Profiling temporary code until optimised

Unused statements generating warnings removed

.github/ISSUE_TEMPLATE/bug-report.yml

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

Grid/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/LinearOperator.h

@@ -542,6 +542,7 @@ public:
       (*this)(in[i], out[i]);
     }
   }
+  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {

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;
+      precisionChange(ps_f[s],src_d);
+    }
+    // r and p for primary
+    p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
+    r_d = p_d;
+    
+    //MdagM+m[0]
+    precisionChange(p_f,p_d);
+    Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    precisionChange(tmp_d,mmp_f);
+    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    tmp_d = tmp_d - mmp_d;
+    std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
+    //    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);
+      precisionChange(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();
+      precisionChange(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();
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
+      PrecChangeTimer.Stop();
+      MatrixTimer.Start();  
+      Linop_f.HermOp(p_f,mmp_f);
+      MatrixTimer.Stop();  
+      PrecChangeTimer.Start();
+      precisionChange(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++){
+	  precisionChange(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"
     ////////////////////////////////////////////////////////////////////////
@@ -200,14 +204,14 @@ 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;
-    //    assert(norm2(tmp_d)< 1.0e-4);
+    assert(norm2(tmp_d)< 1.0);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);
@@ -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 +514,24 @@ 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;
+  
+  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 +541,14 @@ 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 +559,15 @@ 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 +585,8 @@ 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

@@ -12,8 +12,10 @@ 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(void){};
+void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
@@ -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/Communicator_none.cc

@@ -128,7 +128,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int recv_from_rank,int dor,
 							 int xbytes,int rbytes, int dir)
 {
-  return 2.0*bytes;
+  return xbytes+rbytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {

Grid/communicator/SharedMemory.cc

@@ -91,6 +91,59 @@ void *SharedMemory::ShmBufferSelf(void)
   //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
   return ShmCommBufs[ShmRank];
 }
+static inline int divides(int a,int b)
+{
+  return ( b == ( (b/a)*a ) );
+}
+void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
+{
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
+  ////////////////////////////////////////////////////////////////
+  // Powers of 2,3,5 only in prime decomposition for now
+  ////////////////////////////////////////////////////////////////
+  int ndimension = WorldDims.size();
+  ShmDims=Coordinate(ndimension,1);
+
+  std::vector<int> primes({2,3,5});
+
+  int dim = 0;
+  int last_dim = ndimension - 1;
+  int AutoShmSize = 1;
+  while(AutoShmSize != WorldShmSize) {
+    int p;
+    for(p=0;p<primes.size();p++) {
+      int prime=primes[p];
+      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
+        && divides(prime,WorldShmSize/AutoShmSize)  ) {
+  AutoShmSize*=prime;
+  ShmDims[dim]*=prime;
+  last_dim = dim;
+  break;
+      }
+    }
+    if (p == primes.size() && last_dim == dim) {
+      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
+      exit(EXIT_FAILURE);
+    }
+    dim=(dim+1) %ndimension;
+  }
+}
 
 NAMESPACE_END(Grid); 
 

Grid/communicator/SharedMemoryMPI.cc

@@ -27,6 +27,8 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
+#define Mheader "SharedMemoryMpi: "
+
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
@@ -36,12 +38,120 @@ 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
+#include <syscall.h>
+#define SHM_SOCKETS 
 #endif
 
+#include <sys/socket.h>
+#include <sys/un.h>
+
 NAMESPACE_BEGIN(Grid); 
-#define header "SharedMemoryMpi: "
+
+#ifdef SHM_SOCKETS
+
+/*
+ * Barbaric extra intranode communication route in case we need sockets to pass FDs
+ * Forced by level_zero not being nicely designed
+ */
+static int sock;
+static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
+static char sock_path[256];
+class UnixSockets {
+public:
+  static void Open(int rank)
+  {
+    int errnum;
+
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
+    unlink(sa_un.sun_path);
+    if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
+      perror("bind failure");
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  static int RecvFileDescriptor(void)
+  {
+    int n;
+    int fd;
+    char buf[1];
+    struct iovec iov;
+    struct msghdr msg;
+    struct cmsghdr *cmsg;
+    char cms[CMSG_SPACE(sizeof(int))];
+
+    iov.iov_base = buf;
+    iov.iov_len = 1;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_name = 0;
+    msg.msg_namelen = 0;
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+
+    msg.msg_control = (caddr_t)cms;
+    msg.msg_controllen = sizeof cms;
+
+    if((n=recvmsg(sock, &msg, 0)) < 0) {
+      perror("recvmsg failed");
+      return -1;
+    }
+    if(n == 0){
+      perror("recvmsg returned 0");
+      return -1;
+    }
+    cmsg = CMSG_FIRSTHDR(&msg);
+
+    memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
+
+    return fd;
+  }
+
+  static void SendFileDescriptor(int fildes,int xmit_to_rank)
+  {
+    struct msghdr msg;
+    struct iovec iov;
+    struct cmsghdr *cmsg = NULL;
+    char ctrl[CMSG_SPACE(sizeof(int))];
+    char data = ' ';
+
+    memset(&msg, 0, sizeof(struct msghdr));
+    memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
+    iov.iov_base = &data;
+    iov.iov_len = sizeof(data);
+    
+    sprintf(sock_path,sock_path_fmt,xmit_to_rank);
+    
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
+
+    msg.msg_name = (void *)&sa_un;
+    msg.msg_namelen = sizeof(sa_un);
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+    msg.msg_controllen =  CMSG_SPACE(sizeof(int));
+    msg.msg_control = ctrl;
+
+    cmsg = CMSG_FIRSTHDR(&msg);
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int));
+
+    *((int *) CMSG_DATA(cmsg)) = fildes;
+
+    sendmsg(sock, &msg, 0);
+  };
+};
+#endif
+
+
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -64,8 +174,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
   if ( WorldRank == 0) {
-    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
-    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+    std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl;
   }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
@@ -168,59 +278,7 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
   else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
-static inline int divides(int a,int b)
-{
-  return ( b == ( (b/a)*a ) );
-}
-void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
-{
-  ////////////////////////////////////////////////////////////////
-  // Allow user to configure through environment variable
-  ////////////////////////////////////////////////////////////////
-  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
-  if ( str ) {
-    std::vector<int> IntShmDims;
-    GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
-    long ShmSize = 1;
-    for (int dim=0;dim<WorldDims.size();dim++) {
-      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
-    }
-    assert(ShmSize == WorldShmSize);
-    return;
-  }
-  
-  ////////////////////////////////////////////////////////////////
-  // Powers of 2,3,5 only in prime decomposition for now
-  ////////////////////////////////////////////////////////////////
-  int ndimension = WorldDims.size();
-  ShmDims=Coordinate(ndimension,1);
 
-  std::vector<int> primes({2,3,5});
-
-  int dim = 0;
-  int last_dim = ndimension - 1;
-  int AutoShmSize = 1;
-  while(AutoShmSize != WorldShmSize) {
-    int p;
-    for(p=0;p<primes.size();p++) {
-      int prime=primes[p];
-      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
-        && divides(prime,WorldShmSize/AutoShmSize)  ) {
-	AutoShmSize*=prime;
-	ShmDims[dim]*=prime;
-	last_dim = dim;
-	break;
-      }
-    }
-    if (p == primes.size() && last_dim == dim) {
-      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
-      exit(EXIT_FAILURE);
-    }
-    dim=(dim+1) %ndimension;
-  }
-}
 void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
@@ -394,7 +452,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -479,7 +537,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
 
-  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+  std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
 
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -522,7 +580,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -530,8 +588,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef SHM_SOCKETS
+  UnixSockets::Open(WorldShmRank);
+#endif
   for(int r=0;r<WorldShmSize;r++){
 
+    MPI_Barrier(WorldShmComm);
+
 #ifndef GRID_MPI3_SHM_NONE
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
@@ -539,24 +602,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void * thisBuf = ShmCommBuf;
     if(!Stencil_force_mpi) {
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+    typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
     auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
     auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
-
+    
     if ( r==WorldShmRank ) { 
       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
+      memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
+#ifdef SHM_SOCKETS
+      for(int rr=0;rr<WorldShmSize;rr++){
+	if(rr!=r){
+	  UnixSockets::SendFileDescriptor(handle.fd,rr);
+	}
+      }
+#endif
     }
 #endif
 #ifdef GRID_CUDA
@@ -584,6 +655,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
     { 
+      MPI_Barrier(WorldShmComm);
       int ierr=MPI_Bcast(&handle,
 			 sizeof(handle),
 			 MPI_BYTE,
@@ -599,6 +671,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     if ( r!=WorldShmRank ) {
       thisBuf = nullptr;
+      int myfd;
+#ifdef SHM_SOCKETS
+      myfd=UnixSockets::RecvFileDescriptor();
+#else
       std::cout<<"mapping seeking remote pid/fd "
 	       <<handle.pid<<"/"
 	       <<handle.fd<<std::endl;
@@ -606,16 +682,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
       std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
-      int myfd  = syscall(438,pidfd,handle.fd,0);
-
-      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-      
+      myfd  = syscall(438,pidfd,handle.fd,0);
+      int err_t = errno;
+      if (myfd < 0) {
+        fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr);
+	perror("pidfd_getfd failed ");
+	assert(0);
+      }
+#endif
+      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
       auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
@@ -650,6 +732,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
+    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
@@ -661,7 +744,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -698,7 +781,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -708,7 +791,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -755,7 +838,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);

Grid/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>
@@ -658,9 +707,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
 
-  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, AcceleratorRead);
-  accelerator_for(idx,Fg->lSites(),1,{
+  autoView( t_v , To, CpuWrite);
+  autoView( f_v , From, CpuRead);
+  thread_for(idx,Fg->lSites(),{
     sobj s;
     Coordinate Fcoor(nd);
     Coordinate Tcoor(nd);
@@ -673,15 +722,20 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
     }
     if (in_region) {
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
-      vector_type * fp = (vector_type *)&f_v[odx_f];
-      vector_type * tp = (vector_type *)&t_v[odx_t];
+#if 0      
+      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
+      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
+      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
+      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
+      scalar_type * fp = (scalar_type *)&f_v[odx_f];
+      scalar_type * tp = (scalar_type *)&t_v[odx_t];
       for(int w=0;w<words;w++){
 	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
       }
+#else
+    peekLocalSite(s,f_v,Fcoor);
+    pokeLocalSite(s,t_v,Tcoor);
+#endif
     }
   });
 }
@@ -792,9 +846,9 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-  }
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
   // the above should guarantee that the operations are local
@@ -1080,6 +1134,7 @@ 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 +1152,9 @@ 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 +1200,128 @@ 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/lattice/PaddedCell.h

@@ -0,0 +1,136 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/PaddedCell.h
+
+    Copyright (C) 2019
+
+Author: Peter Boyle pboyle@bnl.gov
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+class PaddedCell {
+public:
+  GridCartesian * unpadded_grid;
+  int dims;
+  int depth;
+  std::vector<GridCartesian *> grids;
+  ~PaddedCell()
+  {
+    DeleteGrids();
+  }
+  PaddedCell(int _depth,GridCartesian *_grid)
+  {
+    unpadded_grid = _grid;
+    depth=_depth;
+    dims=_grid->Nd();
+    AllocateGrids();
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    for(int d=0;d<dims;d++){
+      assert(local[d]>=depth);
+    }
+  }
+  void DeleteGrids(void)
+  {
+    for(int d=0;d<grids.size();d++){
+      delete grids[d];
+    }
+    grids.resize(0);
+  };
+  void AllocateGrids(void)
+  {
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate simd      =unpadded_grid->_simd_layout;
+    Coordinate processors=unpadded_grid->_processors;
+    Coordinate plocal    =unpadded_grid->LocalDimensions();
+    Coordinate global(dims);
+
+    // expand up one dim at a time
+    for(int d=0;d<dims;d++){
+
+      plocal[d] += 2*depth; 
+
+      for(int d=0;d<dims;d++){
+	global[d] = plocal[d]*processors[d];
+      }
+
+      grids.push_back(new GridCartesian(global,simd,processors));
+    }
+  };
+  template<class vobj>
+  inline Lattice<vobj> Extract(Lattice<vobj> &in)
+  {
+    Lattice<vobj> out(unpadded_grid);
+
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate fll(dims,depth); // depends on the MPI spread
+    Coordinate tll(dims,0); // depends on the MPI spread
+    localCopyRegion(in,out,fll,tll,local);
+    return out;
+  }
+  template<class vobj>
+  inline Lattice<vobj> Exchange(Lattice<vobj> &in)
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = Expand(d,tmp); // rvalue && assignment
+    }
+    return tmp;
+  }
+  // expand up one dim at a time
+  template<class vobj>
+  inline Lattice<vobj> Expand(int dim,Lattice<vobj> &in)
+  {
+    GridBase *old_grid = in.Grid();
+    GridCartesian *new_grid = grids[dim];//These are new grids
+    Lattice<vobj>  padded(new_grid);
+    Lattice<vobj> shifted(old_grid);    
+    Coordinate local     =old_grid->LocalDimensions();
+    Coordinate plocal    =new_grid->LocalDimensions();
+    if(dim==0) conformable(old_grid,unpadded_grid);
+    else       conformable(old_grid,grids[dim-1]);
+
+    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
+    // Middle bit
+    for(int x=0;x<local[dim];x++){
+      InsertSliceLocal(in,padded,x,depth+x,dim);
+    }
+    // High bit
+    shifted = Cshift(in,dim,depth);
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
+    }
+    // Low bit
+    shifted = Cshift(in,dim,-depth);
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,x,x,dim);
+    }
+    return padded;
+  }
+
+};
+ 
+
+NAMESPACE_END(Grid);
+

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

@@ -104,6 +104,7 @@ template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iSca
 template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
 template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
 template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
+template<typename vtype> using iLorentzComplex            = iVector<iScalar<iScalar<vtype> >, Nd > ;
 template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
 template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >;
 template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >;
@@ -178,6 +179,15 @@ typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF;
 typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD;
 typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2;
 
+// LorentzComplex
+typedef iLorentzComplex<Complex  > LorentzComplex;
+typedef iLorentzComplex<ComplexF > LorentzComplexF;
+typedef iLorentzComplex<ComplexD > LorentzComplexD;
+
+typedef iLorentzComplex<vComplex > vLorentzComplex;
+typedef iLorentzComplex<vComplexF> vLorentzComplexF;
+typedef iLorentzComplex<vComplexD> vLorentzComplexD;
+
 // DoubleStored gauge field
 typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
@@ -307,6 +317,10 @@ typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF;
 typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD;
 typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2;
 
+typedef Lattice<vLorentzComplex>  LatticeLorentzComplex;
+typedef Lattice<vLorentzComplexF> LatticeLorentzComplexF;
+typedef Lattice<vLorentzComplexD> LatticeLorentzComplexD;
+
 // DoubleStored gauge field
 typedef Lattice<vDoubleStoredColourMatrix>   LatticeDoubleStoredColourMatrix;
 typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF;
@@ -507,9 +521,20 @@ 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 +546,23 @@ 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 +574,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
+#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/ActionBase.h

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

Grid/qcd/action/ActionCore.h

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

Grid/qcd/action/fermion/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
@@ -484,27 +484,30 @@ 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);
+    accelerator_barrier();
   }
 
 };

Grid/qcd/action/fermion/WilsonKernels.h

@@ -52,13 +52,6 @@ public:
   typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;   
 public:
 
-#ifdef GRID_SYCL
-#define SYCL_HACK
-#endif  
-#ifdef SYCL_HACK
-  static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
-			       int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
-#endif
   
   static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,

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/WilsonFermion5DImplementation.h

@@ -63,6 +63,10 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   _tmp(&FiveDimRedBlackGrid),
   Dirichlet(0)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
+  
   // some assertions
   assert(FiveDimGrid._ndimension==5);
   assert(FourDimGrid._ndimension==4);
@@ -328,8 +332,7 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   /////////////////////////////
   {
     GRID_TRACE("Gather");
-    st.HaloExchangeOptGather(in,compressor);
-    accelerator_barrier();
+    st.HaloExchangeOptGather(in,compressor); // Put the barrier in the routine
   }
   
   std::vector<std::vector<CommsRequest_t> > requests;

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

@@ -60,6 +60,9 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
       _tmp(&Hgrid),
       anisotropyCoeff(anis)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
   // Allocate the required comms buffer
   ImportGauge(_Umu);
   if  (anisotropyCoeff.isAnisotropic){

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

@@ -428,16 +428,29 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
   auto ptr = &st.surface_list[0];					\
   accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
       int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
+      int sU = sF/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
+    });									\
+  accelerator_barrier();
 
 #define ASM_CALL(A)							\
-  thread_for( ss, Nsite, {						\
+  thread_for( sss, Nsite, {						\
+    int ss = st.lo->Reorder(sss);					\
     int sU = ss;							\
     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,
@@ -451,11 +464,7 @@ 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
@@ -466,8 +475,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+     // dependent on result of merge
+     acceleratorFenceComputeStream();
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
@@ -490,21 +501,20 @@ 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
    } else if( exterior ) {
+     // Dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteDagExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
-     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);
 
     /////////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -36,90 +38,73 @@ NAMESPACE_BEGIN(Grid);
     // cf. GeneralEvenOddRational.h for details
     /////////////////////////////////////////////////////////////////////////////////////////////////////////////
       
-    template<class ImplD, class ImplF, class ImplD2>
+    template<class ImplD, class ImplF>
     class GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<ImplD> {
     private:
-      typedef typename ImplD2::FermionField FermionFieldD2;
       typedef typename ImplD::FermionField FermionFieldD;
       typedef typename ImplF::FermionField FermionFieldF;
 
       FermionOperator<ImplD> & NumOpD;
       FermionOperator<ImplD> & DenOpD;
 
-      FermionOperator<ImplD2> & NumOpD2;
-      FermionOperator<ImplD2> & DenOpD2;
-     
       FermionOperator<ImplF> & NumOpF;
       FermionOperator<ImplF> & DenOpF;
 
       Integer ReliableUpdateFreq;
     protected:
 
+      //Action evaluation
       //Allow derived classes to override the multishift CG
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
-#if 0
-	SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOp : DenOp);
+#if 1
+	SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOpD : DenOpD);
 	ConjugateGradientMultiShift<FermionFieldD> msCG(MaxIter, approx);
 	msCG(schurOp,in, out);
 #else
-	SchurDifferentiableOperator<ImplD2> schurOpD2(numerator ? NumOpD2 : DenOpD2);
+	SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
 	SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
-	FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
-	FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
-	
-	ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
-	precisionChange(inD2,in);
-	std::cout << "msCG single solve "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
-	msCG(schurOpD2, inD2, outD2);
-	precisionChange(out,outD2);
+	FermionFieldD inD(NumOpD.FermionRedBlackGrid());
+	FermionFieldD outD(NumOpD.FermionRedBlackGrid());
+
+	// Action better with higher precision?
+	ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	msCG(schurOpD, in, out);
 #endif
       }
+      //Force evaluation
       virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
-	SchurDifferentiableOperator<ImplD2> schurOpD2(numerator ? NumOpD2 : DenOpD2);
-	SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
+	SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
+	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);
-	precisionChange(inD2,in);
-	std::cout << "msCG in "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
-	msCG(schurOpD2, inD2, out_elemsD2, outD2);
-	precisionChange(out,outD2);
-	for(int i=0;i<out_elems.size();i++){
-	  precisionChange(out_elems[i],out_elemsD2[i]);
-	}
+	FermionFieldD inD(NumOpD.FermionRedBlackGrid());
+	FermionFieldD outD(NumOpD.FermionRedBlackGrid());
+	std::vector<FermionFieldD> out_elemsD(out_elems.size(),NumOpD.FermionRedBlackGrid());
+	ConjugateGradientMultiShiftMixedPrecCleanup<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
+	msCG(schurOpD, in, out_elems, out);
       }
       //Allow derived classes to override the gauge import
       virtual void ImportGauge(const typename ImplD::GaugeField &Ud){
 
 	typename ImplF::GaugeField Uf(NumOpF.GaugeGrid());
-	typename ImplD2::GaugeField Ud2(NumOpD2.GaugeGrid());
 	precisionChange(Uf, Ud);
-	precisionChange(Ud2, Ud);
 
-	std::cout << "Importing "<<norm2(Ud)<<" "<< norm2(Uf)<<" " << norm2(Ud2)<<std::endl;
+	std::cout << "Importing "<<norm2(Ud)<<" "<< norm2(Uf)<<" " <<std::endl;
 	
 	NumOpD.ImportGauge(Ud);
 	DenOpD.ImportGauge(Ud);
 
 	NumOpF.ImportGauge(Uf);
 	DenOpF.ImportGauge(Uf);
-
-	NumOpD2.ImportGauge(Ud2);
-	DenOpD2.ImportGauge(Ud2);
       }
       
     public:
       GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction(FermionOperator<ImplD>  &_NumOpD, FermionOperator<ImplD>  &_DenOpD, 
 							      FermionOperator<ImplF>  &_NumOpF, FermionOperator<ImplF>  &_DenOpF, 
-							      FermionOperator<ImplD2>  &_NumOpD2, FermionOperator<ImplD2>  &_DenOpD2, 
 							      const RationalActionParams & p, Integer _ReliableUpdateFreq
 							      ) : GeneralEvenOddRatioRationalPseudoFermionAction<ImplD>(_NumOpD, _DenOpD, p),
 								  ReliableUpdateFreq(_ReliableUpdateFreq),
 								  NumOpD(_NumOpD), DenOpD(_DenOpD),
-								  NumOpF(_NumOpF), DenOpF(_DenOpF),
-								  NumOpD2(_NumOpD2), DenOpD2(_DenOpD2)
+								  NumOpF(_NumOpF), DenOpF(_DenOpF)
       {}
       
       virtual std::string action_name(){return "GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction";}

Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h

@@ -67,9 +67,9 @@ NAMESPACE_BEGIN(Grid);
       virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
     };
 
-    template<class Impl,class ImplF,class ImplD2>
+    template<class Impl,class ImplF>
     class OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction
-      : public GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF,ImplD2> {
+      : public GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF> {
     public:
       typedef OneFlavourRationalParams Params;
     private:
@@ -91,11 +91,9 @@ NAMESPACE_BEGIN(Grid);
 								 FermionOperator<Impl>  &_DenOp, 
 								 FermionOperator<ImplF>  &_NumOpF, 
 								 FermionOperator<ImplF>  &_DenOpF, 
-								 FermionOperator<ImplD2>  &_NumOpD2, 
-								 FermionOperator<ImplD2>  &_DenOpD2, 
 								 const Params & p, Integer ReliableUpdateFreq
 							) : 
-	GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF,ImplD2>(_NumOp, _DenOp,_NumOpF, _DenOpF,_NumOpD2, _DenOpD2, transcribe(p),ReliableUpdateFreq){}
+	GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<Impl,ImplF>(_NumOp, _DenOp,_NumOpF, _DenOpF, transcribe(p),ReliableUpdateFreq){}
 
       virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}      
     };

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;
@@ -220,20 +207,27 @@ NAMESPACE_BEGIN(Grid);
         //X = (Mdag M)^-1 V^dag phi
         //Y = (Mdag)^-1 V^dag  phi
         Vpc.MpcDag(PhiOdd,Y);          // Y= Vdag phi
+	std::cout << GridLogMessage <<" Y "<<norm2(Y)<<std::endl;
         X=Zero();
         DerivativeSolver(Mpc,Y,X);     // X= (MdagM)^-1 Vdag phi
+	std::cout << GridLogMessage <<" X "<<norm2(X)<<std::endl;
         Mpc.Mpc(X,Y);                  // Y=  Mdag^-1 Vdag phi
+	std::cout << GridLogMessage <<" Y "<<norm2(Y)<<std::endl;
 
         // phi^dag V (Mdag M)^-1 dV^dag  phi
         Vpc.MpcDagDeriv(force , X, PhiOdd );   dSdU = force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
   
         // phi^dag dV (Mdag M)^-1 V^dag  phi
         Vpc.MpcDeriv(force , PhiOdd, X );      dSdU = dSdU+force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
 
         //    -    phi^dag V (Mdag M)^-1 Mdag dM   (Mdag M)^-1 V^dag  phi
         //    -    phi^dag V (Mdag M)^-1 dMdag M   (Mdag M)^-1 V^dag  phi
         Mpc.MpcDeriv(force,Y,X);              dSdU = dSdU-force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
         Mpc.MpcDagDeriv(force,X,Y);           dSdU = dSdU-force;
+	std::cout << GridLogMessage <<" deriv "<<norm2(force)<<std::endl;
 
         // FIXME No force contribution from EvenEven assumed here
         // Needs a fix for clover.

Grid/qcd/hmc/HMC.h

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

Grid/qcd/hmc/checkpointers/BaseCheckpointer.h

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

Grid/qcd/hmc/checkpointers/BinaryCheckpointer.h

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

Grid/qcd/hmc/checkpointers/ILDGCheckpointer.h

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

Grid/qcd/hmc/checkpointers/NerscCheckpointer.h

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

Grid/qcd/hmc/checkpointers/ScidacCheckpointer.h

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

Grid/qcd/hmc/integrators/Integrator.h

@@ -66,6 +66,7 @@ public:
 template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
+public:
   typedef FieldImplementation_ FieldImplementation;
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
@@ -96,7 +97,6 @@ protected:
   {
     t_P[level] += ep;
     update_P(P, U, level, ep);
-
     std::cout << GridLogIntegrator << "[" << level << "] P " << " dt " << ep << " : t_P " << t_P[level] << std::endl;
   }
 
@@ -130,30 +130,20 @@ protected:
       Field force(U.Grid());
       conformable(U.Grid(), Mom.Grid());
 
-      Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
 
-      std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] before"<<std::endl;
-      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(Smearer, 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();
-      if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
 
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
-
-      //      DumpSliceNorm("force ",force,Nd-1);
+      
       MomFilter->applyFilter(force);
+
       std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<<" dt "<<ep<<  std::endl;
-      DumpSliceNorm("force filtered ",force,Nd-1);
       
       Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
       Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
@@ -284,6 +274,15 @@ 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;
@@ -370,15 +369,10 @@ public:
 	auto name = as[level].actions.at(actionID)->action_name();
         std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
 
-        Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
-
-	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] before"<<std::endl;
-	AUDIT();
 	as[level].actions.at(actionID)->refresh_timer_start();
-        as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
+        as[level].actions.at(actionID)->refresh(Smearer, sRNG, pRNG);
 	as[level].actions.at(actionID)->refresh_timer_stop();
-	std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] after"<<std::endl;
-	AUDIT();
+
       }
 
       // Refresh the higher representation actions
@@ -415,17 +409,16 @@ 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);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
 	        as[level].actions.at(actionID)->S_timer_start();
-        Hterm = as[level].actions.at(actionID)->S(Us);
+        Hterm = as[level].actions.at(actionID)->S(Smearer);
    	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
-	AUDIT();
+
       }
       as[level].apply(S_hireps, Representations, level, H);
     }
@@ -438,9 +431,9 @@ 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;
 
@@ -462,13 +455,12 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
-        Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
-	        as[level].actions.at(actionID)->S_timer_start();
-	AUDIT();
-        Hterm = as[level].actions.at(actionID)->Sinitial(Us);
-   	        as[level].actions.at(actionID)->S_timer_stop();
-	AUDIT();
+
+	as[level].actions.at(actionID)->S_timer_start();
+        Hterm = as[level].actions.at(actionID)->S(Smearer);
+	as[level].actions.at(actionID)->S_timer_stop();
+
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
       }
@@ -481,7 +473,6 @@ public:
   
   void integrate(Field& U) 
   {
-    AUDIT();
     // reset the clocks
     t_U = 0;
     for (int level = 0; level < as.size(); ++level) {
@@ -499,10 +490,8 @@ 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/qcd/observables/hmc_observable.h

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

Grid/qcd/observables/plaquette.h

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

Grid/qcd/smearing/GaugeConfiguration.h

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

Grid/qcd/smearing/GaugeConfigurationMasked.h

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

Grid/qcd/smearing/JacobianAction.h

@@ -0,0 +1,87 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/gauge/JacobianAction.h
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+////////////////////////////////////////////////////////////////////////
+// Jacobian Action .. 
+////////////////////////////////////////////////////////////////////////
+template <class Gimpl>
+class JacobianAction : public Action<typename Gimpl::GaugeField> {
+public:  
+  INHERIT_GIMPL_TYPES(Gimpl);
+
+  SmearedConfigurationMasked<Gimpl> * smearer;
+  /////////////////////////// constructors
+  explicit JacobianAction(SmearedConfigurationMasked<Gimpl> * _smearer ) { smearer=_smearer;};
+
+  virtual std::string action_name() {return "JacobianAction";}
+
+  virtual std::string LogParameters(){
+    std::stringstream sstream;
+    sstream << GridLogMessage << "[JacobianAction] " << std::endl;
+    return sstream.str();
+  }
+
+  //////////////////////////////////
+  // Usual cases are not used
+  //////////////////////////////////
+  virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG){ assert(0);};
+  virtual RealD S(const GaugeField &U) { assert(0); }
+  virtual void deriv(const GaugeField &U, GaugeField &dSdU) { assert(0);  }
+
+  //////////////////////////////////
+  // Functions of smart configs only
+  //////////////////////////////////
+  virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
+  {
+    return;
+  }
+  virtual RealD S(ConfigurationBase<GaugeField>& U)
+  {
+    // det M = e^{ - ( - logDetM) }
+    assert( &U == smearer );
+    return -smearer->logDetJacobian();
+  }
+  virtual RealD Sinitial(ConfigurationBase<GaugeField>& U) 
+  {
+    return S(U);
+  }
+  virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU)
+  {
+    assert( &U == smearer );
+    smearer->logDetJacobianForce(dSdU);
+  }
+
+private:
+ };
+
+NAMESPACE_END(Grid);
+

Grid/qcd/smearing/StoutSmearing.h

@@ -40,7 +40,9 @@ template <class Gimpl>
 class Smear_Stout : public Smear<Gimpl> {
  private:
   int OrthogDim = -1;
+public:
   const std::vector<double> SmearRho;
+private:
   // Smear<Gimpl>* ownership semantics:
   //    Smear<Gimpl>* passed in to constructor are owned by caller, so we don't delete them here
   //    Smear<Gimpl>* created within constructor need to be deleted as part of the destructor

Grid/qcd/utils/SUn.h

@@ -34,6 +34,59 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+template<int N, class Vec>
+Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
+
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us()()(i,j);
+      }}
+    ComplexD detD  = EigenU.determinant();
+    typename Vec::scalar_type det(detD.real(),detD.imag());
+    pokeLocalSite(det,ret_v,lcoor);
+  });
+  return ret;
+}
+
+template<int N, class Vec>
+static void ProjectSUn(Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
+{
+  Umu      = ProjectOnGroup(Umu);
+  auto det = Determinant(Umu);
+
+  det = conjugate(det);
+
+  for(int i=0;i<N;i++){
+    auto element = PeekIndex<ColourIndex>(Umu,N-1,i);
+    element = element * det;
+    PokeIndex<ColourIndex>(Umu,element,Nc-1,i);
+  }
+}
+template<int N,class Vec>
+static void ProjectSUn(Lattice<iVector<iScalar<iMatrix<Vec, N> >,Nd> > &U)
+{
+  GridBase *grid=U.Grid();
+  // Reunitarise
+  for(int mu=0;mu<Nd;mu++){
+    auto Umu = PeekIndex<LorentzIndex>(U,mu);
+    Umu      = ProjectOnGroup(Umu);
+    ProjectSUn(Umu);
+    PokeIndex<LorentzIndex>(U,Umu,mu);
+  }
+}
+
 template <int ncolour>
 class SU {
 public:
@@ -741,8 +794,14 @@ public:
     typedef Lattice<vMatrixType> LatticeMatrixType;
 
     LatticeMatrixType Umu(out.Grid());
+    LatticeMatrixType tmp(out.Grid());
     for (int mu = 0; mu < Nd; mu++) {
-      LieRandomize(pRNG, Umu, 1.0);
+      //      LieRandomize(pRNG, Umu, 1.0);
+      //      PokeIndex<LorentzIndex>(out, Umu, mu);
+      gaussian(pRNG,Umu);
+      tmp = Ta(Umu);
+      taExp(tmp,Umu);
+      ProjectSUn(Umu);
       PokeIndex<LorentzIndex>(out, Umu, mu);
     }
   }
@@ -799,12 +858,12 @@ public:
 };
 
 template<int N>
-LatticeComplexD Determinant(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
+Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
 {
   GridBase *grid=Umu.Grid();
   auto lvol = grid->lSites();
-  LatticeComplexD ret(grid);
-
+  Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
+  
   autoView(Umu_v,Umu,CpuRead);
   autoView(ret_v,ret,CpuWrite);
   thread_for(site,lvol,{
@@ -812,42 +871,21 @@ LatticeComplexD Determinant(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N>
     Coordinate lcoor;
     grid->LocalIndexToLocalCoor(site, lcoor);
     iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
     peekLocalSite(Us, Umu_v, lcoor);
     for(int i=0;i<N;i++){
       for(int j=0;j<N;j++){
 	EigenU(i,j) = Us()()(i,j);
       }}
-    ComplexD det = EigenU.determinant();
-    pokeLocalSite(det,ret_v,lcoor);
+    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	Ui()()(i,j) = EigenUinv(i,j);
+      }}
+    pokeLocalSite(Ui,ret_v,lcoor);
   });
   return ret;
 }
-template<int N>
-static void ProjectSUn(Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
-{
-  Umu      = ProjectOnGroup(Umu);
-  auto det = Determinant(Umu);
-
-  det = conjugate(det);
-
-  for(int i=0;i<N;i++){
-    auto element = PeekIndex<ColourIndex>(Umu,N-1,i);
-    element = element * det;
-    PokeIndex<ColourIndex>(Umu,element,Nc-1,i);
-  }
-}
-template<int N>
-static void ProjectSUn(Lattice<iVector<iScalar<iMatrix<vComplexD, N> >,Nd> > &U)
-{
-  GridBase *grid=U.Grid();
-  // Reunitarise
-  for(int mu=0;mu<Nd;mu++){
-    auto Umu = PeekIndex<LorentzIndex>(U,mu);
-    Umu      = ProjectOnGroup(Umu);
-    ProjectSUn(Umu);
-    PokeIndex<LorentzIndex>(U,Umu,mu);
-  }
-}
 // Explicit specialisation for SU(3).
 // Explicit specialisation for SU(3).
 static void

Grid/qcd/utils/SUnAdjoint.h

@@ -51,6 +51,7 @@ public:
   typedef Lattice<iVector<iScalar<iMatrix<vComplexF, Dimension> >, Nd> > LatticeAdjFieldF;
   typedef Lattice<iVector<iScalar<iMatrix<vComplexD, Dimension> >, Nd> > LatticeAdjFieldD;
 
+  typedef Lattice<iScalar<iScalar<iVector<vComplex, Dimension> > > >  LatticeAdjVector;
 
   template <class cplx>
   static void generator(int Index, iSUnAdjointMatrix<cplx> &iAdjTa) {

Grid/simd/Grid_neon.h

@@ -320,7 +320,7 @@ struct Conj{
 
 struct TimesMinusI{
   //Complex single
-  inline float32x4_t operator()(float32x4_t in, float32x4_t ret){
+  inline float32x4_t operator()(float32x4_t in){
     // ar ai br bi -> ai -ar ai -br
     float32x4_t r0, r1;
     r0 = vnegq_f32(in);        // -ar -ai -br -bi
@@ -328,7 +328,7 @@ struct TimesMinusI{
     return vtrn1q_f32(r1, r0); //  ar -ai  br -bi
   }
   //Complex double
-  inline float64x2_t operator()(float64x2_t in, float64x2_t ret){
+  inline float64x2_t operator()(float64x2_t in){
     // a ib -> b -ia
     float64x2_t tmp;
     tmp = vnegq_f64(in);
@@ -338,7 +338,7 @@ struct TimesMinusI{
 
 struct TimesI{
   //Complex single
-  inline float32x4_t operator()(float32x4_t in, float32x4_t ret){
+  inline float32x4_t operator()(float32x4_t in){
     // ar ai br bi -> -ai ar -bi br
     float32x4_t r0, r1;
     r0 = vnegq_f32(in);        // -ar -ai -br -bi
@@ -346,7 +346,7 @@ struct TimesI{
     return vtrn1q_f32(r1, in); // -ai  ar -bi  br
   }
   //Complex double
-  inline float64x2_t operator()(float64x2_t in, float64x2_t ret){
+  inline float64x2_t operator()(float64x2_t in){
     // a ib -> -b ia
     float64x2_t tmp;
     tmp = vnegq_f64(in);

Grid/stencil/GeneralLocalStencil.h

@@ -123,7 +123,7 @@ public:
 	  }
 	  if ( permute_slice ) {
 	    int ptype       =grid->PermuteType(d);
-	    uint8_t mask    =grid->Nsimd() >> (ptype + 1);		
+	    uint8_t mask    =0x1<<ptype;
 	    SE._permute    |= mask;
 	  }
 	}	

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
@@ -290,9 +300,9 @@ public:
 
 protected:
   GridBase *                        _grid;
-
 public:
   GridBase *Grid(void) const { return _grid; }
+  LebesgueOrder *lo;
 
   ////////////////////////////////////////////////////////////////////////
   // Needed to conveniently communicate gparity parameters into GPU memory
@@ -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;
 
@@ -328,8 +339,8 @@ public:
   // Vectors that live on the symmetric heap in case of SHMEM
   // These are used; either SHM objects or refs to the above symmetric heap vectors
   // depending on comms target
-  Vector<cobj *> u_simd_send_buf;
-  Vector<cobj *> u_simd_recv_buf;
+  std::vector<cobj *> u_simd_send_buf;
+  std::vector<cobj *> u_simd_recv_buf;
 
   int u_comm_offset;
   int _unified_buffer_size;
@@ -337,6 +348,7 @@ public:
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
+#if 1
   inline int SameNode(int point) {
 
     int dimension    = this->_directions[point];
@@ -356,7 +368,40 @@ public:
     if ( displacement == 0 ) return 1;
     return 0;
   }
+#else
+  // fancy calculation for shm code
+  inline int SameNode(int point) {
 
+    int dimension    = this->_directions[point];
+    int displacement = this->_distances[point];
+
+    int pd              = _grid->_processors[dimension];
+    int fd              = _grid->_fdimensions[dimension];
+    int ld              = _grid->_ldimensions[dimension];
+    int rd              = _grid->_rdimensions[dimension];
+    int simd_layout     = _grid->_simd_layout[dimension];
+    int comm_dim        = _grid->_processors[dimension] >1 ;
+ 
+    int recv_from_rank;
+    int xmit_to_rank;
+
+    if ( ! comm_dim ) return 1;
+
+    int nbr_proc;
+    if (displacement>0) nbr_proc = 1;
+    else                nbr_proc = pd-1;
+
+    // FIXME  this logic needs to be sorted for three link term
+    //    assert( (displacement==1) || (displacement==-1));
+    // Present hack only works for >= 4^4 subvol per node
+    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
+
+    void *shm = (void *) _grid->ShmBufferTranslate(recv_from_rank,this->u_recv_buf_p);
+
+    if ( shm==NULL ) return 0;
+    return 1;
+  }
+#endif
   //////////////////////////////////////////
   // Comms packet queue for asynch thread
   // Use OpenMP Tasks for cleaner ???
@@ -389,7 +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,
@@ -403,6 +447,11 @@ 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();
+    _grid->StencilBarrier(); 
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
@@ -480,7 +529,6 @@ public:
   {
     _grid->StencilBarrier();// Synch shared memory on a single nodes
 
-    // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
 
     u_comm_offset=0;
@@ -491,6 +539,7 @@ public:
       compress.Point(point);
       HaloGatherDir(source,compress,point,face_idx);
     }
+    accelerator_barrier();
     face_table_computed=1;
     assert(u_comm_offset==_unified_buffer_size);
 
@@ -591,7 +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;
@@ -606,8 +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>
@@ -656,6 +705,7 @@ public:
 	}
       }
     }
+    std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
   }
   /// Introduce a block structure and switch off comms on boundaries
   void DirichletBlock(const Coordinate &dirichlet_block)
@@ -731,6 +781,10 @@ 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);
@@ -1056,7 +1110,7 @@ public:
     int comms_recv   = this->_comms_recv[point];
     int comms_partial_send   = this->_comms_partial_send[point] ;
     int comms_partial_recv   = this->_comms_partial_recv[point] ;
-
+    
     assert(rhs.Grid()==_grid);
     //	  conformable(_grid,rhs.Grid());
 
@@ -1127,11 +1181,32 @@ public:
 	  recv_buf=this->u_recv_buf_p;
 	}
 
+	// potential SHM fast path for intranode
+	int shm_send=0;
+	int shm_recv=0;
+#ifdef SHM_FAST_PATH
+	// Put directly in place if we can
+	send_buf = (cobj *)_grid->ShmBufferTranslate(xmit_to_rank,recv_buf);
+	if ( (send_buf==NULL) ) {
+	  shm_send=0;
+	  send_buf = this->u_send_buf_p;
+	} else {
+	  shm_send=1;
+	}
+	void *test_ptr = _grid->ShmBufferTranslate(recv_from_rank,recv_buf);
+	if ( test_ptr != NULL ) shm_recv = 1;
+	//	static int printed;
+	//	if (!printed){
+	  //	  std::cout << " GATHER FAST PATH SHM "<<shm_send<< " "<<shm_recv<<std::endl;
+	//	  printed = 1;
+	//	}
+#else
 	////////////////////////////////////////////////////////
 	// Gather locally
 	////////////////////////////////////////////////////////
 	send_buf = this->u_send_buf_p; // Gather locally, must send
 	assert(send_buf!=NULL);
+#endif
 
 	//	std::cout << " GatherPlaneSimple partial send "<< comms_partial_send<<std::endl;
 	compressor::Gather_plane_simple(face_table[face_idx],rhs,send_buf,compress,comm_off,so,comms_partial_send);
@@ -1143,10 +1218,13 @@ public:
 	  // Build a list of things to do after we synchronise GPUs
 	  // Start comms now???
 	  ///////////////////////////////////////////////////////////
+	  int do_send = (comms_send|comms_partial_send) && (!shm_send );
+	  int do_recv = (comms_send|comms_partial_send) && (!shm_recv );
+	  
 	  AddPacket((void *)&send_buf[comm_off],
 		    (void *)&recv_buf[comm_off],
-		    xmit_to_rank, comms_send|comms_partial_send,
-		    recv_from_rank, comms_recv|comms_partial_recv,
+		    xmit_to_rank, do_send,
+		    recv_from_rank, do_recv,
 		    xbytes,rbytes);
 	}
 
@@ -1210,8 +1288,8 @@ 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
@@ -1288,19 +1366,47 @@ public:
 
 	    int recv_from_rank;
 	    int xmit_to_rank;
+	    int shm_send=0;
 
 	    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
-
+#ifdef SHM_FAST_PATH
+  #warning STENCIL SHM FAST PATH SELECTED
+  	  int shm_recv=0;
+	    // shm == receive pointer         if offnode
+	    // shm == Translate[send pointer] if on node -- my view of his send pointer
+	    cobj *shm = (cobj *) _grid->ShmBufferTranslate(recv_from_rank,sp);
+	    if (shm==NULL) {
+	      shm = rp;
+	      // we found a packet that comes from MPI and contributes to this shift.
+	      // is_same_node is only used in the WilsonStencil, and gets set for this point in the stencil.
+	      // Kernel will add the exterior_terms except if is_same_node.
+	      // leg of stencil
+	      shm_recv=0;
+	    } else {
+	      shm_recv=1;
+	    }
+	    rpointers[i] = shm;
+	    // Test send side
+	    void *test_ptr = (void *) _grid->ShmBufferTranslate(xmit_to_rank,sp);
+	    if ( test_ptr != NULL ) shm_send = 1;
+	    //	    static int printed;
+	    //	    if (!printed){
+	    //	      std::cout << " GATHERSIMD FAST PATH SHM "<<shm_send<< " "<<shm_recv<<std::endl;
+	    //	      printed = 1;
+	    //	    }
+#else
 	    rpointers[i] = rp;
+#endif
 	    
 	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,xbytes,rbytes,cbmask);
 	    if ( !duplicate  ) { 
 	      if ( (bytes != rbytes) && (rbytes!=0) ){
 		acceleratorMemSet(rp,0,bytes); // Zero prefill comms buffer to zero
 	      }
+	      int do_send = (comms_send|comms_partial_send) && (!shm_send );
 	      AddPacket((void *)sp,(void *)rp,
-			xmit_to_rank,comms_send|comms_partial_send,
-			recv_from_rank,comms_recv|comms_partial_recv,
+			xmit_to_rank,do_send,
+			recv_from_rank,do_send,
 			xbytes,rbytes);
 	    }
 
@@ -1310,7 +1416,7 @@ public:
 
 	  }
 	}
-
+	// rpointer may be doing a remote read in the gather over SHM
 	if ( comms_recv|comms_partial_recv ) {
 	  AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
 	}

Grid/tensors/Tensor_extract_merge.h

@@ -133,7 +133,6 @@ typename vobj::scalar_object extractLane(int lane, const vobj & __restrict__ vec
   typedef scalar_type * pointer;
 
   constexpr int words=sizeof(vobj)/sizeof(vector_type);
-  constexpr int Nsimd=vector_type::Nsimd();
 
   scalar_object extracted;
   pointer __restrict__  sp = (pointer)&extracted; // Type pun
@@ -153,7 +152,6 @@ void insertLane(int lane, vobj & __restrict__ vec,const typename vobj::scalar_ob
   typedef scalar_type * pointer;
 
   constexpr int words=sizeof(vobj)/sizeof(vector_type);
-  constexpr int Nsimd=vector_type::Nsimd();
 
   pointer __restrict__ sp = (pointer)&extracted;
   vector_type *vp = (vector_type *)&vec;
@@ -178,8 +176,6 @@ void extract(const vobj &vec,const ExtractPointerArray<sobj> &extracted, int off
   const int s = Nsimd/Nextr;
 
   vector_type * vp = (vector_type *)&vec;
-  scalar_type      vtmp;
-  sobj_scalar_type stmp;
   for(int w=0;w<words;w++){
     for(int i=0;i<Nextr;i++){
       sobj_scalar_type * pointer = (sobj_scalar_type *)& extracted[i][offset];
@@ -205,7 +201,6 @@ void merge(vobj &vec,const ExtractPointerArray<sobj> &extracted, int offset)
 
   vector_type * vp = (vector_type *)&vec;
   scalar_type      vtmp;
-  sobj_scalar_type stmp;
   for(int w=0;w<words;w++){
     for(int i=0;i<Nextr;i++){
       sobj_scalar_type * pointer = (sobj_scalar_type *)& extracted[i][offset];
@@ -226,7 +221,7 @@ 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;  
@@ -242,18 +237,15 @@ void copyLane(vobjOut & __restrict__ vecOut, int lane_out, const vobjIn & __rest
   typedef oextract_type * opointer;
   typedef iextract_type * ipointer;
 
-  constexpr int oNsimd=ovector_type::Nsimd();
-  constexpr int iNsimd=ivector_type::Nsimd();
-
   iscalar_type itmp;
   oscalar_type otmp;
 
   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

@@ -248,17 +248,23 @@ inline int  acceleratorIsCommunicable(void *ptr)
 //////////////////////////////////////////////
 // SyCL acceleration
 //////////////////////////////////////////////
-#ifdef GRID_SYCL
-NAMESPACE_END(Grid);
-#include <CL/sycl.hpp>
-#include <CL/sycl/usm.hpp>
 
+#ifdef GRID_SYCL
 #define GRID_SYCL_LEVEL_ZERO_IPC
 
-#ifdef GRID_SYCL_LEVEL_ZERO_IPC
+NAMESPACE_END(Grid);
+#if 0
+#include <CL/sycl.hpp>
+#include <CL/sycl/usm.hpp>
 #include <level_zero/ze_api.h>
 #include <CL/sycl/backend/level_zero.hpp>
+#else
+#include <sycl/CL/sycl.hpp>
+#include <sycl/usm.hpp>
+#include <level_zero/ze_api.h>
+#include <sycl/ext/oneapi/backend/level_zero.hpp>
 #endif
+
 NAMESPACE_BEGIN(Grid);
 
 extern cl::sycl::queue *theGridAccelerator;
@@ -299,14 +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();}
@@ -452,7 +458,8 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
 // Common on all GPU targets
 //////////////////////////////////////////////
 #if defined(GRID_SYCL) || defined(GRID_CUDA) || defined(GRID_HIP)
-#define accelerator_forNB( iter1, num1, nsimd, ... ) accelerator_for2dNB( iter1, num1, iter2, 1, nsimd, {__VA_ARGS__} );
+// FIXME -- the non-blocking nature got broken March 30 2023 by PAB
+#define accelerator_forNB( iter1, num1, nsimd, ... ) accelerator_for2dNB( iter1, num1, iter2, 1, nsimd, {__VA_ARGS__} );  
 
 #define accelerator_for( iter, num, nsimd, ... )		\
   accelerator_forNB(iter, num, nsimd, { __VA_ARGS__ } );	\
@@ -519,7 +526,7 @@ inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 //////////////////////////////////////////////
 
 #ifdef GRID_SYCL
-inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
+inline void acceleratorFenceComputeStream(void){ theGridAccelerator->ext_oneapi_submit_barrier(); };
 #else
 // Ordering within a stream guaranteed on Nvidia & AMD
 inline void acceleratorFenceComputeStream(void){ };

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=9000;
+  
+  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;
@@ -366,12 +362,12 @@ int main(int argc, char **argv) {
 
   // Probably dominates the force - back to EOFA.
   OneFlavourRationalParams SFRp;
-  SFRp.lo       = 0.25;
+  SFRp.lo       = 0.1;
   SFRp.hi       = 25.0;
   SFRp.MaxIter  = 10000;
-  SFRp.tolerance= 1.0e-5;
+  SFRp.tolerance= 1.0e-8;
   SFRp.mdtolerance= 2.0e-4;
-  SFRp.degree   = 8;
+  SFRp.degree   = 12;
   SFRp.precision= 50;
   
   MobiusEOFAFermionD Strange_Op_L (U , *FGrid , *FrbGrid , *GridPtr , *GridRBPtr , strange_mass, strange_mass, pv_mass, 0.0, -1, M5, b, c);
@@ -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> *> Bdys;
 #else
-  std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
+  std::vector<GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
 #endif
 
   typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
@@ -532,31 +526,17 @@ 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>(
 			   *Numerators[h],*Denominators[h],
 			   *NumeratorsF[h],*DenominatorsF[h],
-			   *NumeratorsD2[h],*DenominatorsD2[h],
-			   OFRp, 400) );
-      Bdys.push_back( new OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
+			   OFRp, SP_iters) );
+      Bdys.push_back( new GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF>(
 			   *Numerators[h],*Denominators[h],
 			   *NumeratorsF[h],*DenominatorsF[h],
-			   *NumeratorsD2[h],*DenominatorsD2[h],
-			   OFRp, 400) );
+			   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_DD_RHMC_96I.cc

@@ -329,7 +329,6 @@ int main(int argc, char **argv) {
 
     
     auto grid4= GridPtr;
-    auto rbgrid4= GridRBPtr;
     auto rbgrid = StrangeOp.FermionRedBlackGrid();
     auto grid = StrangeOp.FermionGrid();
     if(1){

HMC/Mobius2p1f_EOFA_96I_hmc.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;
 
   //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
@@ -183,7 +178,7 @@ 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 +195,8 @@ 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 +223,7 @@ 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();
@@ -249,11 +244,6 @@ int main(int argc, char **argv) {
   Coordinate shm;
 
   GlobalSharedMemory::GetShmDims(mpi,shm);
-  
-  Coordinate CommDim(Nd);
-  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
-
-  Coordinate NonDirichlet(Nd+1,0);
 
   //////////////////////////
   // Fermion Grids
@@ -272,7 +262,6 @@ 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
@@ -283,8 +272,6 @@ int main(int argc, char **argv) {
   std::vector<Complex> boundary = {1,1,1,-1};
   FermionAction::ImplParams Params(boundary);
   FermionActionF::ImplParams ParamsF(boundary);
-  Params.dirichlet=NonDirichlet;
-  ParamsF.dirichlet=NonDirichlet;
 
   //  double StoppingCondition = 1e-14;
   //  double MDStoppingCondition = 1e-9;
@@ -299,8 +286,8 @@ 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);
 
   ////////////////////////////////////
@@ -311,12 +298,12 @@ int main(int argc, char **argv) {
 
   // Probably dominates the force - back to EOFA.
   OneFlavourRationalParams SFRp;
-  SFRp.lo       = 0.25;
-  SFRp.hi       = 25.0;
+  SFRp.lo       = 0.1;
+  SFRp.hi       = 30.0;
   SFRp.MaxIter  = 10000;
-  SFRp.tolerance= 1.0e-5;
-  SFRp.mdtolerance= 2.0e-4;
-  SFRp.degree   = 8;
+  SFRp.tolerance= 1.0e-8;
+  SFRp.mdtolerance= 2.0e-6;
+  SFRp.degree   = 10;
   SFRp.precision= 50;
   
   MobiusEOFAFermionD Strange_Op_L (U , *FGrid , *FrbGrid , *GridPtr , *GridRBPtr , strange_mass, strange_mass, pv_mass, 0.0, -1, M5, b, c);
@@ -369,40 +356,36 @@ int main(int argc, char **argv) {
 	 ActionCGL, ActionCGR,
 	 DerivativeCGL, DerivativeCGR,
 	 SFRp, true);
-  //  Level2.push_back(&EOFA);
+  Level2.push_back(&EOFA);
 
   ////////////////////////////////////
   // up down action
   ////////////////////////////////////
   std::vector<Real> light_den;
   std::vector<Real> light_num;
-  std::vector<int> dirichlet_den;
-  std::vector<int> dirichlet_num;
 
   int n_hasenbusch = hasenbusch.size();
-  light_den.push_back(light_mass);  dirichlet_den.push_back(0);
+  light_den.push_back(light_mass); 
   for(int h=0;h<n_hasenbusch;h++){
-    light_den.push_back(hasenbusch[h]); dirichlet_den.push_back(0);
+    light_den.push_back(hasenbusch[h]);
   }
 
   for(int h=0;h<n_hasenbusch;h++){
-    light_num.push_back(hasenbusch[h]); dirichlet_num.push_back(0);
+    light_num.push_back(hasenbusch[h]);
   }
-  light_num.push_back(pv_mass);  dirichlet_num.push_back(0);
+  light_num.push_back(pv_mass);
 
   std::vector<FermionAction *> Numerators;
   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<OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF> *> Bdys;
 #else
   std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
 #endif
@@ -416,9 +399,7 @@ int main(int argc, char **argv) {
     std::cout << GridLogMessage
 	      << " 2f quotient Action ";
     std::cout << "det D("<<light_den[h]<<")";
-    if ( dirichlet_den[h] ) std::cout << "^dirichlet    ";
     std::cout << "/ det D("<<light_num[h]<<")";
-    if ( dirichlet_num[h] ) std::cout << "^dirichlet    ";
     std::cout << std::endl;
 
     FermionAction::ImplParams ParamsNum(boundary);
@@ -426,21 +407,11 @@ int main(int argc, char **argv) {
     FermionActionF::ImplParams ParamsDenF(boundary);
     FermionActionF::ImplParams ParamsNumF(boundary);
     
-    ParamsNum.dirichlet = NonDirichlet;
-    ParamsDen.dirichlet = NonDirichlet;
-
-    ParamsNum.partialDirichlet = 0;
-    ParamsDen.partialDirichlet = 0;
-    
     Numerators.push_back  (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, ParamsNum));
     Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, ParamsDen));
 
-    ParamsDenF.dirichlet = ParamsDen.dirichlet;
-    ParamsDenF.partialDirichlet = ParamsDen.partialDirichlet;
     DenominatorsF.push_back(new FermionActionF(UF,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_den[h],M5,b,c, ParamsDenF));
 
-    ParamsNumF.dirichlet = ParamsNum.dirichlet;
-    ParamsNumF.partialDirichlet = ParamsNum.partialDirichlet;
     NumeratorsF.push_back  (new FermionActionF(UF,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_num[h],M5,b,c, ParamsNumF));
 
     LinOpD.push_back(new LinearOperatorD(*Denominators[h]));
@@ -477,7 +448,6 @@ int main(int argc, char **argv) {
   // Gauge action
   /////////////////////////////////////////////////////////////
   Level3.push_back(&GaugeAction);
-  TheHMC.TheAction.push_back(Level1);
   TheHMC.TheAction.push_back(Level2);
   TheHMC.TheAction.push_back(Level3);
   std::cout << GridLogMessage << " Action complete "<< std::endl;

README.md

@@ -1,7 +1,8 @@
-# Grid [![Teamcity status](http://ci.cliath.ph.ed.ac.uk/app/rest/builds/aggregated/strob:(buildType:(affectedProject(id:GridBasedSoftware_Grid)),branch:name:develop)/statusIcon.svg)](http://ci.cliath.ph.ed.ac.uk/project.html?projectId=GridBasedSoftware_Grid&tab=projectOverview) 
-
+# Grid 
 **Data parallel C++ mathematical object library.**
 
+[![Teamcity status](https://ci.dev.dirac.ed.ac.uk/guestAuth/app/rest/builds/aggregated/strob:(buildType:(affectedProject(id:GridBasedSoftware_Grid)),branch:default:true)/statusIcon.svg)](https://ci.dev.dirac.ed.ac.uk/project/GridBasedSoftware_Grid?mode=builds) 
+
 License: GPL v2.
 
 Last update June 2017.

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.cc

@@ -425,7 +425,7 @@ void Benchmark(int Ls, Coordinate Dirichlet)
 
   err = r_eo-result;
   n2e= norm2(err);
-  std::cout<<GridLogMessage << "norm diff   "<< n2e<< "  Line "<<__LINE__ <<std::endl;
+  std::cout<<GridLogMessage << "norm diff   "<< n2e<<std::endl;
   assert(n2e<1.0e-4);
 
   pickCheckerboard(Even,src_e,err);

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();
+}

configure.ac

@@ -646,6 +646,14 @@ case ${ac_SHM_FORCE_MPI} in
       ;;
      *) ;;
 esac
+############### force MPI in SMP
+AC_ARG_ENABLE([shm-fast-path],[AS_HELP_STRING([--enable-shm-fast-path],[Allow kernels to remote copy over intranode])],[ac_SHM_FAST_PATH=${enable_shm_fast_path}],[ac_SHM_FAST_PATH=no])
+case ${ac_SHM_FAST_PATH} in
+     yes)
+        AC_DEFINE([SHM_FAST_PATH],[1],[SHM_FAST_PATH] )
+      ;;
+     *) ;;
+esac
 
 ############### communication type selection
 AC_ARG_ENABLE([comms-threads],[AS_HELP_STRING([--enable-comms-threads | --disable-comms-threads],[Use multiple threads in MPI calls])],[ac_COMMS_THREADS=${enable_comms_threads}],[ac_COMMS_THREADS=yes])

documentation/David_notes.txt

@@ -0,0 +1,90 @@
+Branch: develop
+
+Files:
+
+Grid/lattice/PaddedCell.h -- Halo exchange
+tests/Test_general_stencil.cc -- test local off axis stencil addressing
+tests/debug/Test_padded_cell.cc -- test PaddedCell halo exchange and the General local stencil  by computing ALL plaquettes on lattice
+
+Functionality:
+
+-- extend a lattice field:
+Grid/lattice/PaddedCell.h
+
+// Constructor
+  PaddedCell(int _depth,GridCartesian *_grid)
+
+// Expand a field "in" to depth "d"
+  template<class vobj>
+  inline Lattice<vobj> Exchange(Lattice<vobj> &in)
+  
+// Take the "apple core" of in to a smaller local volume
+  template<class vobj>
+  inline Lattice<vobj> Extract(Lattice<vobj> &in)
+
+-- Plaquette test:
+tests/debug/Test_padded_cell.cc
+  /////////////////////////////////////////////////
+  // Create a padded cell of extra padding depth=1
+  /////////////////////////////////////////////////
+  int depth = 1;
+  PaddedCell Ghost(depth,&GRID);
+  LatticeGaugeField Ughost = Ghost.Exchange(Umu);
+
+///// Array for the site plaquette
+  GridBase *GhostGrid = Ughost.Grid();
+  LatticeComplex gplaq(GhostGrid); 
+
+  std::vector<Coordinate> shifts;
+  for(int mu=0;mu<Nd;mu++){
+    for(int nu=mu+1;nu<Nd;nu++){
+  
+      //    Umu(x) Unu(x+mu) Umu^dag(x+nu) Unu^dag(x)
+      Coordinate shift_0(Nd,0);
+      Coordinate shift_mu(Nd,0); shift_mu[mu]=1;
+      Coordinate shift_nu(Nd,0); shift_nu[nu]=1;
+      shifts.push_back(shift_0);
+      shifts.push_back(shift_mu);
+      shifts.push_back(shift_nu);
+      shifts.push_back(shift_0);
+    }
+  }
+  GeneralLocalStencil gStencil(GhostGrid,shifts);
+
+  gplaq=Zero();
+  {
+    autoView( gp_v , gplaq, CpuWrite);
+    autoView( t_v , trplaq, CpuRead);
+    autoView( U_v , Ughost, CpuRead);
+    for(int ss=0;ss<gp_v.size();ss++){
+      int s=0;
+      for(int mu=0;mu<Nd;mu++){
+	for(int nu=mu+1;nu<Nd;nu++){
+
+	  auto SE0 = gStencil.GetEntry(s+0,ss);
+	  auto SE1 = gStencil.GetEntry(s+1,ss);
+	  auto SE2 = gStencil.GetEntry(s+2,ss);
+	  auto SE3 = gStencil.GetEntry(s+3,ss);
+	
+	  int o0 = SE0->_offset;
+	  int o1 = SE1->_offset;
+	  int o2 = SE2->_offset;
+	  int o3 = SE3->_offset;
+	  
+	  auto U0 = U_v[o0](mu);
+	  auto U1 = U_v[o1](nu);
+	  auto U2 = adj(U_v[o2](mu));
+	  auto U3 = adj(U_v[o3](nu));
+
+	  gpermute(U0,SE0->_permute);
+	  gpermute(U1,SE1->_permute);
+	  gpermute(U2,SE2->_permute);
+	  gpermute(U3,SE3->_permute);
+	  
+	  gp_v[ss]() =gp_v[ss]() + trace( U0*U1*U2*U3 );
+	  s=s+4;
+	}
+      }
+    }
+  }
+  cplaq = Ghost.Extract(gplaq);

examples/socket_grid.cc

@@ -0,0 +1,133 @@
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <unistd.h>
+#include <stdio.h>
+#include <err.h>
+#include <fcntl.h>
+#include <assert.h>
+#include <string.h>
+#include <stdlib.h>
+
+static int sock;
+static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
+static char sock_path[256];
+
+class UnixSockets {
+public:
+  static void Open(int rank)
+  {
+    int errnum;
+
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+    printf("allocated socket %d\n",sock);
+
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
+    unlink(sa_un.sun_path);
+    if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
+      perror("bind failure");
+      exit(EXIT_FAILURE);
+    }
+    printf("bound socket %d to %s\n",sock,sa_un.sun_path);
+  }
+
+  static int RecvFileDescriptor(void)
+  {
+    int n;
+    int fd;
+    char buf[1];
+    struct iovec iov;
+    struct msghdr msg;
+    struct cmsghdr *cmsg;
+    char cms[CMSG_SPACE(sizeof(int))];
+
+    iov.iov_base = buf;
+    iov.iov_len = 1;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_name = 0;
+    msg.msg_namelen = 0;
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+
+    msg.msg_control = (caddr_t)cms;
+    msg.msg_controllen = sizeof cms;
+
+    if((n=recvmsg(sock, &msg, 0)) < 0) {
+      perror("recvmsg failed");
+      return -1;
+    }
+    if(n == 0){
+      perror("recvmsg returned 0");
+      return -1;
+    }
+    cmsg = CMSG_FIRSTHDR(&msg);
+    memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
+    printf("received fd %d from socket %d\n",fd,sock);
+    return fd;
+  }
+
+  static void SendFileDescriptor(int fildes,int xmit_to_rank)
+  {
+    struct msghdr msg;
+    struct iovec iov;
+    struct cmsghdr *cmsg = NULL;
+    char ctrl[CMSG_SPACE(sizeof(int))];
+    char data = ' ';
+
+    memset(&msg, 0, sizeof(struct msghdr));
+    memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
+    iov.iov_base = &data;
+    iov.iov_len = sizeof(data);
+    
+    sprintf(sock_path,sock_path_fmt,xmit_to_rank);
+    printf("sending FD %d over socket %d to rank %d AF_UNIX path %s\n",fildes,sock,xmit_to_rank,sock_path);fflush(stdout);
+    
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
+
+    msg.msg_name = (void *)&sa_un;
+    msg.msg_namelen = sizeof(sa_un);
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+    msg.msg_controllen =  CMSG_SPACE(sizeof(int));
+    msg.msg_control = ctrl;
+
+    cmsg = CMSG_FIRSTHDR(&msg);
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int));
+
+    *((int *) CMSG_DATA(cmsg)) = fildes;
+
+    if ( sendmsg(sock, &msg, 0) == -1 ) perror("sendmsg failed");
+  };
+};
+
+int main(int argc, char **argv)
+{
+  int me = fork()?0:1;
+  
+  UnixSockets::Open(me);
+  
+  // need MPI barrier
+  sleep(10);
+  const char * message = "Hello, World\n";
+  if( me ) {
+    int fd = open("foo",O_RDWR|O_CREAT,0666);
+    if ( fd < 0 ) {
+      perror("failed to open file");
+      exit(EXIT_FAILURE);
+    }
+    // rank 1 sends ot rank 0
+    UnixSockets::SendFileDescriptor(fd,0);
+    close(fd);
+  } else {
+    // rank 0 sends receives frmo rank 1
+    int fd = UnixSockets::RecvFileDescriptor();
+    write(fd,(const void *)message,strlen(message));
+    close(fd);
+  }
+}