Merge branch 'release/0.10.0'

Initial code change review complete

Grid/DisableWarnings.h

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

Grid/Grid_Eigen_Dense.h

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

Grid/algorithms/Algorithms.h

@@ -55,6 +55,7 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h>
 #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>

Grid/algorithms/iterative/ConjugateGradient.h

@@ -191,7 +191,7 @@ public:
 	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
-	std::cout << GridLogMessage << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
+	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -108,7 +108,10 @@ NAMESPACE_BEGIN(Grid);
     GridStopWatch PrecChangeTimer;
     
     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
-      
+
+    precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
+    
     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
@@ -123,7 +126,7 @@ NAMESPACE_BEGIN(Grid);
       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d);
+      precisionChange(src_f, src_d, pc_wk_dp_to_sp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -142,7 +145,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f);
+      precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -0,0 +1,213 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
+
+    Copyright (C) 2015
+
+    Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+
+NAMESPACE_BEGIN(Grid);
+
+//Mixed precision restarted defect correction CG
+template<class FieldD,class FieldF, 
+  typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+  typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
+public:
+  using LinearFunction<FieldD>::operator();
+  RealD   Tolerance;
+  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
+  Integer MaxInnerIterations;
+  Integer MaxOuterIterations;
+  Integer MaxPatchupIterations;
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+  LinearOperatorBase<FieldF> &Linop_f;
+  LinearOperatorBase<FieldD> &Linop_d;
+
+  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+  LinearFunction<FieldF> *guesser;
+  bool updateResidual;
+  
+  MixedPrecisionConjugateGradientBatched(RealD tol, 
+          Integer maxinnerit, 
+          Integer maxouterit, 
+          Integer maxpatchit,
+          GridBase* _sp_grid, 
+          LinearOperatorBase<FieldF> &_Linop_f, 
+          LinearOperatorBase<FieldD> &_Linop_d,
+          bool _updateResidual=true) :
+    Linop_f(_Linop_f), Linop_d(_Linop_d),
+    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
+    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
+
+  void useGuesser(LinearFunction<FieldF> &g){
+    guesser = &g;
+  }
+  
+  void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::vector<FieldD> srcs_d_in{src_d_in};
+    std::vector<FieldD> sols_d{sol_d};
+
+    (*this)(srcs_d_in,sols_d);
+
+    sol_d = sols_d[0];
+  }
+
+  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
+    assert(src_d_in.size() == sol_d.size());
+    int NBatch = src_d_in.size();
+
+    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
+
+    Integer TotalOuterIterations = 0; //Number of restarts
+    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
+    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
+  
+    GridStopWatch TotalTimer;
+    TotalTimer.Start();
+
+    GridStopWatch InnerCGtimer;
+    GridStopWatch PrecChangeTimer;
+    
+    int cb = src_d_in[0].Checkerboard();
+    
+    std::vector<RealD> src_norm;
+    std::vector<RealD> norm;
+    std::vector<RealD> stop;
+    
+    GridBase* DoublePrecGrid = src_d_in[0].Grid();
+    FieldD tmp_d(DoublePrecGrid);
+    tmp_d.Checkerboard() = cb;
+    
+    FieldD tmp2_d(DoublePrecGrid);
+    tmp2_d.Checkerboard() = cb;
+
+    std::vector<FieldD> src_d;
+    std::vector<FieldF> src_f;
+    std::vector<FieldF> sol_f;
+
+    for (int i=0; i<NBatch; i++) {
+      sol_d[i].Checkerboard() = cb;
+
+      src_norm.push_back(norm2(src_d_in[i]));
+      norm.push_back(0.);
+      stop.push_back(src_norm[i] * Tolerance*Tolerance);
+
+      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
+
+      src_f.push_back(SinglePrecGrid);
+      src_f[i].Checkerboard() = cb;
+
+      sol_f.push_back(SinglePrecGrid);
+      sol_f[i].Checkerboard() = cb;
+    }
+    
+    RealD inner_tol = InnerTolerance;
+    
+    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+    CG_f.ErrorOnNoConverge = false;
+    
+    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      std::cout << GridLogMessage << std::endl;
+      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      
+      bool allConverged = true;
+      
+      for (int i=0; i<NBatch; i++) {
+        //Compute double precision rsd and also new RHS vector.
+        Linop_d.HermOp(sol_d[i], tmp_d);
+        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        
+        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
+
+        PrecChangeTimer.Start();
+        precisionChange(src_f[i], src_d[i]);
+        PrecChangeTimer.Stop();
+        
+        sol_f[i] = Zero();
+      
+        if(norm[i] > OuterLoopNormMult * stop[i]) {
+          allConverged = false;
+        }
+      }
+      if (allConverged) break;
+
+      if (updateResidual) {
+        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
+        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
+        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+        CG_f.Tolerance = inner_tol;
+      }
+
+      //Optionally improve inner solver guess (eg using known eigenvectors)
+      if(guesser != NULL) {
+        (*guesser)(src_f, sol_f);
+      }
+
+      for (int i=0; i<NBatch; i++) {
+        //Inner CG
+        InnerCGtimer.Start();
+        CG_f(Linop_f, src_f[i], sol_f[i]);
+        InnerCGtimer.Stop();
+        TotalInnerIterations[i] += CG_f.IterationsToComplete;
+        
+        //Convert sol back to double and add to double prec solution
+        PrecChangeTimer.Start();
+        precisionChange(tmp_d, sol_f[i]);
+        PrecChangeTimer.Stop();
+        
+        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
+      }
+
+    }
+    
+    //Final trial CG
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
+    
+    for (int i=0; i<NBatch; i++) {
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
+      CG_d(Linop_d, src_d_in[i], sol_d[i]);
+      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
+    }
+
+    TotalTimer.Stop();
+
+    std::cout << GridLogMessage << std::endl;
+    for (int i=0; i<NBatch; i++) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
+    }
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

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

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -81,6 +81,7 @@ public:
   using OperatorFunction<FieldD>::operator();
 
   RealD   Tolerance;
+  Integer MaxIterationsMshift;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -95,9 +96,9 @@ public:
 
   ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
 				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
-				       int _ReliableUpdateFreq
-				       ) : 
-    MaxIterations(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
+				       int _ReliableUpdateFreq) : 
+    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+    MaxIterations(20000)
   { 
     verbose=1;
     IterationsToCompleteShift.resize(_shifts.order);
@@ -130,6 +131,9 @@ public:
     GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
     GridBase *DoublePrecGrid = src_d.Grid();
 
+    precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid);
+    
     ////////////////////////////////////////////////////////////////////////
     // Convenience references to the info stored in "MultiShiftFunction"
     ////////////////////////////////////////////////////////////////////////
@@ -200,10 +204,10 @@ public:
     r_d = p_d;
     
     //MdagM+m[0]
-    precisionChangeFast(p_f,p_d);
+    precisionChange(p_f, p_d, pc_wk_d_to_s);
 
     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    precisionChangeFast(tmp_d,mmp_f);
+    precisionChange(tmp_d, mmp_f, pc_wk_s_to_d);
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
@@ -244,7 +248,7 @@ public:
     // Iteration loop
     int k;
   
-    for (k=1;k<=MaxIterations;k++){    
+    for (k=1;k<=MaxIterationsMshift;k++){    
 
       a = c /cp;
       AXPYTimer.Start();
@@ -263,7 +267,7 @@ public:
       AXPYTimer.Stop();
 
       PrecChangeTimer.Start();
-      precisionChangeFast(p_f, p_d); //get back single prec search direction for linop
+      precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop
       PrecChangeTimer.Stop();
 
       cp=c;
@@ -272,7 +276,7 @@ public:
       MatrixTimer.Stop();  
 
       PrecChangeTimer.Start();
-      precisionChangeFast(mmp_d, mmp_f); // From Float to Double
+      precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double
       PrecChangeTimer.Stop();
 
       AXPYTimer.Start();
@@ -350,12 +354,17 @@ public:
 	}
       }
 
-      if ( all_converged ){
+      if ( all_converged || k == MaxIterationsMshift-1){
 
 	SolverTimer.Stop();
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
-	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
-      
+
+	if ( all_converged ){
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
+	} else {
+	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
+	}
+	
 	// Check answers 
 	for(int s=0; s < nshift; s++) { 
 	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
@@ -396,12 +405,10 @@ public:
 
 	return;
       }
-
    
     }
-    // ugly hack
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
+    assert(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -48,7 +48,7 @@ public:
   LinearOperatorBase<FieldF> &Linop_f;
   LinearOperatorBase<FieldD> &Linop_d;
   GridBase* SinglePrecGrid;
-  RealD Delta; //reliable update parameter
+  RealD Delta; //reliable update parameter. A reliable update is performed when the residual drops by a factor of Delta relative to its value at the last update
 
   //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
   LinearOperatorBase<FieldF> *Linop_fallback;
@@ -65,7 +65,9 @@ public:
       ErrorOnNoConverge(err_on_no_conv),
       DoFinalCleanup(true),
       Linop_fallback(NULL)
-  {};
+  {
+    assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
+  };
 
   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
     Linop_fallback = &_Linop_fallback;
@@ -116,9 +118,12 @@ public:
     }
 
     //Single prec initialization
+    precisionChangeWorkspace pc_wk_sp_to_dp(src.Grid(), SinglePrecGrid);
+    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, src.Grid());
+    
     FieldF r_f(SinglePrecGrid);
     r_f.Checkerboard() = r.Checkerboard();
-    precisionChange(r_f, r);
+    precisionChange(r_f, r, pc_wk_dp_to_sp);
 
     FieldF psi_f(r_f);
     psi_f = Zero();
@@ -134,7 +139,8 @@ public:
     GridStopWatch LinalgTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
-
+    GridStopWatch PrecChangeTimer;
+    
     SolverTimer.Start();
     int k = 0;
     int l = 0;
@@ -173,7 +179,9 @@ public:
       // Stopping condition
       if (cp <= rsq) {
 	//Although not written in the paper, I assume that I have to add on the final solution
-	precisionChange(mmp, psi_f);
+	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
+	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 	
 	
@@ -194,7 +202,10 @@ public:
 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPrecChange " << PrecChangeTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPrecChange avg time " << PrecChangeTimer.Elapsed()/(2*l+1) <<std::endl;
 
+	
 	IterationsToComplete = k;	
 	ReliableUpdatesPerformed = l;
 	  
@@ -214,14 +225,21 @@ public:
       else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
 		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
-	precisionChange(mmp, psi_f);
+	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
+	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 
+	MatrixTimer.Start();
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	MatrixTimer.Stop();
+	
 	r = src - mmp;
 
 	psi_f = Zero();
-	precisionChange(r_f, r);
+	PrecChangeTimer.Start();
+	precisionChange(r_f, r, pc_wk_dp_to_sp);
+	PrecChangeTimer.Stop();
 	cp = norm2(r);
 	MaxResidSinceLastRelUp = cp;
 

Grid/allocator/MemoryManager.cc

@@ -4,11 +4,14 @@ NAMESPACE_BEGIN(Grid);
 
 /*Allocation types, saying which pointer cache should be used*/
 #define Cpu      (0)
-#define CpuSmall (1)
-#define Acc      (2)
-#define AccSmall (3)
-#define Shared   (4)
-#define SharedSmall (5)
+#define CpuHuge  (1)
+#define CpuSmall (2)
+#define Acc      (3)
+#define AccHuge  (4)
+#define AccSmall (5)
+#define Shared   (6)
+#define SharedHuge  (7)
+#define SharedSmall (8)
 #undef GRID_MM_VERBOSE 
 uint64_t total_shared;
 uint64_t total_device;
@@ -35,12 +38,15 @@ void MemoryManager::PrintBytes(void)
   
 }
 
+uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
+uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
+
 //////////////////////////////////////////////////////////////////////
 // Data tables for recently freed pooiniter caches
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 8, 16, 8, 16 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils
@@ -170,6 +176,16 @@ void MemoryManager::Init(void)
     }
   }
 
+  str= getenv("GRID_ALLOC_NCACHE_HUGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuHuge]=Nc;
+      Ncache[AccHuge]=Nc;
+      Ncache[SharedHuge]=Nc;
+    }
+  }
+
   str= getenv("GRID_ALLOC_NCACHE_SMALL");
   if ( str ) {
     Nc = atoi(str);
@@ -190,7 +206,9 @@ void MemoryManager::InitMessage(void) {
   
   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
 #ifdef ALLOCATION_CACHE
-  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
 #endif
   
 #ifdef GRID_UVM
@@ -222,8 +240,11 @@ void MemoryManager::InitMessage(void) {
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  int cache = type + small;
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
@@ -232,11 +253,12 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 
 void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
 
+  if (ncache == 0) return ptr;
+
   void * ret = NULL;
   int v = -1;
 
@@ -271,8 +293,11 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 void *MemoryManager::Lookup(size_t bytes,int type)
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  int cache = type+small;
+  int cache;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
   return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
@@ -281,7 +306,6 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 
 void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 

Grid/allocator/MemoryManager.h

@@ -35,6 +35,7 @@ NAMESPACE_BEGIN(Grid);
 // Move control to configure.ac and Config.h?
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
+#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
 
 #define STRINGIFY(x) #x
 #define TOSTRING(x) STRINGIFY(x)
@@ -70,6 +71,21 @@ enum ViewMode {
   CpuWriteDiscard = 0x10 // same for now
 };
 
+struct MemoryStatus {
+  uint64_t     DeviceBytes;
+  uint64_t     DeviceLRUBytes;
+  uint64_t     DeviceMaxBytes;
+  uint64_t     HostToDeviceBytes;
+  uint64_t     DeviceToHostBytes;
+  uint64_t     HostToDeviceXfer;
+  uint64_t     DeviceToHostXfer;
+  uint64_t     DeviceEvictions;
+  uint64_t     DeviceDestroy;
+  uint64_t     DeviceAllocCacheBytes;
+  uint64_t     HostAllocCacheBytes;
+};
+
+
 class MemoryManager {
 private:
 
@@ -83,7 +99,7 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=6;
+  static const int NallocType=9;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
@@ -97,8 +113,8 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
   static void Audit(std::string s);
   static void Init(void);
   static void InitMessage(void);
@@ -119,7 +135,28 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
- 
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
+  
+  static uint64_t     DeviceCacheBytes();
+  static uint64_t     HostCacheBytes();
+
+  static MemoryStatus GetFootprint(void) {
+    MemoryStatus stat;
+    stat.DeviceBytes       = DeviceBytes;
+    stat.DeviceLRUBytes    = DeviceLRUBytes;
+    stat.DeviceMaxBytes    = DeviceMaxBytes;
+    stat.HostToDeviceBytes = HostToDeviceBytes;
+    stat.DeviceToHostBytes = DeviceToHostBytes;
+    stat.HostToDeviceXfer  = HostToDeviceXfer;
+    stat.DeviceToHostXfer  = DeviceToHostXfer;
+    stat.DeviceEvictions   = DeviceEvictions;
+    stat.DeviceDestroy     = DeviceDestroy;
+    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
+    stat.HostAllocCacheBytes   = HostCacheBytes();
+    return stat;
+  };
+  
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////
@@ -176,6 +213,7 @@ private:
 
  public:
   static void Print(void);
+  static void PrintAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -28,6 +28,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -115,8 +117,10 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t) NULL;
     dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
@@ -126,8 +130,14 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove entry
-  // Cannot be locked. If allocated must be in LRU pool.
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
@@ -139,15 +149,17 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
     dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  uint64_t CpuPtr = AccCache.CpuPtr;
-  EntryErase(CpuPtr);
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -221,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back();
+      uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
+    } else {
+      return;
     }
   }
 }
@@ -322,7 +337,8 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
     assert(0);
   }
 
-  // If view is opened on device remove from LRU
+  assert(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
     dprintf("AccCache entry removed from LRU \n");
@@ -388,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  if (!AccCache.AccPtr) {
-     EvictVictims(bytes);
-  }
+  // CPU doesn't need to free space
+  //  if (!AccCache.AccPtr) {
+  //    EvictVictims(bytes);
+  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -444,20 +461,28 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -467,13 +492,13 @@ void  MemoryManager::Print(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -489,6 +514,25 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
 }
 void MemoryManager::Audit(std::string s)
 {
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  uint64_t LockedBytes=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    assert(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    assert(AccCache.LRU_valid==1);
+    assert(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -498,7 +542,14 @@ void MemoryManager::Audit(std::string s)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    if ( AccCache.cpuLock || AccCache.accLock ) { 
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      assert(AccCache.LRU_valid==0);
+
       std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 		<< "\t cpuLock  " << AccCache.cpuLock
@@ -509,6 +560,15 @@ void MemoryManager::Audit(std::string s)
     assert( AccCache.cpuLock== 0 ) ;
     assert( AccCache.accLock== 0 ) ;
   }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  assert(LruBytes1==LruBytes2);
+  assert(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  assert(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  assert(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+
 }
 
 void MemoryManager::PrintState(void* _CpuPtr)
@@ -526,8 +586,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
     if ( AccCache.state==EvictNext) str = std::string("EvictNext");
 
     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
     << "\t" << AccCache.cpuLock
     << "\t" << AccCache.accLock
     << "\t" << AccCache.LRU_valid<<std::endl;

Grid/allocator/MemoryManagerShared.cc

@@ -12,6 +12,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
@@ -22,6 +24,7 @@ void  MemoryManager::PrintState(void* CpuPtr)
 std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
 };
 void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_mpi3.cc

@@ -400,9 +400,6 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
-  acceleratorCopySynchronise();
-  StencilBarrier();// Synch shared memory on a single nodes
-
   int nreq=list.size();
 
   if (nreq==0) return;

Grid/communicator/SharedMemoryMPI.cc

@@ -37,10 +37,11 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
 #endif
-#ifdef GRID_SYCl
-
+#ifdef GRID_SYCL
+#define GRID_SYCL_LEVEL_ZERO_IPC
 #endif
 
+
 NAMESPACE_BEGIN(Grid); 
 #define header "SharedMemoryMpi: "
 /*Construct from an MPI communicator*/

Grid/cshift/Cshift_common.h

@@ -297,6 +297,30 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   }
 }
 
+#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
+
+template <typename T>
+T iDivUp(T a, T b) // Round a / b to nearest higher integer value
+{ return (a % b != 0) ? (a / b + 1) : (a / b); }
+
+template <typename T>
+__global__ void populate_Cshift_table(T* vector, T lo, T ro, T e1, T e2, T stride)
+{
+    int idx = blockIdx.x*blockDim.x + threadIdx.x;
+    if (idx >= e1*e2) return;
+
+    int n, b, o;
+
+    n = idx / e2;
+    b = idx % e2;
+    o = n*stride + b;
+
+    vector[2*idx + 0] = lo + o;
+    vector[2*idx + 1] = ro + o;
+}
+
+#endif
+
 //////////////////////////////////////////////////////
 // local to node block strided copies
 //////////////////////////////////////////////////////
@@ -321,12 +345,20 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int ent=0;
 
   if(cbmask == 0x3 ){
+#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
+    ent = e1*e2;
+    dim3 blockSize(acceleratorThreads());
+    dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
+    populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
+    accelerator_barrier();
+#else
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int o =n*stride+b;
 	Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
       }
     }
+#endif
   } else { 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
@@ -377,11 +409,19 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int ent=0;
 
   if ( cbmask == 0x3 ) {
+#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
+    ent = e1*e2;
+    dim3 blockSize(acceleratorThreads());
+    dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
+    populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
+    accelerator_barrier();
+#else
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
       Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
+#endif
   } else {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){

Grid/lattice/Lattice_base.h

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

Grid/lattice/Lattice_reduction.h

@@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
+#if defined(GRID_SYCL)
+#include <Grid/lattice/Lattice_reduction_sycl.h>
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -124,7 +127,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -133,7 +136,7 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -142,7 +145,7 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   return sumD_gpu_large(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -150,33 +153,44 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 }
 
 template<class vobj>
-inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
-  autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu(&arg_v[0],osites);
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+  autoView( arg_v, arg, AcceleratorRead);
+  return sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
-  Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
+  return sum_cpu(&arg_v[0],osites);
 #endif  
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSum(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
 
 template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)
+#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu_large(&arg_v[0],osites);
+  return sum_gpu_large(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
+  return sum_cpu(&arg_v[0],osites);
 #endif
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSumLarge(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -232,11 +246,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-    
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
-
+    // This code could read coalesce
     // GPU - SIMT lane compliance...
     accelerator_for( ss, sites, nsimd,{
 	auto x_l = left_v(ss);

Grid/lattice/Lattice_reduction_gpu.h

@@ -211,13 +211,25 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   assert(ok);
 
   Integer smemSize = numThreads * sizeof(sobj);
-
+  // Move out of UVM
+  // Turns out I had messed up the synchronise after move to compute stream
+  // as running this on the default stream fools the synchronise
+#undef UVM_BLOCK_BUFFER  
+#ifndef UVM_BLOCK_BUFFER  
+  commVector<sobj> buffer(numBlocks);
+  sobj *buffer_v = &buffer[0];
+  sobj result;
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
+  accelerator_barrier();
+  acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
+#else
   Vector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
-  
-  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  sobj result;
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
-  auto result = buffer_v[0];
+  result = *buffer_v;
+#endif
   return result;
 }
 

Grid/lattice/Lattice_reduction_sycl.h

@@ -0,0 +1,125 @@
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Possibly promote to double and sum
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_objectD sobjD;
+  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
+  sobj identity; zeroit(identity);
+  sobj ret ; 
+
+  Integer nsimd= vobj::Nsimd();
+  
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{osites},
+		      Reduction,
+		      [=] (cl::sycl::id<1> item, auto &sum) {
+      auto osite   = item[0];
+      sum +=Reduce(lat[osite]);
+     });
+   });
+  theGridAccelerator->wait();
+  ret = mysum[0];
+  free(mysum,*theGridAccelerator);
+  sobjD dret; convertType(dret,ret);
+  return dret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_tensor(lat,osites);
+}
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
+{
+  return sumD_gpu_large(lat,osites);
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Return as same precision as input performing reduction in double precision though
+/////////////////////////////////////////////////////////////////////////////////////////////////////////
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu(lat,osites);
+  return result;
+}
+
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu_large(lat,osites);
+  return result;
+}
+
+NAMESPACE_END(Grid);
+
+/*
+template<class Double> Double svm_reduce(Double *vec,uint64_t L)
+{
+  Double sumResult; zeroit(sumResult);
+  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
+  Double identity;  zeroit(identity);
+  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
+     cgh.parallel_for(cl::sycl::range<1>{L},
+		      Reduction,
+		      [=] (cl::sycl::id<1> index, auto &sum) {
+	 sum +=vec[index];
+     });
+   });
+  theGridAccelerator->wait();
+  Double ret = d_sum[0];
+  free(d_sum,*theGridAccelerator);
+  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
+  return ret;
+}
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_type  scalar;
+
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobjD;
+
+  sobjD ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int nsimd = vobj::Nsimd();
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<scalar> buffer(osites*nsimd);
+  scalar *buf = &buffer[0];
+  vector *dat = (vector *)lat;
+
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,nsimd,{
+	int lane = acceleratorSIMTlane(nsimd);
+	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
+    });
+    //Precision change at this point is to late to gain precision
+    ret_p[w] = svm_reduce(buf,nsimd*osites);
+  }
+  return ret;
+}
+*/

Grid/lattice/Lattice_rng.h

@@ -440,6 +440,7 @@ public:
 	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
 
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
+
 	assert(rank == _grid->ThisRank() );
 	
 	int l_idx=generator_idx(o_idx,i_idx);

Grid/lattice/Lattice_transfer.h

@@ -288,7 +288,36 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
   }
 }
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               const std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = fineData.size();
+  assert(coarseData.size() == NBatch);
 
+  GridBase * fine  = fineData[0].Grid();
+  GridBase * coarse= coarseData[0].Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  std::vector<Lattice<vobj>> fineDataCopy = fineData;
+
+  autoView(ip_, ip, AcceleratorWrite);
+  for(int v=0;v<nbasis;v++) {
+    for (int k=0; k<NBatch; k++) {
+      autoView( coarseData_ , coarseData[k], AcceleratorWrite);
+      blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
+      accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+        convertType(coarseData_[sc](v),ip_[sc]);
+      });
+
+      // improve numerical stability of projection
+      // |fine> = |fine> - <basis|fine> |basis>
+      ip=-ip;
+      blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]); 
+    }
+  }
+}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -590,6 +619,26 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 }
 #endif
 
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = coarseData.size();
+  assert(fineData.size() == NBatch);
+
+  GridBase * fine   = fineData[0].Grid();
+  GridBase * coarse = coarseData[0].Grid();
+  for (int k=0; k<NBatch; k++)
+    fineData[k]=Zero();
+  for (int i=0;i<nbasis;i++) {
+    for (int k=0; k<NBatch; k++) {
+      Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
+      blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
+    }
+  }
+}
+
 // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
 // Simd layouts need not match since we use peek/poke Local
 template<class vobj,class vvobj>
@@ -1080,6 +1129,7 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
+//Very fast precision change. Requires in/out objects to reside on same Grid (e.g. by using double2 for the double-precision field)
 template<class VobjOut, class VobjIn>
 void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 {
@@ -1097,9 +1147,9 @@ void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
       precisionChange(vout,vin,N);
   });
 }
-//Convert a Lattice from one precision to another
+//Convert a Lattice from one precision to another (original, slow implementation)
 template<class VobjOut, class VobjIn>
-void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
+void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 {
   assert(out.Grid()->Nd() == in.Grid()->Nd());
   for(int d=0;d<out.Grid()->Nd();d++){
@@ -1145,6 +1195,128 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
   });
 }
 
+//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
+class precisionChangeWorkspace{
+  std::pair<Integer,Integer>* fmap_device; //device pointer
+  //maintain grids for checking
+  GridBase* _out_grid;
+  GridBase* _in_grid;
+public:
+  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
+    //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
+    assert(out_grid->Nd() == in_grid->Nd());
+    for(int d=0;d<out_grid->Nd();d++){
+      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
+    }
+    int Nsimd_out = out_grid->Nsimd();
+
+    std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
+    for(int lane=0; lane < out_grid->Nsimd(); lane++)
+      out_grid->iCoorFromIindex(out_icorrs[lane], lane);
+  
+    std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
+    thread_for(out_oidx,out_grid->oSites(),{
+	Coordinate out_ocorr; 
+	out_grid->oCoorFromOindex(out_ocorr, out_oidx);
+      
+	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
+	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
+	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
+	
+	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
+	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
+	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
+	  int in_oidx = 0, in_lane = 0;
+	  for(int d=0;d<in_grid->_ndimension;d++){
+	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
+	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
+	  }
+	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
+	}
+      });
+
+    //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
+    size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
+    fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
+    acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes); 
+  }
+
+  //Prevent moving or copying
+  precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
+  precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
+  precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
+  
+  std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
+
+  void checkGrids(GridBase* out, GridBase* in) const{
+    conformable(out, _out_grid);
+    conformable(in, _in_grid);
+  }
+  
+  ~precisionChangeWorkspace(){
+    acceleratorFreeDevice(fmap_device);
+  }
+};
+
+
+//We would like to use precisionChangeFast when possible. However usage of this requires the Grids to be the same (runtime check)
+//*and* the precisionChange(VobjOut::vector_type, VobjIn, int) function to be defined for the types; this requires an extra compile-time check which we do using some SFINAE trickery
+template<class VobjOut, class VobjIn>
+auto _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, int dummy)->decltype( precisionChange( ((typename VobjOut::vector_type*)0), ((typename VobjIn::vector_type*)0), 1), int()){
+  if(out.Grid() == in.Grid()){
+    precisionChangeFast(out,in);
+    return 1;
+  }else{
+    return 0;
+  }
+}
+template<class VobjOut, class VobjIn>
+int _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, long dummy){ //note long here is intentional; it means the above is preferred if available
+  return 0;
+}
+
+
+//Convert a lattice of one precision to another. Much faster than original implementation but requires a pregenerated workspace
+//which contains the mapping data.
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
+  if(_precisionChangeFastWrap(out,in,0)) return;
+  
+  static_assert( std::is_same<typename VobjOut::scalar_typeD, typename VobjIn::scalar_typeD>::value == 1, "precisionChange: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+
+  out.Checkerboard() = in.Checkerboard();
+  constexpr int Nsimd_out = VobjOut::Nsimd();
+
+  workspace.checkGrids(out.Grid(),in.Grid());
+  std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
+
+  //Do the copy/precision change
+  autoView( out_v , out, AcceleratorWrite);
+  autoView( in_v , in, AcceleratorRead);
+
+  accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
+      std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
+      for(int out_lane=0; out_lane < Nsimd_out; out_lane++){      
+	int in_oidx = fmap_osite[out_lane].first;
+	int in_lane = fmap_osite[out_lane].second;
+	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
+      }
+    });
+}
+
+//Convert a Lattice from one precision to another. Much faster than original implementation but slower than precisionChangeFast
+//or precisionChange called with pregenerated workspace, as it needs to internally generate the workspace on the host and copy to device
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
+  if(_precisionChangeFastWrap(out,in,0)) return;   
+  precisionChangeWorkspace workspace(out.Grid(), in.Grid());
+  precisionChange(out, in, workspace);
+}
+
+
+
+
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////

Grid/perfmon/PerfCount.h

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

Grid/pugixml/pugixml.cc

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

Grid/qcd/QCD.h

@@ -507,9 +507,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
+#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuWrite);
+  autoView(f_v,f,CpuRead);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -521,12 +532,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
+#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
+#ifdef FAST_FERM_TO_PROP
+  autoView(p_v,p,CpuRead);
+  autoView(f_v,f,CpuWrite);
+  thread_for(idx,p_v.oSites(),{
+      for(int ss = 0; ss < Ns; ++ss) {
+      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
+	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
+      }}
+    });
+#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -538,6 +560,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/fermion/CloverHelpers.h

@@ -205,15 +205,18 @@ public:
   typedef WilsonCloverHelpers<Impl> Helpers;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
     Clover += diag_mass;
   }
 
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
-                          CloverTriangleField& Triangle,
-                          RealD csw_t, RealD diag_mass) {
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
 
-    // Do nothing
+    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
   }
 
   // TODO: implement Cmunu for better performances with compact layout, but don't do it
@@ -238,9 +241,17 @@ public:
   template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-    // do nothing!
-    // mass term is multiplied to exp(Clover) below
+  // Can this be avoided?
+  static void IdentityTimesC(const CloverField& in, RealD c) {
+    int DimRep = Impl::Dimension;
+
+    autoView(in_v, in, AcceleratorWrite);
+
+    accelerator_for(ss, in.Grid()->oSites(), 1, {
+      for (int sa=0; sa<Ns; sa++)
+        for (int ca=0; ca<DimRep; ca++)
+          in_v[ss]()(sa,sa)(ca,ca) = c;
+    });
   }
 
   static int getNMAX(RealD prec, RealD R) {
@@ -255,175 +266,62 @@ public:
     return NMAX;
   }
 
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
-  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
+  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
 
-  	  typedef iMatrix<ComplexD,6> mat;
+    GridBase* grid = Clover.Grid();
+    CloverField ExpClover(grid);
 
-  	  RealD qn[6];
-  	  RealD qnold[6];
-  	  RealD p[5];
-  	  RealD trA2, trA3, trA4;
+    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
 
-  	  mat A2, A3, A4, A5;
-  	  A2 = alpha * alpha * arg * arg;
-  	  A3 = alpha * arg * A2;
-  	  A4 = A2 * A2;
-  	  A5 = A2 * A3;
+    Clover *= (1.0/diag_mass);
 
-  	  trA2 = toReal( trace(A2) );
-  	  trA3 = toReal( trace(A3) );
-  	  trA4 = toReal( trace(A4));
-
-  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
-  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
-  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
-  	  p[3] = trA3 / 3.0;
-  	  p[4] = 0.5 * trA2;
-
-  	  qnold[0] = cN[Niter];
-  	  qnold[1] = 0.0;
-  	  qnold[2] = 0.0;
-  	  qnold[3] = 0.0;
-  	  qnold[4] = 0.0;
-  	  qnold[5] = 0.0;
-
-  	  for(int i = Niter-1; i >= 0; i--)
-  	  {
-  	   qn[0] = p[0] * qnold[5] + cN[i];
-  	   qn[1] = p[1] * qnold[5] + qnold[0];
-  	   qn[2] = p[2] * qnold[5] + qnold[1];
-  	   qn[3] = p[3] * qnold[5] + qnold[2];
-  	   qn[4] = p[4] * qnold[5] + qnold[3];
-  	   qn[5] = qnold[4];
-
-  	   qnold[0] = qn[0];
-  	   qnold[1] = qn[1];
-  	   qnold[2] = qn[2];
-  	   qnold[3] = qn[3];
-  	   qnold[4] = qn[4];
-  	   qnold[5] = qn[5];
-  	  }
-
-  	  mat unit(1.0);
-
-  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
-
-    }
-
-  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Diagonal.Grid();
-    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
-
-    //
-    // Implementation completely in Daniel's layout
-    //
-
-    // Taylor expansion with Cayley-Hamilton recursion
-    // underlying Horner scheme as above
+    // Taylor expansion, slow but generic
+    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+    // qN = cN
+    // qn = cn + qn+1 X
     std::vector<RealD> cn(NMAX+1);
     cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++){
+    for (int i=1; i<=NMAX; i++)
       cn[i] = cn[i-1] / RealD(i);
-    }
 
-      // Taken over from Daniel's implementation
-      conformable(Diagonal, Triangle);
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * Clover + cn[i];
 
-      long lsites = grid->lSites();
-    {
-      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
-      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
-      typedef iMatrix<ComplexD,6> mat;
+    // prepare inverse
+    CloverInv = (-1.0)*Clover;
 
-      autoView(diagonal_v,  Diagonal,  CpuRead);
-      autoView(triangle_v,  Triangle,  CpuRead);
-      autoView(diagonalExp_v, Diagonal, CpuWrite);
-      autoView(triangleExp_v, Triangle, CpuWrite);
+    Clover = ExpClover * diag_mass;
 
-      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
+    ExpClover = Zero();
+    IdentityTimesC(ExpClover, cn[NMAX]);
+    for (int i=NMAX-1; i>=0; i--)
+      ExpClover = ExpClover * CloverInv + cn[i];
 
-    	  mat srcCloverOpUL(0.0); // upper left block
-    	  mat srcCloverOpLR(0.0); // lower right block
-    	  mat ExpCloverOp;
+    CloverInv = ExpClover * (1.0/diag_mass);
 
-        scalar_object_diagonal diagonal_tmp     = Zero();
-        scalar_object_diagonal diagonal_exp_tmp = Zero();
-        scalar_object_triangle triangle_tmp     = Zero();
-        scalar_object_triangle triangle_exp_tmp = Zero();
-
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-
-        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
-        peekLocalSite(triangle_tmp, triangle_v, lcoor);
-
-        int block;
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-        		if (i == j){
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-        		}
-        		else{
-        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-        		}
-        	}
-        }
-        block = 1;
-        for(int i = 0; i < 6; i++){
-          	for(int j = 0; j < 6; j++){
-           		if (i == j){
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
-           		}
-           		else{
-           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-           		}
-            }
-        }
-
-        // exp(Clover)
-
-        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 0;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-            	if (i == j){
-            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-            	}
-            	else if(i < j){
-            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-            	}
-           	}
-        }
-
-        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
-
-        block = 1;
-        for(int i = 0; i < 6; i++){
-        	for(int j = 0; j < 6; j++){
-              	if (i == j){
-              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
-               	}
-               	else if(i < j){
-               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
-               	}
-            }
-        }
-
-        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
-        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
-      });
-    }
-
-    Diagonal *= diag_mass;
-    Triangle *= diag_mass;
   }
 
+  static void InvertClover(CloverField& InvClover,
+                            const CloverDiagonalField& diagonal,
+                            const CloverTriangleField& triangle,
+                            CloverDiagonalField&       diagonalInv,
+                            CloverTriangleField&       triangleInv,
+                            bool fixedBoundaries) {
+
+    if (fixedBoundaries)
+    {
+      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+    }
+    else
+    {
+      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
+    }
+  }
 
   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
     assert(0);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -36,7 +36,7 @@ NAMESPACE_BEGIN(Grid);
 // Wilson compressor will need FaceGather policies for:
 // Periodic, Dirichlet, and partial Dirichlet for DWF
 ///////////////////////////////////////////////////////////////
-const int dwf_compressor_depth=1;
+const int dwf_compressor_depth=2;
 #define DWF_COMPRESS
 class FaceGatherPartialDWF
 {
@@ -110,7 +110,7 @@ public:
   ////////////////////////////////////////////////////////////////////////////////////////////
   template<class vobj,class cobj,class compressor>
   static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				    Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
     GridBase *Grid = rhs.Grid();
@@ -209,7 +209,7 @@ public:
   }
   template<class vobj,class cobj,class compressor>
   static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				    Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
     //    std::cout << " face gather exch DWF partial "<<partial <<std::endl;
@@ -320,7 +320,7 @@ public:
     typedef decltype(coalescedRead(in0))    sobj;
     typedef decltype(coalescedRead(out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
@@ -484,24 +484,26 @@ public:
 
     int dag = compress.dag;
     int face_idx=0;
+#define vet_same_node(a,b) \
+      { auto tmp = b;  }
     if ( dag ) { 
-      assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
-      assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
-      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
-      assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
-      assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
-      assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
-      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
-      assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
+      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
+      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
+      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
+      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
+      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
+      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
+      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
+      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
     } else {
-      assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
-      assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
-      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
-      assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
-      assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
-      assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
-      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
-      assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
+      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
+      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
+      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
+      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
+      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
+      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
+      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
+      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);

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

@@ -48,7 +48,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , csw_r(_csw_r)
   , csw_t(_csw_t)
   , cF(_cF)
-  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
   , Diagonal(&Fgrid),        Triangle(&Fgrid)
   , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
   , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
@@ -67,7 +67,7 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (open_boundaries) {
+  if (fixedBoundaries) {
     this->BoundaryMaskEven.Checkerboard() = Even;
     this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
@@ -77,31 +77,31 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->open_boundaries) ApplyBoundaryMask(out);
+  if(this->fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
-  if(this->open_boundaries) {
+  if(this->fixedBoundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
@@ -112,7 +112,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in,
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -121,19 +121,19 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& i
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -147,7 +147,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -166,7 +166,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(open_boundaries) ApplyBoundaryMask(out);
+  if(fixedBoundaries) ApplyBoundaryMask(out);
 }
 
 template<class Impl, class CloverHelpers>
@@ -186,7 +186,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField
 
 template<class Impl, class CloverHelpers>
 void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!open_boundaries); // TODO check for changes required for open bc
+  assert(!fixedBoundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -305,6 +305,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   GridBase* grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
   CloverField TmpOriginal(grid);
+  CloverField TmpInverse(grid);
 
   // Compute the field strength terms mu>nu
   double t2 = usecond();
@@ -324,24 +325,27 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-  // Handle mass term based on clover policy
-  CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
-  
-  // Convert the data layout of the clover term
+
+  // Instantiate the clover term
+  // - In case of the standard clover the mass term is added
+  // - In case of the exponential clover the clover term is exponentiated
   double t4 = usecond();
+  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
+
+  // Convert the data layout of the clover term
+  double t5 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Exponentiate the clover (nothing happens in case of the standard clover)
-  double t5 = usecond();
-  CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
-
-  // Possible modify the boundary values
+  // Modify the clover term at the temporal boundaries in case of open boundary conditions
   double t6 = usecond();
-  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
+  // Invert the Clover term
+  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
+  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
+  // TODO: For now this inversion is explictly done on the CPU
   double t7 = usecond();
-  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
+  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
 
   // Fill the remaining clover fields
   double t8 = usecond();
@@ -362,10 +366,10 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
   std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
   std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "convert =                    " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "exponentiation =             " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "boundaries =                 " << (t7 - t6) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "inversions =                 " << (t8 - t7) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
+  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
   std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
   std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
 }

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

@@ -439,6 +439,17 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
     int sF = ss*Ls;							\
     WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v);		\
   });
+#define ASM_CALL_SLICE(A)						\
+  auto grid = in.Grid() ;						\
+  int nt = grid->LocalDimensions()[4];					\
+  int nxyz = Nsite/nt ;							\
+  for(int t=0;t<nt;t++){						\
+  thread_for( sss, nxyz, {						\
+    int ss = t*nxyz+sss;						\
+    int sU = ss;							\
+    int sF = ss*Ls;							\
+    WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v);		\
+    });}
 
 template <class Impl>
 void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
@@ -452,11 +463,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteSycl);    return; }
-#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
-#endif     
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
@@ -467,6 +474,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
@@ -491,10 +499,9 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
-     acceleratorFenceComputeStream();
    } else if( interior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteDagInt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteDagInt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
@@ -505,7 +512,6 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
-     acceleratorFenceComputeStream();
    }
    assert(0 && " Kernel optimisation case not covered ");
   }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h

@@ -127,6 +127,8 @@ NAMESPACE_BEGIN(Grid);
 	  ApproxNegPowerAction.tolerances[i]    = action_tolerance[i];
 	  ApproxHalfPowerAction.tolerances[i]   = action_tolerance[i];
 	  ApproxNegHalfPowerAction.tolerances[i]= action_tolerance[i];
+	}
+	for(int i=0;i<ApproxPowerMD.tolerances.size();i++){
 	  ApproxPowerMD.tolerances[i]       = md_tolerance[i];
 	  ApproxNegPowerMD.tolerances[i]    = md_tolerance[i];
 	  ApproxHalfPowerMD.tolerances[i]   = md_tolerance[i];

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h

@@ -29,6 +29,8 @@
 #ifndef QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
 #define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
 
+#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h>
+
 NAMESPACE_BEGIN(Grid);
 
     /////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -58,7 +60,7 @@ NAMESPACE_BEGIN(Grid);
       //Allow derived classes to override the multishift CG
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
 #if 0
-	SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOp : DenOp);
+	SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOpD : DenOpD);
 	ConjugateGradientMultiShift<FermionFieldD> msCG(MaxIter, approx);
 	msCG(schurOp,in, out);
 #else
@@ -66,7 +68,8 @@ NAMESPACE_BEGIN(Grid);
 	SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
 	FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
 	FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
-	
+
+	// Action better with higher precision?
 	ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
 	precisionChange(inD2,in);
 	std::cout << "msCG single solve "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
@@ -76,12 +79,12 @@ NAMESPACE_BEGIN(Grid);
       }
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
 	SchurDifferentiableOperator<ImplD2> schurOpD2(numerator ? NumOpD2 : DenOpD2);
-	SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
+	SchurDifferentiableOperator<ImplF>  schurOpF (numerator ? NumOpF  : DenOpF);
 
 	FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
 	FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
 	std::vector<FermionFieldD2> out_elemsD2(out_elems.size(),NumOpD2.FermionRedBlackGrid());
-	ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	ConjugateGradientMultiShiftMixedPrecCleanup<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
 	precisionChange(inD2,in);
 	std::cout << "msCG in "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
 	msCG(schurOpD2, inD2, out_elemsD2, outD2);

Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h

@@ -112,40 +112,27 @@ NAMESPACE_BEGIN(Grid);
         // NumOp == V
         // DenOp == M
         //
-    AUDIT();
         FermionField etaOdd (NumOp.FermionRedBlackGrid());
         FermionField etaEven(NumOp.FermionRedBlackGrid());
         FermionField tmp    (NumOp.FermionRedBlackGrid());
 
-    AUDIT();
         pickCheckerboard(Even,etaEven,eta);
-    AUDIT();
         pickCheckerboard(Odd,etaOdd,eta);
 
-    AUDIT();
         NumOp.ImportGauge(U);
-    AUDIT();
         DenOp.ImportGauge(U);
 	std::cout << " TwoFlavourRefresh:  Imported gauge "<<std::endl;
-    AUDIT();
 
         SchurDifferentiableOperator<Impl> Mpc(DenOp);
-    AUDIT();
         SchurDifferentiableOperator<Impl> Vpc(NumOp);
-    AUDIT();
 
 	std::cout << " TwoFlavourRefresh: Diff ops "<<std::endl;
-    AUDIT();
         // Odd det factors
         Mpc.MpcDag(etaOdd,PhiOdd);
-    AUDIT();
 	std::cout << " TwoFlavourRefresh: MpcDag "<<std::endl;
         tmp=Zero();
-    AUDIT();
 	std::cout << " TwoFlavourRefresh: Zero() guess "<<std::endl;
-    AUDIT();
         HeatbathSolver(Vpc,PhiOdd,tmp);
-    AUDIT();
 	std::cout << " TwoFlavourRefresh: Heatbath solver "<<std::endl;
         Vpc.Mpc(tmp,PhiOdd);            
 	std::cout << " TwoFlavourRefresh: Mpc "<<std::endl;

Grid/qcd/hmc/integrators/Integrator.h

@@ -134,14 +134,12 @@ protected:
       double start_force = usecond();
 
       std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] before"<<std::endl;
-      AUDIT();
       
       as[level].actions.at(a)->deriv_timer_start();
       as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
       as[level].actions.at(a)->deriv_timer_stop();
 
       std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] after"<<std::endl;
-      AUDIT();
 
       std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
       auto name = as[level].actions.at(a)->action_name();
@@ -284,6 +282,15 @@ public:
 		  << as[level].actions.at(actionID)->deriv_us*1.0e-6<<" s"<< std::endl;
       }
     }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Dslash counts "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    uint64_t full, partial, dirichlet;
+    DslashGetCounts(dirichlet,partial,full);
+    std::cout << GridLogMessage << " Full BCs               : "<<full<<std::endl;
+    std::cout << GridLogMessage << " Partial dirichlet BCs  : "<<partial<<std::endl;
+    std::cout << GridLogMessage << " Dirichlet BCs          : "<<dirichlet<<std::endl;
+
     std::cout << GridLogMessage << "--------------------------- "<<std::endl;
     std::cout << GridLogMessage << " Force average size "<<std::endl;
     std::cout << GridLogMessage << "------------------------- "<<std::endl;
@@ -373,12 +380,12 @@ public:
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
 
 	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] before"<<std::endl;
-	AUDIT();
+
 	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
 	as[level].actions.at(actionID)->refresh_timer_stop();
 	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] after"<<std::endl;
-	AUDIT();
+
       }
 
       // Refresh the higher representation actions
@@ -415,7 +422,7 @@ public:
     // Actions
     for (int level = 0; level < as.size(); ++level) {
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
-	AUDIT();
+
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
@@ -425,7 +432,7 @@ public:
    	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
-	AUDIT();
+
       }
       as[level].apply(S_hireps, Representations, level, H);
     }
@@ -438,9 +445,9 @@ public:
     void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, int level, RealD& H) {
       
       for (int a = 0; a < repr_set.size(); ++a) {
-	AUDIT();
+
         RealD Hterm = repr_set.at(a)->Sinitial(Rep.U);
-	AUDIT();
+
         std::cout << GridLogMessage << "Sinitial Level " << level << " term " << a << " H Hirep = " << Hterm << std::endl;
         H += Hterm;
 
@@ -465,10 +472,10 @@ public:
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
 	        as[level].actions.at(actionID)->S_timer_start();
-	AUDIT();
+
         Hterm = as[level].actions.at(actionID)->Sinitial(Us);
    	        as[level].actions.at(actionID)->S_timer_stop();
-	AUDIT();
+
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
       }
@@ -481,7 +488,6 @@ public:
   
   void integrate(Field& U) 
   {
-    AUDIT();
     // reset the clocks
     t_U = 0;
     for (int level = 0; level < as.size(); ++level) {
@@ -499,10 +505,8 @@ public:
       assert(fabs(t_U - t_P[level]) < 1.0e-6);  // must be the same
       std::cout << GridLogIntegrator << " times[" << level << "]= " << t_P[level] << " " << t_U << std::endl;
     }
-    AUDIT();
 
     FieldImplementation::Project(U);
-    AUDIT();
 
     // and that we indeed got to the end of the trajectory
     assert(fabs(t_U - Params.trajL) < 1.0e-6);

Grid/stencil/SimpleCompressor.h

@@ -36,7 +36,7 @@ public:
   }
   template<class vobj,class cobj,class compressor>
   static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				    Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
     assert( (table.size()&0x1)==0);

Grid/stencil/Stencil.cc

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

Grid/stencil/Stencil.h

@@ -91,11 +91,14 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
 ///////////////////////////////////////////////////////////////////
 template<class cobj,class vobj,class compressor>
 void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
-				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
+				 commVector<cobj *> pointers,
+				 int dimension,int plane,
+				 int cbmask,compressor &compress,int type) __attribute__((noinline));
 
 template<class cobj,class vobj,class compressor>
-void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
+				 const Lattice<vobj> &rhs,
+				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -103,19 +106,26 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const La
   int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
   auto rhs_v = rhs.View(AcceleratorRead);
+  auto rhs_p = &rhs_v[0];
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
   accelerator_forNB(j, num, vobj::Nsimd(), {
-      compress.CompressExchange(p0,p1, &rhs_v[0], j,
-			      so+tp[2*j  ].second,
-			      so+tp[2*j+1].second,
-			      type);
+      compress.CompressExchange(p0,p1, rhs_p, j,
+				so+tp[2*j  ].second,
+				so+tp[2*j+1].second,
+				type);
   });
   rhs_v.ViewClose();
 }
 */
 
+void DslashResetCounts(void);
+void DslashGetCounts(uint64_t &dirichlet,uint64_t &partial,uint64_t &full);
+void DslashLogFull(void);
+void DslashLogPartial(void);
+void DslashLogDirichlet(void);
+
 struct StencilEntry {
 #ifdef GRID_CUDA
   uint64_t _byte_offset;       // 8 bytes
@@ -257,8 +267,8 @@ public:
   struct Merge {
     static constexpr int Nsimd = vobj::Nsimd();
     cobj * mpointer;
-    Vector<scalar_object *> rpointers;
-    Vector<cobj *> vpointers;
+    //    std::vector<scalar_object *> rpointers;
+    std::vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
     Integer partial; // partial dirichlet BCs
@@ -308,6 +318,7 @@ public:
 
   int face_table_computed;
   int partialDirichlet;
+  int fullDirichlet;
   std::vector<commVector<std::pair<int,int> > > face_table ;
   Vector<int> surface_list;
 
@@ -358,7 +369,7 @@ public:
     return 0;
   }
 #else
-  // 
+  // fancy calculation for shm code
   inline int SameNode(int point) {
 
     int dimension    = this->_directions[point];
@@ -378,7 +389,7 @@ public:
 
     int nbr_proc;
     if (displacement>0) nbr_proc = 1;
-    else                 nbr_proc = pd-1;
+    else                nbr_proc = pd-1;
 
     // FIXME  this logic needs to be sorted for three link term
     //    assert( (displacement==1) || (displacement==-1));
@@ -423,7 +434,6 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
-    accelerator_barrier();
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromBegin(MpiReqs,
 					Packets[i].send_buf,
@@ -437,6 +447,12 @@ public:
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
     _grid->StencilSendToRecvFromComplete(MpiReqs,0);
+    if   ( this->partialDirichlet ) DslashLogPartial();
+    else if ( this->fullDirichlet ) DslashLogDirichlet();
+    else DslashLogFull();
+    acceleratorCopySynchronise();
+    // Everyone agrees we are all done
+    _grid->StencilBarrier(); 
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
@@ -514,7 +530,6 @@ public:
   {
     _grid->StencilBarrier();// Synch shared memory on a single nodes
 
-    // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
 
     u_comm_offset=0;
@@ -625,7 +640,7 @@ public:
     d.buffer_size = buffer_size;
     dv.push_back(d);
   }
-  void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
+  void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
     Merge m;
     m.partial  = this->partialDirichlet;
     m.dims     = _grid->_fdimensions;
@@ -640,8 +655,8 @@ public:
     CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    _grid->StencilBarrier();// Synch shared memory on a single nodes
-    CommsMerge(decompress,MergersSHM,DecompressionsSHM);
+    assert(MergersSHM.size()==0);
+    assert(DecompressionsSHM.size()==0);
   }
 
   template<class decompressor>
@@ -650,9 +665,11 @@ public:
     for(int i=0;i<mm.size();i++){
       decompressor::MergeFace(decompress,mm[i]);
     }
+    if ( mm.size() )    acceleratorFenceComputeStream();
     for(int i=0;i<dd.size();i++){
       decompressor::DecompressFace(decompress,dd[i]);
     }
+    if ( dd.size() )    acceleratorFenceComputeStream();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -765,6 +782,10 @@ public:
     if ( p.dirichlet.size() ==0 ) p.dirichlet.resize(grid->Nd(),0);
     partialDirichlet = p.partialDirichlet;
     DirichletBlock(p.dirichlet); // comms send/recv set up
+    fullDirichlet=0;
+    for(int d=0;d<p.dirichlet.size();d++){
+      if (p.dirichlet[d]) fullDirichlet=1;
+    }
 
     _unified_buffer_size=0;
     surface_list.resize(0);
@@ -1268,8 +1289,8 @@ public:
     
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    Vector<cobj *> rpointers(maxl);
-    Vector<cobj *> spointers(maxl);
+    std::vector<cobj *> rpointers(maxl);
+    std::vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where

Grid/tensors/Tensor_extract_merge.h

@@ -226,7 +226,7 @@ template<class vobjOut, class vobjIn>
 accelerator_inline 
 void copyLane(vobjOut & __restrict__ vecOut, int lane_out, const vobjIn & __restrict__ vecIn, int lane_in)
 {
-  static_assert( std::is_same<typename vobjOut::DoublePrecision, typename vobjIn::DoublePrecision>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
+  static_assert( std::is_same<typename vobjOut::scalar_typeD, typename vobjIn::scalar_typeD>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
 
   typedef typename vobjOut::vector_type ovector_type;  
   typedef typename vobjIn::vector_type ivector_type;  
@@ -251,9 +251,9 @@ void copyLane(vobjOut & __restrict__ vecOut, int lane_out, const vobjIn & __rest
   ovector_type * __restrict__ op = (ovector_type *)&vecOut;
   ivector_type * __restrict__ ip = (ivector_type *)&vecIn;
   for(int w=0;w<owords;w++){
-    itmp = ip[iNsimd*w].getlane(lane_in);
+    itmp = ip[w].getlane(lane_in);
     otmp = itmp; //potential precision change
-    op[oNsimd*w].putlane(otmp,lane_out);
+    op[w].putlane(otmp,lane_out);
   }
 }
 

Grid/threads/Accelerator.h

@@ -305,14 +305,14 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
      });	   			              \
     });
 
-#define accelerator_barrier(dummy) { printf(" theGridAccelerator::wait()\n");  theGridAccelerator->wait(); }
+#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
 
-inline void acceleratorCopySynchronise(void) {  printf(" theCopyAccelerator::wait()\n"); theCopyAccelerator->wait(); }
+inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {  theCopyAccelerator->memcpy(to,from,bytes);}
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}

Grid/util/Init.cc

@@ -167,14 +167,13 @@ void GridCmdOptionInt(std::string &str,int & val)
   return;
 }
 
-void GridCmdOptionFloat(std::string &str,float & val)
+void GridCmdOptionFloat(std::string &str,double & val)
 {
   std::stringstream ss(str);
   ss>>val;
   return;
 }
 
-
 void GridParseLayout(char **argv,int argc,
 		     Coordinate &latt_c,
 		     Coordinate &mpi_c)

Grid/util/Init.h

@@ -57,7 +57,7 @@ void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec);
 template<class VectorInt>
 void GridCmdOptionIntVector(const std::string &str,VectorInt & vec);
 void GridCmdOptionInt(std::string &str,int & val);
-void GridCmdOptionFloat(std::string &str,float & val);
+void GridCmdOptionFloat(std::string &str,double & val);
 
 
 void GridParseLayout(char **argv,int argc,

HMC/Mobius2p1f_DD_EOFA_96I_mshift.cc

@@ -164,11 +164,6 @@ int main(int argc, char **argv) {
   typedef MobiusEOFAFermionF FermionEOFAActionF;
   typedef typename FermionActionF::FermionField FermionFieldF;
 
-  typedef WilsonImplD2 FermionImplPolicyD2;
-  typedef MobiusFermionD2 FermionActionD2;
-  typedef MobiusEOFAFermionD2 FermionEOFAActionD2;
-  typedef typename FermionActionD2::FermionField FermionFieldD2;
-
   typedef Grid::XmlReader       Serialiser;
 
   //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
@@ -232,31 +227,34 @@ int main(int argc, char **argv) {
   //  std::vector<Real> hasenbusch({ light_mass, 0.005, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass }); // Updated
   //  std::vector<Real> hasenbusch({ light_mass, 0.0145, 0.045, 0.108, 0.25, 0.51 , 0.75 , pv_mass });
 
-  OneFlavourRationalParams OFRp; // Up/down
-  OFRp.lo       = 4.0e-5;
+  int SP_iters=10000;
+  
+  RationalActionParams OFRp; // Up/down
+  OFRp.lo       = 6.0e-5;
   OFRp.hi       = 90.0;
-  OFRp.MaxIter  = 60000;
-  OFRp.tolerance= 1.0e-5;
-  OFRp.mdtolerance= 1.0e-3;
+  OFRp.inv_pow  = 2;
+  OFRp.MaxIter  = SP_iters; // get most shifts by 2000, stop sharing space
+  OFRp.action_tolerance= 1.0e-8;
+  OFRp.action_degree   = 18;
+  OFRp.md_tolerance= 1.0e-5;
+  OFRp.md_degree   = 14;
   //  OFRp.degree   = 20; converges
   //  OFRp.degree   = 16;
-  OFRp.degree   = 18;
   OFRp.precision= 80;
   OFRp.BoundsCheckFreq=0;
   std::vector<RealD> ActionTolByPole({
-      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-7,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8
     });
   std::vector<RealD> MDTolByPole({
-      1.0e-5,5.0e-6,1.0e-6,1.0e-7, // soften convergence more more
+      1.6e-5,5.0e-6,1.0e-6,3.0e-7, // soften convergence more more
       //      1.0e-6,3.0e-7,1.0e-7,1.0e-7,
       //      3.0e-6,1.0e-6,1.0e-7,1.0e-7, // soften convergence
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
-      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8
     });
 
@@ -265,10 +263,8 @@ int main(int argc, char **argv) {
 
   typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
   typedef SchurDiagMooeeOperator<FermionAction ,FermionField > LinearOperatorD;
-  typedef SchurDiagMooeeOperator<FermionActionD2,FermionFieldD2 > LinearOperatorD2;
   typedef SchurDiagMooeeOperator<FermionEOFAActionF,FermionFieldF> LinearOperatorEOFAF;
   typedef SchurDiagMooeeOperator<FermionEOFAAction ,FermionField > LinearOperatorEOFAD;
-  typedef SchurDiagMooeeOperator<FermionEOFAActionD2,FermionFieldD2 > LinearOperatorEOFAD2;
   typedef MixedPrecisionConjugateGradientOperatorFunction<MobiusFermionD,MobiusFermionF,LinearOperatorD,LinearOperatorF> MxPCG;
   typedef MixedPrecisionConjugateGradientOperatorFunction<MobiusEOFAFermionD,MobiusEOFAFermionF,LinearOperatorEOFAD,LinearOperatorEOFAF> MxPCG_EOFA;
 
@@ -321,7 +317,6 @@ int main(int argc, char **argv) {
   // temporarily need a gauge field
   LatticeGaugeFieldD  U(GridPtr); U=Zero();
   LatticeGaugeFieldF  UF(GridPtrF); UF=Zero();
-  LatticeGaugeFieldD2 UD2(GridPtrF); UD2=Zero();
 
   std::cout << GridLogMessage << " Running the HMC "<< std::endl;
   TheHMC.ReadCommandLine(argc,argv);  // params on CML or from param file
@@ -340,6 +335,7 @@ int main(int argc, char **argv) {
   ParamsDirF.dirichlet=Dirichlet;
   ParamsDir.partialDirichlet=1;
   ParamsDirF.partialDirichlet=1;
+  std::cout << GridLogMessage<< "Partial Dirichlet depth is "<<dwf_compressor_depth<<std::endl;
 
   //  double StoppingCondition = 1e-14;
   //  double MDStoppingCondition = 1e-9;
@@ -424,7 +420,7 @@ int main(int argc, char **argv) {
 	 ActionCGL, ActionCGR,
 	 DerivativeCGL, DerivativeCGR,
 	 SFRp, true);
-  //  Level2.push_back(&EOFA);
+  Level2.push_back(&EOFA);
 
   ////////////////////////////////////
   // up down action
@@ -449,17 +445,15 @@ int main(int argc, char **argv) {
   std::vector<FermionAction *> Denominators;
   std::vector<FermionActionF *> NumeratorsF;
   std::vector<FermionActionF *> DenominatorsF;
-  std::vector<FermionActionD2 *> NumeratorsD2;
-  std::vector<FermionActionD2 *> DenominatorsD2;
   std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
   std::vector<MxPCG *> ActionMPCG;
   std::vector<MxPCG *> MPCG;
   
 #define MIXED_PRECISION
 #ifdef MIXED_PRECISION
-  std::vector<OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2> *> Bdys;
+  std::vector<GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicy> *> Bdys;
 #else
-  std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
+  std::vector<GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
 #endif
 
   typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
@@ -532,31 +526,19 @@ int main(int argc, char **argv) {
       Quotients.push_back (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],*MPCG[h],*ActionMPCG[h],CG));
     } else {
 #ifdef MIXED_PRECISION
-      // Use the D2 data types and make them use same grid as single
-      FermionActionD2::ImplParams ParamsDenD2(boundary);
-      FermionActionD2::ImplParams ParamsNumD2(boundary);
-
-      ParamsDenD2.dirichlet = ParamsDen.dirichlet;
-      ParamsDenD2.partialDirichlet = ParamsDen.partialDirichlet;
-      DenominatorsD2.push_back(new FermionActionD2(UD2,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_den[h],M5,b,c, ParamsDenD2));
-
-      ParamsNumD2.dirichlet = ParamsNum.dirichlet;
-      ParamsNumD2.partialDirichlet = ParamsNum.partialDirichlet;
-      NumeratorsD2.push_back  (new FermionActionD2(UD2,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_num[h],M5,b,c, ParamsNumD2));
-    
-      Bdys.push_back( new OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
+      Bdys.push_back( new GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicy>(
 			   *Numerators[h],*Denominators[h],
 			   *NumeratorsF[h],*DenominatorsF[h],
-			   *NumeratorsD2[h],*DenominatorsD2[h],
-			   OFRp, 400) );
-      Bdys.push_back( new OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
+			   *Numerators[h],*Denominators[h],
+			   OFRp, SP_iters) );
+      Bdys.push_back( new GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicy>(
 			   *Numerators[h],*Denominators[h],
 			   *NumeratorsF[h],*DenominatorsF[h],
-			   *NumeratorsD2[h],*DenominatorsD2[h],
-			   OFRp, 400) );
+			   *Numerators[h],*Denominators[h],
+			   OFRp, SP_iters) );
 #else
-      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
-      Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+      Bdys.push_back( new GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
+      Bdys.push_back( new GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
 #endif
     }
   }

HMC/Mobius2p1f_EOFA_96I_hmc.cc

@@ -183,7 +183,7 @@ int main(int argc, char **argv) {
   // 4/2 => 0.6 dH
   // 3/3 => 0.8 dH .. depth 3, slower
   //MD.MDsteps =  4;
-  MD.MDsteps =  3;
+  MD.MDsteps =  12;
   MD.trajL   = 0.5;
 
   HMCparameters HMCparams;
@@ -200,8 +200,8 @@ int main(int argc, char **argv) {
   TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
 
   CheckpointerParameters CPparams;
-  CPparams.config_prefix = "ckpoint_DDHMC_lat";
-  CPparams.rng_prefix    = "ckpoint_DDHMC_rng";
+  CPparams.config_prefix = "ckpoint_HMC_lat";
+  CPparams.rng_prefix    = "ckpoint_HMC_rng";
   CPparams.saveInterval  = 1;
   CPparams.format        = "IEEE64BIG";
   TheHMC.Resources.LoadNerscCheckpointer(CPparams);
@@ -228,7 +228,7 @@ int main(int argc, char **argv) {
   Real pv_mass      = 1.0;
   //  std::vector<Real> hasenbusch({ 0.01, 0.045, 0.108, 0.25, 0.51 , pv_mass });
   //  std::vector<Real> hasenbusch({ light_mass, 0.01, 0.045, 0.108, 0.25, 0.51 , pv_mass });
-  std::vector<Real> hasenbusch({ light_mass, 0.005, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass }); // Updated
+  std::vector<Real> hasenbusch({ 0.005, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass }); // Updated
   //  std::vector<Real> hasenbusch({ light_mass, 0.0145, 0.045, 0.108, 0.25, 0.51 , 0.75 , pv_mass });
 
   auto GridPtr   = TheHMC.Resources.GetCartesian();
@@ -299,8 +299,8 @@ int main(int argc, char **argv) {
   ////////////////////////////////////
   // Collect actions
   ////////////////////////////////////
-  ActionLevel<HMCWrapper::Field> Level1(1);
-  ActionLevel<HMCWrapper::Field> Level2(3);
+  //  ActionLevel<HMCWrapper::Field> Level1(1);
+  ActionLevel<HMCWrapper::Field> Level2(1);
   ActionLevel<HMCWrapper::Field> Level3(15);
 
   ////////////////////////////////////
@@ -369,7 +369,7 @@ int main(int argc, char **argv) {
 	 ActionCGL, ActionCGR,
 	 DerivativeCGL, DerivativeCGR,
 	 SFRp, true);
-  //  Level2.push_back(&EOFA);
+  Level2.push_back(&EOFA);
 
   ////////////////////////////////////
   // up down action
@@ -477,7 +477,7 @@ int main(int argc, char **argv) {
   // Gauge action
   /////////////////////////////////////////////////////////////
   Level3.push_back(&GaugeAction);
-  TheHMC.TheAction.push_back(Level1);
+  //  TheHMC.TheAction.push_back(Level1);
   TheHMC.TheAction.push_back(Level2);
   TheHMC.TheAction.push_back(Level3);
   std::cout << GridLogMessage << " Action complete "<< std::endl;

TODO

@@ -1,3 +1,12 @@
+- - Slice sum optimisation & A2A - atomic addition
+- - Also faster non-atomic reduction
+- - Remaining PRs
+- - DDHMC
+  - - MixedPrec is the action eval, high precision
+  - - MixedPrecCleanup is the force eval, low precision
+
+=================
+=================
 Lattice_basis.h -- > HIP and SYCL GPU code
 
 
@@ -8,6 +17,7 @@ DDHMC
 -- Multishift Mixed Precision - DONE
 -- Pole dependent residual  - DONE
 
+
 =======
 -- comms threads issue??
 -- Part done: Staggered kernel performance on GPU

benchmarks/Benchmark_dwf_fp32_paranoid.cc

@@ -0,0 +1,387 @@
+ /*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid
+    Source file: ./benchmarks/Benchmark_dwf.cc
+    Copyright (C) 2015
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#ifdef GRID_CUDA
+#define CUDA_PROFILE
+#endif
+
+#ifdef CUDA_PROFILE
+#include <cuda_profiler_api.h>
+#endif
+
+using namespace std;
+using namespace Grid;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
+  };
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+
+  int threads = GridThread::GetThreads();
+
+  Coordinate latt4 = GridDefaultLatt();
+  int Ls=16;
+  for(int i=0;i<argc;i++)
+    if(std::string(argv[i]) == "-Ls"){
+      std::stringstream ss(argv[i+1]); ss >> Ls;
+    }
+
+  GridLogLayout();
+
+  long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
+
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
+  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
+  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+  LatticeFermionF src   (FGrid); random(RNG5,src);
+  LatticeFermionF src1   (FGrid); random(RNG5,src1);
+#if 0
+  src = Zero();
+  {
+    Coordinate origin({0,0,0,latt4[2]-1,0});
+    SpinColourVectorF tmp;
+    tmp=Zero();
+    tmp()(0)(0)=Complex(-2.0,0.0);
+    std::cout << " source site 0 " << tmp<<std::endl;
+    pokeSite(tmp,src,origin);
+  }
+#else
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+#endif
+
+
+  LatticeFermionF result(FGrid); result=Zero();
+  LatticeFermionF    ref(FGrid);    ref=Zero();
+  LatticeFermionF    tmp(FGrid);
+  LatticeFermionF    err(FGrid);
+
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
+  LatticeGaugeFieldF Umu(UGrid);
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+#if 0
+  Umu=1.0;
+  for(int mu=0;mu<Nd;mu++){
+    LatticeColourMatrixF ttmp(UGrid);
+    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
+    //    if (mu !=2 ) ttmp = 0;
+    //    ttmp = ttmp* pow(10.0,mu);
+    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
+  }
+  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
+#endif
+
+  ////////////////////////////////////
+  // Naive wilson implementation
+  ////////////////////////////////////
+  // replicate across fifth dimension
+  //  LatticeGaugeFieldF Umu5d(FGrid);
+  std::vector<LatticeColourMatrixF> U(4,UGrid);
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+  std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
+
+  if (1)
+  {
+    ref = Zero();
+    for(int mu=0;mu<Nd;mu++){
+
+      tmp = Cshift(src,mu+1,1);
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = U_v[ss]*tmp_v[Ls*ss+s];
+	  }
+	}
+      }
+      ref=ref + tmp - Gamma(Gmu[mu])*tmp;
+
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	autoView( src_v, src    , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
+	  }
+	}
+      }
+      tmp =Cshift(tmp,mu+1,-1);
+      ref=ref + tmp + Gamma(Gmu[mu])*tmp;
+    }
+    ref = -0.5*ref;
+  }
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  RealD NP = UGrid->_Nprocessors;
+  RealD NN = UGrid->NodeCount();
+
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionR::Dhop                  "<<std::endl;
+  std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
+  std::cout << GridLogMessage<< "* VComplexF size is "<<sizeof(vComplexF)<< " B"<<std::endl;
+  if ( sizeof(RealF)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+  if ( sizeof(RealF)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  int ncall =100;
+
+  if (1) {
+    FGrid->Barrier();
+    Dw.Dhop(src,result,0);
+    std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      Dw.Dhop(src1,result,0);
+      Dw.Dhop(src,result,0);
+      err = ref-result;
+      std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+      assert (norm2(err)< 1.0e-4 );
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=single_site_flops*volume*ncall;
+
+    auto nsimd = vComplex::Nsimd();
+    auto simdwidth = sizeof(vComplex);
+
+    // RF: Nd Wilson * Ls, Nd gauge * Ls, Nc colors
+    double data_rf = volume * ((2*Nd+1)*Nd*Nc + 2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
+
+    // mem: Nd Wilson * Ls, Nd gauge, Nc colors
+    double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
+
+    std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
+    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
+    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
+    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
+    err = ref-result;
+    std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+    //exit(0);
+
+    if(( norm2(err)>1.0e-4) ) {
+
+      /*
+      std::cout << "RESULT\n " << result<<std::endl;
+      std::cout << "REF   \n " << ref   <<std::endl;
+      std::cout << "ERR   \n " << err   <<std::endl;
+      */
+      std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
+      FGrid->Barrier();
+      exit(-1);
+    }
+    assert (norm2(err)< 1.0e-4 );
+  }
+
+  if (1)
+  { // Naive wilson dag implementation
+    ref = Zero();
+    for(int mu=0;mu<Nd;mu++){
+
+      //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
+      tmp = Cshift(src,mu+1,1);
+      {
+	autoView( ref_v, ref, CpuWrite);
+	autoView( tmp_v, tmp, CpuRead);
+	autoView( U_v  , U[mu]  , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    int i=s+Ls*ss;
+	    ref_v[i]+= U_v[ss]*(tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]); ;
+	  }
+	}
+      }
+      
+      {
+	autoView( tmp_v  , tmp  , CpuWrite);
+	autoView( U_v  , U[mu]  , CpuRead);
+	autoView( src_v, src    , CpuRead);
+	for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
+	  for(int s=0;s<Ls;s++){
+	    tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
+	  }
+	}
+      }
+      //      tmp =adj(U[mu])*src;
+      tmp =Cshift(tmp,mu+1,-1);
+      {
+	autoView( ref_v, ref, CpuWrite);
+	autoView( tmp_v, tmp, CpuRead);
+	for(int i=0;i<ref_v.size();i++){
+	  ref_v[i]+= tmp_v[i] - Gamma(Gmu[mu])*tmp_v[i]; ;
+	}
+      }
+    }
+    ref = -0.5*ref;
+  }
+  //  dump=1;
+  Dw.Dhop(src,result,1);
+  std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
+  std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
+  std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
+  std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
+  err = ref-result;
+  std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
+  if((norm2(err)>1.0e-4)){
+/*
+	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
+	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
+	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
+*/
+  }
+  LatticeFermionF src_e (FrbGrid);
+  LatticeFermionF src_o (FrbGrid);
+  LatticeFermionF r_e   (FrbGrid);
+  LatticeFermionF r_o   (FrbGrid);
+  LatticeFermionF r_eo  (FGrid);
+
+  std::cout<<GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"<<std::endl;
+  pickCheckerboard(Even,src_e,src);
+  pickCheckerboard(Odd,src_o,src);
+
+  std::cout<<GridLogMessage << "src_e"<<norm2(src_e)<<std::endl;
+  std::cout<<GridLogMessage << "src_o"<<norm2(src_o)<<std::endl;
+
+
+  // S-direction is INNERMOST and takes no part in the parity.
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionF::DhopEO                "<<std::endl;
+  std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
+  if ( sizeof(RealF)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+  if ( sizeof(RealF)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  {
+    FGrid->Barrier();
+    Dw.DhopEO(src_o,r_e,DaggerNo);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+#ifdef CUDA_PROFILE
+      if(i==10) cudaProfilerStart();
+#endif
+      Dw.DhopEO(src_o,r_e,DaggerNo);
+#ifdef CUDA_PROFILE
+      if(i==20) cudaProfilerStop();
+#endif
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=(single_site_flops*volume*ncall)/2.0;
+
+    std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
+    std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
+  }
+  Dw.DhopEO(src_o,r_e,DaggerNo);
+  Dw.DhopOE(src_e,r_o,DaggerNo);
+  Dw.Dhop  (src  ,result,DaggerNo);
+
+  std::cout<<GridLogMessage << "r_e"<<norm2(r_e)<<std::endl;
+  std::cout<<GridLogMessage << "r_o"<<norm2(r_o)<<std::endl;
+  std::cout<<GridLogMessage << "res"<<norm2(result)<<std::endl;
+
+  setCheckerboard(r_eo,r_o);
+  setCheckerboard(r_eo,r_e);
+
+  err = r_eo-result;
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+  if((norm2(err)>1.0e-4)){
+    /*
+	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
+	std::cout<< "Deo REF\n " <<result  << std::endl;
+	std::cout<< "Deo ERR   \n " << err <<std::endl;
+    */
+  }
+
+  pickCheckerboard(Even,src_e,err);
+  pickCheckerboard(Odd,src_o,err);
+  std::cout<<GridLogMessage << "norm diff even  "<< norm2(src_e)<<std::endl;
+  std::cout<<GridLogMessage << "norm diff odd   "<< norm2(src_o)<<std::endl;
+
+  assert(norm2(src_e)<1.0e-4);
+  assert(norm2(src_o)<1.0e-4);
+  Grid_finalize();
+  exit(0);
+}

benchmarks/Benchmark_halo.cc

@@ -0,0 +1,131 @@
+ /*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid
+    Source file: ./benchmarks/Benchmark_dwf.cc
+    Copyright (C) 2015
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#ifdef GRID_CUDA
+#define CUDA_PROFILE
+#endif
+
+#ifdef CUDA_PROFILE
+#include <cuda_profiler_api.h>
+#endif
+
+using namespace std;
+using namespace Grid;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
+  };
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  Coordinate latt4= GridDefaultLatt();
+  Coordinate mpi  = GridDefaultMpi();
+  Coordinate simd = GridDefaultSimd(Nd,vComplexF::Nsimd());
+
+  GridLogLayout();
+
+  int Ls=16;
+  for(int i=0;i<argc;i++)
+    if(std::string(argv[i]) == "-Ls"){
+      std::stringstream ss(argv[i+1]); ss >> Ls;
+    }
+
+  
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,simd ,mpi);
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
+  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
+  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+  LatticeFermionF src   (FGrid); random(RNG5,src);
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
+  LatticeGaugeFieldF Umu(UGrid);
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  RealD NP = UGrid->_Nprocessors;
+  RealD NN = UGrid->NodeCount();
+
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+
+  const int ncall = 500;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionF::HaloGatherOpt         "<<std::endl;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  {
+    typename DomainWallFermionF::Compressor compressor(0);
+    FGrid->Barrier();
+    Dw.Stencil.HaloExchangeOptGather(src,compressor);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      Dw.Stencil.HaloExchangeOptGather(src,compressor);
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+
+    double bytes=0.0;
+    if(mpi[0]) bytes+=latt4[1]*latt4[2]*latt4[3];
+    if(mpi[1]) bytes+=latt4[0]*latt4[2]*latt4[3];
+    if(mpi[2]) bytes+=latt4[0]*latt4[1]*latt4[3];
+    if(mpi[3]) bytes+=latt4[0]*latt4[1]*latt4[2];
+    bytes = bytes * Ls * 8.* (24.+12.)* 2.0;
+
+    std::cout<<GridLogMessage << "Gather us /call =   "<< (t1-t0)/ncall<<std::endl;
+    std::cout<<GridLogMessage << "Gather MBs /call =   "<< bytes*ncall/(t1-t0)<<std::endl;
+
+  }
+
+  Grid_finalize();
+  exit(0);
+}

benchmarks/Benchmark_prec_change.cc

@@ -0,0 +1,189 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_prec_change.cc
+
+    Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int Ls = 12;
+  Coordinate latt4 = GridDefaultLatt();
+
+  GridCartesian         * UGridD   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridD = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridD);
+  GridCartesian         * FGridD   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridD);
+  GridRedBlackCartesian * FrbGridD = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridD);
+
+  GridCartesian         * UGridF   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian         * FGridF   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridF);
+  GridRedBlackCartesian * FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridF);
+
+  
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  
+  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
+  GridParallelRNG          RNG4(UGridD);  RNG4.SeedFixedIntegers(seeds4);
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
+  GridParallelRNG          RNG5(FGridD);  RNG5.SeedFixedIntegers(seeds5);
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+  LatticeFermionD field_d(FGridD), tmp_d(FGridD);
+  random(RNG5,field_d); tmp_d = field_d;
+
+  LatticeFermionD2 field_d2(FGridF), tmp_d2(FGridF);
+  precisionChange(field_d2, field_d); tmp_d2 = field_d2;
+
+  LatticeFermionF field_f(FGridF), tmp_f(FGridF);
+  precisionChange(field_f, field_d); tmp_f = field_f;
+
+  int N = 500;
+
+  double time_ds = 0, time_sd = 0;
+
+  std::cout<<GridLogMessage << "Benchmarking single<->double original implementation (fields initially device-resident)" << std::endl;
+  for(int i=0;i<N;i++){
+    //We want to benchmark the typical scenario of both fields being device resident
+    //To do this, invoke an operation that will open a device view and touch all sites
+    //with a write operation that invalidates the CPU copy
+    field_d = tmp_d;
+    field_f = tmp_f;
+
+    double start=usecond();
+    precisionChangeOrig(field_d,field_f);
+    double stop=usecond();
+    time_sd += stop - start;
+
+    field_d = tmp_d;
+    field_f = tmp_f;
+
+    start=usecond();
+    precisionChangeOrig(field_f,field_d);
+    stop=usecond();
+    time_ds += stop - start;   
+  }
+  std::cout << "d->s " << time_ds/N << "us" << " s->d " << time_sd/N << "us" << std::endl;
+
+
+  precisionChangeWorkspace wk_sp_to_dp(field_d.Grid(),field_f.Grid());
+  precisionChangeWorkspace wk_dp_to_sp(field_f.Grid(),field_d.Grid());
+  
+  std::cout<<GridLogMessage << "Benchmarking single<->double with pregenerated workspace(fields initially device-resident)" << std::endl;
+  time_sd = time_ds = 0;
+  for(int i=0;i<N;i++){
+    field_d = tmp_d;
+    field_f = tmp_f;
+
+    double start=usecond();
+    precisionChange(field_d,field_f, wk_sp_to_dp);
+    double stop=usecond();
+    time_sd += stop - start;
+
+    field_d = tmp_d;
+    field_f = tmp_f;
+
+    start=usecond();
+    precisionChange(field_f,field_d, wk_dp_to_sp);
+    stop=usecond();
+    time_ds += stop - start;   
+  }
+  std::cout << "d->s " << time_ds/N << "us" << " s->d " << time_sd/N << "us" << std::endl;
+  
+  std::cout<<GridLogMessage << "Benchmarking single<->double with workspace generated on-the-fly (fields initially device-resident)" << std::endl;
+  time_sd = time_ds = 0;
+  for(int i=0;i<N;i++){
+    field_d = tmp_d;
+    field_f = tmp_f;
+
+    double start=usecond();
+    precisionChange(field_d,field_f);
+    double stop=usecond();
+    time_sd += stop - start;
+
+    field_d = tmp_d;
+    field_f = tmp_f;
+
+    start=usecond();
+    precisionChange(field_f,field_d);
+    stop=usecond();
+    time_ds += stop - start;
+
+  }
+  std::cout << "d->s " << time_ds/N << "us" << " s->d " << time_sd/N << "us" << std::endl;
+
+
+  std::cout<<GridLogMessage << "Benchmarking single<->double2 (fields initially device-resident)" << std::endl;
+  time_sd = time_ds = 0;
+  for(int i=0;i<N;i++){
+    field_d2 = tmp_d2;
+    field_f = tmp_f;
+
+    double start=usecond();
+    precisionChangeFast(field_d2,field_f);
+    double stop=usecond();
+    time_sd += stop - start;
+
+    field_d2 = tmp_d2;
+    field_f = tmp_f;
+
+    start=usecond();
+    precisionChangeFast(field_f,field_d2);
+    stop=usecond();
+    time_ds += stop - start;
+  }
+  std::cout << "d->s " << time_ds/N << "us" << " s->d " << time_sd/N << "us" << std::endl;
+
+
+  std::cout<<GridLogMessage << "Benchmarking single<->double2 through standard precisionChange call(fields initially device-resident) [NB: perf should be the same as the previous test!]" << std::endl;
+  time_sd = time_ds = 0;
+  for(int i=0;i<N;i++){
+    field_d2 = tmp_d2;
+    field_f = tmp_f;
+
+    double start=usecond();
+    precisionChange(field_d2,field_f);
+    double stop=usecond();
+    time_sd += stop - start;
+
+    field_d2 = tmp_d2;
+    field_f = tmp_f;
+
+    start=usecond();
+    precisionChange(field_f,field_d2);
+    stop=usecond();
+    time_ds += stop - start;
+  }
+  std::cout << "d->s " << time_ds/N << "us" << " s->d " << time_sd/N << "us" << std::endl;
+
+  Grid_finalize();
+}

systems/Crusher/config-command

@@ -1,12 +1,13 @@
 CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
 ../../configure --enable-comms=mpi-auto \
 --with-lime=$CLIME \
---enable-unified=no \
+--enable-unified=yes \
 --enable-shm=nvlink \
 --enable-tracing=timer \
 --enable-accelerator=hip \
 --enable-gen-simd-width=64 \
 --enable-simd=GPU \
+--disable-accelerator-cshift \
 --with-gmp=$OLCF_GMP_ROOT \
 --with-fftw=$FFTW_DIR/.. \
 --with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \

systems/PVC/benchmarks/run-1tile.sh

@@ -4,9 +4,9 @@
 #SBATCH -p QZ1J-ICX-PVC
 ##SBATCH -p QZ1J-SPR-PVC-2C
 
-source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
+#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
 
-export NT=16
+export NT=8
 
 export I_MPI_OFFLOAD=1
 export I_MPI_OFFLOAD_TOPOLIB=level_zero
@@ -21,7 +21,7 @@ export I_MPI_OFFLOAD_CELL=tile
 export EnableImplicitScaling=0
 export EnableWalkerPartition=0
 export ZE_AFFINITY_MASK=0.0
-mpiexec -launcher ssh -n 1 -host localhost  ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 --cacheblocking 8.8.8.8
+mpiexec -launcher ssh -n 1 -host localhost  ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 --device-mem 32768
 
 export ZE_AFFINITY_MASK=0
 export I_MPI_OFFLOAD_CELL=device

systems/PVC/benchmarks/run-2tile-mpi.sh

@@ -1,13 +1,13 @@
 #!/bin/bash
 ##SBATCH -p PVC-SPR-QZEH 
 ##SBATCH -p PVC-ICX-QZNW
+
 #SBATCH -p QZ1J-ICX-PVC
 
-source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
+#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
 
 export NT=16
 
-
 # export IGC_EnableLSCFenceUGMBeforeEOT=0
 # export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file=False"
 #export IGC_ShaderDumpEnable=1 
@@ -19,8 +19,15 @@ export SYCL_DEVICE_FILTER=gpu,level_zero
 export I_MPI_OFFLOAD_CELL=tile
 export EnableImplicitScaling=0
 export EnableWalkerPartition=0
+#export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
+#export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
 
-mpiexec -launcher ssh -n 1 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 > 1tile.log
+for i in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
+do
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 0  --device-mem 32768 > 1.1.1.2.log$i
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 0  --device-mem 32768 > 2.1.1.1.log$i 
+done
 
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 > 2tile.log
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0
 

systems/PVC/benchmarks/wrap.sh

@@ -5,10 +5,5 @@ export ZE_AFFINITY_MASK=0.$MPI_LOCALRANKID
 echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
 
 
-if [ $MPI_LOCALRANKID = "0" ] 
-then
-#  ~psteinbr/build_pti/ze_tracer -c $@
-  onetrace --chrome-kernel-timeline $@
-else
   $@
-fi
+

systems/PVC/config-command

@@ -1,15 +1,16 @@
-INSTALL=/nfs/site/home/azusayax/install
+INSTALL=/nfs/site/home/paboylx/prereqs/
 ../../configure \
 	--enable-simd=GPU \
 	--enable-gen-simd-width=64 \
-	--enable-comms=mpi \
+	--enable-comms=mpi-auto \
 	--disable-accelerator-cshift \
 	--disable-gparity \
 	--disable-fermion-reps \
 	--enable-shm=nvlink \
 	--enable-accelerator=sycl \
-	--enable-unified=yes \
-	CXX=mpicxx \
+	--enable-unified=no \
+	MPICXX=mpicxx \
+	CXX=dpcpp \
 	LDFLAGS="-fsycl-device-code-split=per_kernel -fsycl-device-lib=all -lze_loader -L$INSTALL/lib" \
-	CXXFLAGS="-cxx=icpx -fsycl-unnamed-lambda -fsycl -I$INSTALL/include -Wno-tautological-constant-compare"
+	CXXFLAGS="-fsycl-unnamed-lambda -fsycl -no-fma -I$INSTALL/include -Wno-tautological-compare"
 

systems/PVC/setup.sh

@@ -1,5 +1,6 @@
 export https_proxy=http://proxy-chain.intel.com:911
-export LD_LIBRARY_PATH=/nfs/site/home/azusayax/install/lib:$LD_LIBRARY_PATH
+#export LD_LIBRARY_PATH=/nfs/site/home/azusayax/install/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$HOME/prereqs/lib/:$LD_LIBRARY_PATH
 
 module load intel-release
 source /opt/intel/oneapi/PVC_setup.sh

systems/mac-arm/config-command-mpi

@@ -1,3 +1,2 @@
-CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi
-#CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GPU-RRII --enable-comms=mpi
-#CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GPU --enable-debug --enable-comms=mpi
+CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi --enable-unified=yes
+

tests/Test_cayley_even_odd_vec.cc

@@ -73,12 +73,12 @@ int main (int argc, char ** argv)
   RealD M5  =1.8;
 
   std::cout<<GridLogMessage<<"**************************************************************"<<std::endl;
-  std::cout<<GridLogMessage <<"DomainWallFermion vectorised test"<<std::endl;
+  std::cout<<GridLogMessage <<"DomainWallFermion test"<<std::endl;
   std::cout<<GridLogMessage<<"**************************************************************"<<std::endl;
   std::vector<Complex> boundary = {1,1,1,-1};
   DomainWallFermionD::ImplParams Params(boundary);
-  Coordinate Dirichlet({0,8,8,16,32});
-  Params.dirichlet=Dirichlet;
+  //  Coordinate Dirichlet({0,8,8,16,32});
+  //  Params.dirichlet=Dirichlet;
 
   DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,Params);
   TestWhat<DomainWallFermionD>(Ddwf,FGrid,FrbGrid,UGrid,mass,M5,&RNG4,&RNG5);

tests/Test_dwf_mixedcg_prec.cc

@@ -101,7 +101,7 @@ int main (int argc, char ** argv)
   std:: cout << " MdagM site flops = "<< 4*MdagMsiteflops<<std::endl;
   std:: cout << " CG    site flops = "<< CGsiteflops <<std::endl;
   int iters;
-  for(int i=0;i<200;i++){
+  for(int i=0;i<10;i++){
     result_o = Zero();
     t1=usecond();
     mCG(src_o,result_o);

tests/core/Test_fft_matt.cc

@@ -1,35 +1,12 @@
     /*************************************************************************************
 
-    grid` physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./tests/Test_cshift.cc
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/Grid.h>
-
-using namespace Grid;
- ;
+Gamma::Algebra Gmu [] = {
+  Gamma::Algebra::GammaX,
+  Gamma::Algebra::GammaY,
+  Gamma::Algebra::GammaZ,
+  Gamma::Algebra::GammaT,
+  Gamma::Algebra::Gamma5
+};
 
 int main (int argc, char ** argv)
 {
@@ -49,22 +26,8 @@ int main (int argc, char ** argv)
   GridCartesian         GRID(latt_size,simd_layout,mpi_layout);
   GridRedBlackCartesian RBGRID(&GRID);
 
-  LatticeComplexD     one(&GRID);
-  LatticeComplexD      zz(&GRID);
-  LatticeComplexD       C(&GRID);
-  LatticeComplexD  Ctilde(&GRID);
-  LatticeComplexD  Cref  (&GRID);
-  LatticeComplexD  Csav  (&GRID);
   LatticeComplexD    coor(&GRID);
 
-  LatticeSpinMatrixD    S(&GRID);
-  LatticeSpinMatrixD    Stilde(&GRID);
-  
-  Coordinate p({1,3,2,3});
-
-  one = ComplexD(1.0,0.0);
-  zz  = ComplexD(0.0,0.0);
-
   ComplexD ci(0.0,1.0);
 
   std::vector<int> seeds({1,2,3,4});
@@ -73,7 +36,6 @@ int main (int argc, char ** argv)
   pRNG.SeedFixedIntegers(seeds);
 
   LatticeGaugeFieldD Umu(&GRID);
-
   SU<Nc>::ColdConfiguration(pRNG,Umu); // Unit gauge
 
   ////////////////////////////////////////////////////
@@ -81,17 +43,79 @@ int main (int argc, char ** argv)
   ////////////////////////////////////////////////////
   {
     LatticeFermionD    src(&GRID); gaussian(pRNG,src);
+    LatticeFermionD    src_p(&GRID);
     LatticeFermionD    tmp(&GRID);
     LatticeFermionD    ref(&GRID);
+    LatticeFermionD    result(&GRID);
     
-    RealD mass=0.01;
+    RealD mass=0.1;
     WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
     
-    Dw.M(src,tmp);
+    Dw.M(src,ref);
+    std::cout << "Norm src "<<norm2(src)<<std::endl;
+    std::cout << "Norm Dw x src "<<norm2(ref)<<std::endl;
+    {
+      FFT theFFT(&GRID);
 
+      ////////////////
+      // operator in Fourier space
+      ////////////////
+      tmp =ref;
+      theFFT.FFT_all_dim(result,tmp,FFT::forward);
+      std::cout<<"FFT[ Dw x src ]  "<< norm2(result)<<std::endl;    
+
+      tmp = src;
+      theFFT.FFT_all_dim(src_p,tmp,FFT::forward);
+      std::cout<<"FFT[ src      ]  "<< norm2(src_p)<<std::endl;
+      
+      /////////////////////////////////////////////////////////////////
+      // work out the predicted FT from Fourier
+      /////////////////////////////////////////////////////////////////
+      auto FGrid = &GRID;
+      LatticeFermionD    Kinetic(FGrid); Kinetic = Zero();
+      LatticeComplexD    kmu(FGrid); 
+      LatticeInteger     scoor(FGrid); 
+      LatticeComplexD    sk (FGrid); sk = Zero();
+      LatticeComplexD    sk2(FGrid); sk2= Zero();
+      LatticeComplexD    W(FGrid); W= Zero();
+      LatticeComplexD    one(FGrid); one =ComplexD(1.0,0.0);
+      ComplexD ci(0.0,1.0);
+    
+      for(int mu=0;mu<Nd;mu++) {
+	
+	RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
+
+	LatticeCoordinate(kmu,mu);
+
+	kmu = TwoPiL * kmu;
+      
+	sk2 = sk2 + 2.0*sin(kmu*0.5)*sin(kmu*0.5);
+	sk  = sk  +     sin(kmu)    *sin(kmu); 
+      
+	// -1/2 Dw ->  1/2 gmu (eip - emip) = i sinp gmu
+	Kinetic = Kinetic + sin(kmu)*ci*(Gamma(Gmu[mu])*src_p);
+	
+      }
+    
+      W = mass + sk2; 
+      Kinetic = Kinetic + W * src_p;
+    
+      std::cout<<"Momentum space src         "<< norm2(src_p)<<std::endl;
+      std::cout<<"Momentum space Dw x src    "<< norm2(Kinetic)<<std::endl;
+      std::cout<<"FT[Coordinate space Dw]    "<< norm2(result)<<std::endl;
+    
+      result = result - Kinetic;
+      std::cout<<"diff "<< norm2(result)<<std::endl;
+      
+    }
+
+    std::cout << " =======================================" <<std::endl;
+    std::cout << " Checking FourierFreePropagator x Dw = 1" <<std::endl;
+    std::cout << " =======================================" <<std::endl;
     std::cout << "Dw src = " <<norm2(src)<<std::endl;
     std::cout << "Dw tmp = " <<norm2(tmp)<<std::endl;
-    
+    Dw.M(src,tmp);
+
     Dw.FreePropagator(tmp,ref,mass);
 
     std::cout << "Dw ref = " <<norm2(ref)<<std::endl;
@@ -122,7 +146,8 @@ int main (int argc, char ** argv)
     ferm()(0)(0) = ComplexD(1.0);
     pokeSite(ferm,src,point);
 
-    RealD mass=0.01;
+    RealD mass=0.1;
+
     WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
 
     // Momentum space prop
@@ -155,6 +180,65 @@ int main (int argc, char ** argv)
     DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
   }
 
+  ////////////////////////////////////////////////////
+  //Gauge invariance test
+  ////////////////////////////////////////////////////
+  {
+    std::cout<<"****************************************"<<std::endl;
+    std::cout << "Gauge invariance test \n";
+    std::cout<<"****************************************"<<std::endl;
+    LatticeGaugeField     U_GT(&GRID); // Gauge transformed field
+    LatticeColourMatrix   g(&GRID);    // local Gauge xform matrix
+    U_GT = Umu;
+    // Make a random xform to teh gauge field
+    SU<Nc>::RandomGaugeTransform(pRNG,U_GT,g); // Unit gauge
+
+    LatticeFermionD    src(&GRID);
+    LatticeFermionD    tmp(&GRID);
+    LatticeFermionD    ref(&GRID);
+    LatticeFermionD    diff(&GRID);
+
+    // could loop over colors
+    src=Zero();
+    Coordinate point(4,0); // 0,0,0,0
+    SpinColourVectorD ferm;
+    ferm=Zero();
+    ferm()(0)(0) = ComplexD(1.0);
+    pokeSite(ferm,src,point);
+
+    RealD mass=0.1;
+    WilsonFermionD Dw(U_GT,GRID,RBGRID,mass);
+
+    // Momentum space prop
+    std::cout << " Solving by FFT and Feynman rules" <<std::endl;
+    Dw.FreePropagator(src,ref,mass) ;
+
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    LatticeFermionD    result(&GRID); 
+    const int sdir=0;
+    
+    ////////////////////////////////////////////////////////////////////////
+    // Conjugate gradient on normal equations system
+    ////////////////////////////////////////////////////////////////////////
+    std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
+    Dw.Mdag(src,tmp);
+    src=tmp;
+    MdagMLinearOperator<WilsonFermionD,LatticeFermionD> HermOp(Dw);
+    ConjugateGradient<LatticeFermionD> CG(1.0e-10,10000);
+    CG(HermOp,src,result);
+    
+    ////////////////////////////////////////////////////////////////////////
+    std::cout << " Taking difference" <<std::endl;
+    std::cout << "Dw result "<<norm2(result)<<std::endl;
+    std::cout << "Dw ref     "<<norm2(ref)<<std::endl;
+    
+    diff = ref - result;
+    std::cout << "result - ref     "<<norm2(diff)<<std::endl;
+
+    DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
+  }
+  
   
   Grid_finalize();
 }

tests/core/Test_memory_manager.cc

@@ -0,0 +1,110 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_memory_manager.cc
+
+    Copyright (C) 2022
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+void  MemoryTest(GridCartesian         * FGrid,int N);
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+
+  int N=100;
+  for(int i=0;i<N;i++){
+    std::cout << "============================"<<std::endl;
+    std::cout << "Epoch "<<i<<"/"<<N<<std::endl;
+    std::cout << "============================"<<std::endl;
+    MemoryTest(UGrid,256);
+    MemoryManager::Print();
+    AUDIT();
+  }
+  Grid_finalize();
+}
+
+void  MemoryTest(GridCartesian         * FGrid, int N)
+{
+  LatticeComplexD zero(FGrid); zero=Zero();
+  std::vector<LatticeComplexD> A(N,zero);//FGrid);
+
+  std::vector<ComplexD> B(N,ComplexD(0.0)); // Update sequentially on host
+
+  for(int v=0;v<N;v++) A[v] = Zero();
+
+  uint64_t counter = 0;
+  for(int epoch = 0;epoch<10000;epoch++){
+
+    int v  = random() %N; // Which vec
+    int w  = random() %2; // Write or read
+    int e  = random() %3; // expression or for loop
+    int dev= random() %2; // On device?
+    //    int e=1;
+    ComplexD zc = counter++;
+    
+    if ( w ) {
+      B[v] = B[v] + zc;
+      if ( e == 0 ) {
+	A[v] = A[v] + zc - A[v] + A[v];
+      } else {
+	if ( dev ) { 
+	  autoView(A_v,A[v],AcceleratorWrite);
+	  accelerator_for(ss,FGrid->oSites(),1,{
+	    A_v[ss] = A_v[ss] + zc;
+	    });
+	} else {
+	  autoView(A_v,A[v],CpuWrite);
+	  thread_for(ss,FGrid->oSites(),{
+	      A_v[ss] = A_v[ss] + zc;
+	    });
+	}
+      }
+    } else {
+      if ( e == 0 ) {
+	A[v] = A[v] + A[v] - A[v];
+      } else { 
+	if ( dev ) { 
+	  autoView(A_v,A[v],AcceleratorRead);
+	  accelerator_for(ss,FGrid->oSites(),1,{
+	      assert(B[v]==A_v[ss]()()().getlane(0));
+	    });
+	  //	std::cout << "["<<v<<"] checked on GPU"<<B[v]<<std::endl;
+	} else {
+	  autoView(A_v,A[v],CpuRead);
+	  thread_for(ss,FGrid->oSites(),{
+	      assert(B[v]==A_v[ss]()()().getlane(0));
+	    });
+	  //	std::cout << "["<<v<<"] checked on CPU"<<B[v]<<std::endl;
+	}
+      }    
+    }
+  }
+
+}

tests/core/Test_prec_change.cc

@@ -0,0 +1,124 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/core/Test_prec_change.cc
+
+    Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int Ls = 12;
+  Coordinate latt4 = GridDefaultLatt();
+
+  GridCartesian         * UGridD   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridD = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridD);
+  GridCartesian         * FGridD   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridD);
+  GridRedBlackCartesian * FrbGridD = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridD);
+
+  GridCartesian         * UGridF   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian         * FGridF   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridF);
+  GridRedBlackCartesian * FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridF);
+
+  
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
+  GridParallelRNG          RNG5(FGridD);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG5F(FGridF);  RNG5F.SeedFixedIntegers(seeds5);
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+  LatticeFermionD field_d(FGridD), tmp_d(FGridD);
+  random(RNG5,field_d);
+  RealD norm2_field_d = norm2(field_d);
+  
+  LatticeFermionD2 field_d2(FGridF), tmp_d2(FGridF);
+  random(RNG5F,field_d2);
+  RealD norm2_field_d2 = norm2(field_d2);
+  
+  LatticeFermionF field_f(FGridF);
+  
+  //Test original implementation
+  {
+    std::cout << GridLogMessage << "Testing original implementation" << std::endl;
+    field_f = Zero();
+    precisionChangeOrig(field_f,field_d);
+    RealD Ndiff = (norm2_field_d - norm2(field_f))/norm2_field_d;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of single and double prec fields differs by " << Ndiff << std::endl;
+    tmp_d = Zero();
+    precisionChangeOrig(tmp_d, field_f);
+    Ndiff = norm2( LatticeFermionD(tmp_d-field_d) ) / norm2_field_d;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of back-converted and original double prec fields differs by " << Ndiff << std::endl;
+  }
+  //Test new implementation with pregenerated workspace
+  {
+    std::cout << GridLogMessage << "Testing new implementation with pregenerated workspace" << std::endl;
+    precisionChangeWorkspace wk_sp_to_dp(field_d.Grid(),field_f.Grid());
+    precisionChangeWorkspace wk_dp_to_sp(field_f.Grid(),field_d.Grid());
+    
+    field_f = Zero();
+    precisionChange(field_f,field_d,wk_dp_to_sp);
+    RealD Ndiff = (norm2_field_d - norm2(field_f))/norm2_field_d;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of single and double prec fields differs by " << Ndiff << std::endl;
+    tmp_d = Zero();
+    precisionChange(tmp_d, field_f,wk_sp_to_dp);
+    Ndiff = norm2( LatticeFermionD(tmp_d-field_d) ) / norm2_field_d;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of back-converted and original double prec fields differs by " << Ndiff << std::endl;
+  }
+  //Test new implementation without pregenerated workspace
+  {
+    std::cout << GridLogMessage << "Testing new implementation without pregenerated workspace" << std::endl;
+    field_f = Zero();
+    precisionChange(field_f,field_d);
+    RealD Ndiff = (norm2_field_d - norm2(field_f))/norm2_field_d;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of single and double prec fields differs by " << Ndiff << std::endl;
+    tmp_d = Zero();
+    precisionChange(tmp_d, field_f);
+    Ndiff = norm2( LatticeFermionD(tmp_d-field_d) ) / norm2_field_d;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of back-converted and original double prec fields differs by " << Ndiff << std::endl;
+  } 
+  //Test fast implementation
+  {
+    std::cout << GridLogMessage << "Testing fast (double2) implementation" << std::endl;
+    field_f = Zero();
+    precisionChangeFast(field_f,field_d2);
+    RealD Ndiff = (norm2_field_d2 - norm2(field_f))/norm2_field_d2;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of single and double prec fields differs by " << Ndiff << std::endl;
+    tmp_d2 = Zero();
+    precisionChangeFast(tmp_d2, field_f);
+    Ndiff = norm2( LatticeFermionD2(tmp_d2-field_d2) ) / norm2_field_d2;
+    std::cout << GridLogMessage << (fabs(Ndiff) > 1e-05 ? "!!FAIL" : "Pass") << ": relative norm2 of back-converted and original double prec fields differs by " << Ndiff << std::endl;
+  }
+  std::cout << "Done" << std::endl;
+  
+  Grid_finalize();
+}

tests/forces/Test_bdy.cc

@@ -0,0 +1,305 @@
+/*
+
+  2f Full det MdagM 10^6 force ~ 1.3e7
+rid : Message : 1767.283471 s : +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 1767.283476 s : S1 : 1.52885e+09
+Grid : Message : 1767.283480 s : S2 : 1.52886e+09
+Grid : Message : 1767.283482 s : dS : 8877.34
+Grid : Message : 1767.283483 s : dSpred : 8877.7
+Grid : Message : 1767.283484 s : diff : -0.360484
+Grid : Message : 1767.283485 s : *********************************************************
+
+  2f Full det MpcdagMpc 10^6 force ~ 1.8e6
+Grid : Message : 2399.576962 s : +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 2399.576968 s : S1 : 1.52885e+09
+Grid : Message : 2399.576972 s : S2 : 1.52886e+09
+Grid : Message : 2399.576974 s : dS : 9728.49
+Grid : Message : 2399.576975 s : dSpred : 9726.58
+Grid : Message : 2399.576976 s : diff : 1.90683
+Grid : Message : 2399.576977 s : *********************************************************
+
+  2f bdy MdagM 1500 force Force ~ 2800
+Grid : Message : 4622.385061 s : +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 4622.385067 s : S1 : 1.52885e+09
+Grid : Message : 4622.385071 s : S2 : 1.52885e+09
+Grid : Message : 4622.385072 s : dS : 25.4944
+Grid : Message : 4622.385073 s : dSpred : 25.4672
+Grid : Message : 4622.385074 s : diff : 0.0271414
+Grid : Message : 4622.385075 s : *********************************************************
+
+  2f bdy MpcdagMpc 10^6 force   Force ~ 2200
+Grid : Message : 4622.385061 s : +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+Grid : Message : 4622.385067 s : S1 : 1.52885e+09
+Grid : Message : 4622.385071 s : S2 : 1.52885e+09
+Grid : Message : 4622.385072 s : dS : 25.4944
+Grid : Message : 4622.385073 s : dSpred : 25.4672
+Grid : Message : 4622.385074 s : diff : 0.0271414
+Grid : Message : 4622.385075 s : *********************************************************
+  
+  1f Bdy Det
+  Optimisation log:  looser rational AND MD tolerances sloppy
+MobiusForce.221179 -- same as HMC. dS is mispredicted Forece  ~2.8
+Grid : Message : 6582.258991 s : dS : 0.024478
+Grid : Message : 6582.258992 s : dSpred : 0.00791876
+Grid : Message : 6582.258994 s : diff : 0.0165592
+
+MobiusForce.221193 -- tight rational AND MD tolerances to 1e-8 ~ 2.8 same
+Grid : Message : 1964.939209 s : S1 : 7.64404e+08
+Grid : Message : 1964.939213 s : S2 : 7.64404e+08
+Grid : Message : 1964.939215 s : dS : -0.00775838 <--- too loose even on action
+Grid : Message : 1964.939216 s : dSpred : -0.00416793 
+Grid : Message : 1964.939217 s : diff : -0.00359045
+
+MobiusForce.221394 -- looser rational, MD tol 1e-8 ~ 2.8 same
+Grid : Message : 1198.346720 s : S1 : 764404649.48886
+Grid : Message : 1198.346760 s : S2 : 764404649.5133
+Grid : Message : 1198.346780 s : dS : 0.024440884590149
+Grid : Message : 1198.346800 s : dSpred : 0.0079145154465184
+Grid : Message : 1198.346810 s : diff : 0.016526369143631
+
+MobiusForce.221394 -- tight rational, MD tol sloppy Force ~ 2.8
+Grid : Message : 2376.921950 s : S1 : 764404436.44069
+Grid : Message : 2376.921954 s : S2 : 764404436.43299
+Grid : Message : 2376.921956 s : dS : -0.0076971054077148
+Grid : Message : 2376.921958 s : dSpred : -0.0041610472282526
+Grid : Message : 2376.921959 s : diff : -0.0035360581794623
+
+*/
+
+//
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_double_ratio.cc
+
+    Copyright (C) 2022
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+typedef MobiusFermionD FermionAction;
+typedef WilsonImplD FimplD;
+typedef WilsonImplD FermionImplPolicy;
+
+template<class Gimpl>
+void ForceTest(Action<LatticeGaugeField> &action,LatticeGaugeField & U,MomentumFilterBase<LatticeGaugeField> &Filter)
+{
+  GridBase *UGrid = U.Grid();
+
+  std::vector<int> seeds({1,2,3,5});
+  GridSerialRNG            sRNG;         sRNG.SeedFixedIntegers(seeds);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds);
+
+  LatticeColourMatrix Pmu(UGrid); 
+  LatticeGaugeField P(UGrid); 
+  LatticeGaugeField UdSdU(UGrid); 
+
+  std::cout << GridLogMessage << "*********************************************************"<<std::endl;
+  std::cout << GridLogMessage << " Force test for "<<action.action_name()<<std::endl;
+  std::cout << GridLogMessage << "*********************************************************"<<std::endl;
+  
+  RealD eps=0.005;
+
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  std::cout << GridLogMessage << " Refresh "<<action.action_name()<<std::endl;
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  
+  Gimpl::generate_momenta(P,sRNG,RNG4);
+  Filter.applyFilter(P);
+
+#if  0
+  FieldMetaData header;
+  std::string file("./ckpoint_lat.2000");
+  NerscIO::readConfiguration(U,header,file);
+#else
+  U = 1.0;
+#endif
+  action.refresh(U,sRNG,RNG4);
+
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  std::cout << GridLogMessage << " Action "<<action.action_name()<<std::endl;
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+
+  RealD S1 = action.S(U);
+
+  Gimpl::update_field(P,U,eps);
+
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  std::cout << GridLogMessage << " Derivative "<<action.action_name()<<std::endl;
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  action.deriv(U,UdSdU);
+  UdSdU = Ta(UdSdU);
+  Filter.applyFilter(UdSdU);
+
+  DumpSliceNorm("Force",UdSdU,Nd-1);
+  
+  Gimpl::update_field(P,U,eps);
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  std::cout << GridLogMessage << " Action "<<action.action_name()<<std::endl;
+  std::cout << GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  
+  RealD S2 = action.S(U);
+
+  // Use the derivative
+  LatticeComplex dS(UGrid); dS = Zero();
+  for(int mu=0;mu<Nd;mu++){
+    auto UdSdUmu = PeekIndex<LorentzIndex>(UdSdU,mu);
+    Pmu= PeekIndex<LorentzIndex>(P,mu);
+    dS = dS - trace(Pmu*UdSdUmu)*eps*2.0*2.0;
+  }
+  ComplexD dSpred    = sum(dS);
+  RealD diff =  S2-S1-dSpred.real();
+
+  std::cout<< GridLogMessage << "+++++++++++++++++++++++++++++++++++++++++++++++++++++++++"<<std::endl;
+  std::cout<< GridLogMessage << "S1 : "<< S1    <<std::endl;
+  std::cout<< GridLogMessage << "S2 : "<< S2    <<std::endl;
+  std::cout<< GridLogMessage << "dS : "<< S2-S1 <<std::endl;
+  std::cout<< GridLogMessage << "dSpred : "<< dSpred.real() <<std::endl;
+  std::cout<< GridLogMessage << "diff : "<< diff<<std::endl;
+  std::cout<< GridLogMessage << "*********************************************************"<<std::endl;
+  //  assert(diff<1.0);
+  std::cout<< GridLogMessage << "Done" <<std::endl;
+  std::cout << GridLogMessage << "*********************************************************"<<std::endl;
+}
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::cout << std::setprecision(14);
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate mpi_layout  = GridDefaultMpi();
+  Coordinate simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  Coordinate shm;
+  GlobalSharedMemory::GetShmDims(mpi_layout,shm);
+
+  const int Ls=12;
+  const int Nt = latt_size[3];
+  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);
+ 
+  ////////////////////////////////////////////////////////////////
+  // Domain decomposed operator
+  ////////////////////////////////////////////////////////////////
+  Coordinate CommDim(Nd);
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi_layout[d]/shm[d])>1 ? 1 : 0;
+
+  Coordinate NonDirichlet(Nd+1,0);
+  Coordinate Dirichlet(Nd+1,0);
+  Dirichlet[1] = CommDim[0]*latt_size[0]/mpi_layout[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt_size[1]/mpi_layout[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt_size[2]/mpi_layout[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt_size[3]/mpi_layout[3] * shm[3];
+
+  Coordinate Block4(Nd);
+  Block4[0] = Dirichlet[1];
+  Block4[1] = Dirichlet[2];
+  Block4[2] = Dirichlet[3];
+  Block4[3] = Dirichlet[4];
+
+  std::vector<Complex> boundary = {1,1,1,-1};
+  FermionAction::ImplParams Params(boundary);
+  FermionAction::ImplParams ParamsDir(boundary);
+  Params.dirichlet=NonDirichlet;
+  ParamsDir.dirichlet=Dirichlet;
+  ParamsDir.partialDirichlet=1;
+
+  ///////////////////// Gauge Field and Gauge Forces ////////////////////////////
+  LatticeGaugeField U(UGrid);
+
+  RealD beta=6.0;
+  WilsonGaugeActionR PlaqAction(beta);
+  IwasakiGaugeActionR RectAction(beta);
+
+  MomentumFilterNone<LatticeGaugeField> FilterNone;
+  ForceTest<GimplTypesR>(PlaqAction,U,FilterNone);
+  ForceTest<GimplTypesR>(RectAction,U,FilterNone);
+
+  ////////////////////////////////////
+  // Action
+  ////////////////////////////////////
+  RealD mass=0.00078; 
+  RealD pvmass=1.0; 
+  RealD M5=1.8; 
+  RealD b=1.5;
+  RealD c=0.5;
+  
+  // Double versions
+  FermionAction DdwfPeriodic(U,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c,Params);
+  FermionAction PVPeriodic  (U,*FGrid,*FrbGrid,*UGrid,*UrbGrid,pvmass,M5,b,c,Params);
+  FermionAction DdwfDirichlet(U,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c,ParamsDir);
+
+  double StoppingCondition = 1.0e-8;
+  double MaxCGIterations = 50000;
+  ConjugateGradient<LatticeFermion>  CG(StoppingCondition,MaxCGIterations);
+  
+  //////////////////// Two Flavour Determinant Ratio ///////////////////////////////
+  TwoFlavourRatioPseudoFermionAction<FimplD> Nf2(PVPeriodic, DdwfPeriodic,CG,CG);
+  //  ForceTest<GimplTypesR>(Nf2,U,FilterNone);
+
+  //////////////////// Two Flavour Determinant force test Even Odd ///////////////////////////////
+  TwoFlavourEvenOddRatioPseudoFermionAction<FimplD> Nf2eo(PVPeriodic, DdwfPeriodic,CG,CG);
+  //  ForceTest<GimplTypesR>(Nf2eo,U,FilterNone);
+
+  //////////////////// Domain forces ////////////////////
+  int Width=4;
+  DDHMCFilter<WilsonImplD::Field> DDHMCFilter(Block4,Width);
+  
+  //////////////////// Two flavour boundary det  ////////////////////
+  TwoFlavourRatioPseudoFermionAction<FimplD> BdyNf2(DdwfDirichlet, DdwfPeriodic,CG,CG);
+  //  ForceTest<GimplTypesR>(BdyNf2,U,DDHMCFilter);
+
+  //////////////////// Two flavour eo boundary det  ////////////////////
+  TwoFlavourEvenOddRatioPseudoFermionAction<FimplD> BdyNf2eo(DdwfDirichlet, DdwfPeriodic,CG,CG);
+  //  ForceTest<GimplTypesR>(BdyNf2eo,U,DDHMCFilter);
+
+  //////////////////// One flavour boundary det  ////////////////////
+  OneFlavourRationalParams OFRp; // Up/down
+  OFRp.lo       = 4.0e-5;
+  OFRp.hi       = 90.0;
+  OFRp.MaxIter  = 60000;
+  OFRp.tolerance= 1.0e-8;
+  OFRp.mdtolerance= 1.0e-6;
+  OFRp.degree   = 18;
+  OFRp.precision= 80;
+  OFRp.BoundsCheckFreq=0;
+  std::vector<RealD> ActionTolByPole({
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8
+    });
+  std::vector<RealD> MDTolByPole({
+      1.0e-6,3.0e-7,1.0e-7,1.0e-7,  // Orig sloppy
+      //      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8
+    });
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> BdySqrt(DdwfDirichlet,DdwfPeriodic,OFRp);
+  ForceTest<GimplTypesR>(BdySqrt,U,DDHMCFilter);
+
+  Grid_finalize();
+}

tests/forces/Test_double_ratio.cc

@@ -476,6 +476,20 @@ int main (int argc, char ** argv)
   //  ForceTest<GimplTypesR>(BdyNf2eo,U,DDHMCFilter);
 
   //////////////////// One flavour boundary det  ////////////////////
+  RationalActionParams OFRp; // Up/down
+  OFRp.lo       = 6.0e-5;
+  OFRp.hi       = 90.0;
+  OFRp.inv_pow  = 2;
+  OFRp.MaxIter  = SP_iters; // get most shifts by 2000, stop sharing space
+  OFRp.action_tolerance= 1.0e-8;
+  OFRp.action_degree   = 18;
+  OFRp.md_tolerance= 1.0e-5;
+  OFRp.md_degree   = 14;
+  //  OFRp.degree   = 20; converges
+  //  OFRp.degree   = 16;
+  OFRp.precision= 80;
+  OFRp.BoundsCheckFreq=0;
+  /*
   OneFlavourRationalParams OFRp; // Up/down
   OFRp.lo       = 4.0e-5;
   OFRp.hi       = 90.0;
@@ -485,6 +499,23 @@ int main (int argc, char ** argv)
   OFRp.degree   = 18;
   OFRp.precision= 80;
   OFRp.BoundsCheckFreq=0;
+  */
+  std::vector<RealD> ActionTolByPole({
+      1.0e-7,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8
+    });
+  std::vector<RealD> MDTolByPole({
+      1.6e-5,5.0e-6,1.0e-6,3.0e-7, // soften convergence more more
+      //      1.0e-6,3.0e-7,1.0e-7,1.0e-7,
+      //      3.0e-6,1.0e-6,1.0e-7,1.0e-7, // soften convergence
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
+      1.0e-8,1.0e-8
+    });
+  /*
   std::vector<RealD> ActionTolByPole({
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
@@ -499,9 +530,9 @@ int main (int argc, char ** argv)
       //      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8,1.0e-8,1.0e-8,
-      1.0e-8,1.0e-8,1.0e-8,1.0e-8,
       1.0e-8,1.0e-8
     });
+  */
   OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> BdySqrt(DdwfDirichlet,DdwfPeriodic,OFRp);
   BdySqrt.SetTolerances(ActionTolByPole,MDTolByPole);
   ForceTest<GimplTypesR>(BdySqrt,U,DDHMCFilter);

tests/lanczos/Test_dwf_block_lanczos.README

@@ -0,0 +1,73 @@
+#Example script 
+DIR=/gpfs/alpine/phy157/proj-shared/phy157dwf/chulwoo/Grid/BL/build/tests/lanczos
+BIN=${DIR}/Test_dwf_block_lanczos
+
+VOL='--grid 16.16.16.32 '
+GRID='--mpi 1.1.1.4 ' 
+CONF='--gconf ckpoint_lat.IEEE64BIG.2000 '
+OPT='--mass 0.01 --M5 1.8 --phase in.params --omega in.params --shm 4096' 
+#BL='--rbl 16.1024.128.1000.10 --split 1.1.4.4 --check_int 100 --resid 1.0e-5 --cheby_l 0.007 --cheby_u 7 --cheby_n 51'
+BL='--rbl 4.128.16.100.10 --split 1.1.1.4 --check_int 25 --resid 1.0e-5 --cheby_l 0.007 --cheby_u 7 --cheby_n 51'
+
+ARGS=${CONF}" "${OPT}" "${BL}" "${VOL}" "${GRID}
+export APP="${BIN}  ${ARGS}"
+echo APP=${APP}
+#export JS="jsrun --nrs 32 -a4 -g4 -c42 -dpacked -b packed:7 --smpiargs="-gpu" "
+export JS="jsrun --nrs 1 -a4 -g4 -c42 -dpacked -b  packed:10  --smpiargs="-gpu" "
+$JS  $APP
+
+#sample in.param 
+
+boundary_phase 0 1 0
+boundary_phase 1 1 0
+boundary_phase 2 1 0
+boundary_phase 3 -1 0
+
+omega 0 0.5 0
+omega 1 0.5 0
+omega 2 0.5 0
+omega 3 0.5 0
+omega 4 0.5 0
+omega 5 0.5 0
+omega 6 0.5 0
+omega 7 0.5 0
+omega 8 0.5 0
+omega 9 0.5 0
+omega 10 0.5 0
+omega 11 0.5 0
+
+
+#output 
+
+Grid : Message : 1.717474 s :  Gauge Configuration ckpoint_lat.IEEE64BIG.2000
+Grid : Message : 1.717478 s :  boundary_phase[0] = (1,0)
+Grid : Message : 1.717497 s :  boundary_phase[1] = (1,0)
+Grid : Message : 1.717500 s :  boundary_phase[2] = (1,0)
+Grid : Message : 1.717503 s :  boundary_phase[3] = (-1,0)
+Grid : Message : 1.717506 s :  Ls 12
+Grid : Message : 1.717507 s :  mass 0.01
+Grid : Message : 1.717510 s :  M5 1.8
+Grid : Message : 1.717512 s :  mob_b 1.5
+Grid : Message : 1.717514 s :  omega[0] = (0.5,0)
+Grid : Message : 1.717517 s :  omega[1] = (0.5,0)
+Grid : Message : 1.717520 s :  omega[2] = (0.5,0)
+Grid : Message : 1.717523 s :  omega[3] = (0.5,0)
+Grid : Message : 1.717526 s :  omega[4] = (0.5,0)
+Grid : Message : 1.717529 s :  omega[5] = (0.5,0)
+Grid : Message : 1.717532 s :  omega[6] = (0.5,0)
+Grid : Message : 1.717535 s :  omega[7] = (0.5,0)
+Grid : Message : 1.717538 s :  omega[8] = (0.5,0)
+Grid : Message : 1.717541 s :  omega[9] = (0.5,0)
+Grid : Message : 1.717544 s :  omega[10] = (0.5,0)
+Grid : Message : 1.717547 s :  omega[11] = (0.5,0)
+Grid : Message : 1.717550 s :  Nu 4
+Grid : Message : 1.717551 s :  Nk 128
+Grid : Message : 1.717552 s :  Np 16
+Grid : Message : 1.717553 s :  Nm 288
+Grid : Message : 1.717554 s :  Nstop 100
+Grid : Message : 1.717555 s :  Ntest 25
+Grid : Message : 1.717557 s :  MaxIter 10
+Grid : Message : 1.717558 s :  resid 1e-05
+Grid : Message : 1.717560 s :  Cheby Poly 0.007,7,51
+
+

tests/lanczos/Test_dwf_block_lanczos.cc

@@ -0,0 +1,410 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_block_lanczos.cc
+
+    Copyright (C) 2022
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Yong-Chull Jang <ypj@quark.phy.bnl.gov>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/util/Init.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h>
+
+using namespace std;
+using namespace Grid;
+//using namespace Grid::QCD;
+
+//typedef typename GparityDomainWallFermionR::FermionField FermionField;
+typedef typename ZMobiusFermionF::FermionField FermionField;
+
+RealD AllZero(RealD x){ return 0.;}
+
+class CmdJobParams 
+{
+  public:
+    std::string gaugefile;
+
+    int Ls;
+    double mass;
+    double M5;
+    double mob_b;
+    std::vector<ComplexD> omega;
+    std::vector<Complex> boundary_phase;
+    std::vector<int> mpi_split;
+    
+    LanczosType Impl;
+    int Nu;
+    int Nk;
+    int Np;
+    int Nm;
+    int Nstop;
+    int Ntest;
+    int MaxIter;
+    double resid;
+    
+    double low;
+    double high;
+    int order;
+
+    CmdJobParams()
+      : gaugefile("Hot"),
+        Ls(8), mass(0.01), M5(1.8), mob_b(1.5),
+        Impl(LanczosType::irbl),mpi_split(4,1),
+        Nu(4), Nk(200), Np(200), Nstop(100), Ntest(1), MaxIter(10), resid(1.0e-8), 
+        low(0.2), high(5.5), order(11)
+    {Nm=Nk+Np;};
+    
+    void Parse(char **argv, int argc);
+};
+
+
+void CmdJobParams::Parse(char **argv,int argc)
+{
+  std::string arg;
+  std::vector<int> vi;
+  double re,im;
+  int expect, idx;
+  std::string vstr;
+  std::ifstream pfile;
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--gconf") ){
+    gaugefile = GridCmdOptionPayload(argv,argv+argc,"--gconf");
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--phase") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--phase");
+    pfile.open(arg);
+    assert(pfile);
+    expect = 0;
+    while( pfile >> vstr ) {
+      if ( vstr.compare("boundary_phase") == 0 ) {
+        pfile >> vstr;
+        GridCmdOptionInt(vstr,idx);
+        assert(expect==idx);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,re);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,im);
+        boundary_phase.push_back({re,im});
+        expect++;
+      }
+    }
+    pfile.close();
+  } else {
+    for (int i=0; i<4; ++i) boundary_phase.push_back({1.,0.});
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--omega") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--omega");
+    pfile.open(arg);
+    assert(pfile);
+    Ls = 0;
+    while( pfile >> vstr ) {
+      if ( vstr.compare("omega") == 0 ) {
+        pfile >> vstr;
+        GridCmdOptionInt(vstr,idx);
+        assert(Ls==idx);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,re);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,im);
+        omega.push_back({re,im});
+        Ls++;
+      }
+    }
+    pfile.close();
+  } else {
+    if( GridCmdOptionExists(argv,argv+argc,"--Ls") ){
+      arg = GridCmdOptionPayload(argv,argv+argc,"--Ls");
+      GridCmdOptionInt(arg,Ls);
+    }
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--mass") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mass");
+    GridCmdOptionFloat(arg,mass);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--M5") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--M5");
+    GridCmdOptionFloat(arg,M5);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--mob_b") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mob_b");
+    GridCmdOptionFloat(arg,mob_b);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--irbl") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--irbl");
+    GridCmdOptionIntVector(arg,vi);
+    Nu = vi[0];
+    Nk = vi[1];
+    Np = vi[2];
+    Nstop = vi[3];
+    MaxIter = vi[4];
+    // ypj[fixme] mode overriding message is needed.
+    Impl = LanczosType::irbl;
+    Nm = Nk+Np;
+  }
+  
+  // block Lanczos with explicit extension of its dimensions
+  if( GridCmdOptionExists(argv,argv+argc,"--rbl") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--rbl");
+    GridCmdOptionIntVector(arg,vi);
+    Nu = vi[0];
+    Nk = vi[1];
+    Np = vi[2]; // vector space is enlarged by adding Np vectors
+    Nstop = vi[3];
+    MaxIter = vi[4];
+    // ypj[fixme] mode overriding message is needed.
+    Impl = LanczosType::rbl;
+    Nm = Nk+Np*MaxIter;
+  }
+  
+#if 1
+  // block Lanczos with explicit extension of its dimensions
+  if( GridCmdOptionExists(argv,argv+argc,"--split") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--split");
+    GridCmdOptionIntVector(arg,vi);
+    for(int i=0;i<mpi_split.size();i++)
+    mpi_split[i] = vi[i];
+  }
+#endif
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--check_int") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--check_int");
+    GridCmdOptionInt(arg,Ntest);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--resid") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--resid");
+    GridCmdOptionFloat(arg,resid);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_l") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_l");
+    GridCmdOptionFloat(arg,low);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_u") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_u");
+    GridCmdOptionFloat(arg,high);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_n") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_n");
+    GridCmdOptionInt(arg,order);
+  }
+  
+  if ( CartesianCommunicator::RankWorld() == 0 ) {
+    std::streamsize ss = std::cout.precision();
+    std::cout << GridLogMessage <<" Gauge Configuration "<< gaugefile << '\n';
+    std::cout.precision(15);
+    for ( int i=0; i<4; ++i ) std::cout << GridLogMessage <<" boundary_phase["<< i << "] = " << boundary_phase[i] << '\n';
+    std::cout.precision(ss);
+    std::cout << GridLogMessage <<" Ls "<< Ls << '\n';
+    std::cout << GridLogMessage <<" mass "<< mass << '\n';
+    std::cout << GridLogMessage <<" M5 "<< M5 << '\n';
+    std::cout << GridLogMessage <<" mob_b "<< mob_b << '\n';
+    std::cout.precision(15);
+    for ( int i=0; i<Ls; ++i ) std::cout << GridLogMessage <<" omega["<< i << "] = " << omega[i] << '\n';
+    std::cout.precision(ss);
+    std::cout << GridLogMessage <<" Nu "<< Nu << '\n'; 
+    std::cout << GridLogMessage <<" Nk "<< Nk << '\n'; 
+    std::cout << GridLogMessage <<" Np "<< Np << '\n'; 
+    std::cout << GridLogMessage <<" Nm "<< Nm << '\n'; 
+    std::cout << GridLogMessage <<" Nstop "<< Nstop << '\n'; 
+    std::cout << GridLogMessage <<" Ntest "<< Ntest << '\n'; 
+    std::cout << GridLogMessage <<" MaxIter "<< MaxIter << '\n'; 
+    std::cout << GridLogMessage <<" resid "<< resid << '\n'; 
+    std::cout << GridLogMessage <<" Cheby Poly "<< low << "," << high << "," << order << std::endl; 
+  }
+}
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+  
+  CmdJobParams JP;
+  JP.Parse(argv,argc);
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,UGrid);
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n",UGrid,UrbGrid,FGrid,FrbGrid);
+  GridCartesian         * UGridF   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian         * FGridF   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,UGridF);
+  GridRedBlackCartesian * FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,UGridF);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGridF);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  // ypj [note] why seed RNG5 again? bug? In this case, run with a default seed().
+  GridParallelRNG          RNG5rb(FrbGridF);  RNG5rb.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid); 
+  std::vector<LatticeColourMatrix> U(4,UGrid);
+  LatticeGaugeFieldF UmuF(UGridF); 
+  std::vector<LatticeColourMatrix> UF(4,UGridF);
+  
+  if ( JP.gaugefile.compare("Hot") == 0 ) {
+    SU3::HotConfiguration(RNG4, Umu);
+  } else {
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu,header,JP.gaugefile);
+    // ypj [fixme] additional checks for the loaded configuration?
+  }
+  precisionChange (UmuF,Umu);
+  
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+  
+  RealD mass = JP.mass;
+  RealD M5 = JP.M5;
+
+// ypj [fixme] flexible support for a various Fermions
+//  RealD mob_b = JP.mob_b;      // Gparity
+//  std::vector<ComplexD> omega; // ZMobius
+  
+//  GparityMobiusFermionD ::ImplParams params;
+//  std::vector<int> twists({1,1,1,0});
+//  params.twists = twists;
+//  GparityMobiusFermionR  Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,params);
+//  SchurDiagTwoOperator<GparityMobiusFermionR,FermionField> HermOp(Ddwf);
+
+
+//  int mrhs = JP.Nu;
+  int Ndir=4;
+  auto mpi_layout  = GridDefaultMpi();
+  std::vector<int> mpi_split (Ndir,1);
+#if 0
+    int tmp=mrhs, dir=0;
+    std::cout << GridLogMessage  << "dir= "<<dir <<"tmp= "<<tmp<<"mpi_split= "<<mpi_split[dir]<<"mpi_layout= "<<mpi_split[dir]<<std::endl;
+    while ( tmp> 1) {
+    if ((mpi_split[dir]*2) <= mpi_layout[dir]){
+        mpi_split[dir] *=2;
+        tmp = tmp/2;
+    }
+    std::cout << GridLogMessage  << "dir= "<<dir <<"tmp= "<<tmp<<"mpi_split= "<<mpi_split[dir]<<"mpi_layout= "<<mpi_layout[dir]<<std::endl;
+        dir = (dir+1)%Ndir;
+    }
+#endif
+    int mrhs=1;
+    for(int i =0;i<Ndir;i++){
+      mpi_split[i] = mpi_layout[i] / JP.mpi_split[i] ;
+      mrhs *= JP.mpi_split[i];
+    }
+    std::cout << GridLogMessage  << "mpi_layout= " << mpi_layout << std::endl;
+    std::cout << GridLogMessage  << "mpi_split= " << mpi_split << std::endl;
+    std::cout << GridLogMessage  << "mrhs= " << mrhs << std::endl;
+//    assert(JP.Nu==tmp);
+
+  /////////////////////////////////////////////
+  // Split into 1^4 mpi communicators, keeping it explicitly single
+  /////////////////////////////////////////////
+  GridCartesian         * SGrid = new GridCartesian(GridDefaultLatt(),
+                                                    GridDefaultSimd(Nd,vComplexF::Nsimd()),
+                                                    mpi_split,
+                                                    *UGrid);
+
+  GridCartesian         * SFGrid   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,SGrid);
+  GridRedBlackCartesian * SrbGrid  = SpaceTimeGrid::makeFourDimRedBlackGrid(SGrid);
+  GridRedBlackCartesian * SFrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,SGrid);
+
+  LatticeGaugeFieldF s_Umu(SGrid);
+  Grid_split  (UmuF,s_Umu);
+
+  //WilsonFermionR::ImplParams params;
+  ZMobiusFermionF::ImplParams params;
+  params.overlapCommsCompute = true;
+  params.boundary_phases = JP.boundary_phase;
+  ZMobiusFermionF  DdwfF(UmuF,*FGridF,*FrbGridF,*UGridF,*UrbGridF,mass,M5,JP.omega,1.,0.,params);
+//  SchurDiagTwoOperator<ZMobiusFermionF,FermionField> HermOp(Ddwf);
+  SchurDiagOneOperator<ZMobiusFermionF,FermionField> HermOp(DdwfF);
+  ZMobiusFermionF  Dsplit(s_Umu,*SFGrid,*SFrbGrid,*SGrid,*SrbGrid,mass,M5,JP.omega,1.,0.,params);
+//  SchurDiagTwoOperator<ZMobiusFermionF,FermionField> SHermOp(Dsplit);
+  SchurDiagOneOperator<ZMobiusFermionF,FermionField> SHermOp(Dsplit);
+
+  //std::vector<double> Coeffs { 0.,-1.}; 
+  // ypj [note] this may not be supported by some compilers
+  std::vector<double> Coeffs({ 0.,-1.}); 
+  Polynomial<FermionField> PolyX(Coeffs);
+  //Chebyshev<FermionField> Cheb(0.2,5.5,11);
+  Chebyshev<FermionField> Cheb(JP.low,JP.high,JP.order);
+//  Cheb.csv(std::cout);
+  ImplicitlyRestartedBlockLanczos<FermionField> IRBL(HermOp, SHermOp,
+						     FrbGridF,SFrbGrid,mrhs,
+                                                     Cheb,
+                                                     JP.Nstop, JP.Ntest,
+                                                     JP.Nu, JP.Nk, JP.Nm,
+                                                     JP.resid,
+                                                     JP.MaxIter,
+						     IRBLdiagonaliseWithEigen);
+//						     IRBLdiagonaliseWithLAPACK);
+  IRBL.split_test=1;
+  
+  std::vector<RealD> eval(JP.Nm);
+  
+  std::vector<FermionField> src(JP.Nu,FrbGridF);
+if (0)
+{
+// in case RNG is too slow
+  std::cout << GridLogMessage << "Using RNG5"<<std::endl;
+  FermionField src_tmp(FGrid);
+  for ( int i=0; i<JP.Nu; ++i ){
+//    gaussian(RNG5,src_tmp);
+     ComplexD rnd;
+     RealD re;
+     fillScalar(re,RNG5._gaussian[0],RNG5._generators[0]);
+    std::cout << i <<" / "<< JP.Nm  <<" re "<< re  << std::endl;
+// printf("%d / %d re %e\n",i,FGrid->_processor,re);
+    src_tmp=re;
+    pickCheckerboard(Odd,src[i],src_tmp);
+  }
+  RNG5.Report();
+} else {
+  std::cout << GridLogMessage << "Using RNG5rb"<<std::endl;
+  for ( int i=0; i<JP.Nu; ++i )
+    gaussian(RNG5rb,src[i]);
+  RNG5rb.Report();
+
+}
+  
+  std::vector<FermionField> evec(JP.Nm,FrbGridF);
+  for(int i=0;i<1;++i){
+    std::cout << GridLogMessage << i <<" / "<< JP.Nm <<" grid pointer "<< evec[i].Grid() << std::endl;
+  };
+
+  int Nconv;
+  IRBL.calc(eval,evec,src,Nconv,JP.Impl);
+
+
+  Grid_finalize();
+}

tests/lanczos/Test_dwf_block_lanczos.cc.double

@@ -0,0 +1,401 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_block_lanczos.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/util/Init.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h>
+
+using namespace std;
+using namespace Grid;
+//using namespace Grid::QCD;
+
+//typedef typename GparityDomainWallFermionR::FermionField FermionField;
+typedef typename ZMobiusFermionR::FermionField FermionField;
+
+RealD AllZero(RealD x){ return 0.;}
+
+class CmdJobParams 
+{
+  public:
+    std::string gaugefile;
+
+    int Ls;
+    double mass;
+    double M5;
+    double mob_b;
+    std::vector<ComplexD> omega;
+    std::vector<Complex> boundary_phase;
+    std::vector<int> mpi_split;
+    
+    LanczosType Impl;
+    int Nu;
+    int Nk;
+    int Np;
+    int Nm;
+    int Nstop;
+    int Ntest;
+    int MaxIter;
+    double resid;
+    
+    double low;
+    double high;
+    int order;
+
+    CmdJobParams()
+      : gaugefile("Hot"),
+        Ls(8), mass(0.01), M5(1.8), mob_b(1.5),
+        Impl(LanczosType::irbl),mpi_split(4,1),
+        Nu(4), Nk(200), Np(200), Nstop(100), Ntest(1), MaxIter(10), resid(1.0e-8), 
+        low(0.2), high(5.5), order(11)
+    {Nm=Nk+Np;};
+    
+    void Parse(char **argv, int argc);
+};
+
+
+void CmdJobParams::Parse(char **argv,int argc)
+{
+  std::string arg;
+  std::vector<int> vi;
+  double re,im;
+  int expect, idx;
+  std::string vstr;
+  std::ifstream pfile;
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--gconf") ){
+    gaugefile = GridCmdOptionPayload(argv,argv+argc,"--gconf");
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--phase") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--phase");
+    pfile.open(arg);
+    assert(pfile);
+    expect = 0;
+    while( pfile >> vstr ) {
+      if ( vstr.compare("boundary_phase") == 0 ) {
+        pfile >> vstr;
+        GridCmdOptionInt(vstr,idx);
+        assert(expect==idx);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,re);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,im);
+        boundary_phase.push_back({re,im});
+        expect++;
+      }
+    }
+    pfile.close();
+  } else {
+    for (int i=0; i<4; ++i) boundary_phase.push_back({1.,0.});
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--omega") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--omega");
+    pfile.open(arg);
+    assert(pfile);
+    Ls = 0;
+    while( pfile >> vstr ) {
+      if ( vstr.compare("omega") == 0 ) {
+        pfile >> vstr;
+        GridCmdOptionInt(vstr,idx);
+        assert(Ls==idx);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,re);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,im);
+        omega.push_back({re,im});
+        Ls++;
+      }
+    }
+    pfile.close();
+  } else {
+    if( GridCmdOptionExists(argv,argv+argc,"--Ls") ){
+      arg = GridCmdOptionPayload(argv,argv+argc,"--Ls");
+      GridCmdOptionInt(arg,Ls);
+    }
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--mass") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mass");
+    GridCmdOptionFloat(arg,mass);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--M5") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--M5");
+    GridCmdOptionFloat(arg,M5);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--mob_b") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mob_b");
+    GridCmdOptionFloat(arg,mob_b);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--irbl") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--irbl");
+    GridCmdOptionIntVector(arg,vi);
+    Nu = vi[0];
+    Nk = vi[1];
+    Np = vi[2];
+    Nstop = vi[3];
+    MaxIter = vi[4];
+    // ypj[fixme] mode overriding message is needed.
+    Impl = LanczosType::irbl;
+    Nm = Nk+Np;
+  }
+  
+  // block Lanczos with explicit extension of its dimensions
+  if( GridCmdOptionExists(argv,argv+argc,"--rbl") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--rbl");
+    GridCmdOptionIntVector(arg,vi);
+    Nu = vi[0];
+    Nk = vi[1];
+    Np = vi[2]; // vector space is enlarged by adding Np vectors
+    Nstop = vi[3];
+    MaxIter = vi[4];
+    // ypj[fixme] mode overriding message is needed.
+    Impl = LanczosType::rbl;
+    Nm = Nk+Np*MaxIter;
+  }
+  
+#if 1
+  // block Lanczos with explicit extension of its dimensions
+  if( GridCmdOptionExists(argv,argv+argc,"--split") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--split");
+    GridCmdOptionIntVector(arg,vi);
+    for(int i=0;i<mpi_split.size();i++)
+    mpi_split[i] = vi[i];
+  }
+#endif
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--check_int") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--check_int");
+    GridCmdOptionInt(arg,Ntest);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--resid") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--resid");
+    GridCmdOptionFloat(arg,resid);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_l") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_l");
+    GridCmdOptionFloat(arg,low);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_u") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_u");
+    GridCmdOptionFloat(arg,high);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_n") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_n");
+    GridCmdOptionInt(arg,order);
+  }
+  
+  if ( CartesianCommunicator::RankWorld() == 0 ) {
+    std::streamsize ss = std::cout.precision();
+    std::cout << GridLogMessage <<" Gauge Configuration "<< gaugefile << '\n';
+    std::cout.precision(15);
+    for ( int i=0; i<4; ++i ) std::cout << GridLogMessage <<" boundary_phase["<< i << "] = " << boundary_phase[i] << '\n';
+    std::cout.precision(ss);
+    std::cout << GridLogMessage <<" Ls "<< Ls << '\n';
+    std::cout << GridLogMessage <<" mass "<< mass << '\n';
+    std::cout << GridLogMessage <<" M5 "<< M5 << '\n';
+    std::cout << GridLogMessage <<" mob_b "<< mob_b << '\n';
+    std::cout.precision(15);
+    for ( int i=0; i<Ls; ++i ) std::cout << GridLogMessage <<" omega["<< i << "] = " << omega[i] << '\n';
+    std::cout.precision(ss);
+    std::cout << GridLogMessage <<" Nu "<< Nu << '\n'; 
+    std::cout << GridLogMessage <<" Nk "<< Nk << '\n'; 
+    std::cout << GridLogMessage <<" Np "<< Np << '\n'; 
+    std::cout << GridLogMessage <<" Nm "<< Nm << '\n'; 
+    std::cout << GridLogMessage <<" Nstop "<< Nstop << '\n'; 
+    std::cout << GridLogMessage <<" Ntest "<< Ntest << '\n'; 
+    std::cout << GridLogMessage <<" MaxIter "<< MaxIter << '\n'; 
+    std::cout << GridLogMessage <<" resid "<< resid << '\n'; 
+    std::cout << GridLogMessage <<" Cheby Poly "<< low << "," << high << "," << order << std::endl; 
+  }
+}
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+  
+  CmdJobParams JP;
+  JP.Parse(argv,argc);
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,UGrid);
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n",UGrid,UrbGrid,FGrid,FrbGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  // ypj [note] why seed RNG5 again? bug? In this case, run with a default seed().
+  GridParallelRNG          RNG5rb(FrbGrid);  RNG5rb.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid); 
+  std::vector<LatticeColourMatrix> U(4,UGrid);
+  
+  if ( JP.gaugefile.compare("Hot") == 0 ) {
+    SU3::HotConfiguration(RNG4, Umu);
+  } else {
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu,header,JP.gaugefile);
+    // ypj [fixme] additional checks for the loaded configuration?
+  }
+  
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+  
+  RealD mass = JP.mass;
+  RealD M5 = JP.M5;
+
+// ypj [fixme] flexible support for a various Fermions
+//  RealD mob_b = JP.mob_b;      // Gparity
+//  std::vector<ComplexD> omega; // ZMobius
+  
+//  GparityMobiusFermionD ::ImplParams params;
+//  std::vector<int> twists({1,1,1,0});
+//  params.twists = twists;
+//  GparityMobiusFermionR  Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,params);
+//  SchurDiagTwoOperator<GparityMobiusFermionR,FermionField> HermOp(Ddwf);
+
+
+//  int mrhs = JP.Nu;
+  int Ndir=4;
+  auto mpi_layout  = GridDefaultMpi();
+  std::vector<int> mpi_split (Ndir,1);
+#if 0
+    int tmp=mrhs, dir=0;
+    std::cout << GridLogMessage  << "dir= "<<dir <<"tmp= "<<tmp<<"mpi_split= "<<mpi_split[dir]<<"mpi_layout= "<<mpi_split[dir]<<std::endl;
+    while ( tmp> 1) {
+    if ((mpi_split[dir]*2) <= mpi_layout[dir]){
+        mpi_split[dir] *=2;
+        tmp = tmp/2;
+    }
+    std::cout << GridLogMessage  << "dir= "<<dir <<"tmp= "<<tmp<<"mpi_split= "<<mpi_split[dir]<<"mpi_layout= "<<mpi_layout[dir]<<std::endl;
+        dir = (dir+1)%Ndir;
+    }
+#endif
+    int mrhs=1;
+    for(int i =0;i<Ndir;i++){
+      mpi_split[i] = mpi_layout[i] / JP.mpi_split[i] ;
+      mrhs *= JP.mpi_split[i];
+    }
+    std::cout << GridLogMessage  << "mpi_layout= " << mpi_layout << std::endl;
+    std::cout << GridLogMessage  << "mpi_split= " << mpi_split << std::endl;
+    std::cout << GridLogMessage  << "mrhs= " << mrhs << std::endl;
+//    assert(JP.Nu==tmp);
+
+  /////////////////////////////////////////////
+  // Split into 1^4 mpi communicators
+  /////////////////////////////////////////////
+  GridCartesian         * SGrid = new GridCartesian(GridDefaultLatt(),
+                                                    GridDefaultSimd(Nd,vComplex::Nsimd()),
+                                                    mpi_split,
+                                                    *UGrid);
+
+  GridCartesian         * SFGrid   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,SGrid);
+  GridRedBlackCartesian * SrbGrid  = SpaceTimeGrid::makeFourDimRedBlackGrid(SGrid);
+  GridRedBlackCartesian * SFrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,SGrid);
+
+  LatticeGaugeField s_Umu(SGrid);
+  Grid_split  (Umu,s_Umu);
+
+  //WilsonFermionR::ImplParams params;
+  ZMobiusFermionR::ImplParams params;
+  params.overlapCommsCompute = true;
+  params.boundary_phases = JP.boundary_phase;
+  ZMobiusFermionR  Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,JP.omega,1.,0.,params);
+//  SchurDiagTwoOperator<ZMobiusFermionR,FermionField> HermOp(Ddwf);
+  SchurDiagOneOperator<ZMobiusFermionR,FermionField> HermOp(Ddwf);
+  ZMobiusFermionR  Dsplit(s_Umu,*SFGrid,*SFrbGrid,*SGrid,*SrbGrid,mass,M5,JP.omega,1.,0.,params);
+//  SchurDiagTwoOperator<ZMobiusFermionR,FermionField> SHermOp(Dsplit);
+  SchurDiagOneOperator<ZMobiusFermionR,FermionField> SHermOp(Dsplit);
+
+  //std::vector<double> Coeffs { 0.,-1.}; 
+  // ypj [note] this may not be supported by some compilers
+  std::vector<double> Coeffs({ 0.,-1.}); 
+  Polynomial<FermionField> PolyX(Coeffs);
+  //Chebyshev<FermionField> Cheb(0.2,5.5,11);
+  Chebyshev<FermionField> Cheb(JP.low,JP.high,JP.order);
+//  Cheb.csv(std::cout);
+  ImplicitlyRestartedBlockLanczos<FermionField> IRBL(HermOp, SHermOp,
+						     FrbGrid,SFrbGrid,mrhs,
+                                                     Cheb,
+                                                     JP.Nstop, JP.Ntest,
+                                                     JP.Nu, JP.Nk, JP.Nm,
+                                                     JP.resid,
+                                                     JP.MaxIter,
+						     IRBLdiagonaliseWithEigen);
+//						     IRBLdiagonaliseWithLAPACK);
+  IRBL.split_test=0;
+  
+  std::vector<RealD> eval(JP.Nm);
+  
+  std::vector<FermionField> src(JP.Nu,FrbGrid);
+if (0)
+{
+// in case RNG is too slow
+  std::cout << GridLogMessage << "Using RNG5"<<std::endl;
+  FermionField src_tmp(FGrid);
+  for ( int i=0; i<JP.Nu; ++i ){
+//    gaussian(RNG5,src_tmp);
+     ComplexD rnd;
+     RealD re;
+     fillScalar(re,RNG5._gaussian[0],RNG5._generators[0]);
+    std::cout << i <<" / "<< JP.Nm  <<" re "<< re  << std::endl;
+// printf("%d / %d re %e\n",i,FGrid->_processor,re);
+    src_tmp=re;
+    pickCheckerboard(Odd,src[i],src_tmp);
+  }
+  RNG5.Report();
+} else {
+  std::cout << GridLogMessage << "Using RNG5rb"<<std::endl;
+  for ( int i=0; i<JP.Nu; ++i )
+    gaussian(RNG5rb,src[i]);
+  RNG5rb.Report();
+
+}
+  
+  std::vector<FermionField> evec(JP.Nm,FrbGrid);
+  for(int i=0;i<1;++i){
+    std::cout << GridLogMessage << i <<" / "<< JP.Nm <<" grid pointer "<< evec[i].Grid() << std::endl;
+  };
+
+  int Nconv;
+  IRBL.calc(eval,evec,src,Nconv,JP.Impl);
+
+
+  Grid_finalize();
+}

tests/lanczos/Test_dwf_block_lanczos.cc.single

@@ -0,0 +1,408 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_block_lanczos.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/util/Init.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h>
+
+using namespace std;
+using namespace Grid;
+//using namespace Grid::QCD;
+
+//typedef typename GparityDomainWallFermionR::FermionField FermionField;
+typedef typename ZMobiusFermionF::FermionField FermionField;
+
+RealD AllZero(RealD x){ return 0.;}
+
+class CmdJobParams 
+{
+  public:
+    std::string gaugefile;
+
+    int Ls;
+    double mass;
+    double M5;
+    double mob_b;
+    std::vector<ComplexD> omega;
+    std::vector<Complex> boundary_phase;
+    std::vector<int> mpi_split;
+    
+    LanczosType Impl;
+    int Nu;
+    int Nk;
+    int Np;
+    int Nm;
+    int Nstop;
+    int Ntest;
+    int MaxIter;
+    double resid;
+    
+    double low;
+    double high;
+    int order;
+
+    CmdJobParams()
+      : gaugefile("Hot"),
+        Ls(8), mass(0.01), M5(1.8), mob_b(1.5),
+        Impl(LanczosType::irbl),mpi_split(4,1),
+        Nu(4), Nk(200), Np(200), Nstop(100), Ntest(1), MaxIter(10), resid(1.0e-8), 
+        low(0.2), high(5.5), order(11)
+    {Nm=Nk+Np;};
+    
+    void Parse(char **argv, int argc);
+};
+
+
+void CmdJobParams::Parse(char **argv,int argc)
+{
+  std::string arg;
+  std::vector<int> vi;
+  double re,im;
+  int expect, idx;
+  std::string vstr;
+  std::ifstream pfile;
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--gconf") ){
+    gaugefile = GridCmdOptionPayload(argv,argv+argc,"--gconf");
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--phase") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--phase");
+    pfile.open(arg);
+    assert(pfile);
+    expect = 0;
+    while( pfile >> vstr ) {
+      if ( vstr.compare("boundary_phase") == 0 ) {
+        pfile >> vstr;
+        GridCmdOptionInt(vstr,idx);
+        assert(expect==idx);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,re);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,im);
+        boundary_phase.push_back({re,im});
+        expect++;
+      }
+    }
+    pfile.close();
+  } else {
+    for (int i=0; i<4; ++i) boundary_phase.push_back({1.,0.});
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--omega") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--omega");
+    pfile.open(arg);
+    assert(pfile);
+    Ls = 0;
+    while( pfile >> vstr ) {
+      if ( vstr.compare("omega") == 0 ) {
+        pfile >> vstr;
+        GridCmdOptionInt(vstr,idx);
+        assert(Ls==idx);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,re);
+        pfile >> vstr;
+        GridCmdOptionFloat(vstr,im);
+        omega.push_back({re,im});
+        Ls++;
+      }
+    }
+    pfile.close();
+  } else {
+    if( GridCmdOptionExists(argv,argv+argc,"--Ls") ){
+      arg = GridCmdOptionPayload(argv,argv+argc,"--Ls");
+      GridCmdOptionInt(arg,Ls);
+    }
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--mass") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mass");
+    GridCmdOptionFloat(arg,mass);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--M5") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--M5");
+    GridCmdOptionFloat(arg,M5);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--mob_b") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mob_b");
+    GridCmdOptionFloat(arg,mob_b);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--irbl") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--irbl");
+    GridCmdOptionIntVector(arg,vi);
+    Nu = vi[0];
+    Nk = vi[1];
+    Np = vi[2];
+    Nstop = vi[3];
+    MaxIter = vi[4];
+    // ypj[fixme] mode overriding message is needed.
+    Impl = LanczosType::irbl;
+    Nm = Nk+Np;
+  }
+  
+  // block Lanczos with explicit extension of its dimensions
+  if( GridCmdOptionExists(argv,argv+argc,"--rbl") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--rbl");
+    GridCmdOptionIntVector(arg,vi);
+    Nu = vi[0];
+    Nk = vi[1];
+    Np = vi[2]; // vector space is enlarged by adding Np vectors
+    Nstop = vi[3];
+    MaxIter = vi[4];
+    // ypj[fixme] mode overriding message is needed.
+    Impl = LanczosType::rbl;
+    Nm = Nk+Np*MaxIter;
+  }
+  
+#if 1
+  // block Lanczos with explicit extension of its dimensions
+  if( GridCmdOptionExists(argv,argv+argc,"--split") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--split");
+    GridCmdOptionIntVector(arg,vi);
+    for(int i=0;i<mpi_split.size();i++)
+    mpi_split[i] = vi[i];
+  }
+#endif
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--check_int") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--check_int");
+    GridCmdOptionInt(arg,Ntest);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--resid") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--resid");
+    GridCmdOptionFloat(arg,resid);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_l") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_l");
+    GridCmdOptionFloat(arg,low);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_u") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_u");
+    GridCmdOptionFloat(arg,high);
+  }
+  
+  if( GridCmdOptionExists(argv,argv+argc,"--cheby_n") ){
+    arg = GridCmdOptionPayload(argv,argv+argc,"--cheby_n");
+    GridCmdOptionInt(arg,order);
+  }
+  
+  if ( CartesianCommunicator::RankWorld() == 0 ) {
+    std::streamsize ss = std::cout.precision();
+    std::cout << GridLogMessage <<" Gauge Configuration "<< gaugefile << '\n';
+    std::cout.precision(15);
+    for ( int i=0; i<4; ++i ) std::cout << GridLogMessage <<" boundary_phase["<< i << "] = " << boundary_phase[i] << '\n';
+    std::cout.precision(ss);
+    std::cout << GridLogMessage <<" Ls "<< Ls << '\n';
+    std::cout << GridLogMessage <<" mass "<< mass << '\n';
+    std::cout << GridLogMessage <<" M5 "<< M5 << '\n';
+    std::cout << GridLogMessage <<" mob_b "<< mob_b << '\n';
+    std::cout.precision(15);
+    for ( int i=0; i<Ls; ++i ) std::cout << GridLogMessage <<" omega["<< i << "] = " << omega[i] << '\n';
+    std::cout.precision(ss);
+    std::cout << GridLogMessage <<" Nu "<< Nu << '\n'; 
+    std::cout << GridLogMessage <<" Nk "<< Nk << '\n'; 
+    std::cout << GridLogMessage <<" Np "<< Np << '\n'; 
+    std::cout << GridLogMessage <<" Nm "<< Nm << '\n'; 
+    std::cout << GridLogMessage <<" Nstop "<< Nstop << '\n'; 
+    std::cout << GridLogMessage <<" Ntest "<< Ntest << '\n'; 
+    std::cout << GridLogMessage <<" MaxIter "<< MaxIter << '\n'; 
+    std::cout << GridLogMessage <<" resid "<< resid << '\n'; 
+    std::cout << GridLogMessage <<" Cheby Poly "<< low << "," << high << "," << order << std::endl; 
+  }
+}
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+  
+  CmdJobParams JP;
+  JP.Parse(argv,argc);
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,UGrid);
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n",UGrid,UrbGrid,FGrid,FrbGrid);
+  GridCartesian         * UGridF   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian         * FGridF   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,UGridF);
+  GridRedBlackCartesian * FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,UGridF);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGridF);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  // ypj [note] why seed RNG5 again? bug? In this case, run with a default seed().
+  GridParallelRNG          RNG5rb(FrbGridF);  RNG5rb.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid); 
+  std::vector<LatticeColourMatrix> U(4,UGrid);
+  LatticeGaugeFieldF UmuF(UGridF); 
+  std::vector<LatticeColourMatrix> UF(4,UGridF);
+  
+  if ( JP.gaugefile.compare("Hot") == 0 ) {
+    SU3::HotConfiguration(RNG4, Umu);
+  } else {
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu,header,JP.gaugefile);
+    // ypj [fixme] additional checks for the loaded configuration?
+  }
+  precisionChange (UmuF,Umu);
+  
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+  
+  RealD mass = JP.mass;
+  RealD M5 = JP.M5;
+
+// ypj [fixme] flexible support for a various Fermions
+//  RealD mob_b = JP.mob_b;      // Gparity
+//  std::vector<ComplexD> omega; // ZMobius
+  
+//  GparityMobiusFermionD ::ImplParams params;
+//  std::vector<int> twists({1,1,1,0});
+//  params.twists = twists;
+//  GparityMobiusFermionR  Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,params);
+//  SchurDiagTwoOperator<GparityMobiusFermionR,FermionField> HermOp(Ddwf);
+
+
+//  int mrhs = JP.Nu;
+  int Ndir=4;
+  auto mpi_layout  = GridDefaultMpi();
+  std::vector<int> mpi_split (Ndir,1);
+#if 0
+    int tmp=mrhs, dir=0;
+    std::cout << GridLogMessage  << "dir= "<<dir <<"tmp= "<<tmp<<"mpi_split= "<<mpi_split[dir]<<"mpi_layout= "<<mpi_split[dir]<<std::endl;
+    while ( tmp> 1) {
+    if ((mpi_split[dir]*2) <= mpi_layout[dir]){
+        mpi_split[dir] *=2;
+        tmp = tmp/2;
+    }
+    std::cout << GridLogMessage  << "dir= "<<dir <<"tmp= "<<tmp<<"mpi_split= "<<mpi_split[dir]<<"mpi_layout= "<<mpi_layout[dir]<<std::endl;
+        dir = (dir+1)%Ndir;
+    }
+#endif
+    int mrhs=1;
+    for(int i =0;i<Ndir;i++){
+      mpi_split[i] = mpi_layout[i] / JP.mpi_split[i] ;
+      mrhs *= JP.mpi_split[i];
+    }
+    std::cout << GridLogMessage  << "mpi_layout= " << mpi_layout << std::endl;
+    std::cout << GridLogMessage  << "mpi_split= " << mpi_split << std::endl;
+    std::cout << GridLogMessage  << "mrhs= " << mrhs << std::endl;
+//    assert(JP.Nu==tmp);
+
+  /////////////////////////////////////////////
+  // Split into 1^4 mpi communicators, keeping it explicitly single
+  /////////////////////////////////////////////
+  GridCartesian         * SGrid = new GridCartesian(GridDefaultLatt(),
+                                                    GridDefaultSimd(Nd,vComplexF::Nsimd()),
+                                                    mpi_split,
+                                                    *UGrid);
+
+  GridCartesian         * SFGrid   = SpaceTimeGrid::makeFiveDimGrid(JP.Ls,SGrid);
+  GridRedBlackCartesian * SrbGrid  = SpaceTimeGrid::makeFourDimRedBlackGrid(SGrid);
+  GridRedBlackCartesian * SFrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(JP.Ls,SGrid);
+
+  LatticeGaugeFieldF s_Umu(SGrid);
+  Grid_split  (UmuF,s_Umu);
+
+  //WilsonFermionR::ImplParams params;
+  ZMobiusFermionF::ImplParams params;
+  params.overlapCommsCompute = true;
+  params.boundary_phases = JP.boundary_phase;
+  ZMobiusFermionF  DdwfF(UmuF,*FGridF,*FrbGridF,*UGridF,*UrbGridF,mass,M5,JP.omega,1.,0.,params);
+//  SchurDiagTwoOperator<ZMobiusFermionF,FermionField> HermOp(Ddwf);
+  SchurDiagOneOperator<ZMobiusFermionF,FermionField> HermOp(DdwfF);
+  ZMobiusFermionF  Dsplit(s_Umu,*SFGrid,*SFrbGrid,*SGrid,*SrbGrid,mass,M5,JP.omega,1.,0.,params);
+//  SchurDiagTwoOperator<ZMobiusFermionF,FermionField> SHermOp(Dsplit);
+  SchurDiagOneOperator<ZMobiusFermionF,FermionField> SHermOp(Dsplit);
+
+  //std::vector<double> Coeffs { 0.,-1.}; 
+  // ypj [note] this may not be supported by some compilers
+  std::vector<double> Coeffs({ 0.,-1.}); 
+  Polynomial<FermionField> PolyX(Coeffs);
+  //Chebyshev<FermionField> Cheb(0.2,5.5,11);
+  Chebyshev<FermionField> Cheb(JP.low,JP.high,JP.order);
+//  Cheb.csv(std::cout);
+  ImplicitlyRestartedBlockLanczos<FermionField> IRBL(HermOp, SHermOp,
+						     FrbGridF,SFrbGrid,mrhs,
+                                                     Cheb,
+                                                     JP.Nstop, JP.Ntest,
+                                                     JP.Nu, JP.Nk, JP.Nm,
+                                                     JP.resid,
+                                                     JP.MaxIter,
+						     IRBLdiagonaliseWithEigen);
+//						     IRBLdiagonaliseWithLAPACK);
+  IRBL.split_test=1;
+  
+  std::vector<RealD> eval(JP.Nm);
+  
+  std::vector<FermionField> src(JP.Nu,FrbGridF);
+if (0)
+{
+// in case RNG is too slow
+  std::cout << GridLogMessage << "Using RNG5"<<std::endl;
+  FermionField src_tmp(FGrid);
+  for ( int i=0; i<JP.Nu; ++i ){
+//    gaussian(RNG5,src_tmp);
+     ComplexD rnd;
+     RealD re;
+     fillScalar(re,RNG5._gaussian[0],RNG5._generators[0]);
+    std::cout << i <<" / "<< JP.Nm  <<" re "<< re  << std::endl;
+// printf("%d / %d re %e\n",i,FGrid->_processor,re);
+    src_tmp=re;
+    pickCheckerboard(Odd,src[i],src_tmp);
+  }
+  RNG5.Report();
+} else {
+  std::cout << GridLogMessage << "Using RNG5rb"<<std::endl;
+  for ( int i=0; i<JP.Nu; ++i )
+    gaussian(RNG5rb,src[i]);
+  RNG5rb.Report();
+
+}
+  
+  std::vector<FermionField> evec(JP.Nm,FrbGridF);
+  for(int i=0;i<1;++i){
+    std::cout << GridLogMessage << i <<" / "<< JP.Nm <<" grid pointer "<< evec[i].Grid() << std::endl;
+  };
+
+  int Nconv;
+  IRBL.calc(eval,evec,src,Nconv,JP.Impl);
+
+
+  Grid_finalize();
+}

tests/lanczos/Test_dwf_lanczos.cc

@@ -35,26 +35,45 @@ template<typename Action>
 struct Setup{};
 
 template<>
-struct Setup<GparityMobiusFermionD>{
-  static GparityMobiusFermionD* getAction(LatticeGaugeField &Umu,
+struct Setup<GparityMobiusFermionF>{
+  static GparityMobiusFermionF* getAction(LatticeGaugeFieldF &Umu,
 					  GridCartesian* FGrid, GridRedBlackCartesian* FrbGrid, GridCartesian* UGrid, GridRedBlackCartesian* UrbGrid){
-    RealD mass=0.01;
+    RealD mass=0.00054;
     RealD M5=1.8;
     RealD mob_b=1.5;
     GparityMobiusFermionD ::ImplParams params;
     std::vector<int> twists({1,1,1,0});
     params.twists = twists;
-    return new GparityMobiusFermionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,params);
+    return new GparityMobiusFermionF(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,params);
   }
 };
 
 template<>
+struct Setup<DomainWallFermionF>{
+  static DomainWallFermionF* getAction(LatticeGaugeFieldF &Umu,
 struct Setup<DomainWallFermionD>{
   static DomainWallFermionD* getAction(LatticeGaugeField &Umu,
 					  GridCartesian* FGrid, GridRedBlackCartesian* FrbGrid, GridCartesian* UGrid, GridRedBlackCartesian* UrbGrid){
-    RealD mass=0.01;
+    RealD mass=0.00054;
     RealD M5=1.8;
-    return new DomainWallFermionD(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    return new DomainWallFermionF(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  }
+};
+
+template<>
+struct Setup<MobiusFermionF>{
+  static MobiusFermionF* getAction(LatticeGaugeFieldF &Umu,
+					  GridCartesian* FGrid, GridRedBlackCartesian* FrbGrid, GridCartesian* UGrid, GridRedBlackCartesian* UrbGrid){
+    RealD mass=0.00054;
+    RealD M5=1.8;
+    RealD mob_b=1.5;
+    std::vector<Complex> boundary = {1,1,1,-1};
+    MobiusFermionF::ImplParams Params(boundary);
+
+  std::cout << GridLogMessage << "mass "<<mass<<std::endl;
+  std::cout << GridLogMessage << "M5 "<<M5<<std::endl;
+  std::cout << GridLogMessage << "mob_b "<<mob_b<<std::endl;
+    return new MobiusFermionF(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,mob_b,mob_b-1.,Params);
   }
 };
 
@@ -63,48 +82,70 @@ struct Setup<DomainWallFermionD>{
 template<typename Action>
 void run(){
   typedef typename Action::FermionField FermionField;
-  const int Ls=8;
+  const int Ls=12;
 
   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);
-  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n",UGrid,UrbGrid,FGrid,FrbGrid);
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n",UGrid,UrbGrid,FGrid,FrbGrid);
+  
+  GridCartesian* UGridF = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian* UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian* FGridF = SpaceTimeGrid::makeFiveDimGrid(Ls, UGridF);
+  GridRedBlackCartesian* FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, UGridF);
+
 
   std::vector<int> seeds4({1,2,3,4});
   std::vector<int> seeds5({5,6,7,8});
-  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
-  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
-  GridParallelRNG          RNG5rb(FrbGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG5(FGridF);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGridF);  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          RNG5rb(FrbGridF);  RNG5.SeedFixedIntegers(seeds5);
 
   LatticeGaugeField Umu(UGrid); 
-  SU<Nc>::HotConfiguration(RNG4, Umu);
+//  SU<Nc>::HotConfiguration(RNG4, Umu);
+  FieldMetaData header;
+  std::string file("./config");
 
-  Action *action = Setup<Action>::getAction(Umu,FGrid,FrbGrid,UGrid,UrbGrid);
+//  int precision32 = 0;
+//  int tworow      = 0;
+//  NerscIO::writeConfiguration(Umu,file,tworow,precision32);
+  NerscIO::readConfiguration(Umu,header,file);
+
+  LatticeGaugeFieldF UmuF(UGridF); 
+  precisionChange(UmuF, Umu);
+
+  Action *action = Setup<Action>::getAction(UmuF,FGridF,FrbGridF,UGridF,UrbGridF);
  
   //MdagMLinearOperator<Action,FermionField> HermOp(Ddwf);
-  SchurDiagTwoOperator<Action,FermionField> HermOp(*action);
+//  SchurDiagTwoOperator<Action,FermionField> HermOp(*action);
+  SchurDiagOneOperator<Action,FermionField> HermOp(*action);
 
-  const int Nstop = 30;
-  const int Nk = 40;
+  const int Nstop = 150;
+  const int Nk = 160;
   const int Np = 40;
   const int Nm = Nk+Np;
   const int MaxIt= 10000;
-  RealD resid = 1.0e-8;
+  RealD resid = 1.0e-6;
+  std::cout << GridLogMessage << "Nstop "<<Nstop<<std::endl;
+  std::cout << GridLogMessage << "Nk "<<Nk<<std::endl;
+  std::cout << GridLogMessage << "Np "<<Np<<std::endl;
+  std::cout << GridLogMessage << "resid "<<resid<<std::endl;
 
   std::vector<double> Coeffs { 0.,-1.};
   Polynomial<FermionField> PolyX(Coeffs);
-  Chebyshev<FermionField> Cheby(0.2,5.,11);
+  Chebyshev<FermionField> Cheby(0.0000006,5.5,4001);
+  std::cout << GridLogMessage << "Cheby(0.0000006,5.5,4001) "<<std::endl;
 
   FunctionHermOp<FermionField> OpCheby(Cheby,HermOp);
-     PlainHermOp<FermionField> Op     (HermOp);
+  PlainHermOp<FermionField> Op     (HermOp);
 
   ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby,Op,Nstop,Nk,Nm,resid,MaxIt);
  
   std::vector<RealD>          eval(Nm);
-  FermionField    src(FrbGrid); 
+  FermionField    src(FrbGridF); 
   gaussian(RNG5rb,src);
-  std::vector<FermionField> evec(Nm,FrbGrid);
+  std::vector<FermionField> evec(Nm,FrbGridF);
   for(int i=0;i<1;i++){
     std::cout << GridLogMessage <<i<<" / "<< Nm<< " grid pointer "<<evec[i].Grid()<<std::endl;
   };
@@ -119,7 +160,7 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::string action = "GparityMobius";
+  std::string action = "Mobius";
   for(int i=1;i<argc;i++){
     if(std::string(argv[i]) == "-action"){
       action = argv[i+1];
@@ -127,9 +168,11 @@ int main (int argc, char ** argv)
   }
 
   if(action == "GparityMobius"){
-    run<GparityMobiusFermionD>();
+    run<GparityMobiusFermionF>();
   }else if(action == "DWF"){
-    run<DomainWallFermionD>();
+    run<DomainWallFermionF>();
+  }else if(action == "Mobius"){
+    run<MobiusFermionF>();
   }else{
     std::cout << "Unknown action" << std::endl;
     exit(1);

tests/solver/Test_dwf_mixedcg_prec.cc

@@ -0,0 +1,122 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_cg_prec.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+//using namespace std;
+using namespace Grid;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=12;
+
+  std::cout << GridLogMessage << "::::: NB: to enable a quick bit reproducibility check use the --checksums flag. " << std::endl;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  GridCartesian         * UGrid_f   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_f);
+  GridCartesian         * FGrid_f   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_f);
+  GridRedBlackCartesian * FrbGrid_f = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_f);
+  
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermionD    src(FGrid); random(RNG5,src);
+  LatticeFermionD result(FGrid); result=Zero();
+  LatticeGaugeFieldD Umu(UGrid);
+  LatticeGaugeFieldF Umu_f(UGrid_f); 
+  
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+
+  precisionChange(Umu_f,Umu);
+  
+  RealD mass=0.1;
+  RealD M5=1.8;
+  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionF Ddwf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+
+  LatticeFermionD    src_o(FrbGrid);
+  LatticeFermionD result_o(FrbGrid);
+  LatticeFermionD result_o_2(FrbGrid);
+  pickCheckerboard(Odd,src_o,src);
+  result_o.Checkerboard() = Odd;
+  result_o = Zero();
+  result_o_2.Checkerboard() = Odd;
+  result_o_2 = Zero();
+
+  SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+  SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
+
+  std::cout << GridLogMessage << "::::::::::::: Starting mixed CG" << std::endl;
+  MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(1.0e-8, 10000, 50, FrbGrid_f, HermOpEO_f, HermOpEO);
+  double t1,t2,flops;
+  double MdagMsiteflops = 1452; // Mobius (real coeffs)
+  // CG overhead: 8 inner product, 4+8 axpy_norm, 4+4 linear comb (2 of)
+  double CGsiteflops = (8+4+8+4+4)*Nc*Ns ;
+  std:: cout << " MdagM site flops = "<< 4*MdagMsiteflops<<std::endl;
+  std:: cout << " CG    site flops = "<< CGsiteflops <<std::endl;
+
+  result_o = Zero();
+  t1=usecond();
+  mCG(src_o,result_o);
+  t2=usecond();
+  int iters = mCG.TotalInnerIterations; //Number of inner CG iterations
+  flops = MdagMsiteflops*4*FrbGrid->gSites()*iters;
+  flops+= CGsiteflops*FrbGrid->gSites()*iters;
+  std::cout << " SinglePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
+  std::cout << " SinglePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
+
+  std::cout << GridLogMessage << "::::::::::::: Starting regular CG" << std::endl;
+  ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+  result_o_2 = Zero();
+  t1=usecond();
+  CG(HermOpEO,src_o,result_o_2);
+  t2=usecond();
+  iters = CG.IterationsToComplete;
+  flops = MdagMsiteflops*4*FrbGrid->gSites()*iters; 
+  flops+= CGsiteflops*FrbGrid->gSites()*iters;
+  
+  std::cout << " DoublePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
+  std::cout << " DoublePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
+
+  LatticeFermionD diff_o(FrbGrid);
+  RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
+
+  std::cout << GridLogMessage << "::::::::::::: Diff between mixed and regular CG: " << diff << std::endl;
+
+  MemoryManager::Print();
+
+  Grid_finalize();
+}

tests/solver/Test_dwf_relupcg_prec.cc

@@ -0,0 +1,143 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/solver/Test_dwf_relupcg_prec.cc
+
+    Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@bnl.gov>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  double relup_delta = 0.2;
+  for(int i=1;i<argc-1;i++){
+    std::string sarg = argv[i];
+    if(sarg == "--relup_delta"){
+      std::stringstream ss; ss << argv[i+1]; ss >> relup_delta;
+      std::cout << GridLogMessage << "Set reliable update Delta to " << relup_delta << std::endl;
+    }
+  }   
+  
+  const int Ls=12;
+
+  { 
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  GridCartesian         * UGrid_f   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_f);
+  GridCartesian         * FGrid_f   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_f);
+  GridRedBlackCartesian * FrbGrid_f = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_f);
+  
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermionD    src(FGrid); random(RNG5,src);
+  LatticeFermionD result(FGrid); result=Zero();
+  LatticeGaugeFieldD Umu(UGrid);
+  LatticeGaugeFieldF Umu_f(UGrid_f); 
+  
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+
+  precisionChange(Umu_f,Umu);
+  
+  RealD mass=0.1;
+  RealD M5=1.8;
+  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionF Ddwf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+
+  LatticeFermionD    src_o(FrbGrid);
+  LatticeFermionD result_o(FrbGrid);
+  LatticeFermionD result_o_2(FrbGrid);
+  pickCheckerboard(Odd,src_o,src);
+  result_o.Checkerboard() = Odd;
+  result_o = Zero();
+  result_o_2.Checkerboard() = Odd;
+  result_o_2 = Zero();
+
+  SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+  SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
+
+  std::cout << GridLogMessage << "::::::::::::: Starting mixed CG" << std::endl;
+  ConjugateGradientReliableUpdate<LatticeFermionD,LatticeFermionF> mCG(1e-8, 10000, relup_delta, FrbGrid_f, HermOpEO_f, HermOpEO);
+  double t1,t2,flops;
+  double MdagMsiteflops = 1452; // Mobius (real coeffs)
+  // CG overhead: 8 inner product, 4+8 axpy_norm, 4+4 linear comb (2 of)
+  double CGsiteflops = (8+4+8+4+4)*Nc*Ns ;
+  std:: cout << " MdagM site flops = "<< 4*MdagMsiteflops<<std::endl;
+  std:: cout << " CG    site flops = "<< CGsiteflops <<std::endl;
+  int iters, iters_cleanup, relups, tot_iters;
+  for(int i=0;i<10;i++){
+    result_o = Zero();
+    t1=usecond();
+    mCG(src_o,result_o);
+    t2=usecond();
+    iters = mCG.IterationsToComplete; //Number of single prec CG iterations
+    iters_cleanup = mCG.IterationsToCleanup;
+    relups = mCG.ReliableUpdatesPerformed;
+    tot_iters  = iters + iters_cleanup + relups; //relup cost MdagM application in double
+    
+    flops = MdagMsiteflops*4*FrbGrid->gSites()*tot_iters;
+    flops+= CGsiteflops*FrbGrid->gSites()*tot_iters;
+    std::cout << " SinglePrecision single prec iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
+    std::cout << " SinglePrecision double prec cleanup iterations/sec "<< iters_cleanup/(t2-t1)*1000.*1000.<<std::endl;
+    std::cout << " SinglePrecision reliable updates/sec "<< relups/(t2-t1)*1000.*1000.<<std::endl;
+    std::cout << " SinglePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
+  }
+  std::cout << GridLogMessage << "::::::::::::: Starting regular CG" << std::endl;
+  ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+  for(int i=0;i<1;i++){
+    result_o_2 = Zero();
+    t1=usecond();
+    CG(HermOpEO,src_o,result_o_2);
+    t2=usecond();
+    iters = CG.IterationsToComplete;
+    flops = MdagMsiteflops*4*FrbGrid->gSites()*iters; 
+    flops+= CGsiteflops*FrbGrid->gSites()*iters;
+    
+    std::cout << " DoublePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
+    std::cout << " DoublePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
+  }
+  
+  //  MemoryManager::Print();
+
+  LatticeFermionD diff_o(FrbGrid);
+  RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
+
+  std::cout << GridLogMessage << "::::::::::::: Diff between mixed and regular CG: " << diff << std::endl;
+  }
+  
+  MemoryManager::Print();
+
+  Grid_finalize();
+}