
+// Let each rank print in turn: flush, then a barrier so rank r is on the page before r+1 starts.
+static void ordered(void (*print)(void))
+{
+  for(int r=0;r<size;r++) {
+    if ( r == rank ) { print(); fflush(stdout); }
+    MPI_Barrier(MPI_COMM_WORLD);
+  }
+}
+
+static const char *test_name;
+static int         test_sendoff, test_recvoff;
+static void        print_report(void) { report(test_name,test_sendoff,test_recvoff); }
+static void        print_setup (void)
+{
+  printf("rank %d  A %p  B %p  message %d words (%d bytes)  offset %d words (%d bytes)\n",
+	 rank,(void *)A,(void *)B,NMSG,(int)(NMSG*sizeof(uint64_t)),OFF,(int)(OFF*sizeof(uint64_t)));
+}
+
+int main(int argc,char **argv)
+{
+  MPI_Init(&argc,&argv);
+  MPI_Comm_rank(MPI_COMM_WORLD,&rank);
+  MPI_Comm_size(MPI_COMM_WORLD,&size);
+  if ( size != 2 ) {
+    if ( rank == 0 ) printf("run with two ranks, one per node\n");
+    MPI_Finalize();
+    return 1;
+  }
+  peer = 1-rank;
+
+  if ( argc > 1 ) NMSG = atoi(argv[1]);
+  if ( argc > 2 ) OFF  = atoi(argv[2]);
+  if ( NMSG < 1 || OFF < 0 || OFF+NMSG > NWORDS ) {
+    if ( rank == 0 ) printf("message %d and offset %d words do not fit in %d\n",NMSG,OFF,NWORDS);
+    MPI_Finalize();
+    return 1;
+  }
+
+  int ndev;
+  HIP(hipGetDeviceCount(&ndev));
+  if ( ndev < 1 ) {
+    printf("rank %d sees no GPU\n",rank);
+    MPI_Abort(MPI_COMM_WORLD,1);
+  }
+  HIP(hipSetDevice(rank%ndev));
+  HIP(hipMalloc(&A,NWORDS*sizeof(uint64_t)));
+  HIP(hipMalloc(&B,NWORDS*sizeof(uint64_t)));
+
+  ordered(print_setup);
+
+  const int  offsets[4][2] = { {0,0}, {OFF,0}, {0,OFF}, {OFF,OFF} };
+  const char *names[4]     = { "test 1","test 2","test 3","test 4" };
+
+  for(int t=0;t<4;t++) {
+    refresh();
+    exchange(offsets[t][0],offsets[t][1]);
+    test_name = names[t]; test_sendoff = offsets[t][0]; test_recvoff = offsets[t][1];
+    ordered(print_report);
+  }
+
+  HIP(hipFree(A));
+  HIP(hipFree(B));
+  MPI_Finalize();
+  return 0;
+}

examples/Example_mdagm.cc

@@ -87,7 +87,7 @@ int main (int argc, char ** argv)
   HermOpAdaptor<LatticeFermionD> HermFineOp(MdagMOp);
 
   // ── Coarse geometry ────────────────────────────────────────────────────
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
 
@@ -171,7 +171,7 @@ int main (int argc, char ** argv)
   LatticeFermionD src(FGrid);    random(RNG5, src);
   LatticeFermionD result(FGrid); result = Zero();
 
-  TwoLevelADEF2<LatticeFermionD, CoarseVector, Subspace>
+  DeprecatedTwoLevelADEF2<LatticeFermionD, CoarseVector, Subspace>
     HDCG(1.0e-8, 1000,
          HermFineOp,
          Smoother,

examples/Example_mdagm_cg.cc

@@ -31,7 +31,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 //
 // Operator hierarchy:
 //   Fine:   M†M, acted on via HermOpAdaptor so Op = HermOp = M†M
-//   Coarse: 33-point GeneralCoarsenedMatrix (NextToNearest, 2-hop M†M)
+//   Coarse: 33-point DeprecatedGeneralCoarsenedMatrix (NextToNearest, 2-hop M†M)
 //
 // Setup:
 //   1. Chebyshev filter (T_600 x T_2500) on M†M to build near-null subspace {ψᵢ}
@@ -198,7 +198,7 @@ int main(int argc, char **argv)
   ///////////////////////////////////////////////////////////
   // Coarse operator: 33-point stencil for 2-hop M†M
   ///////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector, vTComplex, nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector, vTComplex, nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNearestStencilGeometry5D geom(Coarse5d);

examples/Example_pvdagm.cc

@@ -364,7 +364,7 @@ void runMG(
 ) {
 
   // typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
-  // typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  // typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   // typedef LittleDiracOperator::CoarseVector CoarseVector;
   ParseEnvironment();
 
@@ -620,7 +620,7 @@ int main (int argc, char ** argv)
 
   // assert(nbasis <= Nevecs);     // need to have enough evecs
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNearestStencilGeometry5D geom(Coarse5d);

examples/Example_pvdagm_3level.cc

@@ -190,8 +190,8 @@ public:
   }
 };
 
-// Lüscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
-// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi† src.
+// Luscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
+// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi^dag src.
 template<class Field>
 class LuscherGuesser : public LinearFunction<Field> {
   const std::vector<Field> &psi;
@@ -342,7 +342,7 @@ void runMG(
   TrivialPrecon<LatticeFermionD>   simple_fine;
 
   //////////////////////////////////////////////////////////////////////
-  // Level 0→1: coarsen PVdagM, build LinOpCoarse
+  // Level 0->1: coarsen PVdagM, build LinOpCoarse
   //////////////////////////////////////////////////////////////////////
   LittleDiracOperator LittleDiracOpPV(geom, FGrid, Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM, AggregatesPD);
@@ -366,7 +366,7 @@ void runMG(
   //////////////////////////////////////////////////////////////////////
   // psi_coarse: coarse projections of pre-GS fine null vectors.
   // These are the Level 1 near-null vectors, promoted from Level 0.
-  // Used as the aggregation basis for Level 1→2 coarsening.
+  // Used as the aggregation basis for Level 1->2 coarsening.
   //////////////////////////////////////////////////////////////////////
   std::vector<CoarseVector> psi_coarse(nbasis, Coarse5d);
   for (int k = 0; k < nbasis; k++)
@@ -396,22 +396,22 @@ void runMG(
     RealD normC      = C.norm();
     RealD normCmCdag = (C - C.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse null matrix ||C||            = " << normC << std::endl;
-    std::cout << GridLogMessage << "Coarse null matrix ||C - C†||/||C|| = " << normCmCdag/normC << std::endl;
+    std::cout << GridLogMessage << "Coarse null matrix ||C - C^dag||/||C|| = " << normCmCdag/normC << std::endl;
     std::cout << GridLogMessage << "Galerkin check ||C||/||W||          = " << normC/normW << std::endl;
   }
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2: set up aggregation using psi_coarse as subspace.
+  // Level 1->2: set up aggregation using psi_coarse as subspace.
   // Block factor 2,2,3,2 (removes odd local sublattice in z given MPI
-  // geometry 3×6×4×4 where z-local at Level 1 is 6).
+  // geometry 3x6x4x4 where z-local at Level 1 is 6).
   // psi_coarse are assigned directly; CoarsenOperator performs
   // block-GS orthogonalisation before building LinOpCoarseCoarse.
   //////////////////////////////////////////////////////////////////////
   // innerProduct(CoarseSiteObj, CoarseSiteObj) returns iScalar<vTComplex>, so CComplex
-  // for the L1→L2 level must be iScalar<vTComplex>, not vTComplex.
+  // for the L1->L2 level must be iScalar<vTComplex>, not vTComplex.
   typedef typename CoarseVector::vector_object                            CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
   typedef typename LittleDiracOperatorL2::CoarseVector                   CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,NB>                       SubspaceL2;
   typedef MGPreconditioner<CoarseSiteObj,vTTComplex,NB>                  L1to2MG;
@@ -429,12 +429,12 @@ void runMG(
   TrivialPrecon<CoarseCoarseVector> simpleCC;
 
   //////////////////////////////////////////////////////////////////////
-  // Lüscher deflation guesser for L3PGCR.
+  // Luscher deflation guesser for L3PGCR.
   // Step 1: project psi_coarse[k] (promoted fine null vectors) to
-  //         CoarseCoarseVector space — these cover the zero-momentum
+  //         CoarseCoarseVector space -- these cover the zero-momentum
   //         component of the near-null space of LinOpCC.
   // Step 2: breed Nextra additional null vectors directly on LinOpCC
-  //         using GCR with random sources — these pick up near-null
+  //         using GCR with random sources -- these pick up near-null
   //         modes at all spatial frequencies not spanned by step 1.
   // Step 3: build C_{st} = <psi_cc[s]|LinOpCC|psi_cc[t]> over the
   //         full augmented basis and invert directly via Eigen LU.
@@ -476,7 +476,7 @@ void runMG(
     RealD normCcc      = Ccc.norm();
     RealD normCccmCdag = (Ccc - Ccc.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc||              = " << normCcc << std::endl;
-    std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc-Ccc†||/||Ccc|| = " << normCccmCdag/normCcc << std::endl;
+    std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc-Ccc^dag||/||Ccc|| = " << normCccmCdag/normCcc << std::endl;
   }
   Eigen::MatrixXcd Ccc_inv = Ccc.inverse();
   LuscherGuesser<CoarseCoarseVector> CCDeflGuesser(psi_cc, Ccc_inv);
@@ -489,7 +489,7 @@ void runMG(
   L3PGCR.Name("CCouter");
 
   //////////////////////////////////////////////////////////////////////
-  // Coarse-level GCR smoother for Level 1→2 V-cycle.
+  // Coarse-level GCR smoother for Level 1->2 V-cycle.
   // Mirrors fine-grid SmootherGCR: shifted operator + fixed step count.
   // coarse_smoother_shift and coarse_smoother_nstep are the tuning knobs.
   //////////////////////////////////////////////////////////////////////
@@ -505,7 +505,7 @@ void runMG(
   CoarseSmootherGCR.Name("Csmoother");
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2 V-cycle preconditioner.
+  // Level 1->2 V-cycle preconditioner.
   //////////////////////////////////////////////////////////////////////
   L1to2MG L1to2Precon(AggregatesL2,
                        LinOpCoarse,
@@ -513,7 +513,7 @@ void runMG(
                        CoarseSmootherGCR,  // post-smoother: 12 GCR steps
                        LinOpCC,
                        L3PGCR,
-                       CCDeflGuesser);     // Lüscher guesser: psi_cc C^{-1} psi_cc†
+                       CCDeflGuesser);     // Luscher guesser: psi_cc C^{-1} psi_cc^dag
 
   //////////////////////////////////////////////////////////////////////
   // Standalone Level 1 two-level solve test.
@@ -548,7 +548,7 @@ void runMG(
   f_src = one;
 
   // Pre-smoother: none (TrivialPrecon); post-smoother: shifted PGCR.
-  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1→2 V-cycle).
+  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1->2 V-cycle).
   TwoLevelMG ThreeLevelPrecon(AggregatesPD,
                                PVdagM,
                                simple_fine,
@@ -592,7 +592,7 @@ int main (int argc, char ** argv)
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  // Level 1 coarse grid: block 2^4 from fine (48×48×48×96 → 24×24×24×48, Ls=1)
+  // Level 1 coarse grid: block 2^4 from fine (48x48x48x96 -> 24x24x24x48, Ls=1)
   Coordinate clatt = lat_size;
   for (int d = 0; d < 4; d++) clatt[d] /= 2;
   std::cout << GridLogMessage << "Level 1 coarse lattice: " << clatt << std::endl;
@@ -600,13 +600,13 @@ int main (int argc, char ** argv)
   GridCartesian *Coarse4d  = SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridCartesian *Coarse5d  = SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
 
-  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24×24×24×48 → 12×12×8×16, Ls=1).
+  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24x24x24x48 -> 12x12x8x16, Ls=1).
   // MPI geometry 3.6.4.4 (288 ranks): fine local {16,8,12,24}.
   // Level 1 local {8,4,6,12}; Level 2 local {4,2,2,4}.
-  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) → must use 3.
+  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) -> must use 3.
   // t blocked by 3: t-Level1-local=12; 12/3=4 divisible by Nsimd=4 (gen-simd-width=64).
-  //   t-block=2 gives t2-local=6, 6 mod 4 ≠ 0, fails Grid SIMD assertion. ✓
-  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4). ✓
+  //   t-block=2 gives t2-local=6, 6 mod 4 != 0, fails Grid SIMD assertion.
+  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4).
   Coordinate clatt2 = clatt;
   clatt2[0] /= 2;
   clatt2[1] /= 2;
@@ -635,7 +635,7 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
   typedef MGPreconditioner<vSpinColourVector,vTComplex,nbasis>        TwoLevelMG;

examples/Example_pvdagm_3level_DenseCoarseMatrix.cc

@@ -2,7 +2,7 @@
 
     Grid physics library, www.github.com/paboyle/Grid
 
-    Source file: ./examples/Example_pvdagm_v2_3level_DenseCoarseMatrix.cc
+    Source file: ./examples/Example_pvdagm_3level_DenseCoarseMatrix.cc
 
     Copyright (C) 2026
 
@@ -27,7 +27,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 /*  END LEGAL */
 
 //
-// PVdagM three level multigrid on the V2 coarse operator.
+// PVdagM three level multigrid on MultiGeneralCoarsenedOperator.
 //
 // STAGES ONE AND TWO: grids, types, subspace, and the L1 and L2 coarsenings.
 // The dense bottom and the solves are not here yet.
@@ -37,12 +37,13 @@ Author: Peter Boyle <pboyle@bnl.gov>
 //  * The coarse space is UNVECTORISED (sComplexD). The fine space stays
 //    vectorised. MultiRHSBlockProject carries the mixed layout.
 //
-//  * One operator, not two. V1 needed GeneralCoarsenedMatrix to coarsen and
-//    MultiGeneralCoarsenedMatrix to apply, bridged by CopyMatrix. V2 does
-//    both, and single versus multiRHS is SetGrid on the same object with the
-//    matrix elements built once.
+//  * One operator, not two. The deprecated path needed
+//    DeprecatedGeneralCoarsenedMatrix to coarsen and
+//    DeprecatedMultiGeneralCoarsenedMatrix to apply, bridged by CopyMatrix.
+//    This operator does both, and single versus multiRHS is SetGrid on the
+//    same object with the matrix elements built once.
 //
-//  * Nrhs is unconstrained. V1 required nrhs % vComplex::Nsimd() == 0 because
+//  * Nrhs is unconstrained. The deprecated path required nrhs % vComplex::Nsimd() == 0 because
 //    its multiRHS grid carried the SIMD in the rhs direction.
 //
 //  * CoarsenOperator takes the subspace vectors, not an Aggregation. It block
@@ -54,7 +55,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 //    ||<psi|psi> - I||_F guard below is what catches that.
 //
 // Env: LATT LS MASS NBASIS(compile time) NRHS BLOCK BLOCK2 COARSEN_BATCH
-//      HOT_START CONFIG SUBSPACE_FILE V1_CHECK MRHS_COARSEN
+//      HOT_START CONFIG SUBSPACE_FILE DEPRECATED_CHECK MRHS_COARSEN
 //
 
 #include <Grid/Grid.h>
@@ -205,7 +206,7 @@ void ParseEnvironment(void)
   P("PolyRecordSelect",PolyRecordSelect);     P("PolyRefresh",PolyRefresh);               P("PolyVerbose",PolyVerbose);
   P("SmootherCoeffLog",SmootherCoeffLog);
   // library-side knobs, read straight from the environment as the library will
-  const char *envs[] = {"BLOCK","BLOCK2","SUBSPACE_FILE","CONFIG","HOT_START","MRHS_COARSEN","V1_CHECK",
+  const char *envs[] = {"BLOCK","BLOCK2","SUBSPACE_FILE","CONFIG","HOT_START","MRHS_COARSEN","DEPRECATED_CHECK",
                         "DENSE_CC","DENSE_SCHUR","DENSE_SCHUR2D","DENSE_DEVICE_SUM","DENSE_SPLITK","DENSE_NB",
                         "SCHUR2D_LEAF_SPAN","SCHUR2D_LEAF_LU","SCHUR2D_PROBE","SUMMA_HANDSHAKE","DENSE_APPLY_PROFILE","SLAB_FILE",
                         "SOLVE_SRHS"};
@@ -433,7 +434,7 @@ int main (int argc, char ** argv)
   GridCartesian *Coarse5d = new GridCartesian(c5latt,c5simd,c5mpi);
 
   // 6D coarse multiRHS grid: rhs is dim 0, undistributed and unvectorised.
-  // No divisibility constraint on nrhs, unlike V1.
+  // No divisibility constraint on nrhs, unlike the deprecated operators.
   Coordinate cmlatt({nrhs,1,clatt[0],clatt[1],clatt[2],clatt[3]});
   Coordinate cmsimd({1,1,1,1,1,1});
   Coordinate cmmpi ({1,1,mpi[0],mpi[1],mpi[2],mpi[3]});
@@ -495,7 +496,7 @@ int main (int argc, char ** argv)
   // Level 1 types: unvectorised coarse scalar
   //////////////////////////////////////////////////////////////////////
   typedef sTComplexD                                                          CComplexS;
-  typedef MultiGeneralCoarsenedOperatorV2<vSpinColourVector,CComplexS,nbasis> CoarseOperator;
+  typedef MultiGeneralCoarsenedOperator<vSpinColourVector,CComplexS,nbasis> CoarseOperator;
   typedef CoarseOperator::CoarseVector                                        CoarseVector;
   typedef Aggregation<vSpinColourVector,CComplexS,nbasis>                     Subspace;
 
@@ -556,23 +557,23 @@ int main (int argc, char ** argv)
   //  the projector imports it below -- the 60 fine vectors are 20 GB of host memory per
   //  rank and 60 LRU-eligible device fields competing with the solver's working set)
   //////////////////////////////////////////////////////////////////////
-  // Optional cross check of the coarse matrix elements against the V1
-  // path, which needs a vectorised coarse space. Block Gram-Schmidt is
-  // idempotent, so V1 may re-orthonormalise the same vectors in place
+  // Optional cross check of the coarse matrix elements against the
+  // deprecated path, which needs a vectorised coarse space. Block Gram-Schmidt
+  // is idempotent, so the deprecated path may re-orthonormalise the same vectors in place
   // without a second copy of the subspace.
   //////////////////////////////////////////////////////////////////////
-  if ( getenv("V1_CHECK") ) {
+  if ( getenv("DEPRECATED_CHECK") ) {
 
-    typedef GeneralCoarsenedMatrix     <vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
-    typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MrhsLittleDiracOperator;
+    typedef DeprecatedGeneralCoarsenedMatrix     <vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+    typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MrhsLittleDiracOperator;
     typedef Aggregation                <vSpinColourVector,vTComplex,nbasis> SubspaceV;
 
     Coordinate v5latt({1,clatt[0],clatt[1],clatt[2],clatt[3]});
     Coordinate v5simd({1,fsimd[0],fsimd[1],fsimd[2],fsimd[3]});
     GridCartesian *Coarse5dV = new GridCartesian(v5latt,v5simd,c5mpi);
 
-    int nrhs_v1 = vComplex::Nsimd();
-    Coordinate vmlatt({nrhs_v1,1,clatt[0],clatt[1],clatt[2],clatt[3]});
+    int nrhs_dep = vComplex::Nsimd();
+    Coordinate vmlatt({nrhs_dep,1,clatt[0],clatt[1],clatt[2],clatt[3]});
     Coordinate vmsimd({vComplex::Nsimd(),1,1,1,1,1});
     GridCartesian *CoarseMrhsV = new GridCartesian(vmlatt,vmsimd,cmmpi);
 
@@ -582,25 +583,30 @@ int main (int argc, char ** argv)
     for(int k=0;k<nbasis;k++) AggV.subspace[k]=AggregatesGCR.subspace[k];
 
     LittleDiracOperator LittleDiracOpPV(geomV,FGrid,Coarse5dV);
-    std::cout << GridLogMessage << "*** V1 CoarsenOperator (cross check) ***" << std::endl;
+    std::cout << GridLogMessage << "*** deprecated CoarsenOperator (cross check) ***" << std::endl;
     SetFineSloppy(FineSloppyComms);
     LittleDiracOpPV.CoarsenOperator(PVdagM,AggV);
     SetFineSloppy(0);
 
-    MrhsLittleDiracOperator mrhsV1(geomV,CoarseMrhsV);
-    mrhsV1.CopyMatrix(LittleDiracOpPV);
+    MrhsLittleDiracOperator mrhsDep(geomV,CoarseMrhsV);
+    mrhsDep.CopyMatrix(LittleDiracOpPV);
 
     // BLAS_A is written by GridtoBLAS in lSite order and both sides carry
     // the same scalar_object, so the two are directly comparable.
     typedef MrhsLittleDiracOperator::calcMatrix calcMatrix;
     int npoint = geom.npoint;
     RealD num=0.0, den=0.0;
+    // The operator holds ONE site-major matrix buffer; MatrixPointOut hands back
+    // stencil point p in the same lSite order as the deprecated per-point
+    // array, and sizes a2.
+    deviceVector<calcMatrix> a2;
     for(int p=0;p<npoint;p++){
-      int64_t sites = mrhsV1.BLAS_A[p].size();
-      GRID_ASSERT(sites == (int64_t)CoarseOpPV.BLAS_A[p].size());
+      int64_t sites = mrhsDep.BLAS_A[p].size();
+      CoarseOpPV.MatrixPointOut(p,a2);
+      GRID_ASSERT(sites == (int64_t)a2.size());
       std::vector<calcMatrix> h1(sites),h2(sites);
-      acceleratorCopyFromDevice(&mrhsV1.BLAS_A[p][0],    &h1[0],sites*sizeof(calcMatrix));
-      acceleratorCopyFromDevice(&CoarseOpPV.BLAS_A[p][0],&h2[0],sites*sizeof(calcMatrix));
+      acceleratorCopyFromDevice(&mrhsDep.BLAS_A[p][0],&h1[0],sites*sizeof(calcMatrix));
+      acceleratorCopyFromDevice(&a2[0],              &h2[0],sites*sizeof(calcMatrix));
       ComplexD *w1=(ComplexD *)&h1[0];
       ComplexD *w2=(ComplexD *)&h2[0];
       int64_t words = sites*sizeof(calcMatrix)/sizeof(ComplexD);
@@ -610,8 +616,8 @@ int main (int argc, char ** argv)
         den += real(w1[i])*real(w1[i])+imag(w1[i])*imag(w1[i]);
       }
     }
-    std::cout << GridLogMessage << "V1_CHECK: |A_V1|^2 = " << den << std::endl;
-    std::cout << GridLogMessage << "V1_CHECK: |A_V1 - A_V2|^2 / |A_V1|^2 = " << num/den << std::endl;
+    std::cout << GridLogMessage << "DEPRECATED_CHECK: |A_dep|^2 = " << den << std::endl;
+    std::cout << GridLogMessage << "DEPRECATED_CHECK: |A_dep - A|^2 / |A_dep|^2 = " << num/den << std::endl;
     GRID_ASSERT( den > 0.0 );
     GRID_ASSERT( num/den < 1.0e-18 );
   }
@@ -627,7 +633,7 @@ int main (int argc, char ** argv)
   MrhsProjector.Allocate(nbasis,FGrid,Coarse5d);
   MrhsProjector.ImportBasis(AggregatesGCR.subspace);   // block orthonormal basis
   // The basis now lives in the projector's BLAS_V (device); nothing after this
-  // point reads AggregatesGCR.subspace (V1_CHECK moved above).  Free it: 60 fine
+  // point reads AggregatesGCR.subspace (DEPRECATED_CHECK moved above).  Free it: 60 fine
   // fields = 20 GB host per rank, and 60 LRU-eligible device copies.
   AggregatesGCR.subspace.clear(); AggregatesGCR.subspace.shrink_to_fit();
 
@@ -640,7 +646,7 @@ int main (int argc, char ** argv)
   //////////////////////////////////////////////////////////////////////
   // STAGE TWO: L2 -> L3.
   //
-  // The fine operator here is V2 at L1, which is natively multiRHS, so the
+  // The fine operator here is the L1 operator, which is natively multiRHS, so the
   // multiRHS driver applies with no promotion adapter: its D+1 grid IS the
   // batch grid the L1 operator is currently set to.
   //////////////////////////////////////////////////////////////////////
@@ -662,7 +668,7 @@ int main (int argc, char ** argv)
   // Coarsening deepens the tensor nest by one iScalar
   typedef CoarseVector::vector_object                                         CoarseSiteObj;
   typedef iScalar<CComplexS>                                                  CComplexS2;
-  typedef MultiGeneralCoarsenedOperatorV2<CoarseSiteObj,CComplexS2,nbasis>    CoarseCoarseOperator;
+  typedef MultiGeneralCoarsenedOperator<CoarseSiteObj,CComplexS2,nbasis>    CoarseCoarseOperator;
   typedef CoarseCoarseOperator::CoarseVector                                  CoarseCoarseVector;
 
   NextToNearestStencilGeometry5D geom2(CoarseCoarse5d);
@@ -730,7 +736,7 @@ int main (int argc, char ** argv)
   // STAGE THREE (part one): the dense bottom on L2.
   //
   // DenseCoarseMatrix is bilingual: it takes the elements through
-  // Geometry()/ExtractMatrix(), so the V2 operator serves directly. It does
+  // Geometry()/ExtractMatrix(), so the L2 operator serves directly. It does
   // detect that a multiRHS op cannot apply on the D dimensional grid and
   // skips its own certificate and VERIFY, so the equivalent check is done
   // here instead, driving the L2 operator at Nrhs 1 through a slice.
@@ -740,7 +746,7 @@ int main (int argc, char ** argv)
 
   if ( getenv("DENSE_CC")==nullptr || atoi(getenv("DENSE_CC")) ) {
 
-    std::cout << GridLogMessage << "*** L3 dense bottom: import from the V2 L2 operator ***" << std::endl;
+    std::cout << GridLogMessage << "*** L3 dense bottom: import from the L2 operator ***" << std::endl;
     DenseCC.reset(new DenseCC_t(CoarseCoarse5d));
     DenseCC->Import(CoarseOpL2);
 
@@ -790,7 +796,7 @@ int main (int argc, char ** argv)
   auto RunSolve = [&](int nr)
   {
     std::cout << GridLogMessage << "**********************************************" << std::endl;
-    std::cout << GridLogMessage << " V2 THREE-level solve, Nrhs = " << nr << std::endl;
+    std::cout << GridLogMessage << " THREE-level solve, Nrhs = " << nr << std::endl;
     // Device-memory budget BEFORE the solve (2026-08-28: NRHS=6 died in hipMalloc at the
     // first fine-smoother history allocation -- reported as an asynchronous "memory
     // access fault" unless AMD_SERIALIZE_KERNEL/COPY made it a clean OOM).  The outer
@@ -952,7 +958,7 @@ int main (int argc, char ** argv)
     GridStopWatch w; w.Start();
     L1PGCR(src,sol);
     w.Stop();
-    std::cout << GridLogMessage << "V2 3-level solve Nrhs "<<nr<<" total " << w.Elapsed()
+    std::cout << GridLogMessage << "3-level solve Nrhs "<<nr<<" total " << w.Elapsed()
               << "  (per RHS: " << w.useconds()/1.0e6/nr << " s)" << std::endl;
 
     // The outer operator is exact by policy; assert the state rather than

examples/Example_pvdagm_3level_dense.cc

@@ -212,8 +212,8 @@ public:
   }
 };
 
-// Lüscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
-// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi† src.
+// Luscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
+// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi^dag src.
 template<class Field>
 class LuscherGuesser : public LinearFunction<Field> {
   const std::vector<Field> &psi;
@@ -704,7 +704,7 @@ void runMG(
   TrivialPrecon<LatticeFermionD>   simple_fine;
 
   //////////////////////////////////////////////////////////////////////
-  // Level 0→1: coarsen PVdagM, build LinOpCoarse
+  // Level 0->1: coarsen PVdagM, build LinOpCoarse
   //////////////////////////////////////////////////////////////////////
   LittleDiracOperator LittleDiracOpPV(geom, FGrid, Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM, AggregatesPD);
@@ -736,7 +736,7 @@ void runMG(
   // Optional sigma-ordering of psi_coarse (SVD_REORDER set): replace the crude
   // first-NB_CC slice with the NB_CC most-null directions of span(psi_coarse)
   // under LinOpCoarse.  Nullness measure = singular values (eig of the
-  // Gram-whitened Psi†A†APsi), NOT the numerical range Q†AQ which
+  // Gram-whitened Psi^dag A^dag APsi), NOT the numerical range Q^dag AQ which
   // non-normality contaminates.  The printed sigma spectrum shows where the
   // truncation cliff sits.  Unset => raw first-30 (crude GS-ordered slice).
   //////////////////////////////////////////////////////////////////////
@@ -787,7 +787,7 @@ void runMG(
   }
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2: SUPERCOARSE aggregation using psi_coarse as subspace.
+  // Level 1->2: SUPERCOARSE aggregation using psi_coarse as subspace.
   // Maximal block {8,4,3,6}: CC = [3,6,8,8] = the dense-invertible floor.
   //
   // UNBLOCKING TRUNCATION NB_CC = 30 (first-30 slice of psi_coarse):
@@ -811,7 +811,7 @@ void runMG(
 
   typedef typename CoarseVector::vector_object                            CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB_CC>         LittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB_CC>         LittleDiracOperatorL2;
   typedef typename LittleDiracOperatorL2::CoarseVector                   CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,NB_CC>                    SubspaceL2;
   typedef MGPreconditioner<CoarseSiteObj,vTTComplex,NB_CC>               L1to2MG;
@@ -830,7 +830,7 @@ void runMG(
 
   //////////////////////////////////////////////////////////////////////
   // CC solve: DENSE (exact, non-iterative) by default; DENSE_CC=0 gives
-  // the previous iterative L3PGCR + Lüscher-guesser path for A/B.
+  // the previous iterative L3PGCR + Luscher-guesser path for A/B.
   //////////////////////////////////////////////////////////////////////
   int use_dense = 1;
   if (getenv("DENSE_CC")) use_dense = atoi(getenv("DENSE_CC"));
@@ -859,7 +859,7 @@ void runMG(
     ccGuess = &simpleCC;      // exact solve ignores/overwrites any guess
   } else {
     ////////////////////////////////////////////////////////////////////
-    // Lüscher deflation guesser (arXiv:0706.2298 A.3) for the iterative CC
+    // Luscher deflation guesser (arXiv:0706.2298 A.3) for the iterative CC
     // solve, as in the earlier supercoarse configuration.
     ////////////////////////////////////////////////////////////////////
     psi_cc.resize(nbasis, CoarseCoarse5d);
@@ -900,7 +900,7 @@ void runMG(
   }
 
   //////////////////////////////////////////////////////////////////////
-  // Coarse-level GCR smoother for Level 1→2 V-cycle.
+  // Coarse-level GCR smoother for Level 1->2 V-cycle.
   //////////////////////////////////////////////////////////////////////
   RealD coarse_smoother_shift = 0.1;
   int   coarse_smoother_nstep = 2;
@@ -914,7 +914,7 @@ void runMG(
   CoarseSmootherGCR.SetZeroGuess(1);  // post-smoother slot: caller zeroes vec2 (NOT L2MGsolver: it takes the Luscher guess)
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2 V-cycle preconditioner.
+  // Level 1->2 V-cycle preconditioner.
   //////////////////////////////////////////////////////////////////////
   L1to2MG L1to2Precon(AggregatesL2,
                        LinOpCoarse,
@@ -998,7 +998,7 @@ int main (int argc, char ** argv)
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  // Level 1 coarse grid: block 2^4 from fine (48×48×48×96 → 24×24×24×48, Ls=1)
+  // Level 1 coarse grid: block 2^4 from fine (48x48x48x96 -> 24x24x24x48, Ls=1)
   Coordinate clatt = lat_size;
   //  Coordinate Block1({2,2,2,2});
   //  Coordinate Block2({8,4,3,6});
@@ -1047,7 +1047,7 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
   typedef MGPreconditioner<vSpinColourVector,vTComplex,nbasis>        TwoLevelMG;

examples/Example_pvdagm_3level_madj.cc

@@ -323,7 +323,7 @@ void runMG(
   TrivialPrecon<LatticeFermionD>   simple_fine;
 
   //////////////////////////////////////////////////////////////////////
-  // Level 0→1: coarsen PVdagM, build LinOpCoarse
+  // Level 0->1: coarsen PVdagM, build LinOpCoarse
   //////////////////////////////////////////////////////////////////////
   LittleDiracOperator LittleDiracOpPV(geom, FGrid, Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM, AggregatesPD);
@@ -348,7 +348,7 @@ void runMG(
   //////////////////////////////////////////////////////////////////////
   // psi_coarse: coarse projections of pre-GS fine null vectors.
   // These are the Level 1 near-null vectors, promoted from Level 0.
-  // Used as the aggregation basis for Level 1→2 coarsening.
+  // Used as the aggregation basis for Level 1->2 coarsening.
   //////////////////////////////////////////////////////////////////////
   std::vector<CoarseVector> psi_coarse(nbasis, Coarse5d);
   for (int k = 0; k < nbasis; k++)
@@ -378,22 +378,22 @@ void runMG(
     RealD normC      = C.norm();
     RealD normCmCdag = (C - C.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse null matrix ||C||            = " << normC << std::endl;
-    std::cout << GridLogMessage << "Coarse null matrix ||C - C†||/||C|| = " << normCmCdag/normC << std::endl;
+    std::cout << GridLogMessage << "Coarse null matrix ||C - C^dag||/||C|| = " << normCmCdag/normC << std::endl;
     std::cout << GridLogMessage << "Galerkin check ||C||/||W||          = " << normC/normW << std::endl;
   }
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2: set up aggregation using psi_coarse as subspace.
+  // Level 1->2: set up aggregation using psi_coarse as subspace.
   // Block factor 2,2,3,2 (removes odd local sublattice in z given MPI
-  // geometry 3×6×4×4 where z-local at Level 1 is 6).
+  // geometry 3x6x4x4 where z-local at Level 1 is 6).
   // psi_coarse are assigned directly; CoarsenOperator performs
   // block-GS orthogonalisation before building LinOpCoarseCoarse.
   //////////////////////////////////////////////////////////////////////
   // innerProduct(CoarseSiteObj, CoarseSiteObj) returns iScalar<vTComplex>, so CComplex
-  // for the L1→L2 level must be iScalar<vTComplex>, not vTComplex.
+  // for the L1->L2 level must be iScalar<vTComplex>, not vTComplex.
   typedef typename CoarseVector::vector_object                            CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
   typedef typename LittleDiracOperatorL2::CoarseVector                   CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,NB>                       SubspaceL2;
   typedef MGPreconditioner<CoarseSiteObj,vTTComplex,NB>                  L1to2MG;
@@ -412,14 +412,14 @@ void runMG(
   // Level 2 solver: plain GCR, no further coarsening
   //////////////////////////////////////////////////////////////////////
   TrivialPrecon<CoarseCoarseVector> simpleCC;
-  // L3PGCR is an inner solver inside the L1→2 V-cycle; does not need to converge
+  // L3PGCR is an inner solver inside the L1->2 V-cycle; does not need to converge
   // to fine-grid precision. Loose tolerance (3e-2) and large restart (64) to allow
   // the Krylov space to span enough of the near-null spectrum of LinOpCC per cycle.
   PrecGeneralisedConjugateResidualNonHermitian<CoarseCoarseVector> L3PGCR(1.0e-4,5,LinOpCC,simpleCC,64,64);
   L3PGCR.Level(3);
 
   //////////////////////////////////////////////////////////////////////
-  // Coarse-level GCR smoother for Level 1→2 V-cycle.
+  // Coarse-level GCR smoother for Level 1->2 V-cycle.
   // Mirrors fine-grid SmootherGCR: shifted operator + fixed step count.
   // coarse_smoother_shift and coarse_smoother_nstep are the tuning knobs.
   //////////////////////////////////////////////////////////////////////
@@ -433,7 +433,7 @@ void runMG(
   CoarseSmootherGCR.Level(2);
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2 V-cycle preconditioner.
+  // Level 1->2 V-cycle preconditioner.
   //////////////////////////////////////////////////////////////////////
   L1to2MG L1to2Precon(AggregatesL2,
 		      LinOpCoarse,
@@ -474,7 +474,7 @@ void runMG(
   f_src = one;
 
   // Pre-smoother: none (TrivialPrecon); post-smoother: shifted PGCR.
-  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1→2 V-cycle).
+  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1->2 V-cycle).
   TwoLevelMG ThreeLevelPrecon(AggregatesPD,
 			      PVdagM,
 			      simple_fine,
@@ -518,7 +518,7 @@ int main (int argc, char ** argv)
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  // Level 1 coarse grid: block 2^4 from fine (48×48×48×96 → 24×24×24×48, Ls=1)
+  // Level 1 coarse grid: block 2^4 from fine (48x48x48x96 -> 24x24x24x48, Ls=1)
   Coordinate clatt = lat_size;
   for (int d = 0; d < 4; d++) clatt[d] /= 2;
   std::cout << GridLogMessage << "Level 1 coarse lattice: " << clatt << std::endl;
@@ -526,13 +526,13 @@ int main (int argc, char ** argv)
   GridCartesian *Coarse4d  = SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridCartesian *Coarse5d  = SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
 
-  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24×24×24×48 → 12×12×8×16, Ls=1).
+  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24x24x24x48 -> 12x12x8x16, Ls=1).
   // MPI geometry 3.6.4.4 (288 ranks): fine local {16,8,12,24}.
   // Level 1 local {8,4,6,12}; Level 2 local {4,2,2,4}.
-  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) → must use 3.
+  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) -> must use 3.
   // t blocked by 3: t-Level1-local=12; 12/3=4 divisible by Nsimd=4 (gen-simd-width=64).
-  //   t-block=2 gives t2-local=6, 6 mod 4 ≠ 0, fails Grid SIMD assertion. ✓
-  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4). ✓
+  //   t-block=2 gives t2-local=6, 6 mod 4 != 0, fails Grid SIMD assertion.
+  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4).
   Coordinate clatt2 = clatt;
   clatt2[0] /= 2;
   clatt2[1] /= 2;
@@ -566,7 +566,7 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
   typedef MGPreconditioner<vSpinColourVector,vTComplex,nbasis>        TwoLevelMG;

examples/Example_pvdagm_4level.cc

@@ -190,8 +190,8 @@ public:
   }
 };
 
-// Lüscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
-// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi† src.
+// Luscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
+// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi^dag src.
 template<class Field>
 class LuscherGuesser : public LinearFunction<Field> {
   const std::vector<Field> &psi;
@@ -343,7 +343,7 @@ void runMG(
   TrivialPrecon<LatticeFermionD>   simple_fine;
 
   //////////////////////////////////////////////////////////////////////
-  // Level 0→1: coarsen PVdagM, build LinOpCoarse
+  // Level 0->1: coarsen PVdagM, build LinOpCoarse
   //////////////////////////////////////////////////////////////////////
   LittleDiracOperator LittleDiracOpPV(geom, FGrid, Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM, AggregatesPD);
@@ -367,7 +367,7 @@ void runMG(
   //////////////////////////////////////////////////////////////////////
   // psi_coarse: coarse projections of pre-GS fine null vectors.
   // These are the Level 1 near-null vectors, promoted from Level 0.
-  // Used as the aggregation basis for Level 1→2 coarsening.
+  // Used as the aggregation basis for Level 1->2 coarsening.
   //////////////////////////////////////////////////////////////////////
   std::vector<CoarseVector> psi_coarse(nbasis, Coarse5d);
   for (int k = 0; k < nbasis; k++)
@@ -397,22 +397,22 @@ void runMG(
     RealD normC      = C.norm();
     RealD normCmCdag = (C - C.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse null matrix ||C||            = " << normC << std::endl;
-    std::cout << GridLogMessage << "Coarse null matrix ||C - C†||/||C|| = " << normCmCdag/normC << std::endl;
+    std::cout << GridLogMessage << "Coarse null matrix ||C - C^dag||/||C|| = " << normCmCdag/normC << std::endl;
     std::cout << GridLogMessage << "Galerkin check ||C||/||W||          = " << normC/normW << std::endl;
   }
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2: set up aggregation using psi_coarse as subspace.
+  // Level 1->2: set up aggregation using psi_coarse as subspace.
   // Block factor 2,2,3,2 (removes odd local sublattice in z given MPI
-  // geometry 3×6×4×4 where z-local at Level 1 is 6).
+  // geometry 3x6x4x4 where z-local at Level 1 is 6).
   // psi_coarse are assigned directly; CoarsenOperator performs
   // block-GS orthogonalisation before building LinOpCoarseCoarse.
   //////////////////////////////////////////////////////////////////////
   // innerProduct(CoarseSiteObj, CoarseSiteObj) returns iScalar<vTComplex>, so CComplex
-  // for the L1→L2 level must be iScalar<vTComplex>, not vTComplex.
+  // for the L1->L2 level must be iScalar<vTComplex>, not vTComplex.
   typedef typename CoarseVector::vector_object                            CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
   typedef typename LittleDiracOperatorL2::CoarseVector                   CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,NB>                       SubspaceL2;
   typedef MGPreconditioner<CoarseSiteObj,vTTComplex,NB>                  L1to2MG;
@@ -430,12 +430,12 @@ void runMG(
   TrivialPrecon<CoarseCoarseVector> simpleCC;
 
   //////////////////////////////////////////////////////////////////////
-  // Lüscher deflation guesser for L3PGCR.
+  // Luscher deflation guesser for L3PGCR.
   // Step 1: project psi_coarse[k] (promoted fine null vectors) to
-  //         CoarseCoarseVector space — these cover the zero-momentum
+  //         CoarseCoarseVector space -- these cover the zero-momentum
   //         component of the near-null space of LinOpCC.
   // Step 2: breed Nextra additional null vectors directly on LinOpCC
-  //         using GCR with random sources — these pick up near-null
+  //         using GCR with random sources -- these pick up near-null
   //         modes at all spatial frequencies not spanned by step 1.
   // Step 3: build C_{st} = <psi_cc[s]|LinOpCC|psi_cc[t]> over the
   //         full augmented basis and invert directly via Eigen LU.
@@ -477,13 +477,13 @@ void runMG(
     RealD normCcc      = Ccc.norm();
     RealD normCccmCdag = (Ccc - Ccc.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc||              = " << normCcc << std::endl;
-    std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc-Ccc†||/||Ccc|| = " << normCccmCdag/normCcc << std::endl;
+    std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc-Ccc^dag||/||Ccc|| = " << normCccmCdag/normCcc << std::endl;
   }
   Eigen::MatrixXcd Ccc_inv = Ccc.inverse();
   LuscherGuesser<CoarseCoarseVector> CCDeflGuesser(psi_cc, Ccc_inv);
 
   //////////////////////////////////////////////////////////////////////
-  // Level 2→3: coarsen LinOpCC using the RAW promoted psi_cc as aggregation
+  // Level 2->3: coarsen LinOpCC using the RAW promoted psi_cc as aggregation
   // to build the Level 4 (coarse-coarse-coarse) operator.
   //   psi_cc[0..nbasis-1] are the coarse-coarse near-null vectors, projected
   //   from the RAW psi_coarse (themselves projected from the RAW fine null
@@ -492,11 +492,11 @@ void runMG(
   //   so assign COPIES of psi_cc and keep psi_cc itself raw.
   //
   // Tensor depth deepens once more: innerProduct(CoarseCoarseSiteObj,...) returns
-  // iScalar<vTTComplex>, so CComplex for the L2→L3 level is iScalar<iScalar<vTComplex>>.
+  // iScalar<vTTComplex>, so CComplex for the L2->L3 level is iScalar<iScalar<vTComplex>>.
   //////////////////////////////////////////////////////////////////////
   typedef typename CoarseCoarseVector::vector_object                     CoarseCoarseSiteObj;
   typedef iScalar<vTTComplex>                                            vTTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseCoarseSiteObj,vTTTComplex,NB>     LittleDiracOperatorL3;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseCoarseSiteObj,vTTTComplex,NB>     LittleDiracOperatorL3;
   typedef typename LittleDiracOperatorL3::CoarseVector                   CoarseCoarseCoarseVector;
   typedef Aggregation<CoarseCoarseSiteObj,vTTTComplex,NB>               SubspaceL3;
   typedef MGPreconditioner<CoarseCoarseSiteObj,vTTTComplex,NB>          L2to3MG;
@@ -531,7 +531,7 @@ void runMG(
   L4PGCR.Name("CCCouter");
 
   //////////////////////////////////////////////////////////////////////
-  // Level 2→3 V-cycle: depth-2 SHIFTED smoother on LinOpCC + L4 bottom.
+  // Level 2->3 V-cycle: depth-2 SHIFTED smoother on LinOpCC + L4 bottom.
   // The shift slides the coarse-coarse field of values off the origin so a
   // 2-step smoother has something to bite on a non-normal operator (IRS idea).
   //////////////////////////////////////////////////////////////////////
@@ -556,7 +556,7 @@ void runMG(
                       simpleCCC);                // trivial guesser at the bottom
 
   //////////////////////////////////////////////////////////////////////
-  // Level 3 (coarse-coarse) solve: GCR preconditioned by the L2→L3 V-cycle.
+  // Level 3 (coarse-coarse) solve: GCR preconditioned by the L2->L3 V-cycle.
   // Replaces the plain L3PGCR of the 3-level build -- the coarse-coarse level
   // is now smoothed shallowly and recursed rather than solved deeply.
   //////////////////////////////////////////////////////////////////////
@@ -565,7 +565,7 @@ void runMG(
   L3MGsolver.Name("CCouter");
 
   //////////////////////////////////////////////////////////////////////
-  // Coarse-level GCR smoother for Level 1→2 V-cycle.
+  // Coarse-level GCR smoother for Level 1->2 V-cycle.
   // Mirrors fine-grid SmootherGCR: shifted operator + fixed step count.
   // coarse_smoother_shift and coarse_smoother_nstep are the tuning knobs.
   //////////////////////////////////////////////////////////////////////
@@ -581,15 +581,15 @@ void runMG(
   CoarseSmootherGCR.Name("Csmoother");
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2 V-cycle preconditioner.
+  // Level 1->2 V-cycle preconditioner.
   //////////////////////////////////////////////////////////////////////
   L1to2MG L1to2Precon(AggregatesL2,
                        LinOpCoarse,
                        simpleC,            // no pre-smoother (matches fine-grid setup)
                        CoarseSmootherGCR,  // post-smoother: depth-2 shifted GCR
                        LinOpCC,
-                       L3MGsolver,         // coarse-coarse solve is now the L2→L3 V-cycle
-                       CCDeflGuesser);     // Lüscher guesser: psi_cc C^{-1} psi_cc†
+                       L3MGsolver,         // coarse-coarse solve is now the L2->L3 V-cycle
+                       CCDeflGuesser);     // Luscher guesser: psi_cc C^{-1} psi_cc^dag
 
   //////////////////////////////////////////////////////////////////////
   // Standalone Level 1 two-level solve test.
@@ -624,7 +624,7 @@ void runMG(
   f_src = one;
 
   // Pre-smoother: none (TrivialPrecon); post-smoother: shifted PGCR.
-  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1→2 V-cycle).
+  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1->2 V-cycle).
   TwoLevelMG ThreeLevelPrecon(AggregatesPD,
                                PVdagM,
                                simple_fine,
@@ -668,7 +668,7 @@ int main (int argc, char ** argv)
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  // Level 1 coarse grid: block 2^4 from fine (48×48×48×96 → 24×24×24×48, Ls=1)
+  // Level 1 coarse grid: block 2^4 from fine (48x48x48x96 -> 24x24x24x48, Ls=1)
   Coordinate clatt = lat_size;
   for (int d = 0; d < 4; d++) clatt[d] /= 2;
   std::cout << GridLogMessage << "Level 1 coarse lattice: " << clatt << std::endl;
@@ -676,13 +676,13 @@ int main (int argc, char ** argv)
   GridCartesian *Coarse4d  = SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridCartesian *Coarse5d  = SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
 
-  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24×24×24×48 → 12×12×8×16, Ls=1).
+  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24x24x24x48 -> 12x12x8x16, Ls=1).
   // MPI geometry 3.6.4.4 (288 ranks): fine local {16,8,12,24}.
   // Level 1 local {8,4,6,12}; Level 2 local {4,2,2,4}.
-  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) → must use 3.
+  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) -> must use 3.
   // t blocked by 3: t-Level1-local=12; 12/3=4 divisible by Nsimd=4 (gen-simd-width=64).
-  //   t-block=2 gives t2-local=6, 6 mod 4 ≠ 0, fails Grid SIMD assertion. ✓
-  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4). ✓
+  //   t-block=2 gives t2-local=6, 6 mod 4 != 0, fails Grid SIMD assertion.
+  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4).
   Coordinate clatt2 = clatt;
   clatt2[0] /= 2;
   clatt2[1] /= 2;
@@ -728,7 +728,7 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
   typedef MGPreconditioner<vSpinColourVector,vTComplex,nbasis>        TwoLevelMG;

examples/Example_pvdagm_5level.cc

@@ -190,8 +190,8 @@ public:
   }
 };
 
-// Lüscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
-// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi† src.
+// Luscher deflated guesser (arXiv:0706.2298 Sec A.3) for a non-Hermitian solve.
+// C_{st} = <psi[s] | LinOp | psi[t]>;  guess = sum_s c_s psi[s]  where c = C^{-1} psi^dag src.
 template<class Field>
 class LuscherGuesser : public LinearFunction<Field> {
   const std::vector<Field> &psi;
@@ -344,7 +344,7 @@ void runMG(
   TrivialPrecon<LatticeFermionD>   simple_fine;
 
   //////////////////////////////////////////////////////////////////////
-  // Level 0→1: coarsen PVdagM, build LinOpCoarse
+  // Level 0->1: coarsen PVdagM, build LinOpCoarse
   //////////////////////////////////////////////////////////////////////
   LittleDiracOperator LittleDiracOpPV(geom, FGrid, Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM, AggregatesPD);
@@ -368,7 +368,7 @@ void runMG(
   //////////////////////////////////////////////////////////////////////
   // psi_coarse: coarse projections of pre-GS fine null vectors.
   // These are the Level 1 near-null vectors, promoted from Level 0.
-  // Used as the aggregation basis for Level 1→2 coarsening.
+  // Used as the aggregation basis for Level 1->2 coarsening.
   //////////////////////////////////////////////////////////////////////
   std::vector<CoarseVector> psi_coarse(nbasis, Coarse5d);
   for (int k = 0; k < nbasis; k++)
@@ -398,22 +398,22 @@ void runMG(
     RealD normC      = C.norm();
     RealD normCmCdag = (C - C.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse null matrix ||C||            = " << normC << std::endl;
-    std::cout << GridLogMessage << "Coarse null matrix ||C - C†||/||C|| = " << normCmCdag/normC << std::endl;
+    std::cout << GridLogMessage << "Coarse null matrix ||C - C^dag||/||C|| = " << normCmCdag/normC << std::endl;
     std::cout << GridLogMessage << "Galerkin check ||C||/||W||          = " << normC/normW << std::endl;
   }
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2: set up aggregation using psi_coarse as subspace.
+  // Level 1->2: set up aggregation using psi_coarse as subspace.
   // Block factor 2,2,3,2 (removes odd local sublattice in z given MPI
-  // geometry 3×6×4×4 where z-local at Level 1 is 6).
+  // geometry 3x6x4x4 where z-local at Level 1 is 6).
   // psi_coarse are assigned directly; CoarsenOperator performs
   // block-GS orthogonalisation before building LinOpCoarseCoarse.
   //////////////////////////////////////////////////////////////////////
   // innerProduct(CoarseSiteObj, CoarseSiteObj) returns iScalar<vTComplex>, so CComplex
-  // for the L1→L2 level must be iScalar<vTComplex>, not vTComplex.
+  // for the L1->L2 level must be iScalar<vTComplex>, not vTComplex.
   typedef typename CoarseVector::vector_object                            CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,NB>            LittleDiracOperatorL2;
   typedef typename LittleDiracOperatorL2::CoarseVector                   CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,NB>                       SubspaceL2;
   typedef MGPreconditioner<CoarseSiteObj,vTTComplex,NB>                  L1to2MG;
@@ -431,12 +431,12 @@ void runMG(
   TrivialPrecon<CoarseCoarseVector> simpleCC;
 
   //////////////////////////////////////////////////////////////////////
-  // Lüscher deflation guesser for L3PGCR.
+  // Luscher deflation guesser for L3PGCR.
   // Step 1: project psi_coarse[k] (promoted fine null vectors) to
-  //         CoarseCoarseVector space — these cover the zero-momentum
+  //         CoarseCoarseVector space -- these cover the zero-momentum
   //         component of the near-null space of LinOpCC.
   // Step 2: breed Nextra additional null vectors directly on LinOpCC
-  //         using GCR with random sources — these pick up near-null
+  //         using GCR with random sources -- these pick up near-null
   //         modes at all spatial frequencies not spanned by step 1.
   // Step 3: build C_{st} = <psi_cc[s]|LinOpCC|psi_cc[t]> over the
   //         full augmented basis and invert directly via Eigen LU.
@@ -478,13 +478,13 @@ void runMG(
     RealD normCcc      = Ccc.norm();
     RealD normCccmCdag = (Ccc - Ccc.adjoint()).norm();
     std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc||              = " << normCcc << std::endl;
-    std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc-Ccc†||/||Ccc|| = " << normCccmCdag/normCcc << std::endl;
+    std::cout << GridLogMessage << "Coarse-coarse deflation matrix ||Ccc-Ccc^dag||/||Ccc|| = " << normCccmCdag/normCcc << std::endl;
   }
   Eigen::MatrixXcd Ccc_inv = Ccc.inverse();
   LuscherGuesser<CoarseCoarseVector> CCDeflGuesser(psi_cc, Ccc_inv);
 
   //////////////////////////////////////////////////////////////////////
-  // Level 2→3: coarsen LinOpCC using the RAW promoted psi_cc as aggregation
+  // Level 2->3: coarsen LinOpCC using the RAW promoted psi_cc as aggregation
   // to build the Level 4 (coarse-coarse-coarse) operator.
   //   psi_cc[0..nbasis-1] are the coarse-coarse near-null vectors, projected
   //   from the RAW psi_coarse (themselves projected from the RAW fine null
@@ -493,11 +493,11 @@ void runMG(
   //   so assign COPIES of psi_cc and keep psi_cc itself raw.
   //
   // Tensor depth deepens once more: innerProduct(CoarseCoarseSiteObj,...) returns
-  // iScalar<vTTComplex>, so CComplex for the L2→L3 level is iScalar<iScalar<vTComplex>>.
+  // iScalar<vTTComplex>, so CComplex for the L2->L3 level is iScalar<iScalar<vTComplex>>.
   //////////////////////////////////////////////////////////////////////
   typedef typename CoarseCoarseVector::vector_object                     CoarseCoarseSiteObj;
   typedef iScalar<vTTComplex>                                            vTTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseCoarseSiteObj,vTTTComplex,NB>     LittleDiracOperatorL3;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseCoarseSiteObj,vTTTComplex,NB>     LittleDiracOperatorL3;
   typedef typename LittleDiracOperatorL3::CoarseVector                   CoarseCoarseCoarseVector;
   typedef Aggregation<CoarseCoarseSiteObj,vTTTComplex,NB>               SubspaceL3;
   typedef MGPreconditioner<CoarseCoarseSiteObj,vTTTComplex,NB>          L2to3MG;
@@ -514,7 +514,7 @@ void runMG(
   TrivialPrecon<CoarseCoarseCoarseVector> simpleCCC;
 
   //////////////////////////////////////////////////////////////////////
-  // Level 3→4: coarsen LinOpCCC to build the Level 5 operator, using a
+  // Level 3->4: coarsen LinOpCCC to build the Level 5 operator, using a
   // TRUNCATED basis of only the first NB5 (< nbasis) raw promoted null vectors.
   //   psi_ccc[k] = raw psi_cc projected through the (block-GS'd) L3 aggregation
   //   -- the pre-block-GS chain continued one level deeper.  We keep only the
@@ -526,7 +526,7 @@ void runMG(
   //   NB: a positive result is conservative (sigma-ordering can only help); a
   //   negative one is inconclusive until the sigma-ordered NB5 is tried.
   //
-  // Tensor depth deepens once more: CComplex for the L3→L4 level is
+  // Tensor depth deepens once more: CComplex for the L3->L4 level is
   // iScalar<vTTTComplex>.  NB5 (the coarse dimension) is independent of the
   // depth -- it just makes the coarsest site vector NB5-dimensional.
   //////////////////////////////////////////////////////////////////////
@@ -542,10 +542,10 @@ void runMG(
   // Optional sigma-ordering of psi_ccc (SVD_REORDER set): replace the crude
   // first-NB5 slice with the NB5 genuinely-most-null directions of span(psi_ccc)
   // under LinOpCCC.  For a NON-NORMAL operator the nullness measure is the
-  // singular value of A restricted to the span -- eig of Q†A†AQ -- NOT the
-  // numerical range Q†AQ (which non-normality contaminates).  Robust route:
+  // singular value of A restricted to the span -- eig of Q^dag A^dag AQ -- NOT the
+  // numerical range Q^dag AQ (which non-normality contaminates).  Robust route:
   // whiten by the Gram (drop near-dependent directions), Hermitian-eig the
-  // whitened A†A, rotate.  The printed singular spectrum IS the SVD study: where
+  // whitened A^dag A, rotate.  The printed singular spectrum IS the SVD study: where
   // it falls off tells you the natural NB5, and the same numbers illuminate why
   // the earlier singular-subspace deflation re-entered.  Safe here because we
   // ORDER vectors that then feed a Galerkin projection, not REMOVE a subspace.
@@ -603,7 +603,7 @@ void runMG(
 
   typedef typename CoarseCoarseCoarseVector::vector_object              CoarseCoarseCoarseSiteObj;
   typedef iScalar<vTTTComplex>                                          vTTTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseCoarseCoarseSiteObj,vTTTTComplex,NB5> LittleDiracOperatorL4;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseCoarseCoarseSiteObj,vTTTTComplex,NB5> LittleDiracOperatorL4;
   typedef typename LittleDiracOperatorL4::CoarseVector                  CoarseCoarseCoarseCoarseVector;
   typedef Aggregation<CoarseCoarseCoarseSiteObj,vTTTTComplex,NB5>       SubspaceL4;
   typedef MGPreconditioner<CoarseCoarseCoarseSiteObj,vTTTTComplex,NB5>  L3to4MG;
@@ -635,7 +635,7 @@ void runMG(
   L5PGCR.Name("CCCCouter");
 
   //////////////////////////////////////////////////////////////////////
-  // Level 3→4 V-cycle: depth-2 SHIFTED smoother on LinOpCCC + Level 5 bottom.
+  // Level 3->4 V-cycle: depth-2 SHIFTED smoother on LinOpCCC + Level 5 bottom.
   // Level 4 is no longer the bottom -- it is smoothed shallowly and recursed to
   // Level 5, mirroring how Level 3 recurses to Level 4.
   //////////////////////////////////////////////////////////////////////
@@ -660,14 +660,14 @@ void runMG(
                       simpleCCCC);                   // trivial guesser at the bottom
 
   //////////////////////////////////////////////////////////////////////
-  // Level 4 (coarse-coarse-coarse) solve: GCR preconditioned by the L3→L4 V-cycle.
+  // Level 4 (coarse-coarse-coarse) solve: GCR preconditioned by the L3->L4 V-cycle.
   //////////////////////////////////////////////////////////////////////
   PrecGeneralisedConjugateResidualNonHermitian<CoarseCoarseCoarseVector> L4MGsolver(1.0e-1,200,LinOpCCC,L3to4Precon,16,16);
   L4MGsolver.Level(4);
   L4MGsolver.Name("CCCouter");
 
   //////////////////////////////////////////////////////////////////////
-  // Level 2→3 V-cycle: depth-2 SHIFTED smoother on LinOpCC + Level 4 solve.
+  // Level 2->3 V-cycle: depth-2 SHIFTED smoother on LinOpCC + Level 4 solve.
   // The shift slides the coarse-coarse field of values off the origin so a
   // 2-step smoother has something to bite on a non-normal operator (IRS idea).
   //////////////////////////////////////////////////////////////////////
@@ -688,11 +688,11 @@ void runMG(
                       simpleCC,                  // no pre-smoother
                       CoarseCoarseSmootherGCR,   // post-smoother: depth-2 shifted GCR
                       LinOpCCC,
-                      L4MGsolver,                // coarse solve is now the L3→L4 V-cycle
+                      L4MGsolver,                // coarse solve is now the L3->L4 V-cycle
                       simpleCCC);                // trivial guesser
 
   //////////////////////////////////////////////////////////////////////
-  // Level 3 (coarse-coarse) solve: GCR preconditioned by the L2→L3 V-cycle.
+  // Level 3 (coarse-coarse) solve: GCR preconditioned by the L2->L3 V-cycle.
   // Replaces the plain L3PGCR of the 3-level build -- the coarse-coarse level
   // is now smoothed shallowly and recursed rather than solved deeply.
   //////////////////////////////////////////////////////////////////////
@@ -701,7 +701,7 @@ void runMG(
   L3MGsolver.Name("CCouter");
 
   //////////////////////////////////////////////////////////////////////
-  // Coarse-level GCR smoother for Level 1→2 V-cycle.
+  // Coarse-level GCR smoother for Level 1->2 V-cycle.
   // Mirrors fine-grid SmootherGCR: shifted operator + fixed step count.
   // coarse_smoother_shift and coarse_smoother_nstep are the tuning knobs.
   //////////////////////////////////////////////////////////////////////
@@ -717,15 +717,15 @@ void runMG(
   CoarseSmootherGCR.Name("Csmoother");
 
   //////////////////////////////////////////////////////////////////////
-  // Level 1→2 V-cycle preconditioner.
+  // Level 1->2 V-cycle preconditioner.
   //////////////////////////////////////////////////////////////////////
   L1to2MG L1to2Precon(AggregatesL2,
                        LinOpCoarse,
                        simpleC,            // no pre-smoother (matches fine-grid setup)
                        CoarseSmootherGCR,  // post-smoother: depth-2 shifted GCR
                        LinOpCC,
-                       L3MGsolver,         // coarse-coarse solve is now the L2→L3 V-cycle
-                       CCDeflGuesser);     // Lüscher guesser: psi_cc C^{-1} psi_cc†
+                       L3MGsolver,         // coarse-coarse solve is now the L2->L3 V-cycle
+                       CCDeflGuesser);     // Luscher guesser: psi_cc C^{-1} psi_cc^dag
 
   //////////////////////////////////////////////////////////////////////
   // Standalone Level 1 two-level solve test.
@@ -760,7 +760,7 @@ void runMG(
   f_src = one;
 
   // Pre-smoother: none (TrivialPrecon); post-smoother: shifted PGCR.
-  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1→2 V-cycle).
+  // Coarse solver: L2MGsolver (PGCR preconditioned by Level 1->2 V-cycle).
   TwoLevelMG ThreeLevelPrecon(AggregatesPD,
                                PVdagM,
                                simple_fine,
@@ -804,7 +804,7 @@ int main (int argc, char ** argv)
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  // Level 1 coarse grid: block 2^4 from fine (48×48×48×96 → 24×24×24×48, Ls=1)
+  // Level 1 coarse grid: block 2^4 from fine (48x48x48x96 -> 24x24x24x48, Ls=1)
   Coordinate clatt = lat_size;
   for (int d = 0; d < 4; d++) clatt[d] /= 2;
   std::cout << GridLogMessage << "Level 1 coarse lattice: " << clatt << std::endl;
@@ -812,13 +812,13 @@ int main (int argc, char ** argv)
   GridCartesian *Coarse4d  = SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridCartesian *Coarse5d  = SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
 
-  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24×24×24×48 → 12×12×8×16, Ls=1).
+  // Level 2 coarse-coarse grid: block 2,2,3,3 from Level 1 (24x24x24x48 -> 12x12x8x16, Ls=1).
   // MPI geometry 3.6.4.4 (288 ranks): fine local {16,8,12,24}.
   // Level 1 local {8,4,6,12}; Level 2 local {4,2,2,4}.
-  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) → must use 3.
+  // z blocked by 3: z-Level1-local=6; 6/3=2 (even), 6/2=3 (odd) -> must use 3.
   // t blocked by 3: t-Level1-local=12; 12/3=4 divisible by Nsimd=4 (gen-simd-width=64).
-  //   t-block=2 gives t2-local=6, 6 mod 4 ≠ 0, fails Grid SIMD assertion. ✓
-  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4). ✓
+  //   t-block=2 gives t2-local=6, 6 mod 4 != 0, fails Grid SIMD assertion.
+  // With {4,2,2,4}: Nsimd=4 goes into x or t (both =4).
   Coordinate clatt2 = clatt;
   clatt2[0] /= 2;
   clatt2[1] /= 2;
@@ -880,7 +880,7 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
   typedef MGPreconditioner<vSpinColourVector,vTComplex,nbasis>        TwoLevelMG;

examples/Example_pvdagm_census.cc

@@ -49,7 +49,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 //   sigma_min << min|lambda|                       : non-normal near origin
 //   lambda_min(H) < 0                              : half-plane condition violated
 //
-// Requires the dagger code path in GeneralCoarsenedMatrix:
+// Requires the dagger code path in DeprecatedGeneralCoarsenedMatrix:
 //   _Adag allocated, PopulateAdag active, _Adag exchanged, hermitian=0.
 //
 // Env vars:
@@ -308,7 +308,7 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
 

examples/Example_pvdagm_coherence.cc

@@ -275,7 +275,7 @@ int main (int argc, char ** argv)
   MobiusFermionD Dpv (Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0, M5,b,c);
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>       PVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                          CoarseVector;
   // Grid index contraction is positional (colour/spin/lorentz order is meaningful),
   // so each MG projection adds one index to the tensor nest rather than reusing a slot:

examples/Example_pvdagm_mrhs.cc

@@ -38,7 +38,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 //   precon: V-cycle -- per-RHS fine post-smoother (16-step shifted GCR),
 //           batched restriction (MultiRHSBlockProject / GEMM),
 //           ONE coarse PGCR on the 6D mrhs coarse operator
-//           (MultiGeneralCoarsenedMatrix, GEMM mults -- the ~10x win),
+//           (DeprecatedMultiGeneralCoarsenedMatrix, GEMM mults -- the ~10x win),
 //           batched prolongation.
 //
 // The coarse operator is coarsened once with the standard single-RHS
@@ -260,8 +260,8 @@ int main (int argc, char ** argv)
 
   typedef PVdagMLinearOperator<MobiusFermionD,LatticeFermionD>        PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MrhsLittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  LittleDiracOperator;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MrhsLittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                           CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>             Subspace;
 

examples/Example_pvdagm_mrhs_3level.cc

@@ -23,7 +23,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 // with L3_DEFL=0 (NO deflation), applied to the enlarged block-diagonal mRHS system:
 // the coarse and coarse-coarse levels run a SINGLE Krylov (one GCR polynomial, inner
 // products summed over rhs) on the packed 6D mrhs fields, so both coarse levels batch
-// through GEMM (MultiGeneralCoarsenedMatrix) -- the valence throughput win at BOTH levels.
+// through GEMM (DeprecatedMultiGeneralCoarsenedMatrix) -- the valence throughput win at BOTH levels.
 //
 // Level structure (each coarse level is a single-field PGCR on a packed 6D mrhs field):
 //   L1 (fine)         : std::vector<LatticeFermionD>, MrhsPGCRNonHermitian on PVdagM,
@@ -227,16 +227,16 @@ int main (int argc, char ** argv)
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
 
   // Level 1 tensor types
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>       LittleDiracOperator;
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  MrhsLittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>       LittleDiracOperator;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  MrhsLittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                                CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>                  Subspace;
 
   // Level 2 tensor types (coarsening deepens the nest by one iScalar -- see CLAUDE.md)
   typedef CoarseVector::vector_object                                      CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>          LittleDiracOperatorL2;
-  typedef MultiGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>     MrhsLittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>          LittleDiracOperatorL2;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>     MrhsLittleDiracOperatorL2;
   typedef LittleDiracOperatorL2::CoarseVector                              CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,nbasis>                     SubspaceL2;
 

examples/Example_pvdagm_mrhs_3level_DenseCoarseMatrix.cc

@@ -233,16 +233,16 @@ int main (int argc, char ** argv)
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
 
   // Level 1 tensor types
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>       LittleDiracOperator;
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  MrhsLittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>       LittleDiracOperator;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  MrhsLittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                                CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>                  Subspace;
 
   // Level 2 tensor types (coarsening deepens the nest by one iScalar)
   typedef CoarseVector::vector_object                                      CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>          LittleDiracOperatorL2;
-  typedef MultiGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>     MrhsLittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>          LittleDiracOperatorL2;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>     MrhsLittleDiracOperatorL2;
   typedef LittleDiracOperatorL2::CoarseVector                              CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,nbasis>                     SubspaceL2;
 

examples/Example_pvdagm_mrhs_3level_dense.cc

@@ -237,7 +237,7 @@ public:
 #endif
   // NB: the apply GEMM (Y = slab^dag X) is a tiny-output/huge-K shape that
   // under-fills the GPU (~13ms).  The fix is a software split-K via
-  // GridBLAS.gemmBatched (see MultiRHSBlockCGLinalg.h / 2409.03904 Fig 11) —
+  // GridBLAS.gemmBatched (see MultiRHSBlockCGLinalg.h / 2409.03904 Fig 11) --
   // NOT a raw rocblas strided-batched batch, which hung on Frontier and was
   // removed.  TODO: reimplement through GridBLAS when the ~1.4 s/RHS is wanted.
 
@@ -891,16 +891,16 @@ int main (int argc, char ** argv)
   typedef ShiftedPVdagMLinearOperator<MobiusFermionD,LatticeFermionD> ShiftedPVdagM_t;
 
   // Level 1 tensor types
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>       LittleDiracOperator;
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  MrhsLittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>       LittleDiracOperator;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis>  MrhsLittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector                                CoarseVector;
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis>                  Subspace;
 
   // Level 2 tensor types (coarsening deepens the nest by one iScalar)
   typedef CoarseVector::vector_object                                      CoarseSiteObj;
   typedef iScalar<vTComplex>                                              vTTComplex;
-  typedef GeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>          LittleDiracOperatorL2;
-  typedef MultiGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>     MrhsLittleDiracOperatorL2;
+  typedef DeprecatedGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>          LittleDiracOperatorL2;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<CoarseSiteObj,vTTComplex,nbasis>     MrhsLittleDiracOperatorL2;
   typedef LittleDiracOperatorL2::CoarseVector                              CoarseCoarseVector;
   typedef Aggregation<CoarseSiteObj,vTTComplex,nbasis>                     SubspaceL2;
 

examples/Makefile.am

@@ -2,13 +2,15 @@ SUBDIRS = .
 
 include Make.inc
 
-# Precision variants of the two multigrid drivers.  The coarse-sector and
+# Compile-time variants of the two multigrid drivers.  The coarse-sector and
 # dense-inversion precisions are compile-time instantiations, so each is its
 # own binary built from the same source; the fine-level precision is a
-# run-time parameter and needs no variant.  Per-target flags give each
-# variant its own object file.
+# run-time parameter and needs no variant.  The recur variant additionally
+# takes the GCR residual norm from the recurrence rather than a reduction.
+# Per-target flags give each variant its own object file.
 bin_PROGRAMS += Example_pvdagm_multigrid_fp32coarse \
                 Example_pvdagm_multigrid_fp32coarse_fp32dense \
+                Example_pvdagm_multigrid_fp32coarse_fp32dense_recur \
                 Example_hdcg_multigrid_fp32coarse
 
 Example_pvdagm_multigrid_fp32coarse_SOURCES  = Example_pvdagm_multigrid.cc
@@ -19,6 +21,10 @@ Example_pvdagm_multigrid_fp32coarse_fp32dense_SOURCES  = Example_pvdagm_multigri
 Example_pvdagm_multigrid_fp32coarse_fp32dense_CPPFLAGS = -DCOARSE_SINGLE -DGRID_DENSE_INVERSE_SINGLE
 Example_pvdagm_multigrid_fp32coarse_fp32dense_LDADD    = $(top_builddir)/Grid/libGrid.a
 
+Example_pvdagm_multigrid_fp32coarse_fp32dense_recur_SOURCES  = Example_pvdagm_multigrid.cc
+Example_pvdagm_multigrid_fp32coarse_fp32dense_recur_CPPFLAGS = -DCOARSE_SINGLE -DGRID_DENSE_INVERSE_SINGLE -DGRID_GCR_RESIDUAL_RECURRENCE
+Example_pvdagm_multigrid_fp32coarse_fp32dense_recur_LDADD    = $(top_builddir)/Grid/libGrid.a
+
 Example_hdcg_multigrid_fp32coarse_SOURCES  = Example_hdcg_multigrid.cc
 Example_hdcg_multigrid_fp32coarse_CPPFLAGS = -DCOARSE_SINGLE
 Example_hdcg_multigrid_fp32coarse_LDADD    = $(top_builddir)/Grid/libGrid.a

scripts/gcr_polynomial.py

@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+"""
+Reconstruct the GCR smoother polynomial from PGCR coefficient logs
+(SmootherCoeffLog=1 in Example_pvdagm_3level_DenseCoarseMatrix, or
+LogCoefficients(1) on any PrecGeneralisedConjugateResidualNonHermitian).
+
+GCR with z = r (trivial preconditioner) builds  p_k = P_k(A) r0,
+r_k = R_k(A) r0:
+
+    R_0 = P_0 = 1
+    R_{k+1}(l) = R_k(l) - a_k l P_k(l)
+    P_{k+1}(l) = R_{k+1}(l) + sum_j b_{k,j} P_{k-j}(l)
+
+so after m steps  x_m = S(A) r0  with  S = sum_k a_k P_k  and residual
+polynomial  R_m = 1 - l S(l),  R_m(0) = 1.
+
+Usage
+  gcr_polynomial.py LOG [--name Fsmoother] [--lmin 0] [--lmax 3] [--n 600]
+      writes <name>_poly.dat : l  |R_mean(l)|  Re R_mean  Im R_mean  |R|min  |R|max  |S_mean(l)|
+      where mean = polynomial built from the per-step MEAN coefficients over
+      all calls, and min/max are the envelope over the individual calls.
+      Prints the mean coefficients and their relative spread per step.
+  gcr_polynomial.py --selftest
+      runs GCR in numpy on a random non-normal matrix, logs a,b the same way,
+      and checks  R_m(A) r0 == the actual residual  to rounding.
+
+gnuplot:
+  set logscale y; plot 'Fsmoother_poly.dat' u 1:2 w l t '|R_m| mean', \
+                       '' u 1:5 w l t 'min', '' u 1:6 w l t 'max'
+"""
+import re, sys, argparse
+import numpy as np
+from numpy.polynomial import polynomial as Poly   # coefficient arrays, low->high
+
+A_RE = re.compile(r'(\S+)\s+coeff\[(\d+)\]\s+a=\(([^,]+),([^)]+)\)')
+B_RE = re.compile(r'(\S+)\s+coeff\[(\d+)\]((?:\s+b\[\d+\]=\([^)]*\))+)')
+BJ_RE = re.compile(r'b\[(\d+)\]=\(([^,]+),([^)]+)\)')
+
+def parse(path, name):
+    """-> list of calls; each call = list of (a_k, [b_k0, b_k1, ...]) in step order."""
+    calls, cur = [], None
+    for line in open(path, errors='replace'):
+        m = A_RE.search(line)
+        if m and m.group(1) == name:
+            k = int(m.group(2)); a = complex(float(m.group(3)), float(m.group(4)))
+            if k == 0:
+                if cur: calls.append(cur)
+                cur = []
+            if cur is not None:
+                cur.append([a, []])
+            continue
+        m = B_RE.search(line)
+        if m and m.group(1) == name and cur is not None:
+            k = int(m.group(2))
+            bs = [complex(float(x[1]), float(x[2])) for x in BJ_RE.findall(m.group(3))]
+            if k < len(cur): cur[k][1] = bs
+    if cur: calls.append(cur)
+    return calls
+
+def polynomials(coeffs):
+    """coeffs: list of (a_k, [b_kj]) -> (S, R) as low->high complex coefficient arrays."""
+    R = [np.array([1.0+0j])]; P = [np.array([1.0+0j])]
+    S = np.array([0.0+0j])
+    for k, (a, bs) in enumerate(coeffs):
+        S = Poly.polyadd(S, a*P[k])
+        Rn = Poly.polysub(R[k], a*Poly.polymulx(P[k]))          # R_{k+1} = R_k - a l P_k
+        Pn = Rn.copy()
+        for j, b in enumerate(bs):                              # P_{k+1} = R_{k+1} + sum b_kj P_{k-j}
+            Pn = Poly.polyadd(Pn, b*P[k-j])
+        R.append(Rn); P.append(Pn)
+    return S, R[-1]
+
+def mean_coeffs(calls):
+    m = min(len(c) for c in calls)
+    out, spread = [], []
+    for k in range(m):
+        a = np.array([c[k][0] for c in calls])
+        nb = min(len(c[k][1]) for c in calls)
+        bs = [np.array([c[k][1][j] for c in calls]) for j in range(nb)]
+        out.append((a.mean(), [b.mean() for b in bs]))
+        sa = a.std()/max(abs(a.mean()),1e-300)
+        sb = [b.std()/max(abs(b.mean()),1e-300) for b in bs]
+        spread.append((sa, sb))
+    return out, spread
+
+def selftest(n=40, m=6, mmax=6, seed=1):
+    rng = np.random.default_rng(seed)
+    A = np.eye(n)*1.5 + 0.4*rng.standard_normal((n,n)) + 0.2j*rng.standard_normal((n,n))  # non-normal
+    r0 = rng.standard_normal(n) + 1j*rng.standard_normal(n)
+    r = r0.copy(); x = np.zeros(n, complex)
+    p = [r.copy()]; q = [A@r]; qq = [np.vdot(q[0],q[0]).real]
+    log = []
+    for k in range(m):
+        a = np.vdot(q[k], r)/qq[k]                               # <q,r>/<q,q>
+        x = x + a*p[k]; r = r - a*q[k]
+        z = r; Az = A@z; pn = z.copy(); qn = Az.copy(); bs = []
+        for back in range(min(k+1, mmax-1)):
+            b = -(np.vdot(q[k-back], Az)/qq[k-back]).real        # real part, as Grid does
+            pn = pn + b*p[k-back]; qn = qn + b*q[k-back]; bs.append(complex(b))
+        p.append(pn); q.append(qn); qq.append(np.vdot(qn,qn).real)
+        log.append((a, bs))
+    S, R = polynomials(log)
+    def apply(poly, v):
+        out = np.zeros_like(v); Ak = v.copy()
+        for c in poly: out = out + c*Ak; Ak = A@Ak
+        return out
+    err_r = np.linalg.norm(apply(R, r0) - r)/np.linalg.norm(r)
+    err_x = np.linalg.norm(apply(S, r0) - x)/np.linalg.norm(x)
+    print(f"selftest: |R_m(A)r0 - r_m|/|r_m| = {err_r:.2e}   |S(A)r0 - x_m|/|x_m| = {err_x:.2e}")
+    return err_r < 1e-10 and err_x < 1e-10
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser()
+    ap.add_argument("log", nargs="?")
+    ap.add_argument("--name", default="Fsmoother")
+    ap.add_argument("--lmin", type=float, default=0.0)
+    ap.add_argument("--lmax", type=float, default=3.0)
+    ap.add_argument("--n", type=int, default=600)
+    ap.add_argument("--selftest", action="store_true")
+    args = ap.parse_args()
+    if args.selftest:
+        sys.exit(0 if selftest() else 1)
+    calls = parse(args.log, args.name)
+    if not calls:
+        sys.exit(f"no '{args.name} coeff[...]' lines found in {args.log}")
+    mean, spread = mean_coeffs(calls)
+    print(f"{args.name}: {len(calls)} calls, {len(mean)} steps")
+    for k,(a,bs) in enumerate(mean):
+        print(f"  step {k}: a = ({a.real:+.5f},{a.imag:+.5f})  rel spread {spread[k][0]:.3f}   "
+              + "  ".join(f"b[{j}]={b.real:+.5f} (spread {spread[k][1][j]:.3f})" for j,b in enumerate(bs)))
+    S, R = polynomials(mean)
+    print("  S coefficients (low->high):", np.array2string(S, precision=5))
+    lam = np.linspace(args.lmin, args.lmax, args.n)
+    Rm = Poly.polyval(lam, R); Sm = Poly.polyval(lam, S)
+    per = np.array([np.abs(Poly.polyval(lam, polynomials(c)[1])) for c in calls])
+    out = f"{args.name}_poly.dat"
+    with open(out, "w") as f:
+        f.write("# lambda |R_mean| ReR ImR |R|min |R|max |S_mean|\n")
+        for i,l in enumerate(lam):
+            f.write(f"{l:.6f} {abs(Rm[i]):.6e} {Rm[i].real:.6e} {Rm[i].imag:.6e} "
+                    f"{per[:,i].min():.6e} {per[:,i].max():.6e} {abs(Sm[i]):.6e}\n")
+    print(f"  wrote {out}")

skills/multigrid-design-notes.md

@@ -221,11 +221,11 @@ judgement.
   the exact inverse of whatever matrix it was handed. Both certify the import
   and inversion, not the accuracy of `A_c`.
   The instruments that DO see it: coarsen twice and compare `BLAS_A` element by
-  element against an exact run (the comparison `Test_coarse_v2_coarsen` already
-  performs between V1 and V2), or the outer iteration count and time to
+  element against an exact run (the comparison `Test_coarse_coarsen` already
+  performs against the deprecated operator), or the outer iteration count and time to
   solution.
 - **Does not reach the coarse levels.** `SloppyComms` sets a flag on
-  `Stencil`/`StencilEven`/`StencilOdd`; V2's `M` uses `PaddedCell::Face_exchange`,
+  `Stencil`/`StencilEven`/`StencilOdd`; the coarse operator.s `M` uses `PaddedCell::Face_exchange`,
   which sends raw `vobj` bytes with no precision option. Coarse levels are less
   bandwidth sensitive, so this matters less than it sounds.
 - **Object sharing caveat:** the flag is per fermion operator. A single

systems/Frontier/nrhs_fault.job

@@ -66,7 +66,7 @@ export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export MPICH_OFI_NIC_POLICY=GPU
 module load libfabric
 
-BIN=$root/examples/Example_pvdagm_v2_3level_DenseCoarseMatrix
+BIN=$root/examples/Example_pvdagm_3level_DenseCoarseMatrix
 OPTS1="--accelerator-threads 8 --shm 4096 --shm-mpi 1"
 # LOCAL VOLUME MATCHED to production: 48.48.48.96 on 3.6.4.4 is local 16.8.12.24; on 8 GCDs
 # (1.1.2.4) that is global 16.8.24.96 with the production blockings 2.2.3.3 / 4.4.2.4:
@@ -112,7 +112,7 @@ run_cell () {
   logopt=""; [ -n "$GRID_LOG_EXTRA" ] && logopt="--log Error,Warning,Message,$GRID_LOG_EXTRA"
   env "$@" srun -N1 -n8 --kill-on-bad-exit=1 ./select_gpu $BIN --mpi ${MPI_GEOM} --grid $vol $OPTS1 --device-mem $devmem --comms-overlap $logopt > $fname 2>&1
   echo "  exit $?"; sleep 10
-  grep -h "V2 3-level solve\|Memory access fault\|GRID_ASSERT\|Converged on iteration" $fname | sed 's/^Grid : Message : [0-9.]* s : //' | tail -4 | cut -c1-140
+  grep -h "3-level solve\|Memory access fault\|GRID_ASSERT\|Converged on iteration" $fname | sed 's/^Grid : Message : [0-9.]* s : //' | tail -4 | cut -c1-140
   grep -n -B3 "Memory access fault" $fname | tail -5 | cut -c1-160
 }
 
@@ -124,5 +124,5 @@ run_cell   F3_nrhs6_nodense 16000 NRHS=6 DENSE_CC=0
 run_cell   F4_nrhs6_bigmem 48000 NRHS=6                                                    # no MemoryManager eviction
 
 echo "========================================================="
-for f in log.fault.F*; do echo "$f: $(grep -c 'Memory access fault' $f) faults; $(grep -h 'V2 3-level solve' $f | tail -1 | sed 's/.*V2/V2/' | cut -c1-60)"; done
+for f in log.fault.F*; do echo "$f: $(grep -c 'Memory access fault' $f) faults; $(grep -h '3-level solve' $f | tail -1 | sed 's/.*3-level/3-level/' | cut -c1-60)"; done
 echo "========================================================="

systems/Frontier/nrhs_fault_36.job

@@ -67,7 +67,7 @@ export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export MPICH_OFI_NIC_POLICY=GPU
 module load libfabric
 
-BIN=$root/examples/Example_pvdagm_v2_3level_DenseCoarseMatrix
+BIN=$root/examples/Example_pvdagm_3level_DenseCoarseMatrix
 OPTS1="--accelerator-threads 8 --shm 4096 --shm-mpi 1 --device-mem 32000"
 OVERLAP="--comms-overlap"          # set empty for a no-overlap job
 vol=48.48.48.96
@@ -112,7 +112,7 @@ run_cell () {
     tail -15 $(dirname $f)/Grid.stdout.$rank | cut -c1-160
   else
     echo "  no 'Memory access fault' in any Grid.stderr.*"
-    grep -h "V2 3-level solve\|Fouter MrhsPGCR: Converged" $GRID_STDOUT_ROOT/0/Grid.stdout.0 | tail -3 | cut -c1-120
+    grep -h "3-level solve\|Fouter MrhsPGCR: Converged" $GRID_STDOUT_ROOT/0/Grid.stdout.0 | tail -3 | cut -c1-120
   fi
 }
 

systems/Frontier/pvdagm_mixed_precision.job

@@ -0,0 +1,206 @@
+#!/bin/bash -l
+#SBATCH --job-name=pvdagm-mixed-precision
+#SBATCH --nodes=36
+#SBATCH --ntasks-per-node=8
+#SBATCH --cpus-per-task=7
+#SBATCH --gpus-per-node=8
+#SBATCH --time=2:00:00
+#SBATCH --account=phy157_dwf
+#SBATCH --gpu-bind=none
+#SBATCH --exclusive
+#SBATCH --mem=0
+#SBATCH -S 0
+
+##############################################################################
+# The three precision axes of the PVdagM multigrid, on the banked point of
+# pvdagm_multigrid.job (48^3x96, Ls=24, nbasis=60, 288 GCDs).  The outer
+# Krylov is fp64 and exact in every cell, so fp32 below it costs convergence
+# rate only, never correctness -- the FINAL exact-halo true residual is the
+# proof and is printed by every cell.
+#
+#   1. fine level inside the preconditioner   run time  <FinePrecision>
+#   2. coarse + coarse-coarse sector          compile   -DCOARSE_SINGLE
+#   3. distributed dense coarse-coarse invert compile   -DGRID_DENSE_INVERSE_SINGLE
+#
+# The compile-time axes are separate binaries from the same source, all
+# plain `make` targets (examples/Makefile.am):
+#   Example_pvdagm_multigrid                          fp64 coarse, fp64 dense
+#   Example_pvdagm_multigrid_fp32coarse               fp32 coarse, fp64 dense
+#   Example_pvdagm_multigrid_fp32coarse_fp32dense     fp32 coarse, fp32 dense
+# The default basis is 60.  The subspace load reads the first 60 vectors
+# of the file.
+##############################################################################
+
+LUSTRE=/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle
+RUNDIR=$LUSTRE/runs/${SLURM_JOB_NAME}_${SLURM_JOB_ID}
+mkdir -p $RUNDIR && cd $RUNDIR && echo "RUNDIR $RUNDIR"
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3 7 6 5 4)
+export NUMA_MAP=(3 3 1 1 2 2 0 0)
+export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
+export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
+export HIP_VISIBLE_DEVICES=\$GPU
+unset ROCR_VISIBLE_DEVICES
+if [ \$SLURM_PROCID = "0" ]; then echo \$*; fi
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+chmod +x ./select_gpu
+
+root=/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/MGrewrite/Grid/systems/Frontier
+source $root/sourceme-rocm7.2.sh
+# sourceme sets FI_HMEM_ROCR_USE_DMABUF=0 -- the standing avoidance of the CXI
+# NO_TRANSLATION fault at NRHS>=12 (libfabric #12775).
+
+export OMP_NUM_THREADS=7
+ulimit -c 0                             # no 22 GB GPU core dumps
+export FI_MR_CACHE_MONITOR=kdreg2       # site default; device-buffer MPI on Slingshot
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_OFI_NIC_POLICY=GPU
+export GRID_ALLOC_NCACHE_LARGE=64       # allocator large-ring depth, not a solver knob
+module load libfabric
+
+C64D64=$root/examples/Example_pvdagm_multigrid
+C32D64=$root/examples/Example_pvdagm_multigrid_fp32coarse
+C32D32=$root/examples/Example_pvdagm_multigrid_fp32coarse_fp32dense
+# 40000, not 32000: at Nrhs 12 the working set is about 37 GB (outer restart
+# history 24.5, fine smoother history 8.2, sources and solutions 4.1), so a
+# 32 GB cache evicts.  Measured: 808 ms of device-to-host copies in a 20 s
+# window, and 9.706 s per rhs against 8.748 with the larger cache.  The
+# non-evictable total is 6.3 GiB in the fp32 sector, so 40 GB of cache plus
+# comms and shm still fits a 64 GB device.
+OPTS="--accelerator-threads 8 --shm 4096 --shm-mpi 1 --device-mem 40000 --comms-overlap"
+vol=48.48.48.96
+MPI_GEOM=3.6.4.4
+
+SUBSPACE=$LUSTRE/subspace_nb64.scidac
+CONFIG=/ccs/home/poare/ckpoint_lat.1000
+
+##############################################################################
+# write_params <file> <FinePrecision> <FineSloppyComms>
+#
+# Everything else is the banked adaptive operating point of
+# pvdagm_multigrid.job, unchanged, so the cells differ only in precision.
+##############################################################################
+write_params () {
+  local f=$1 fprec=$2 sloppy=$3
+cat << EOF > $f
+<?xml version="1.0"?>
+<grid>
+  <PVdagMDriver>
+    <Ls>24</Ls>
+    <Mass>0.00078</Mass>
+    <M5>1.8</M5>
+    <MobiusB>1.5</MobiusB>
+    <MobiusC>0.5</MobiusC>
+    <Config>$CONFIG</Config>
+    <Nrhs>12</Nrhs>
+    <SolveSingleRHS>1</SolveSingleRHS>
+    <MultiGrid>
+      <Setup>
+        <Block1><elem>2</elem><elem>2</elem><elem>3</elem><elem>3</elem></Block1>
+        <Block2><elem>4</elem><elem>4</elem><elem>2</elem><elem>4</elem></Block2>
+        <CoarsenBatch>9</CoarsenBatch>
+        <SubspaceFile>$SUBSPACE</SubspaceFile>
+        <FineSloppyComms>$sloppy</FineSloppyComms>
+        <FinePrecision>$fprec</FinePrecision>
+        <RetainSubspace>0</RetainSubspace>
+      </Setup>
+      <FineSmoother>
+        <Shift>0.1</Shift><Nstep>6</Nstep><Mmax>4</Mmax>
+      </FineSmoother>
+      <CoarseSmoother>
+        <Shift>2.0</Shift><Nstep>2</Nstep><Mmax>2</Mmax>
+      </CoarseSmoother>
+      <CoarseSolver>
+        <Tol>0.05</Tol><Order>200</Order><Mmax>8</Mmax>
+      </CoarseSolver>
+      <Outer>
+        <Tol>1e-8</Tol><MaxIterations>1000</MaxIterations><Mmax>6</Mmax><Nstep>12</Nstep>
+      </Outer>
+      <Dense>
+        <LeafSpan>9</LeafSpan>
+      </Dense>
+    </MultiGrid>
+  </PVdagMDriver>
+</grid>
+EOF
+}
+
+##############################################################################
+# run_cell <name> <binary> <FinePrecision> <FineSloppyComms>
+##############################################################################
+run_cell () {
+  name=$1; bin=$2; fprec=$3; sloppy=$4
+  if [ ! -x "$bin" ]; then
+    echo "----- $name : SKIPPED, no binary $bin"
+    echo "        build it with a plain make in examples/"
+    return
+  fi
+  write_params params.$name.xml $fprec $sloppy
+  echo "----- $name : fine=$fprec sloppy=$sloppy bin=$(basename $bin) -----"
+  export GRID_STDOUT_ROOT=$RUNDIR/mg_$name
+  srun -N36 -n288 --kill-on-bad-exit=1 ./select_gpu $bin --mpi ${MPI_GEOM} --grid $vol $OPTS \
+       --pvdagm-params params.$name.xml --debug-stdout --log Error,Warning,Message,Performance \
+       > log.mg.$name 2>&1
+  echo "  exit $?"; sleep 30
+  # Device OOM is reported on stderr as "hipMalloc failed for <bytes> out of
+  # memory" and nowhere else, so it must be in this pattern or a cell that
+  # died of it looks like a cell that printed nothing.
+  f=$(grep -l "Memory access fault\|NO_TRANSLATION\|hipMalloc failed\|out of memory\|illegal memory access" $GRID_STDOUT_ROOT/*/Grid.stderr.* 2>/dev/null | head -1)
+  if [ -n "$f" ]; then
+    echo "  FAULT in $f  (rank $(basename $f | sed 's/Grid.stderr.//'))"; tail -20 $f | cut -c1-160
+  else
+    echo "  certificates, convergence, timing:"
+    grep -h "Coarse sector precision\|GALERKIN CERTIFICATE\|IMPORT CERTIFICATE\|inversion-source import\|SCHUR .* distributed invert\|VERIFY\|PVdagMMultiGridSolver: Nrhs\|MrhsPGCR: Converged\|FINAL Nrhs .*worst" \
+         $GRID_STDOUT_ROOT/0/Grid.stdout.0 2>/dev/null | cut -c1-150
+  fi
+}
+
+##############################################################################
+# Cells.  P0 is the regression gate: with fp64 everywhere it must reproduce
+# the banked outer count and residual of pvdagm_multigrid.job.  If it does
+# not, nothing downstream means anything, and the Galerkin certificate
+# (2.3e-15 expected in fp64) separates a coarsening change from a solve one.
+##############################################################################
+
+#           name          binary   fineprec sloppy
+run_cell    P0_fp64       $C64D64  fp64     1
+run_cell    P1_fp32fine   $C64D64  fp32     1
+run_cell    P2_fp32coarse $C32D64  fp64     1
+run_cell    P3_fp32both   $C32D64  fp32     1
+run_cell    P4_denseF32   $C32D32  fp32     1
+
+# Halo wire format is orthogonal to arithmetic: on an fp32 operator
+# FineSloppyComms is bf16 compression, on an fp64 operator it is fp32.  One
+# exact-halo cell to price it inside the preconditioner.
+run_cell    P5_exacthalo  $C32D64  fp32     0
+
+##############################################################################
+# What to read.
+#
+# P0: the gate.  Outer count and s/RHS against the banked numbers.
+#
+# P1 vs P0: fp32 fine level.  Laptop 8^4 saw the same outer count and ~12%
+# less wall; here the fine level is the dominant cost and the halo is real,
+# so this is where the fp32 case is actually made or lost.
+#
+# P2 vs P0: fp32 coarse sector.  Watch the Galerkin certificate move from
+# ~2e-15 to ~3e-7 -- that is fp32 rounding in the coarsening, not an error --
+# and check the outer count does not move with it.  Halves the coarse
+# operator's storage and its comms.
+#
+# P4 vs P3: fp32 dense coarse-coarse inversion.  The reading is the VERIFY
+# certificate ||A Ainv x - x||/||x||, which scales as kappa(A_cc) * eps32 and
+# so measures the conditioning of the coarse-coarse operator at production
+# size.  The laptop's N=128 was too small to say anything (5.3e-6 -> 6.7e-6);
+# at N=69120 it is a real test.  If VERIFY degrades but the outer count holds,
+# the fp32 factorisation is fine and halves that footprint; if the outer count
+# moves, the fallback is fp32 factorise plus one fp64 refinement step, which
+# is two extra dense applies and is not yet written.
+#
+# P5 vs P3: what the reduced-precision halo is worth once the arithmetic is
+# already fp32 (bf16 wire on an fp32 operator).
+##############################################################################

systems/Frontier/pvdagm_multigrid.job

@@ -70,7 +70,7 @@ export GRID_ALLOC_NCACHE_LARGE=0       # allocator large-ring depth (Grid alloca
 module load libfabric
 
 BIN=$root/examples/Example_pvdagm_multigrid
-OPTS="--accelerator-threads 8 --shm 4096 --shm-mpi 0 --device-mem 32000 --comms-overlap"
+OPTS="--accelerator-threads 8 --shm 4096 --shm-mpi 0 --device-mem 40000 --comms-overlap"
 vol=48.48.48.96
 MPI_GEOM=3.6.4.4
 

systems/Frontier/schur2d_ladder.job

@@ -131,14 +131,14 @@ export NRHS=4
 
 echo "----- F4a: 1D baseline (DENSE_SCHUR2D unset) -----"
 unset DENSE_SCHUR2D
-srun -N36 -n288 ./select_gpu $root/examples/Example_pvdagm_v2_3level_DenseCoarseMatrix \
+srun -N36 -n288 ./select_gpu $root/examples/Example_pvdagm_3level_DenseCoarseMatrix \
      --mpi ${MPI_GEOM} --grid $vol $OPTS1 --comms-overlap
 
 if [ $F3RC -eq 0 ]
 then
 echo "----- F4b: 2D block-cyclic (DENSE_SCHUR2D=1) -----"
 export DENSE_SCHUR2D=1
-srun -N36 -n288 ./select_gpu $root/examples/Example_pvdagm_v2_3level_DenseCoarseMatrix \
+srun -N36 -n288 ./select_gpu $root/examples/Example_pvdagm_3level_DenseCoarseMatrix \
      --mpi ${MPI_GEOM} --grid $vol $OPTS1 --comms-overlap
 else
 echo "----- F4b SKIPPED: F3 failed (rc=$F3RC), not spending the setup on it -----"

systems/Frontier/smoother_modes.job

@@ -121,10 +121,10 @@ run_cell () {
   export FineSmootherMode=$5; export CoarseSmootherMode=$6
   echo "----- $name : Fso=$FineSmootherOrder Fss=$FineSmootherShift Csn=$CoarseSmootherNstep fine=$FineSmootherMode coarse=$CoarseSmootherMode -----"
   fname=log.ladder.$name
-  srun -N36 -n288 --kill-on-bad-exit=1 ./select_gpu $root/examples/Example_pvdagm_v2_3level_DenseCoarseMatrix \
+  srun -N36 -n288 --kill-on-bad-exit=1 ./select_gpu $root/examples/Example_pvdagm_3level_DenseCoarseMatrix \
        --mpi ${MPI_GEOM} --grid $vol $OPTS1 --comms-overlap > $fname 2>&1
   echo "  exit $?"; sleep 60
-  echo "  $(grep -h 'V2 3-level solve Nrhs' $fname | sed 's/.*V2/V2/' | tr '\n' ' ')"
+  echo "  $(grep -h '3-level solve Nrhs' $fname | sed 's/.*3-level/3-level/' | tr '\n' ' ')"
   echo "  $(grep -h 'Fouter MrhsPGCR: Converged' $fname | sed 's/.*Converged/Converged/' | cut -c1-60 | tr '\n' ' ')"
   echo "  replay per-call |r|/|r0| (Nrhs=1 solve): $(awk '/THREE-level solve, Nrhs = 1/{s=1} s && /Fsmoother replay \|r\|/{v=$NF; n++; t+=v; if(v>mx)mx=v} END{if(n) printf "mean %.4f max %.4f over %d calls", t/n, mx, n}' $fname)"
   grep -h "SCHUR fp64 distributed invert took\|GB/s/rank\|BIG LEAVES\|ring histogram\|^Grid : Message : [0-9.]* s :    >=" $fname | sed 's/^Grid : Message : [0-9.]* s : //' | cut -c1-150 | head -12

systems/WorkArounds.txt

@@ -9,7 +9,7 @@ Contents:
 3. Profiling
 
 ************************
-* 1. INTERCONNECT + MPI
+* 0. MPI IO
 ************************
 
 --------------------------------------------------------------------
@@ -27,6 +27,89 @@ https://github.com/paboyle/Grid/issues/381
 
 https://github.com/pmodels/mpich/commit/3a479ab0
 
+--------------------------------------------------------------------
+MPI2-IO performance: introduced All to all based aggregation no changes required.
+--------------------------------------------------------------------
+
+https://github.com/pmodels/mpich/issues/7935
+
+MPI_File_write_all / read_all collective buffering does not scale for 4D block-decomposed subarray views (Lustre): 6x slower than a user-level MPI_Alltoallv + POSIX transposition
+
+/* phase 1: redistribute within the row subcommunicator.
+   sendcounts/senddispls derived analytically from the plan; iodata untouched. */
+MPI_Alltoallv(iodata,     sendcounts, senddispls, MPI_BYTE,
+              aggregated, recvcounts, recvdispls, MPI_BYTE, rowcomm);
+
+/* phase 2: a handful of large contiguous writes per rank, no MPI-IO involved */
+for (e = 0; e < nextent; e++) {
+    lseek(fd, offset + extentGsite[e]*sizeof(fobj), SEEK_SET);
+    write(fd, aggregated + extentLocal[e], extentSites[e]*sizeof(fobj));
+}
+
+************************
+* 1. INTERCONNECT + MPI
+************************
+
+============================================================================
+Piz Daint
+============================================================================
+
+https://github.com/ofiwg/libfabric/issues/11451
+
+SEGFAULT in first MPI_Comm_dup on more than one node, after MPI_Init
+ 
+Cause: symptom (ALPS/CSCS, aarch64 Grace/H200, cray-mpich 8.1.32, libfabric 1.22): the
+memhooks monitor intercepts munmap by PATCHING THE PLT at MPI_Init
+(ofi_memhooks_start -> ofi_write_patch with a garbage data_size); the write overruns
+munmap@plt into the NEXT PLT entry (MPI_Comm_dup in Grid's binary)
+leaving
+`br x15` where an adrp belongs ->
+
+Diagnosed with a hardware watchpoint on the PLT entry during MPI_Init; see the issue for the
+gdb transcript.  Reproduce on one node with MPICH_SINGLE_HOST_ENABLED=0.
+
+Fix (either):
+                export FI_MR_CACHE_MONITOR=kdreg2
+OR                                                 
+                export FI_MR_CACHE_MONITOR=disabled
+                export FI_MR_CACHE_MAX_COUNT=0
+
+On ALPS this is a CORRECTNESS requirement for MPI on Slingshot, not a tuning.
+
+
+============================================================================
+Frontier
+============================================================================
+
+a) FI_MR_CACHE_MONITOR=kdreg2 (site default) leads to Runtime MPI errors with "NO_TRANSLATION"
+   userfaultfd and memhooks also fail.
+
+https://github.com/ofiwg/libfabric/issues/12775
+
+Symptom (Frontier, x86, Cray MPICH 8.1.x, ROCm 7.2): device-buffer MPI fails with
+   MPI_Waitall ... MPIDI_OFI_handle_cq_error: OFI poll failed
+   (ofi_events.c:MPIDI_OFI_handle_cq_error: Input/output error - NO_TRANSLATION)
+on a LIVE, never-freed hipMalloc buffer
+
+
+b)              export FI_MR_CACHE_MONITOR=disabled
+                export FI_MR_CACHE_MAX_COUNT=0
+
+Leads to INCORRECT RESULTS, see 
+
+https://github.com/ofiwg/libfabric/issues/12773
+
+Fix:
+
+export FI_HMEM_ROCR_USE_DMABUF=0
+and use site default
+FI_MR_CACHE_MONITOR=kdreg2
+
+Every systems/Frontier job and sourceme now sets kdreg2 on OLCF's recommendation;
+whether it resolves the Frontier NO_TRANSLATION is the pending test above.
+
+
+
 --------------------------------------------------------------------
 Slingshot: Frontier and Perlmutter libfabric slow down 
 and physical memory fragmentation 
@@ -187,7 +270,6 @@ unset MPIR_CVAR_COLL_SELECTION_TUNING_JSON_FILE
 unset MPIR_CVAR_CH4_POSIX_COLL_SELECTION_TUNING_JSON_FILE
 
 
-
 ************************
 * 3. Visual profile tools
 ************************
@@ -196,59 +278,97 @@ unset MPIR_CVAR_CH4_POSIX_COLL_SELECTION_TUNING_JSON_FILE
 Frontier/rocprof
 --------------------------------------------------------------------
 
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3 7 6 5 4)
+export NUMA_MAP=(3 3 1 1 2 2 0 0)
+export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
+export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
+export HIP_VISIBLE_DEVICES=\$GPU
+unset ROCR_VISIBLE_DEVICES
+if [ \$SLURM_PROCID = "0" ]; then echo \$*; fi
+
+if [ "1" = "1" ]
+then
+if [ \$SLURM_PROCID = "0" ]
+then
+  exec numactl -m \$NUMA -N \$NUMA rocprofv3 -d $TRACE -o mgtrace.\$SLURM_JOBID \
+     -r -f csv pftrace --stats -D \
+     -P 140:20:1 \
+     -- \$*
+else 
+  exec numactl -m \$NUMA -N \$NUMA \$*
+fi
+else 
+  exec numactl -m \$NUMA -N \$NUMA \$*
+fi
+
+EOF
+chmod +x ./select_gpu
+
 --------------------------------------------------------------------
 Aurora/unitrace
 --------------------------------------------------------------------
+select_gpu:
 
+#!/bin/bash
+
+export NUMA_PMAP=(0 0 0 1 1 1 0 0 0 1 1 1 );
+export NUMA_HMAP=(2 2 2 3 3 3 3 2 2 2 2 3 3 3 );
+export  GPU_MAP=(0.0 1.0 2.0 3.0 4.0 5.0 0.1 1.1 2.1 3.1 4.1 5.1 )
+
+export NUMAP=${NUMA_PMAP[$PALS_LOCAL_RANKID]}
+export NUMAH=${NUMA_HMAP[$PALS_LOCAL_RANKID]}
+export gpu_id=${GPU_MAP[$PALS_LOCAL_RANKID]}
+  
+unset EnableWalkerPartition
+export EnableImplicitScaling=0
+export ZE_AFFINITY_MASK=$gpu_id
+export ONEAPI_DEVICE_FILTER=gpu,level_zero
+
+export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=0
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:4
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
+echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "
+
+if [ $PALS_RANKID = "0" ]
+then
+    numactl -p $NUMAP -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
+else 
+    numactl -p $NUMAP -N $NUMAP  "$@"
+fi
 
 --------------------------------------------------------------------
 Tursa/nsight-sys
 --------------------------------------------------------------------
 
-============================================================================
-2026-08-28  libfabric memory-registration-cache "memhooks" monitor -- DEFECTIVE
-           (libfabric issue #11451, filed by PB, HPE JIRA opened)
-============================================================================
-STATUS: the ALPS/aarch64 case below is DEMONSTRATED (issue + both fixes verified
-there).  The Frontier case is a HYPOTHESIS as of 2026-08-28: the NO_TRANSLATION
-failure is observed, its attribution to the memhooks MR cache is by mechanism and
-by analogy, and the fix is UNTESTED here.  Test ladder, one knob per run, NRHS=12:
-  (1) FineSloppyComms=0  (2) FI_MR_CACHE_MONITOR=kdreg2  (3) FI_MR_CACHE_MAX_COUNT=0
-  (4) --disable-accelerator-aware-mpi build.
-REVISIT this entry with the result; if (2) does not cure it, remove kdreg2 from the
-job scripts' justification (it stays as OLCF's recommendation regardless).
-  2026-08-28 22:18, job 5372414: step (2) DID NOT cure it -- kdreg2 set (and the
-  environment already carried FI_MR_CACHE_MAX_COUNT=786432), sloppy comms OFF, still
-  NO_TRANSLATION at outer step 60 (Couter, Waitall count=2).
-  CORRECTION 2026-08-28 23:xx (fi_mr(3) man page): FI_MR_CACHE_MONITOR governs SYSTEM
-  memory only; device (HMEM_ROCR) registrations are monitored by
-  FI_MR_ROCR_CACHE_MONITOR_ENABLED=0|1.  Grid's comms window is hipMalloc'd device
-  memory, so the kdreg2 step tested nothing relevant -- the ALPS PLT bug (#11451) and
-  the Frontier NO_TRANSLATION are the same *library* but different monitors.
-  Attributed by gdb on a hung rank (5371826): main thread in MPI_Waitall <- Grid
-  0x5d539c3 (StencilSendToRecvFromComplete/CommsComplete class), i.e. the HALO
-  EXCHANGE; the two aborts' request counts (16 = fine Dhop, 2 = coarse PaddedCell
-  direction) agree.  Next: FI_MR_CACHE_MAX_COUNT=0 (all memory) and
-  FI_MR_ROCR_CACHE_MONITOR_ENABLED=0, one per cell.
+cat << EOF > mpiwrapper.sh
+#!/bin/bash
+lrank=\$OMPI_COMM_WORLD_LOCAL_RANK
+numa1=\$(( 2 * \$lrank ))
+numa2=\$(( 2 * \$lrank + 1 ))
+export CUDA_VISIBLE_DEVICES=\$lrank
+export UCX_NET_DEVICES=mlx5_\${lrank}:1
+if [ \$OMPI_COMM_WORLD_RANK = "0" ]; then echo \$*; fi
 
-Symptom (Frontier, x86, Cray MPICH 8.1.x, ROCm 7.2): device-buffer MPI fails
-with
-   MPI_Waitall ... MPIDI_OFI_handle_cq_error: OFI poll failed
-   (ofi_events.c:MPIDI_OFI_handle_cq_error: Input/output error - NO_TRANSLATION)
-on a LIVE, never-freed hipMalloc buffer (the sloppy-comms compressed halo buffer,
-a static deviceVector) once the process does sustained hipMalloc/hipFree churn
-(MemoryManager eviction at NRHS>=12, EvictAll/DropCache).  Translations cached by
-the provider go stale.
-Symptom (ALPS/CSCS, aarch64 Grace/H200, cray-mpich 8.1.32, libfabric 1.22): the
-memhooks monitor intercepts munmap by PATCHING THE PLT at MPI_Init
-(ofi_memhooks_start -> ofi_write_patch with a garbage data_size); the write overruns
-munmap@plt into the NEXT PLT entry (MPI_Comm_dup in Grid's binary), leaving
-`br x15` where an adrp belongs -> segfault on first MPI_Comm_dup.  Diagnosed with
-a hardware watchpoint on the PLT entry during MPI_Init; see the issue for the
-gdb transcript.  Reproduce on one node with MPICH_SINGLE_HOST_ENABLED=0.
-Fix (either):   export FI_MR_CACHE_MONITOR=kdreg2      (kernel-driven invalidation; OLCF's
-                                                        own recommendation, NOT the default)
-                export FI_MR_CACHE_MAX_COUNT=0         (no registration cache at all)
-On ALPS this is a CORRECTNESS requirement for MPI on Slingshot, not a tuning.
-Every systems/Frontier job and sourceme now sets kdreg2 on OLCF's recommendation;
-whether it resolves the Frontier NO_TRANSLATION is the pending test above.
+if [ "1" = "1" ]
+then
+if [ \$OMPI_COMM_WORLD_RANK = "0" ]
+then
+  exec numactl --interleave=\$numa1,\$numa2 nsys profile \
+     -o $TRACE/nsys.\$SLURM_JOB_ID --force-overwrite=true \
+     --trace=cuda,nvtx,mpi,osrt --mpi-impl=op
+     --sample=none --cpuctxsw=none \
+     --delay=140 --duration=20 --stats=true \
+     \$*
+else
+  exec numactl --interleave=\$numa1,\$numa2 \$*
+fi
+else
+  exec numactl --interleave=\$numa1,\$numa2 \
+fi
+EOF
+chmod +x ./mpiwrapper.sh
+
+mpirun -np $SLURM_NTASKS -x LD_LIBRARY_PATH --bind-to none ./mpiwrapper.sh <binary> <args>

tests/Test_fft_memory.cc

@@ -8,7 +8,7 @@
  then repeatedly applies FFT_all_dim to the same propagator 400 times.
 
  If PlannedFFT is working correctly the RSS should remain flat after the first
- iteration — no new plans, no new deviceVector allocations beyond the per-call
+ iteration -- no new plans, no new deviceVector allocations beyond the per-call
  pencil buffer which is freed at the end of each FFT_dim_execute call.
 
  Build exactly like any other Grid test, e.g.:
@@ -63,7 +63,7 @@ static long getGPUUsedMb()
 }
 
 // ============================================================
-//  Convenience struct — one snapshot of both sides
+//  Convenience struct -- one snapshot of both sides
 // ============================================================
 struct MemSnapshot {
   long cpu_rss_kb;   // host RSS in kB  (-1 if unavailable)
@@ -113,7 +113,7 @@ int main(int argc, char **argv)
             << "Grid is setup to use " << threads << " threads" << std::endl;
 
   // ------------------------------------------------------------------
-  // Grid setup — use whatever lattice/mpi/simd was passed on the CLI,
+  // Grid setup -- use whatever lattice/mpi/simd was passed on the CLI,
   // e.g.  --grid 8.8.8.8 --mpi 1.1.1.1
   // ------------------------------------------------------------------
   Coordinate latt_size   = GridDefaultLatt();
@@ -174,13 +174,13 @@ int main(int argc, char **argv)
             << std::endl;
  
   // ------------------------------------------------------------------
-  // Create the PlannedFFT — plans are allocated here ONCE for all
+  // Create the PlannedFFT -- plans are allocated here ONCE for all
   // dimensions and stored inside the object.
   // ------------------------------------------------------------------
   PlannedFFT<iSpinColourMatrix<vComplexD>> plannedFFT(&GRID);
 
   // ------------------------------------------------------------------
-  // Snapshot AFTER plan construction — this is the true baseline
+  // Snapshot AFTER plan construction -- this is the true baseline
   // for the loop, because cufftPlanMany itself grabs device memory.
   // ------------------------------------------------------------------
   MemSnapshot snap_after_plan = takeSnapshot();
@@ -197,7 +197,7 @@ int main(int argc, char **argv)
   // ------------------------------------------------------------------
   // 400-iteration loop.
   // Each iteration computes the full 4d forward FFT of `prop`.
-  // We deliberately do NOT cache the result — we always start from
+  // We deliberately do NOT cache the result -- we always start from
   // the same `prop` so the FFT is recomputed identically each time.
   // The point is to watch memory, not correctness.
   // ------------------------------------------------------------------
@@ -229,7 +229,7 @@ int main(int argc, char **argv)
     }
  
     // cudaMemGetInfo reflects the state *after* any pooled frees have
-    // been committed, so this is accurate without an explicit sync —
+    // been committed, so this is accurate without an explicit sync --
     // FFT_dim_execute already calls accelerator_barrier() internally.
     MemSnapshot snap_now = takeSnapshot();
     printRow(i, snap_now, snap_prev);

tests/core/Test_planned_fft.cc

@@ -156,7 +156,7 @@ int main (int argc, char ** argv)
   }
 
   ////////////////////////////////////////////////////
-  // Dwf matrix — verify Fourier representation using PlannedFFT<LatticeFermionD>
+  // Dwf matrix -- verify Fourier representation using PlannedFFT<LatticeFermionD>
   ////////////////////////////////////////////////////
   {
     std::cout<<"****************************************"<<std::endl;

tests/debug/Test_blockcyclic.cc

@@ -20,7 +20,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 
 //////////////////////////////////////////////////////////////////////////////
 // Regression gate for BlockCyclicLayout -- stage 1 of the 2D distributed
-// dense inverse (documentation/DistributedDenseInverse2D.tex).
+// dense inverse.
 //
 // The layout is pure index arithmetic, so this test is EXHAUSTIVE rather
 // than statistical: every stage sweeps a battery of (N, nb, Pr, Pc)

tests/debug/Test_coarse.cc

@@ -2,7 +2,7 @@
 
     Grid physics library, www.github.com/paboyle/Grid
 
-    Source file: ./tests/debug/Test_coarse_v2.cc
+    Source file: ./tests/debug/Test_coarse.cc
 
     Copyright (C) 2026
 
@@ -27,17 +27,18 @@ Author: Peter Boyle <pboyle@bnl.gov>
 /*  END LEGAL */
 
 //
-// MultiGeneralCoarsenedOperatorV2 against the existing mrhs coarse operator.
+// MultiGeneralCoarsenedOperator against the existing mrhs coarse operator.
 //
-// V1 is constructed on the D+1 grid as now; V2 on the D dimensional grid,
-// with SetGrid() adopting the caller owned D+1 grid and building its padded
-// cell and neighbour table from the D dimensional stencil, with the Nrhs
-// factor multiplied in.
+// The reference (the deprecated operator) is constructed on the D+1 grid; the
+// operator under test on the D dimensional grid, with SetGrid() adopting the
+// caller owned D+1 grid and building its padded cell and neighbour table from
+// the D dimensional stencil, with the Nrhs factor multiplied in.
 //
 // Both are given identical matrix elements, so any difference in the apply is
 // the restructured neighbour table. The same geometry object is passed to
-// both: V1 adds one to skip for the rhs direction, V2 uses it as is on the D
-// dimensional grid, so both describe the same stencil over the D dimensions.
+// both: the reference adds one to skip for the rhs direction, the operator
+// under test uses it as is on the D dimensional grid, so both describe the
+// same stencil over the D dimensions.
 //
 #include <Grid/Grid.h>
 
@@ -48,8 +49,8 @@ const int nbasis = 8;
 typedef vSpinColourVector FineObj;
 typedef sTComplexD        CComplexT;   // unvectorised coarse space
 
-typedef MultiGeneralCoarsenedMatrix    <FineObj,CComplexT,nbasis> MrhsV1;
-typedef MultiGeneralCoarsenedOperatorV2<FineObj,CComplexT,nbasis> MrhsV2;
+typedef DeprecatedMultiGeneralCoarsenedMatrix    <FineObj,CComplexT,nbasis> RefOperator;
+typedef MultiGeneralCoarsenedOperator<FineObj,CComplexT,nbasis> TestOperator;
 
 ////////////////////////////////////////////////////////////////////////
 // Identical random matrix elements into both operators
@@ -59,11 +60,16 @@ void SeedMatrixElements(OpA &A,OpB &B,int npoint,GridSerialRNG &sRNG)
 {
   typedef typename OpA::calcMatrix calcMatrix;
 
+  deviceVector<calcMatrix> bufA,bufB;
   for(int p=0;p<npoint;p++){
 
-    GRID_ASSERT(A.BLAS_A[p].size() == B.BLAS_A[p].size());
+    // MatrixPointOut/In carry one stencil point in lSite order whichever
+    // internal layout the operator uses.
+    A.MatrixPointOut(p,bufA);
+    B.MatrixPointOut(p,bufB);
+    GRID_ASSERT(bufA.size() == bufB.size());
 
-    int64_t sites = A.BLAS_A[p].size();
+    int64_t sites = bufA.size();
     std::vector<calcMatrix> host(sites);
 
     ComplexD *w = (ComplexD *)&host[0];
@@ -75,8 +81,9 @@ void SeedMatrixElements(OpA &A,OpB &B,int npoint,GridSerialRNG &sRNG)
       w[i] = ComplexD(re-0.5,im-0.5);
     }
 
-    acceleratorCopyToDevice(&host[0],&A.BLAS_A[p][0],sites*sizeof(calcMatrix));
-    acceleratorCopyToDevice(&host[0],&B.BLAS_A[p][0],sites*sizeof(calcMatrix));
+    acceleratorCopyToDevice(&host[0],&bufA[0],sites*sizeof(calcMatrix));
+    A.MatrixPointIn(p,bufA);
+    B.MatrixPointIn(p,bufA);
   }
 }
 
@@ -85,10 +92,12 @@ RealD MatrixChecksum(Op &O,int npoint)
 {
   typedef typename Op::calcMatrix calcMatrix;
   RealD sum=0.0;
+  deviceVector<calcMatrix> buf;
   for(int p=0;p<npoint;p++){
-    int64_t sites = O.BLAS_A[p].size();
+    O.MatrixPointOut(p,buf);
+    int64_t sites = buf.size();
     std::vector<calcMatrix> host(sites);
-    acceleratorCopyFromDevice(&O.BLAS_A[p][0],&host[0],sites*sizeof(calcMatrix));
+    acceleratorCopyFromDevice(&buf[0],&host[0],sites*sizeof(calcMatrix));
     ComplexD *w = (ComplexD *)&host[0];
     int64_t words = sites*sizeof(calcMatrix)/sizeof(ComplexD);
     for(int64_t i=0;i<words;i++) sum += real(w[i])*real(w[i]) + imag(w[i])*imag(w[i]);
@@ -107,7 +116,7 @@ int main (int argc, char ** argv)
   Coordinate cmpi  = GridDefaultMpi();
 
   ////////////////////////////////////////////////
-  // D dimensional coarse grid, and D+1 for V1
+  // D dimensional coarse grid, and D+1 for the reference
   ////////////////////////////////////////////////
   GridCartesian *CoarseD = new GridCartesian(clatt,csimd,cmpi);
 
@@ -129,35 +138,35 @@ int main (int argc, char ** argv)
   }
   GridCartesian *CoarseMulti = new GridCartesian(mlatt,msimd,mmpi);
 
-  MrhsV2 OpV2(geom,CoarseD);
-  MrhsV1 OpV1(geom,CoarseMulti);
-  OpV2.SetGrid(CoarseMulti);
+  TestOperator OpTest(geom,CoarseD);
+  RefOperator OpRef(geom,CoarseMulti);
+  OpTest.SetGrid(CoarseMulti);
 
   std::cout << GridLogMessage << "coarse D+1 grid nrhs " << nrhs
 	    << "  Nsimd " << CoarseMulti->Nsimd() << std::endl;
 
-  std::cout << GridLogMessage << "npoint V1 " << OpV1.geom.npoint
-	    << "   npoint V2 " << OpV2.geom.npoint << std::endl;
-  GRID_ASSERT(OpV1.geom.npoint == OpV2.geom.npoint);
+  std::cout << GridLogMessage << "npoint ref " << OpRef.geom.npoint
+	    << "   npoint test " << OpTest.geom.npoint << std::endl;
+  GRID_ASSERT(OpRef.geom.npoint == OpTest.geom.npoint);
 
-  int npoint = OpV1.geom.npoint;
+  int npoint = OpRef.geom.npoint;
 
   ////////////////////////////////////////////////
   // Identical matrix elements
   ////////////////////////////////////////////////
   GridSerialRNG sRNG; sRNG.SeedFixedIntegers(std::vector<int>({7,8,9,10}));
-  SeedMatrixElements(OpV1,OpV2,npoint,sRNG);
+  SeedMatrixElements(OpRef,OpTest,npoint,sRNG);
 
-  RealD ckV1 = MatrixChecksum(OpV1,npoint);
-  RealD ckV2 = MatrixChecksum(OpV2,npoint);
-  std::cout << GridLogMessage << "matrix element checksum V1 " << ckV1
-	    << "  V2 " << ckV2 << std::endl;
-  GRID_ASSERT( ckV1 == ckV2 );
+  RealD ckRef = MatrixChecksum(OpRef,npoint);
+  RealD ckTest = MatrixChecksum(OpTest,npoint);
+  std::cout << GridLogMessage << "matrix element checksum ref " << ckRef
+	    << "  test " << ckTest << std::endl;
+  GRID_ASSERT( ckRef == ckTest );
 
   ////////////////////////////////////////////////
   // Same input, compare the applies
   ////////////////////////////////////////////////
-  typedef MrhsV1::CoarseVector CoarseVector;
+  typedef RefOperator::CoarseVector CoarseVector;
   // RNG on the D dimensional grid fills any D+1 field: the rhs direction is
   // undistributed and divides cleanly. One RNG serves every Nrhs.
   GridParallelRNG pRNG(CoarseD); pRNG.SeedFixedIntegers(std::vector<int>({1,2,3,4}));
@@ -167,22 +176,22 @@ int main (int argc, char ** argv)
   CoarseVector out2(CoarseMulti);
   CoarseVector err (CoarseMulti);
 
-  OpV1.M(in,out1);
-  OpV2.M(in,out2);
+  OpRef.M(in,out1);
+  OpTest.M(in,out2);
 
   err = out1 - out2;
-  std::cout << GridLogMessage << "|V1 out|^2 = " << norm2(out1)
-	    << "   |V2 out|^2 = " << norm2(out2) << std::endl;
-  std::cout << GridLogMessage << "|V1 - V2|^2 = " << norm2(err) << std::endl;
+  std::cout << GridLogMessage << "|ref out|^2 = " << norm2(out1)
+	    << "   |test out|^2 = " << norm2(out2) << std::endl;
+  std::cout << GridLogMessage << "|ref - test|^2 = " << norm2(err) << std::endl;
   GRID_ASSERT( norm2(out1) > 0.0 );
   GRID_ASSERT( norm2(err) == 0.0 );
 
   ////////////////////////////////////////////////
   // SetGrid is idempotent on pointer identity
   ////////////////////////////////////////////////
-  OpV2.SetGrid(CoarseMulti);
-  GRID_ASSERT( MatrixChecksum(OpV2,npoint) == ckV2 );
-  OpV2.M(in,out2);
+  OpTest.SetGrid(CoarseMulti);
+  GRID_ASSERT( MatrixChecksum(OpTest,npoint) == ckTest );
+  OpTest.M(in,out2);
   err = out1 - out2;
   GRID_ASSERT( norm2(err) == 0.0 );
   std::cout << GridLogMessage << "SetGrid idempotent on identity" << std::endl;
@@ -194,7 +203,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////
   // Nrhs 1 is the single RHS case through the multiRHS path, and each slice
   // of the Nrhs 4 apply must come back unchanged.
-  OpV2.M(in,out2);
+  OpTest.M(in,out2);
   for(int nr=2;nr>=1;nr--){
     Coordinate latt2(1,nr), simd2(1,1), mpi2(1,1);
     for(int d=0;d<Nd;d++){
@@ -204,9 +213,9 @@ int main (int argc, char ** argv)
     }
     GridCartesian *CoarseMulti2 = new GridCartesian(latt2,simd2,mpi2);
 
-    OpV2.SetGrid(CoarseMulti2);
-    GRID_ASSERT( OpV2.Nrhs() == nr );
-    GRID_ASSERT( MatrixChecksum(OpV2,npoint) == ckV2 );
+    OpTest.SetGrid(CoarseMulti2);
+    GRID_ASSERT( OpTest.Nrhs() == nr );
+    GRID_ASSERT( MatrixChecksum(OpTest,npoint) == ckTest );
 
     CoarseVector in2 (CoarseMulti2);
     CoarseVector out(CoarseMulti2);
@@ -215,7 +224,7 @@ int main (int argc, char ** argv)
       ExtractSliceFast(slice,in,r,0);
       InsertSliceFast(slice,in2,r,0);
     }
-    OpV2.M(in2,out);
+    OpTest.M(in2,out);
 
     RealD sdiff=0.0;
     for(int r=0;r<nr;r++){
@@ -231,22 +240,22 @@ int main (int argc, char ** argv)
     GRID_ASSERT( norm2(out) > 0.0 );
     GRID_ASSERT( sdiff/norm2(out) < 1.0e-20 );
 
-    OpV2.ReleaseGrid();
-    GRID_ASSERT( MatrixChecksum(OpV2,npoint) == ckV2 );  // survives release
+    OpTest.ReleaseGrid();
+    GRID_ASSERT( MatrixChecksum(OpTest,npoint) == ckTest );  // survives release
 
     delete CoarseMulti2;
   }
 
-  OpV2.SetGrid(CoarseMulti);
-  GRID_ASSERT( OpV2.Nrhs() == nrhs );
+  OpTest.SetGrid(CoarseMulti);
+  GRID_ASSERT( OpTest.Nrhs() == nrhs );
 
-  OpV2.M(in,out2);
+  OpTest.M(in,out2);
   err = out1 - out2;
-  std::cout << GridLogMessage << "after Nrhs 4 -> 2 -> 1 -> release -> 4, |V1 - V2|^2 = "
+  std::cout << GridLogMessage << "after Nrhs 4 -> 2 -> 1 -> release -> 4, |ref - test|^2 = "
 	    << norm2(err) << std::endl;
   GRID_ASSERT( norm2(err) == 0.0 );
 
-  std::cout << GridLogMessage << "Test_coarse_v2: ALL PASS" << std::endl;
+  std::cout << GridLogMessage << "Test_coarse: ALL PASS" << std::endl;
 
   Grid_finalize();
 }

tests/debug/Test_coarse_coarsen.cc

@@ -2,7 +2,7 @@
 
     Grid physics library, www.github.com/paboyle/Grid
 
-    Source file: ./tests/debug/Test_coarse_v2_coarsen.cc
+    Source file: ./tests/debug/Test_coarse_coarsen.cc
 
     Copyright (C) 2026
 
@@ -29,11 +29,13 @@ Author: Peter Boyle <pboyle@bnl.gov>
 //
 // Coarsen the same fine operator two ways and compare the matrix elements.
 //
-//   V1 : existing mrhs CoarsenOperator, vectorised coarse space matched to
-//        the fine SIMD layout, single RHS fine applications
-//   V2 : D+1 CoarsenOperator, unvectorised sComplexD coarse space, the batch
-//        of phased basis vectors carried in the rhs direction and applied
-//        through MrhsPromotedOperator
+//   ref  : DeprecatedMultiGeneralCoarsenedMatrix::CoarsenOperator, vectorised
+//          coarse space matched to the fine SIMD layout, single RHS fine
+//          applications
+//   test : MultiGeneralCoarsenedOperator::CoarsenOperator on the D+1 grid,
+//          unvectorised sComplexD coarse space, the batch of phased basis
+//          vectors carried in the rhs direction and applied through
+//          MrhsPromotedOperator
 //
 // BLAS_A is written by GridtoBLAS in lSite order, which does not depend on
 // the SIMD layout, so the two are directly comparable.
@@ -46,21 +48,25 @@ const int nbasis = 8;
 const int batch  = 9;
 
 typedef vSpinColourVector FineObj;
-typedef vTComplex         CComplexV;   // vectorised coarse space, V1 reference
-typedef sTComplexD        CComplexS;   // unvectorised coarse space, V2
+typedef vTComplex         CComplexV;   // vectorised coarse space, reference
+typedef sTComplexD        CComplexS;   // unvectorised coarse space, under test
 
-typedef MultiGeneralCoarsenedMatrix    <FineObj,CComplexV,nbasis> MrhsV1;
-typedef MultiGeneralCoarsenedOperatorV2<FineObj,CComplexS,nbasis> MrhsV2;
+typedef DeprecatedMultiGeneralCoarsenedMatrix    <FineObj,CComplexV,nbasis> RefOperator;
+typedef MultiGeneralCoarsenedOperator<FineObj,CComplexS,nbasis> TestOperator;
 
 template<class Op>
 void ReadMatrix(Op &O,int npoint,std::vector<std::vector<typename Op::calcMatrix> > &host)
 {
   typedef typename Op::calcMatrix calcMatrix;
   host.resize(npoint);
+  // MatrixPointOut carries one stencil point in lSite order whichever internal
+  // layout the operator uses.
+  deviceVector<calcMatrix> buf;
   for(int p=0;p<npoint;p++){
-    int64_t sites = O.BLAS_A[p].size();
+    O.MatrixPointOut(p,buf);
+    int64_t sites = buf.size();
     host[p].resize(sites);
-    acceleratorCopyFromDevice(&O.BLAS_A[p][0],&host[p][0],sites*sizeof(calcMatrix));
+    acceleratorCopyFromDevice(&buf[0],&host[p][0],sites*sizeof(calcMatrix));
   }
 }
 
@@ -83,24 +89,24 @@ int main (int argc, char ** argv)
   GridRedBlackCartesian *FrbGrid    = SpaceTimeGrid::makeFourDimRedBlackGrid(FineGrid);
 
   ////////////////////////////////////////////////
-  // V1 coarse space: SIMD layout matched to the fine
+  // Reference coarse space: SIMD layout matched to the fine
   ////////////////////////////////////////////////
   Coordinate cvsimd = GridDefaultSimd(Nd,CComplexV::Nsimd());
   GridCartesian *CoarseV = new GridCartesian(clatt,cvsimd,fmpi);
 
-  // V1 puts all of the SIMD in the rhs direction. CoarsenOperator does not
-  // use the multiRHS grid; it only sizes BLAS_A, so one lane of rhs suffices.
-  int nrhs_v1 = CComplexV::Nsimd();
-  Coordinate v1latt(1,nrhs_v1),v1simd(1,CComplexV::Nsimd()),v1mpi(1,1);
+  // The reference puts all of the SIMD in the rhs direction. CoarsenOperator
+  // does not use the multiRHS grid; it only sizes BLAS_A, so one lane of rhs suffices.
+  int nrhs_ref = CComplexV::Nsimd();
+  Coordinate reflatt(1,nrhs_ref),refsimd(1,CComplexV::Nsimd()),refmpi(1,1);
   for(int d=0;d<Nd;d++){
-    v1latt.push_back(clatt[d]);
-    v1simd.push_back(1);
-    v1mpi .push_back(fmpi[d]);
+    reflatt.push_back(clatt[d]);
+    refsimd.push_back(1);
+    refmpi .push_back(fmpi[d]);
   }
-  GridCartesian *CoarseVMulti = new GridCartesian(v1latt,v1simd,v1mpi);
+  GridCartesian *CoarseVMulti = new GridCartesian(reflatt,refsimd,refmpi);
 
   ////////////////////////////////////////////////
-  // V2 coarse space: unvectorised
+  // Coarse space under test: unvectorised
   ////////////////////////////////////////////////
   Coordinate cssimd(Nd,1);
   GridCartesian *CoarseS = new GridCartesian(clatt,cssimd,fmpi);
@@ -124,8 +130,8 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "fine "<<flatt<<"   coarse "<<clatt<<"   batch "<<batch<<std::endl;
   std::cout << GridLogMessage << "Nsimd   fine "<<FineGrid->Nsimd()
-	    << "   coarse V1 "<<CoarseV->Nsimd()
-	    << "   coarse V2 "<<CoarseS->Nsimd()<<std::endl;
+	    << "   coarse ref "<<CoarseV->Nsimd()
+	    << "   coarse test "<<CoarseS->Nsimd()<<std::endl;
 
   ////////////////////////////////////////////////
   // Fine operator
@@ -151,32 +157,32 @@ int main (int argc, char ** argv)
   NextToNearestStencilGeometry4D geomS(CoarseS);
 
   ////////////////////////////////////////////////
-  // V1 coarsening, matched layouts
+  // Reference coarsening, matched layouts
   ////////////////////////////////////////////////
-  MrhsV1 OpV1(geomV,CoarseVMulti);
-  std::cout << GridLogMessage << "V1 CoarsenOperator" << std::endl;
-  OpV1.CoarsenOperator(HermOp,Subspace,CoarseV);
+  RefOperator OpRef(geomV,CoarseVMulti);
+  std::cout << GridLogMessage << "reference CoarsenOperator" << std::endl;
+  OpRef.CoarsenOperator(HermOp,Subspace,CoarseV);
 
   ////////////////////////////////////////////////
-  // V2 coarsening, D+1 fine applications, unvectorised coarse
+  // Coarsening under test, D+1 fine applications, unvectorised coarse
   ////////////////////////////////////////////////
-  MrhsV2 OpV2(geomS,CoarseS);
-  OpV2.SetGrid(CoarseSMulti);
+  TestOperator OpTest(geomS,CoarseS);
+  OpTest.SetGrid(CoarseSMulti);
 
   MrhsPromotedOperator<LatticeFermionD> MrhsHermOp(HermOp,FineGrid,batch);
 
-  std::cout << GridLogMessage << "V2 CoarsenOperator (D+1)" << std::endl;
-  OpV2.CoarsenOperator(MrhsHermOp,FineGridMulti,subspace,CoarseS);
+  std::cout << GridLogMessage << "test CoarsenOperator (D+1)" << std::endl;
+  OpTest.CoarsenOperator(MrhsHermOp,FineGridMulti,subspace,CoarseS);
 
   ////////////////////////////////////////////////
   // Compare matrix elements
   ////////////////////////////////////////////////
-  int npoint = OpV1.geom.npoint;
-  GRID_ASSERT(npoint == OpV2.geom.npoint);
-  typedef MrhsV1::calcMatrix calcMatrix;
+  int npoint = OpRef.geom.npoint;
+  GRID_ASSERT(npoint == OpTest.geom.npoint);
+  typedef RefOperator::calcMatrix calcMatrix;
   std::vector<std::vector<calcMatrix> > A1,A2;
-  ReadMatrix(OpV1,npoint,A1);
-  ReadMatrix(OpV2,npoint,A2);
+  ReadMatrix(OpRef,npoint,A1);
+  ReadMatrix(OpTest,npoint,A2);
 
   RealD num=0.0, den=0.0;
   for(int p=0;p<npoint;p++){
@@ -190,8 +196,8 @@ int main (int argc, char ** argv)
       den += real(w1[i])*real(w1[i])+imag(w1[i])*imag(w1[i]);
     }
   }
-  std::cout << GridLogMessage << "|A_V1|^2 = " << den << std::endl;
-  std::cout << GridLogMessage << "|A_V1 - A_V2|^2 / |A_V1|^2 = " << num/den << std::endl;
+  std::cout << GridLogMessage << "|A_ref|^2 = " << den << std::endl;
+  std::cout << GridLogMessage << "|A_ref - A_test|^2 / |A_ref|^2 = " << num/den << std::endl;
   GRID_ASSERT( den > 0.0 );
   GRID_ASSERT( num/den < 1.0e-20 );
 
@@ -200,14 +206,14 @@ int main (int argc, char ** argv)
   // batch assembled on the coarse side by the mixed blockProject. Block
   // Gram-Schmidt is idempotent so the subspace may be reused in place.
   ////////////////////////////////////////////////
-  MrhsV2 OpV2s(geomS,CoarseS);
-  OpV2s.SetGrid(CoarseSMulti);
+  TestOperator OpTestSrhs(geomS,CoarseS);
+  OpTestSrhs.SetGrid(CoarseSMulti);
 
-  std::cout << GridLogMessage << "V2 CoarsenOperator (single RHS fine op)" << std::endl;
-  OpV2s.CoarsenOperator(HermOp,subspace,CoarseS,batch);
+  std::cout << GridLogMessage << "test CoarsenOperator (single RHS fine op)" << std::endl;
+  OpTestSrhs.CoarsenOperator(HermOp,subspace,CoarseS,batch);
 
   std::vector<std::vector<calcMatrix> > A3;
-  ReadMatrix(OpV2s,npoint,A3);
+  ReadMatrix(OpTestSrhs,npoint,A3);
 
   RealD nums=0.0;
   for(int p=0;p<npoint;p++){
@@ -220,7 +226,7 @@ int main (int argc, char ** argv)
       nums += real(d)*real(d)+imag(d)*imag(d);
     }
   }
-  std::cout << GridLogMessage << "|A_V1 - A_V2srhs|^2 / |A_V1|^2 = " << nums/den << std::endl;
+  std::cout << GridLogMessage << "|A_ref - A_test_srhs|^2 / |A_ref|^2 = " << nums/den << std::endl;
   GRID_ASSERT( nums/den < 1.0e-20 );
 
   ////////////////////////////////////////////////
@@ -229,15 +235,15 @@ int main (int argc, char ** argv)
   // needs no SetGrid since the matrix elements are Nrhs independent.
   ////////////////////////////////////////////////
   {
-    typedef MrhsV2::CoarseMatrix CoarseMatrixS;
+    typedef TestOperator::CoarseMatrix CoarseMatrixS;
     std::vector<CoarseMatrixS> Aget(npoint,CoarseS);
-    for(int p=0;p<npoint;p++) OpV2.GetMatrix(p,Aget);
+    for(int p=0;p<npoint;p++) OpTest.GetMatrix(p,Aget);
 
-    MrhsV2 OpV2rt(geomS,CoarseS);
-    for(int p=0;p<npoint;p++) OpV2rt.SetMatrix(p,Aget);
+    TestOperator OpTestRoundTrip(geomS,CoarseS);
+    for(int p=0;p<npoint;p++) OpTestRoundTrip.SetMatrix(p,Aget);
 
     std::vector<std::vector<calcMatrix> > A4;
-    ReadMatrix(OpV2rt,npoint,A4);
+    ReadMatrix(OpTestRoundTrip,npoint,A4);
 
     RealD numrt=0.0;
     for(int p=0;p<npoint;p++){
@@ -255,19 +261,19 @@ int main (int argc, char ** argv)
   }
 
   ////////////////////////////////////////////////
-  // and V2 applies the matrix it just built
+  // and the operator under test applies the matrix it just built
   ////////////////////////////////////////////////
-  typedef MrhsV2::CoarseVector CoarseVectorS;
+  typedef TestOperator::CoarseVector CoarseVectorS;
   GridParallelRNG cRNG(CoarseS); cRNG.SeedFixedIntegers(std::vector<int>({5,6,7,8}));
   CoarseVectorS in(CoarseSMulti);  random(cRNG,in);
   CoarseVectorS out(CoarseSMulti);
 
-  OpV2.M(in,out);
+  OpTest.M(in,out);
   std::cout << GridLogMessage << "|in|^2 = " << norm2(in)
-	    << "   |M_V2 in|^2 = " << norm2(out) << std::endl;
+	    << "   |M_test in|^2 = " << norm2(out) << std::endl;
   GRID_ASSERT( norm2(out) > 0.0 );
 
-  std::cout << GridLogMessage << "Test_coarse_v2_coarsen: ALL PASS" << std::endl;
+  std::cout << GridLogMessage << "Test_coarse_coarsen: ALL PASS" << std::endl;
 
   Grid_finalize();
 }

tests/debug/Test_general_coarse.cc

@@ -114,7 +114,7 @@ int main (int argc, char ** argv)
   ///////////////////////////////////////////////////
   std::cout<<GridLogMessage << "Building little Dirac operator"<< std::endl;
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -223,7 +223,7 @@ int main (int argc, char ** argv)
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
 
 #if 0  
-  MultiGeneralCoarsenedMatrix mrhs(LittleDiracOp,CoarseMrhs);
+  DeprecatedMultiGeneralCoarsenedMatrix mrhs(LittleDiracOp,CoarseMrhs);
   typedef decltype(mrhs) MultiGeneralCoarsenedMatrix_t;
   
   //////////////////////////////////////////

tests/debug/Test_general_coarse_hdcg.cc

@@ -96,7 +96,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -140,7 +140,7 @@ int main (int argc, char ** argv)
   Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
     
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
   MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
 
   mrhs.CoarsenOperator(FineHermOp,Aggregates,Coarse5d);

tests/debug/Test_general_coarse_hdcg_phys.cc

@@ -189,7 +189,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -371,7 +371,7 @@ slurm-1482367.out:Grid : Message : 6169.469330 s : HDCG: Pcg converged in 487 it
       //////////////////////////////////////////
       // Build a HDCG solver
       //////////////////////////////////////////
-      TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
+      DeprecatedTwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
 	HDCG(1.0e-8, 700,
 	     FineHermOp,
 	     CGsmooth,

tests/debug/Test_general_coarse_hdcg_phys48.cc

@@ -218,7 +218,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -269,7 +269,7 @@ int main (int argc, char ** argv)
   Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
     
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
   MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
 
   std::cout << "**************************************"<<std::endl;

tests/debug/Test_general_coarse_hdcg_phys48_blockcg.cc

@@ -407,7 +407,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -456,7 +456,7 @@ int main (int argc, char ** argv)
   Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
     
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
   MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
 
   std::cout << "**************************************"<<std::endl;

tests/debug/Test_general_coarse_hdcg_phys48_lanczos.cc

@@ -406,7 +406,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -455,7 +455,7 @@ int main (int argc, char ** argv)
   Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
     
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
   MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
 
   std::cout << "**************************************"<<std::endl;

tests/debug/Test_general_coarse_hdcg_phys48_lanczos_subspace.cc

@@ -215,7 +215,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);

tests/debug/Test_general_coarse_hdcg_phys48_mixed.cc

@@ -161,7 +161,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -210,7 +210,7 @@ int main (int argc, char ** argv)
   Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
     
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
   MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
 
   std::cout << "**************************************"<<std::endl;

tests/debug/Test_general_coarse_hdcg_phys96_mixed.cc

@@ -162,7 +162,7 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
@@ -217,7 +217,7 @@ int main (int argc, char ** argv)
   Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
     
   GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
-  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  typedef DeprecatedMultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
   MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
 
   ///////////////////////

tests/debug/Test_general_coarse_pvdagm.cc

@@ -254,7 +254,7 @@ int main (int argc, char ** argv)
   const int cb = 0 ;
   LatticeFermion prom(FGrid);
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NextToNearestStencilGeometry5D geom(Coarse5d);

tests/debug/Test_general_coarse_pvdagm_svd.cc

@@ -376,10 +376,10 @@ int main (int argc, char ** argv)
     std::cout <<bl<<" "<<br<<"\t"<<innerProduct(U.subspace[bl],src)<<std::endl;
   }}
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperatorV;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperatorV;
   typedef LittleDiracOperatorV::CoarseVector CoarseVectorV;
   
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   V.Orthogonalise();

tests/debug/Test_general_coarse_pvdagm_svd_cg.cc

@@ -375,10 +375,10 @@ int main (int argc, char ** argv)
     std::cout <<bl<<" "<<br<<"\t"<<innerProduct(U.subspace[bl],src)<<std::endl;
   }}
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperatorV;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperatorV;
   typedef LittleDiracOperatorV::CoarseVector CoarseVectorV;
   
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   V.Orthogonalise();

tests/debug/Test_general_coarse_pvdagm_svd_uv.cc

@@ -369,8 +369,8 @@ int main (int argc, char ** argv)
   //  }
   
   
-  //  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  //  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
   LittleDiracOperator LittleDiracOpPV(geom,FGrid,Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM,V,V);

tests/debug/Test_general_coarse_wilson.cc

@@ -185,7 +185,7 @@ int main (int argc, char ** argv)
   const int cb = 0 ;
   LatticeFermion prom(FGrid);
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NearestStencilGeometry4D geom(Coarse4d);

tests/debug/Test_general_coarse_wilson_nog5.cc

@@ -185,7 +185,7 @@ int main (int argc, char ** argv)
   const int cb = 0 ;
   LatticeFermion prom(FGrid);
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NearestStencilGeometry4D geom(Coarse4d);

tests/debug/Test_general_coarse_wilson_svd.cc

@@ -182,7 +182,7 @@ int main (int argc, char ** argv)
   const int cb = 0 ;
   LatticeFermion prom(FGrid);
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,2*nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NearestStencilGeometry4D geom(Coarse4d);

tests/debug/Test_general_coarse_wilson_svd_no5g.cc

@@ -182,7 +182,7 @@ int main (int argc, char ** argv)
   const int cb = 0 ;
   LatticeFermion prom(FGrid);
 
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
 
   NearestStencilGeometry4D geom(Coarse4d);

tests/debug/Test_hipfft_bug_fail.cc

@@ -3,9 +3,9 @@
  *
  * Tests three orderings with an empty rocFFT cache to find which GPU
  * operation before plan creation triggers the failure:
- *   A) hipMalloc only            — hypothesis: passes (no async GPU work)
- *   B) hipMalloc + hipMemset     — hypothesis: fails  (async GPU work in flight)
- *   C) hipMalloc + hipMemset     — hypothesis: passes (work completed before plan)
+ *   A) hipMalloc only            -- hypothesis: passes (no async GPU work)
+ *   B) hipMalloc + hipMemset     -- hypothesis: fails  (async GPU work in flight)
+ *   C) hipMalloc + hipMemset     -- hypothesis: passes (work completed before plan)
  *      + hipDeviceSynchronize
  *
  * Compile:
@@ -59,13 +59,13 @@ int main(void) {
     hipfftResult rvB = makePlan(G, howmany);
     printf("G=%-4d  B) hipMalloc + hipMemset       : %s\n", G, res(rvB));
 
-    // C: hipMalloc + hipMemset + sync — does syncing before plan creation fix it?
+    // C: hipMalloc + hipMemset + sync -- does syncing before plan creation fix it?
     hipMemset(buf, 0, nelems * sizeof(hipfftDoubleComplex));
     hipDeviceSynchronize();
     hipfftResult rvC = makePlan(G, howmany);
     printf("G=%-4d  C) hipMalloc + hipMemset + sync: %s\n", G, res(rvC));
 
-    // A last: hipMalloc only, no async GPU work — should always pass
+    // A last: hipMalloc only, no async GPU work -- should always pass
     hipfftResult rvA = makePlan(G, howmany);
     printf("G=%-4d  A) hipMalloc only             : %s\n\n", G, res(rvA));
 

tests/debug/Test_hipfft_bug_pass.cc

@@ -33,7 +33,7 @@ int main(void) {
     int n[] = {G};
     long nelems = (long)G * howmany;
 
-    // Plan created BEFORE hipMalloc — succeeds for all G
+    // Plan created BEFORE hipMalloc -- succeeds for all G
     hipfftHandle p;
     size_t workSize = 0;
     hipfftCreate(&p);

tests/debug/Test_hipfft_minimal.cc

@@ -38,8 +38,8 @@ static const char *hipfftResultString(hipfftResult r) {
 // Plan creation + execution for (G, howmany).
 // Tests two orderings to isolate whether a prior hipMalloc poisons hipfft
 // plan creation for small G on ROCm 7:
-//   A) plan BEFORE hipMalloc  — hypothesis: succeeds
-//   B) hipMalloc BEFORE plan  — hypothesis: fails for G < 32
+//   A) plan BEFORE hipMalloc  -- hypothesis: succeeds
+//   B) hipMalloc BEFORE plan  -- hypothesis: fails for G < 32
 static void tryPlanAndExec(int G, long howmany) {
   int n[] = {G};
   long nelems = (long)G * howmany;

tests/debug/Test_hipfft_repro.cc

@@ -58,12 +58,12 @@ static void tryPlanAndExec(int G, long howmany) {
   hipfftDoubleComplex *dbuf = nullptr;
   hipError_t herr = hipMalloc(&dbuf, nelems * sizeof(hipfftDoubleComplex));
   if (herr != hipSuccess) {
-    printf("  hipMalloc failed (%d) for %ld elems — skipping\n\n", (int)herr, nelems);
+    printf("  hipMalloc failed (%d) for %ld elems -- skipping\n\n", (int)herr, nelems);
     return;
   }
   hipMemset(dbuf, 0, nelems * sizeof(hipfftDoubleComplex));
 
-  // 1. hipfftPlanMany (one-step, nullptr embed) — current Grid path
+  // 1. hipfftPlanMany (one-step, nullptr embed) -- current Grid path
   {
     hipfftHandle p;
     hipfftResult rv = hipfftPlanMany(&p, 1, n,
@@ -79,7 +79,7 @@ static void tryPlanAndExec(int G, long howmany) {
     }
   }
 
-  // 2. hipfftCreate + hipfftMakePlanMany (two-step) — also current Grid path
+  // 2. hipfftCreate + hipfftMakePlanMany (two-step) -- also current Grid path
   {
     hipfftHandle p;
     size_t workSize = 0;

tests/debug/Test_mrhs_deflation.cc

@@ -0,0 +1,89 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./tests/debug/Test_mrhs_deflation.cc
+
+    Copyright (C) 2026
+
+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.
+
+    See the full license in the file "LICENSE" in the top level distribution
+    directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/Grid.h>
+
+// MultiRHSDeflation: the D+1 multiRHS interface (one permutation pass each
+// way) must reproduce the vector-of-D-fields interface bit for bit up to
+// the summation order of the GEMMs.  Random "eigenvectors" and values: the
+// deflation formula G = E (E^dag R)/lambda does not care that they are not
+// eigenpairs of anything.
+
+using namespace Grid;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int nbasis = 8;
+  const int nev    = 12;
+  const int nrhs   = 5;
+
+  typedef iVector<sTComplexD,nbasis> siteVector;
+  typedef Lattice<siteVector>        CoarseVector;
+
+  // Unvectorised D and D+1 coarse grids, as MGCoarseGrids builds them
+  Coordinate latt({1,4,4,4,4});
+  Coordinate simd({1,1,1,1,1});
+  Coordinate mpi = GridDefaultMpi();
+  Coordinate mpi5({1,mpi[0],mpi[1],mpi[2],mpi[3]});
+  GridCartesian Coarse5d(latt,simd,mpi5);
+
+  Coordinate latt6({nrhs,1,4,4,4,4});
+  Coordinate simd6({1,1,1,1,1,1});
+  Coordinate mpi6({1,1,mpi[0],mpi[1],mpi[2],mpi[3]});
+  GridCartesian CoarseMrhs(latt6,simd6,mpi6);
+
+  GridParallelRNG RNG(&Coarse5d); RNG.SeedFixedIntegers(std::vector<int>({1,2,3,4}));
+
+  std::vector<CoarseVector> evec(nev,&Coarse5d);
+  std::vector<RealD>        eval(nev);
+  for(int e=0;e<nev;e++){ random(RNG,evec[e]); eval[e] = 0.5 + 0.1*e; }
+
+  std::vector<CoarseVector> src(nrhs,&Coarse5d), guess(nrhs,&Coarse5d);
+  for(int r=0;r<nrhs;r++) random(RNG,src[r]);
+
+  MultiRHSDeflation<CoarseVector> Deflator;
+  Deflator.ImportEigenBasis(evec,eval);
+
+  // Reference: the vector interface
+  Deflator.DeflateSources(src,guess);
+
+  // The D+1 interface on the same sources
+  CoarseVector src_mrhs(&CoarseMrhs), guess_mrhs(&CoarseMrhs);
+  for(int r=0;r<nrhs;r++) InsertSliceFast(src[r],src_mrhs,r,0);
+  guess_mrhs = Zero();
+  Deflator.DeflateSources(src_mrhs,guess_mrhs);
+
+  RealD worst = 0.0;
+  CoarseVector g(&Coarse5d), d(&Coarse5d);
+  for(int r=0;r<nrhs;r++){
+    ExtractSliceFast(g,guess_mrhs,r,0);
+    d = g - guess[r];
+    RealD rel = std::sqrt(norm2(d)/norm2(guess[r]));
+    std::cout << GridLogMessage << "rhs " << r << " ||guess_mrhs - guess||/||guess|| = " << rel << std::endl;
+    worst = std::max(worst,rel);
+  }
+  std::cout << GridLogMessage << "MultiRHSDeflation D+1 vs vector interface: worst " << worst << std::endl;
+  GRID_ASSERT( worst < 1.0e-12 );
+  std::cout << GridLogMessage << "PASS" << std::endl;
+
+  Grid_finalize();
+  return 0;
+}

tests/debug/Test_schur_dense_coarse.cc

@@ -30,7 +30,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 // The DenseCoarseMatrix GLUE test, on a real (tiny) lattice coarse
 // operator, CPU laptop build.
 //
-// Builds a genuine GeneralCoarsenedMatrix (DWF MdagM + 0.5 shift for a
+// Builds a genuine DeprecatedGeneralCoarsenedMatrix (DWF MdagM + 0.5 shift for a
 // guaranteed-invertible Galerkin coarse op, random aggregation basis)
 // and runs the whole Import certificate chain through the glue:
 //   - fresh ImportDense + IMPORT CERTIFICATE vs Op.M
@@ -143,7 +143,7 @@ int main (int argc, char ** argv)
   Aggregates.CreateSubspaceRandom(RNG5);
 
   std::cout << GridLogMessage << "Coarsening shifted MdagM" << std::endl;
-  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef DeprecatedGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
   typedef LittleDiracOperator::CoarseVector CoarseVector;
   NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
   LittleDiracOperator LittleDiracOp(geom,FGrid,Coarse5d);

tests/debug/Test_sloppy_dagger.cc

@@ -0,0 +1,207 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./tests/debug/Test_sloppy_dagger.cc
+
+    Copyright (C) 2026
+
+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.
+
+    See the full license in the file "LICENSE" in the top level distribution
+    directory
+*************************************************************************************/
+/*  END LEGAL */
+
+//
+// Isolate a sloppy-comms dagger defect: on a CPU/GEN multi-rank build the
+// sloppy dagger halo can be wrong by ~5% of norm2, deterministically, which
+// Benchmark_dwf's sloppy pass catches as a failed dagger Cshift check.
+//
+// The sloppy compressed-buffer pool (StencilBuffer::DeviceCommBuf) is a
+// SHARED STATIC across all stencil objects, so scenarios contaminate each
+// other inside one process: this program runs exactly ONE scenario per
+// invocation, selected with --seq <name>:
+//
+//   fresh-dag       sloppy dagger is the operator's first call
+//   nodag-dag       sloppy nodag x3 (one source), then sloppy dagger (same source)
+//   nodag-dag-fresh nodag x3 (fresh source each), then dagger (fresh source)
+//   nodag-dag-newsrc nodag x3 (one source), then dagger (different source)
+//   dag-nodag-dag   dagger, nodag x3, dagger  (dagger-first history)
+//
+// Every call is checked against a separate never-sloppy reference
+// operator (exact Dhop == the Cshift construction at 1e-31, certified by
+// Benchmark_dwf's exact pass).  The last dagger's error is fingerprinted
+// by slice along each MPI-decomposed direction.
+//
+//   OMP_NUM_THREADS=1 mpirun -n 2 ./Test_sloppy_dagger --grid 8.8.8.16 \
+//        --mpi 1.1.1.2 --seq nodag-dag
+//
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::string seq("nodag-dag");
+  if( GridCmdOptionExists(argv,argv+argc,"--seq") )
+    seq = GridCmdOptionPayload(argv,argv+argc,"--seq");
+
+  const int Ls=8;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+                                                                   GridDefaultSimd(Nd,vComplex::Nsimd()),
+                                                                   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers({1,2,3,4});
+  GridParallelRNG RNG5(FGrid); RNG5.SeedFixedIntegers({5,6,7,8});
+
+  LatticeGaugeField Umu(UGrid);
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+
+  RealD mass=0.1, M5=1.8;
+  DomainWallFermionD Dref(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);  // never sloppy
+  DomainWallFermionD Dw  (Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);  // under test
+  Dw.SloppyComms(1);
+
+  Coordinate mpi = GridDefaultMpi();
+  LatticeFermionD src(FGrid), exact(FGrid), sloppy(FGrid), err(FGrid);
+
+  // One call: fresh or reused source, dag or not; always checked.
+  LatticeFermionD srcA(FGrid); gaussian(RNG5,srcA);
+  auto call = [&](int dag, int fresh, const char *tag){
+    if ( fresh ) gaussian(RNG5,src); else src = srcA;
+    Dref.Dhop(src,exact,dag);
+    Dw.Dhop(src,sloppy,dag);
+    err = sloppy - exact;
+    std::cout << GridLogMessage << "SEQ[" << seq << "] " << tag
+              << (dag?" DAG   ":" NODAG ") << " norm2(err) " << norm2(err) << std::endl;
+  };
+  // First bad site: coordinate + exact-vs-sloppy hex words (PB: look for
+  // expected/actual mismatching in the high-order half word).
+  auto firstbad = [&](LatticeFermionD &ex, LatticeFermionD &sl){
+    typedef LatticeFermionD::scalar_object sobj;
+    Coordinate gdims = FGrid->GlobalDimensions();
+    int64_t gsites = FGrid->gSites();
+    for(int64_t g=0; g<gsites; g++){
+      Coordinate gcoor(5);
+      Lexicographic::CoorFromIndex(gcoor,g,gdims);
+      sobj se, ss_;
+      peekSite(se,ex,gcoor);
+      peekSite(ss_,sl,gcoor);
+      uint64_t *we=(uint64_t *)&se;
+      uint64_t *ws=(uint64_t *)&ss_;
+      int nw = sizeof(sobj)/8;
+      for(int w=0;w<nw;w++){
+        double de=((double *)we)[w], ds=((double *)ws)[w];
+        if ( fabs(de-ds) > 1.0e-4 ) {
+          std::cout << GridLogMessage << "FIRSTBAD site " << gcoor
+                    << " word " << w << " (spin "<<(w/6)%4<<" col "<<(w/2)%3<<" reim "<<(w%2)<<")"
+                    << "  exact " << std::hex << we[w]
+                    << "  sloppy " << ws[w] << std::dec << std::endl;
+          for(int w2=0;w2<8&&w2<nw;w2++)
+            std::cout << GridLogMessage << "   word["<<w2<<"] exact "<<std::hex<<we[w2]
+                      <<" sloppy "<<ws[w2]<<std::dec<<std::endl;
+          return;
+        }
+      }
+    }
+    std::cout << GridLogMessage << "FIRSTBAD: none above 1e-4" << std::endl;
+  };
+  auto fingerprint = [&](void){
+    for(int mu=0;mu<Nd;mu++){
+      if ( mpi[mu] == 1 ) continue;
+      std::vector<RealD> sn;
+      sliceNorm(sn,err,mu+1);
+      std::cout << GridLogMessage << "SEQ[" << seq << "] err by slice of dim " << mu << ":";
+      for(int t=0;t<(int)sn.size();t++) std::cout << " " << sn[t];
+      std::cout << std::endl;
+    }
+  };
+
+  if        ( seq == "fresh-dag" ) {
+    call(1,0,"call0");
+  } else if ( seq == "nodag-dag" ) {
+    call(0,0,"call0"); call(0,0,"call1"); call(0,0,"call2");
+    call(1,0,"call3");
+  } else if ( seq == "nodag-dag-fresh" ) {
+    call(0,1,"call0"); call(0,1,"call1"); call(0,1,"call2");
+    call(1,1,"call3");
+  } else if ( seq == "nodag-dag-newsrc" ) {
+    call(0,0,"call0"); call(0,0,"call1"); call(0,0,"call2");
+    call(1,1,"call3");
+  } else if ( seq == "dag-nodag-dag" ) {
+    call(1,0,"call0");
+    call(0,0,"call1"); call(0,0,"call2"); call(0,0,"call3");
+    call(1,0,"call4");
+  } else if ( seq == "noleave" ) {
+    // NO exact-operator call between sloppy calls: references precomputed.
+    LatticeFermionD refnodag(FGrid), refdag(FGrid);
+    Dref.Dhop(srcA,refnodag,DaggerNo);
+    Dref.Dhop(srcA,refdag,DaggerYes);
+    for(int i=0;i<2;i++){
+      Dw.Dhop(srcA,sloppy,DaggerNo);
+      err = sloppy - refnodag;
+      std::cout << GridLogMessage << "SEQ[noleave] call" << i << " NODAG  norm2(err) " << norm2(err) << std::endl;
+      if ( i==0 ) firstbad(refnodag,sloppy);
+    }
+    Dw.Dhop(srcA,sloppy,DaggerYes);
+    err = sloppy - refdag;
+    std::cout << GridLogMessage << "SEQ[noleave] call3 DAG    norm2(err) " << norm2(err) << std::endl;
+  } else if ( seq == "twoop" ) {
+    // A SECOND sloppy operator shares the static pool; sequence runs on it
+    // with interleaved exact references (as the honest runs had).
+    DomainWallFermionD Dw2(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    Dw2.SloppyComms(1);
+    Dw.Dhop(srcA,sloppy,DaggerNo);   // prime the FIRST sloppy op's state
+    for(int i=0;i<3;i++){
+      Dref.Dhop(srcA,exact,DaggerNo);
+      Dw2.Dhop(srcA,sloppy,DaggerNo);
+      err = sloppy - exact;
+      std::cout << GridLogMessage << "SEQ[twoop] call" << i << " NODAG  norm2(err) " << norm2(err) << std::endl;
+    }
+    Dref.Dhop(srcA,exact,DaggerYes);
+    Dw2.Dhop(srcA,sloppy,DaggerYes);
+    err = sloppy - exact;
+    std::cout << GridLogMessage << "SEQ[twoop] call3 DAG    norm2(err) " << norm2(err) << std::endl;
+  } else if ( seq == "self-exact" ) {
+    // Exact call on the SAME object (Dw), opposite sense, then sloppy:
+    // distinguishes per-object from shared state.
+    LatticeFermionD refnodag(FGrid);
+    Dref.Dhop(srcA,refnodag,DaggerNo);              // reference only, early
+    Dw.SloppyComms(0);
+    Dw.Dhop(srcA,sloppy,DaggerYes);                 // exact DAG on Dw itself
+    Dw.SloppyComms(1);
+    Dw.Dhop(srcA,sloppy,DaggerNo);                  // sloppy NODAG: sense mismatch
+    err = sloppy - refnodag;
+    std::cout << GridLogMessage << "SEQ[self-exact] sloppy NODAG after own exact DAG: " << norm2(err) << std::endl;
+  } else if ( seq == "cold" ) {
+    // NO exact Dhop anywhere before the sloppy calls.
+    LatticeFermionD outn(FGrid), outd(FGrid);
+    Dw.Dhop(srcA,outn,DaggerNo);
+    Dw.Dhop(srcA,outd,DaggerYes);
+    LatticeFermionD refnodag(FGrid), refdag(FGrid);
+    Dref.Dhop(srcA,refnodag,DaggerNo);
+    Dref.Dhop(srcA,refdag,DaggerYes);
+    err = outn - refnodag;
+    std::cout << GridLogMessage << "SEQ[cold] first-ever sloppy NODAG: " << norm2(err) << std::endl;
+    err = outd - refdag;
+    std::cout << GridLogMessage << "SEQ[cold] then      sloppy DAG:   " << norm2(err) << std::endl;
+  } else {
+    std::cout << GridLogMessage << "unknown --seq " << seq << std::endl;
+  }
+  fingerprint();
+
+  Grid_finalize();
+}

tests/solver/Test_multigrid_common.h

@@ -397,7 +397,7 @@ public:
     MdagMLinearOperator<CoarseDiracMatrix, CoarseVector> coarseMdagMOp(_CoarseMatrix);
 
     std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
-    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == (M - (Mdiag + Σ_μ Mdir_μ)) * v" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == (M - (Mdiag + Sum_mu Mdir_mu)) * v" << std::endl;
     std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
 
     random(_LevelInfo.PRNGs[_CurrentLevel], fineTmps[0]);
@@ -406,22 +406,22 @@ public:
     fineMdagMOp.OpDiag(fineTmps[0], fineTmps[2]); // Mdiag * v
 
     fineTmps[4] = zero;
-    for(int dir = 0; dir < 4; dir++) { //       Σ_μ Mdir_μ * v
+    for(int dir = 0; dir < 4; dir++) { //       Sum_mu Mdir_mu * v
       for(auto disp : {+1, -1}) {
         fineMdagMOp.OpDir(fineTmps[0], fineTmps[3], dir, disp);
         fineTmps[4] = fineTmps[4] + fineTmps[3];
       }
     }
 
-    fineTmps[5] = fineTmps[2] + fineTmps[4]; // (Mdiag + Σ_μ Mdir_μ) * v
+    fineTmps[5] = fineTmps[2] + fineTmps[4]; // (Mdiag + Sum_mu Mdir_mu) * v
 
     fineTmps[6]    = fineTmps[1] - fineTmps[5];
     auto deviation = std::sqrt(norm2(fineTmps[6]) / norm2(fineTmps[1]));
 
     std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(M * v)                    = " << norm2(fineTmps[1]) << std::endl;
     std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(Mdiag * v)                = " << norm2(fineTmps[2]) << std::endl;
-    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(Σ_μ Mdir_μ * v)           = " << norm2(fineTmps[4]) << std::endl;
-    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2((Mdiag + Σ_μ Mdir_μ) * v) = " << norm2(fineTmps[5]) << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(Sum_mu Mdir_mu * v)           = " << norm2(fineTmps[4]) << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2((Mdiag + Sum_mu Mdir_mu) * v) = " << norm2(fineTmps[5]) << std::endl;
     std::cout << GridLogMG << " Level " << _CurrentLevel << ": relative deviation              = " << deviation;
 
     if(deviation > tolerance) {
@@ -664,7 +664,7 @@ createMGInstance(WilsonMGParams &mgParams, LevelInfo &levelInfo, Matrix &FineMat
     CASE_FOR_N_LEVELS(3);
     CASE_FOR_N_LEVELS(4);
     default:
-      std::cout << GridLogError << "We currently only support nLevels ∈ {2, 3, 4}" << std::endl;
+      std::cout << GridLogError << "We currently only support nLevels in {2, 3, 4}" << std::endl;
       exit(EXIT_FAILURE);
       break;
   }