Getting rid of one more non-auto View, comms overlap in Laplace operator

Checking in to move to aurora

.github/ISSUE_TEMPLATE/bug-report.yml

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

Grid/Makefile.am

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

Grid/algorithms/Algorithms.h

@@ -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

@@ -460,6 +460,53 @@ class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Fi
   }
 };
 
+template<class Matrix,class Field>
+class QuadLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+public:
+  RealD a0,a1,a2;
+  QuadLinearOperator(Matrix &Mat): _Mat(Mat),a0(0.),a1(0.),a2(1.) {};
+  QuadLinearOperator(Matrix &Mat, RealD _a0,RealD _a1,RealD _a2): _Mat(Mat),a0(_a0),a1(_a1),a2(_a2) {};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    assert(0);
+    _Mat.Mdiag(in,out);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    assert(0);
+    _Mat.MdirAll(in,out);
+  }
+  void HermOp (const Field &in, Field &out){
+//    _Mat.M(in,out);
+    Field tmp1(in.Grid());
+//    Linop.HermOpAndNorm(psi, mmp, d, b);
+    _Mat.M(in,tmp1);
+    _Mat.M(tmp1,out);
+    out *= a2;
+    axpy(out, a1, tmp1, out);
+    axpy(out, a0, in, out);
+//    d=real(innerProduct(psi,mmp));
+//    b=norm2(mmp);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    assert(0);
+    _Mat.M(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot= innerProduct(in,out); n1=real(dot);
+    n2=norm2(out);
+  }
+  void Op(const Field &in, Field &out){
+    assert(0);
+    _Mat.M(in,out);
+  }
+};
+
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 // Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
 // Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi
@@ -542,6 +589,7 @@ public:
       (*this)(in[i], out[i]);
     }
   }
+  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {

Grid/algorithms/approx/Forecast.h

@@ -36,11 +36,12 @@ NAMESPACE_BEGIN(Grid);
 // Abstract base class.
 // Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
 // and returns a forecasted solution to the system D*psi = phi (psi).
-template<class Matrix, class Field>
+// Changing to operator
+template<class LinearOperatorBase, class Field>
 class Forecast
 {
 public:
-  virtual Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
+  virtual Field operator()(LinearOperatorBase &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
 };
 
 // Implementation of Brower et al.'s chronological inverter (arXiv:hep-lat/9509012),
@@ -54,13 +55,13 @@ public:
   Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& prev_solns)
   {
     int degree = prev_solns.size();
+    std::cout << GridLogMessage << "ChronoForecast: degree= " << degree << std::endl;
     Field chi(phi); // forecasted solution
 
     // Trivial cases
     if(degree == 0){ chi = Zero(); return chi; }
     else if(degree == 1){ return prev_solns[0]; }
 
-    //    RealD dot;
     ComplexD xp;
     Field r(phi); // residual
     Field Mv(phi);
@@ -83,8 +84,9 @@ public:
     // Perform sparse matrix multiplication and construct rhs
     for(int i=0; i<degree; i++){
       b[i] = innerProduct(v[i],phi);
-      Mat.M(v[i],Mv);
-      Mat.Mdag(Mv,MdagMv[i]);
+//      Mat.M(v[i],Mv);
+//      Mat.Mdag(Mv,MdagMv[i]);
+      Mat.HermOp(v[i],MdagMv[i]);
       G[i][i] = innerProduct(v[i],MdagMv[i]);
     }
 

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/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

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

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

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

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -419,14 +419,15 @@ until convergence
 	}
       }
 
-      if ( Nconv < Nstop )
+      if ( Nconv < Nstop ) {
 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-
+	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
+      }
       eval=eval2;
       
       //Keep only converged
-      eval.resize(Nconv);// Nstop?
-      evec.resize(Nconv,grid);// Nstop?
+      eval.resize(Nstop);// was Nconv
+      evec.resize(Nstop,grid);// was Nconv
       basisSortInPlace(evec,eval,reverse);
       
     }

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];
@@ -121,7 +137,26 @@ private:
   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
   //////////////////////////////////////////////////////////////////////

Grid/allocator/MemoryManagerCache.cc

@@ -519,7 +519,6 @@ void MemoryManager::Audit(std::string s)
   uint64_t LruBytes1=0;
   uint64_t LruBytes2=0;
   uint64_t LruCnt=0;
-  uint64_t LockedBytes=0;
   
   std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
   for(auto it=LRU.begin();it!=LRU.end();it++){

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

@@ -27,9 +27,10 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
+#define Mheader "SharedMemoryMpi: "
+
 #include <Grid/GridCore.h>
 #include <pwd.h>
-#include <syscall.h>
 
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
@@ -37,12 +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)
 {
@@ -65,8 +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
 
@@ -169,59 +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)
 {
   ////////////////////////////////////////////////////////////////
@@ -395,7 +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);
 
@@ -480,7 +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);
@@ -523,7 +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);
@@ -531,8 +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
@@ -540,24 +602,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void * thisBuf = ShmCommBuf;
     if(!Stencil_force_mpi) {
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+    typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
-
+    
     if ( r==WorldShmRank ) { 
       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
+      memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
+#ifdef SHM_SOCKETS
+      for(int rr=0;rr<WorldShmSize;rr++){
+	if(rr!=r){
+	  UnixSockets::SendFileDescriptor(handle.fd,rr);
+	}
+      }
+#endif
     }
 #endif
 #ifdef GRID_CUDA
@@ -585,6 +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,
@@ -600,6 +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;
@@ -607,16 +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;
@@ -651,6 +732,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
+    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
@@ -662,7 +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);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -699,7 +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;
@@ -709,7 +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);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -756,7 +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.h

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

Grid/lattice/Lattice_ET.h

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

Grid/lattice/Lattice_reduction.h

@@ -153,33 +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)
 {
   Integer osites = arg.Grid()->oSites();
 #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  typename vobj::scalar_object ssum;
   autoView( arg_v, arg, AcceleratorRead);
-  ssum= sum_gpu(&arg_v[0],osites);
+  return sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
-  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)||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;
 }

Grid/lattice/Lattice_reduction_gpu.h

@@ -211,25 +211,22 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   assert(ok);
 
   Integer smemSize = numThreads * sizeof(sobj);
-  // UVM seems to be buggy under later CUDA drivers
-  // This fails on A100 and driver 5.30.02 / CUDA 12.1
-  // Fails with multiple NVCC versions back to 11.4,
-  // which worked with earlier drivers.
-  // Not sure which driver had first fail and this bears checking
-  // Is awkward as must install multiple driver versions
+  // 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 >>>(lat, buffer_v, size);
+  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];
   sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
   result = *buffer_v;
 #endif

Grid/lattice/Lattice_rng.h

@@ -440,17 +440,8 @@ public:
 	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
 
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-#if 1
-	assert(rank == _grid->ThisRank() );
-#else
-// 
-	if (rank != _grid->ThisRank() ){
-	std::cout <<"rank "<<rank<<" _grid->ThisRank() "<<_grid->ThisRank()<< std::endl;
-//	exit(-42);
-//	assert(0);
-	}
-#endif
 
+	assert(rank == _grid->ThisRank() );
 	
 	int l_idx=generator_idx(o_idx,i_idx);
 	_generators[l_idx] = master_engine;

Grid/lattice/Lattice_trace.h

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

Grid/lattice/Lattice_transfer.h

@@ -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>
@@ -648,8 +697,68 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
-
   // the above should guarantee that the operations are local
+  
+#if 1
+
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
+  
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+ 
+  thread_for(idx, nsite, {
+      Coordinate from_coor, to_coor;
+      size_t rem = idx;
+      for(int i=0;i<nd;i++){
+	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i];
+	from_coor[i] = base_i + FromLowerLeft[i];
+	to_coor[i] = base_i + ToLowerLeft[i];
+      }
+      
+      int foidx = Fg->oIndex(from_coor);
+      int fiidx = Fg->iIndex(from_coor);
+      int toidx = Tg->oIndex(to_coor);
+      int tiidx = Tg->iIndex(to_coor);
+      int* tt = table + 4*idx;
+      tt[0] = foidx;
+      tt[1] = fiidx;
+      tt[2] = toidx;
+      tt[3] = tiidx;
+    });
+  
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+
+#else  
   Coordinate ldf = Fg->_ldimensions;
   Coordinate rdf = Fg->_rdimensions;
   Coordinate isf = Fg->_istride;
@@ -658,9 +767,9 @@ 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);
@@ -673,17 +782,24 @@ 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
     }
   });
+
+#endif
 }
 
 
@@ -776,6 +892,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 }
 
 
+//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
+//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -792,11 +910,70 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-  }
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
+#if 1
+  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+  
+  thread_for(idx,nsite,{
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
+    lcoor[orthog] = slice_lo;
+    hcoor[orthog] = slice_hi;
+    size_t rem = idx;
+    for(int mu=0;mu<nl;mu++){
+      if(mu != orthog){
+	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
+	lcoor[mu] = hcoor[mu] = xmu;
+      }
+    }
+    int loidx = lg->oIndex(lcoor);
+    int liidx = lg->iIndex(lcoor);
+    int hoidx = hg->oIndex(hcoor);
+    int hiidx = hg->iIndex(hcoor);
+    int* tt = table + 4*idx;
+    tt[0] = loidx;
+    tt[1] = liidx;
+    tt[2] = hoidx;
+    tt[3] = hiidx;
+    });
+   
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(lowDim_v,lowDim,AcceleratorRead);
+  autoView(higherDim_v,higherDim,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
+      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+#else
   // the above should guarantee that the operations are local
   autoView(lowDimv,lowDim,CpuRead);
   autoView(higherDimv,higherDim,CpuWrite);
@@ -812,6 +989,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       pokeLocalSite(s,higherDimv,hcoor);
     }
   });
+#endif
 }
 
 

Grid/lattice/PaddedCell.h

@@ -0,0 +1,174 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/PaddedCell.h
+
+    Copyright (C) 2019
+
+Author: Peter Boyle pboyle@bnl.gov
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include<Grid/cshift/Cshift.h>
+
+NAMESPACE_BEGIN(Grid);
+
+//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
+template<typename vobj>
+struct CshiftImplBase{
+  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
+  virtual ~CshiftImplBase(){}
+};
+template<typename vobj>
+struct CshiftImplDefault: public CshiftImplBase<vobj>{
+  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
+};
+template<typename Gimpl>
+struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
+  typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
+};  
+
+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(const Lattice<vobj> &in) const
+  {
+    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(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = Expand(d,tmp,cshift); // rvalue && assignment
+    }
+    return tmp;
+  }
+  // expand up one dim at a time
+  template<class vobj>
+  inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    GridCartesian *new_grid = grids[dim];//These are new grids
+    Lattice<vobj>  padded(new_grid);
+    Lattice<vobj> shifted(old_grid);    
+    Coordinate local     =old_grid->LocalDimensions();
+    Coordinate plocal    =new_grid->LocalDimensions();
+    if(dim==0) conformable(old_grid,unpadded_grid);
+    else       conformable(old_grid,grids[dim-1]);
+
+    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
+
+    double tins=0, tshift=0;
+    
+    // Middle bit
+    double t = usecond();
+    for(int x=0;x<local[dim];x++){
+      InsertSliceLocal(in,padded,x,depth+x,dim);
+    }
+    tins += usecond() - t;
+    
+    // High bit
+    t = usecond();
+    shifted = cshift.Cshift(in,dim,depth);
+    tshift += usecond() - t;
+
+    t=usecond();
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
+    }
+    tins += usecond() - t;
+    
+    // Low bit
+    t = usecond();
+    shifted = cshift.Cshift(in,dim,-depth);
+    tshift += usecond() - t;
+    
+    t = usecond();
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,x,x,dim);
+    }
+    tins += usecond() - t;
+
+    std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
+    
+    return padded;
+  }
+
+};
+ 
+
+NAMESPACE_END(Grid);
+

Grid/qcd/QCD.h

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

Grid/qcd/action/ActionBase.h

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

Grid/qcd/action/ActionCore.h

@@ -30,6 +30,8 @@ directory
 #ifndef QCD_ACTION_CORE
 #define QCD_ACTION_CORE
 
+#include <Grid/qcd/action/gauge/GaugeImplementations.h>
+
 #include <Grid/qcd/action/ActionBase.h>
 NAMESPACE_CHECK(ActionBase);
 #include <Grid/qcd/action/ActionSet.h>
@@ -65,6 +67,7 @@ NAMESPACE_CHECK(Scalar);
 #include <Grid/qcd/utils/Metric.h>
 NAMESPACE_CHECK(Metric);
 #include <Grid/qcd/utils/CovariantLaplacian.h>
+#include <Grid/qcd/utils/CovariantLaplacianRat.h>
 NAMESPACE_CHECK(CovariantLaplacian);
 
 

Grid/qcd/action/ActionParams.h

@@ -65,6 +65,19 @@ struct WilsonImplParams {
   }
 };
 
+struct GaugeImplParams {
+//  bool overlapCommsCompute;
+//  AcceleratorVector<Real,Nd> twist_n_2pi_L;
+  AcceleratorVector<Complex,Nd> boundary_phases;
+  GaugeImplParams()  {
+    boundary_phases.resize(Nd, 1.0);
+//      twist_n_2pi_L.resize(Nd, 0.0);
+  };
+  GaugeImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi) {
+//    twist_n_2pi_L.resize(Nd, 0.0);
+  }
+};
+
 struct StaggeredImplParams {
   Coordinate dirichlet; // Blocksize of dirichlet BCs
   int  partialDirichlet;

Grid/qcd/action/fermion/Fermion.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

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

Grid/qcd/action/fermion/WilsonImpl.h

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

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

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

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

@@ -423,14 +423,14 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
 #define KERNEL_CALL_EXT(A)						\
-  const uint64_t    NN = Nsite*Ls;					\
   const uint64_t    sz = st.surface_list.size();			\
   auto ptr = &st.surface_list[0];					\
   accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
       int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
+      int sU = sF/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
+    });									\
+  accelerator_barrier();
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
@@ -463,11 +463,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteSycl);    return; }
-#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
-#endif     
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
@@ -478,8 +474,10 @@ 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
@@ -502,21 +500,20 @@ 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/SpWilsonImplD/impl.h

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -1 +0,0 @@
-../WilsonFermion5DInstantiation.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/impl.h

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

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

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

Grid/qcd/action/gauge/Gauge.h

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

Grid/qcd/action/gauge/GaugeImplTypes.h

@@ -32,7 +32,7 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
-#define CPS_MD_TIME
+#undef CPS_MD_TIME
 
 #ifdef CPS_MD_TIME
 #define HMC_MOMENTUM_DENOMINATOR (2.0)
@@ -61,7 +61,7 @@ NAMESPACE_BEGIN(Grid);
   typedef typename Impl::Field Field;
 
 // hardcodes the exponential approximation in the template
-template <class S, int Nrepresentation = Nc, int Nexp = 12 > class GaugeImplTypes {
+template <class S, int Nrepresentation = Nc, int Nexp = 12, class Group = SU<Nc> > class GaugeImplTypes {
 public:
   typedef S Simd;
   typedef typename Simd::scalar_type scalar_type;
@@ -78,8 +78,6 @@ public:
   typedef Lattice<SiteLink>    LinkField; 
   typedef Lattice<SiteField>   Field;
 
-  typedef SU<Nrepresentation> Group;
-
   // Guido: we can probably separate the types from the HMC functions
   // this will create 2 kind of implementations
   // probably confusing the users
@@ -119,6 +117,7 @@ public:
     //
     LinkField Pmu(P.Grid());
     Pmu = Zero();
+
     for (int mu = 0; mu < Nd; mu++) {
       Group::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
       RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR) ;
@@ -126,8 +125,12 @@ public:
       PokeIndex<LorentzIndex>(P, Pmu, mu);
     }
   }
-
-  static inline Field projectForce(Field &P) { return Ta(P); }
+    
+  static inline Field projectForce(Field &P) {
+      Field ret(P.Grid());
+      Group::taProj(P, ret);
+      return ret;
+    }
 
   static inline void update_field(Field& P, Field& U, double ep){
     //static std::chrono::duration<double> diff;
@@ -137,14 +140,15 @@ public:
     autoView(P_v,P,AcceleratorRead);
     accelerator_for(ss, P.Grid()->oSites(),1,{
       for (int mu = 0; mu < Nd; mu++) {
-        U_v[ss](mu) = ProjectOnGroup(Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu));
+          U_v[ss](mu) = Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu);
+          U_v[ss](mu) = Group::ProjectOnGeneralGroup(U_v[ss](mu));
       }
     });
    //auto end = std::chrono::high_resolution_clock::now();
    // diff += end - start;
    // std::cout << "Time to exponentiate matrix " << diff.count() << " s\n";
   }
-
+    
   static inline RealD FieldSquareNorm(Field& U){
     LatticeComplex Hloc(U.Grid());
     Hloc = Zero();
@@ -157,7 +161,7 @@ public:
   }
 
   static inline void Project(Field &U) {
-    ProjectSUn(U);
+    Group::ProjectOnSpecialGroup(U);
   }
 
   static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
@@ -171,6 +175,7 @@ public:
   static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
     Group::ColdConfiguration(pRNG, U);
   }
+
 };
 
 
@@ -178,10 +183,17 @@ typedef GaugeImplTypes<vComplex, Nc> GimplTypesR;
 typedef GaugeImplTypes<vComplexF, Nc> GimplTypesF;
 typedef GaugeImplTypes<vComplexD, Nc> GimplTypesD;
 
+typedef GaugeImplTypes<vComplex, Nc, 12, Sp<Nc> > SpGimplTypesR;
+typedef GaugeImplTypes<vComplexF, Nc, 12, Sp<Nc> > SpGimplTypesF;
+typedef GaugeImplTypes<vComplexD, Nc, 12, Sp<Nc> > SpGimplTypesD;
+
 typedef GaugeImplTypes<vComplex, SU<Nc>::AdjointDimension> GimplAdjointTypesR;
 typedef GaugeImplTypes<vComplexF, SU<Nc>::AdjointDimension> GimplAdjointTypesF;
 typedef GaugeImplTypes<vComplexD, SU<Nc>::AdjointDimension> GimplAdjointTypesD;
 
+
+
+
 NAMESPACE_END(Grid);
 
 #endif // GRID_GAUGE_IMPL_TYPES_H

Grid/qcd/action/gauge/GaugeImplementations.h

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

Grid/qcd/action/gauge/PlaqPlusRectangleAction.h

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

Grid/qcd/action/gauge/WilsonGaugeAction.h

@@ -42,9 +42,13 @@ template <class Gimpl>
 class WilsonGaugeAction : public Action<typename Gimpl::GaugeField> {
 public:  
   INHERIT_GIMPL_TYPES(Gimpl);
+  typedef GaugeImplParams ImplParams;
+  ImplParams Params;
 
   /////////////////////////// constructors
-  explicit WilsonGaugeAction(RealD beta_):beta(beta_){};
+  explicit WilsonGaugeAction(RealD beta_,
+		  const ImplParams &p = ImplParams()
+		  ):beta(beta_),Params(p){};
 
   virtual std::string action_name() {return "WilsonGaugeAction";}
 
@@ -56,14 +60,53 @@ public:
 
   virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG){};  // noop as no pseudoferms
 
+// Umu<->U maximally confusing
+  virtual void boundary(const GaugeField &Umu, GaugeField &Ub){
+    typedef typename Simd::scalar_type scalar_type;
+    assert(Params.boundary_phases.size() == Nd);
+    GridBase *GaugeGrid=Umu.Grid();
+    GaugeLinkField U(GaugeGrid);
+    GaugeLinkField tmp(GaugeGrid);
+
+    Lattice<iScalar<vInteger> > coor(GaugeGrid);
+    for (int mu = 0; mu < Nd; mu++) {
+	////////// boundary phase /////////////
+      auto pha = Params.boundary_phases[mu];
+      scalar_type phase( real(pha),imag(pha) );
+      std::cout<< GridLogIterative << "[WilsonGaugeAction] boundary "<<mu<<" "<<phase<< std::endl; 
+
+	int L   = GaugeGrid->GlobalDimensions()[mu];
+        int Lmu = L - 1;
+
+      LatticeCoordinate(coor, mu);
+
+      U = PeekIndex<LorentzIndex>(Umu, mu);
+      tmp = where(coor == Lmu, phase * U, U);
+      PokeIndex<LorentzIndex>(Ub, tmp, mu);
+//      PokeIndex<LorentzIndex>(Ub, U, mu);
+//      PokeIndex<LorentzIndex>(Umu, tmp, mu);
+
+    }
+  };
+
   virtual RealD S(const GaugeField &U) {
-    RealD plaq = WilsonLoops<Gimpl>::avgPlaquette(U);
-    RealD vol = U.Grid()->gSites();
+    GaugeField Ub(U.Grid());
+    this->boundary(U,Ub);
+    static RealD lastG=0.;
+    RealD plaq = WilsonLoops<Gimpl>::avgPlaquette(Ub);
+    RealD vol = Ub.Grid()->gSites();
     RealD action = beta * (1.0 - plaq) * (Nd * (Nd - 1.0)) * vol * 0.5;
+    std::cout << GridLogMessage << "[WilsonGaugeAction] dH: " << action-lastG << std::endl;
+    RealD plaq_o = WilsonLoops<Gimpl>::avgPlaquette(U);
+    RealD action_o = beta * (1.0 - plaq_o) * (Nd * (Nd - 1.0)) * vol * 0.5;
+    std::cout << GridLogMessage << "[WilsonGaugeAction] U: " << action_o <<" Ub: "<< action  << std::endl;
+    lastG=action;
     return action;
   };
 
   virtual void deriv(const GaugeField &U, GaugeField &dSdU) {
+    GaugeField Ub(U.Grid());
+    this->boundary(U,Ub);
     // not optimal implementation FIXME
     // extend Ta to include Lorentz indexes
 
@@ -73,10 +116,9 @@ public:
     GaugeLinkField dSdU_mu(U.Grid());
     for (int mu = 0; mu < Nd; mu++) {
 
-      Umu = PeekIndex<LorentzIndex>(U, mu);
-      
+      Umu = PeekIndex<LorentzIndex>(Ub, mu);
       // Staple in direction mu
-      WilsonLoops<Gimpl>::Staple(dSdU_mu, U, mu);
+      WilsonLoops<Gimpl>::Staple(dSdU_mu, Ub, mu);
       dSdU_mu = Ta(Umu * dSdU_mu) * factor;
       
       PokeIndex<LorentzIndex>(dSdU, dSdU_mu, mu);

Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h

@@ -178,7 +178,10 @@ NAMESPACE_BEGIN(Grid);
         // Use chronological inverter to forecast solutions across poles
         std::vector<FermionField> prev_solns;
         if(use_heatbath_forecasting){ prev_solns.reserve(param.degree); }
-        ChronoForecast<AbstractEOFAFermion<Impl>, FermionField> Forecast;
+	MdagMLinearOperator<AbstractEOFAFermion<Impl> ,FermionField> MdagML(Lop);
+	MdagMLinearOperator<AbstractEOFAFermion<Impl> ,FermionField> MdagMR(Rop);
+//        ChronoForecast<AbstractEOFAFermion<Impl>, FermionField> Forecast;
+	ChronoForecast<MdagMLinearOperator<AbstractEOFAFermion<Impl>, FermionField> , FermionField> Forecast;
 
         // \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta
         RealD N(PowerNegHalf.norm);
@@ -198,7 +201,7 @@ NAMESPACE_BEGIN(Grid);
           heatbathRefreshShiftCoefficients(0, -gamma_l);
           if(use_heatbath_forecasting){ // Forecast CG guess using solutions from previous poles
             Lop.Mdag(CG_src, Forecast_src);
-            CG_soln = Forecast(Lop, Forecast_src, prev_solns);
+            CG_soln = Forecast(MdagML, Forecast_src, prev_solns);
             SolverHBL(Lop, CG_src, CG_soln);
             prev_solns.push_back(CG_soln);
           } else {
@@ -225,7 +228,7 @@ NAMESPACE_BEGIN(Grid);
 	  heatbathRefreshShiftCoefficients(1, -gamma_l*PowerNegHalf.poles[k]);
           if(use_heatbath_forecasting){
             Rop.Mdag(CG_src, Forecast_src);
-            CG_soln = Forecast(Rop, Forecast_src, prev_solns);
+            CG_soln = Forecast(MdagMR, Forecast_src, prev_solns);
             SolverHBR(Rop, CG_src, CG_soln);
             prev_solns.push_back(CG_soln);
           } else {

Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h

@@ -38,91 +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
+#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());
+	FermionFieldD inD(NumOpD.FermionRedBlackGrid());
+	FermionFieldD outD(NumOpD.FermionRedBlackGrid());
 
 	// Action better with higher precision?
-	ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
-	precisionChange(inD2,in);
-	std::cout << "msCG single solve "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
-	msCG(schurOpD2, inD2, outD2);
-	precisionChange(out,outD2);
+	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());
-	ConjugateGradientMultiShiftMixedPrecCleanup<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
-	precisionChange(inD2,in);
-	std::cout << "msCG in "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
-	msCG(schurOpD2, inD2, out_elemsD2, outD2);
-	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/action/scalar/ScalarImpl.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#define CPS_MD_TIME 
+#undef CPS_MD_TIME 
 
 #ifdef CPS_MD_TIME
 #define HMC_MOMENTUM_DENOMINATOR (2.0)

Grid/qcd/hmc/GenericHMCrunner.h

@@ -121,12 +121,19 @@ public:
 
   template <class SmearingPolicy>
   void Run(SmearingPolicy &S) {
-    Runner(S);
+    TrivialMetric<typename Implementation::Field> Mtr;
+    Runner(S,Mtr);
+  }
+
+  template <class SmearingPolicy, class Metric>
+  void Run(SmearingPolicy &S, Metric &Mtr) {
+    Runner(S,Mtr);
   }
 
   void Run(){
     NoSmearing<Implementation> S;
-    Runner(S);
+    TrivialMetric<typename Implementation::Field> Mtr;
+    Runner(S,Mtr);
   }
 
   //Use the checkpointer to initialize the RNGs and the gauge field, writing the resulting gauge field into U.
@@ -176,15 +183,15 @@ public:
   //////////////////////////////////////////////////////////////////
 
 private:
-  template <class SmearingPolicy>
-  void Runner(SmearingPolicy &Smearing) {
+  template <class SmearingPolicy, class Metric>
+  void Runner(SmearingPolicy &Smearing, Metric &Mtr) {
     auto UGrid = Resources.GetCartesian();
     Field U(UGrid);
 
     initializeGaugeFieldAndRNGs(U);
 
     typedef IntegratorType<SmearingPolicy> TheIntegrator;
-    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
+    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing,Mtr);
 
     // Sets the momentum filter
     MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
@@ -225,6 +232,18 @@ template <class RepresentationsPolicy,
 using GenericHMCRunnerHirep =
 				     HMCWrapperTemplate<PeriodicGimplR, Integrator, RepresentationsPolicy>;
 
+// sp2n
+
+template <template <typename, typename, typename> class Integrator>
+using GenericSpHMCRunner = HMCWrapperTemplate<SpPeriodicGimplR, Integrator>;
+
+template <class RepresentationsPolicy,
+          template <typename, typename, typename> class Integrator>
+using GenericSpHMCRunnerHirep =
+                     HMCWrapperTemplate<SpPeriodicGimplR, Integrator, RepresentationsPolicy>;
+
+
+
 template <class Implementation, class RepresentationsPolicy, 
           template <typename, typename, typename> class Integrator>
 using GenericHMCRunnerTemplate = HMCWrapperTemplate<Implementation, Integrator, RepresentationsPolicy>;

Grid/qcd/hmc/HMC.h

@@ -55,6 +55,8 @@ struct HMCparameters: Serializable {
                                   Integer, NoMetropolisUntil,
 				  bool, PerformRandomShift, /* @brief Randomly shift the gauge configuration at the start of a trajectory */
                                   std::string, StartingType,
+				  Integer, SW,
+                                  RealD, Kappa,
                                   IntegratorParameters, MD)
 
   HMCparameters() {
@@ -110,6 +112,8 @@ private:
   IntegratorType &TheIntegrator;
   ObsListType Observables;
 
+  int traj_num;
+
   /////////////////////////////////////////////////////////
   // Metropolis step
   /////////////////////////////////////////////////////////
@@ -200,14 +204,14 @@ private:
 
     std::cout << GridLogMessage << "--------------------------------------------------\n";
     std::cout << GridLogMessage << " Molecular Dynamics evolution ";
-    TheIntegrator.integrate(U);
+    TheIntegrator.integrate(U,traj_num);
     std::cout << GridLogMessage << "--------------------------------------------------\n";
 
     //////////////////////////////////////////////////////////////////////////////////////////////////////
     // updated state action
     //////////////////////////////////////////////////////////////////////////////////////////////////////
     std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Compute final action";
+    std::cout << GridLogMessage << "Compute final action" <<std::endl;
     RealD H1 = TheIntegrator.S(U);  
     std::cout << GridLogMessage << "--------------------------------------------------\n";
 
@@ -242,7 +246,7 @@ public:
   HybridMonteCarlo(HMCparameters _Pams, IntegratorType &_Int,
                    GridSerialRNG &_sRNG, GridParallelRNG &_pRNG, 
                    ObsListType _Obs, Field &_U)
-    : Params(_Pams), TheIntegrator(_Int), sRNG(_sRNG), pRNG(_pRNG), Observables(_Obs), Ucur(_U) {}
+    : Params(_Pams), TheIntegrator(_Int), sRNG(_sRNG), pRNG(_pRNG), Observables(_Obs), Ucur(_U),traj_num(0) {}
   ~HybridMonteCarlo(){};
 
   void evolve(void) {
@@ -257,9 +261,10 @@ public:
     unsigned int FinalTrajectory = Params.Trajectories + Params.NoMetropolisUntil + Params.StartTrajectory;
 
     for (int traj = Params.StartTrajectory; traj < FinalTrajectory; ++traj) {
+    
 
       std::cout << GridLogHMC << "-- # Trajectory = " << traj << "\n";
-
+      traj_num=traj;
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
       	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
       }
@@ -283,12 +288,13 @@ 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

@@ -9,6 +9,7 @@ Copyright (C) 2015
 Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: Guido Cossu <cossu@post.kek.jp>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -33,6 +34,7 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
+#include <Grid/parallelIO/NerscIO.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -41,10 +43,19 @@ public:
   GRID_SERIALIZABLE_CLASS_MEMBERS(IntegratorParameters,
 				  std::string, name,      // name of the integrator
 				  unsigned int, MDsteps,  // number of outer steps
+				  RealD, RMHMCTol,
+                                  RealD, RMHMCCGTol,
+                                  RealD, lambda0,
+                                  RealD, lambda1,
+                                  RealD, lambda2,
 				  RealD, trajL)           // trajectory length
 
   IntegratorParameters(int MDsteps_ = 10, RealD trajL_ = 1.0)
   : MDsteps(MDsteps_),
+   lambda0(0.1931833275037836),
+   lambda1(0.1931833275037836),
+   lambda2(0.1931833275037836),
+   RMHMCTol(1e-8),RMHMCCGTol(1e-8),
     trajL(trajL_) {};
 
   template <class ReaderClass, typename std::enable_if<isReader<ReaderClass>::value, int >::type = 0 >
@@ -66,6 +77,7 @@ 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;
@@ -74,11 +86,14 @@ protected:
   double t_U;  // Track time passing on each level and for U and for P
   std::vector<double> t_P;  
 
-  MomentaField P;
+//  MomentaField P;
+  GeneralisedMomenta<FieldImplementation > P;
   SmearingPolicy& Smearer;
   RepresentationPolicy Representations;
   IntegratorParameters Params;
 
+  RealD Saux,Smom,Sg;
+
   //Filters allow the user to manipulate the conjugate momentum, for example to freeze links in DDHMC
   //It is applied whenever the momentum is updated / refreshed
   //The default filter does nothing
@@ -95,7 +110,15 @@ protected:
   void update_P(Field& U, int level, double ep) 
   {
     t_P[level] += ep;
-    update_P(P, U, level, ep);
+    update_P(P.Mom, U, level, ep);
+
+    std::cout << GridLogIntegrator << "[" << level << "] P " << " dt " << ep << " : t_P " << t_P[level] << std::endl;
+  }
+
+  void update_P2(Field& U, int level, double ep) 
+  {
+    t_P[level] += ep;
+    update_P2(P.Mom, U, level, ep);
 
     std::cout << GridLogIntegrator << "[" << level << "] P " << " dt " << ep << " : t_P " << t_P[level] << std::endl;
   }
@@ -119,72 +142,174 @@ protected:
     }
   } update_P_hireps{};
 
- 
   void update_P(MomentaField& Mom, Field& U, int level, double ep) {
     // input U actually not used in the fundamental case
     // Fundamental updates, include smearing
 
     for (int a = 0; a < as[level].actions.size(); ++a) {
-
       double start_full = usecond();
       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_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;    
-
-      Real force_max   = std::sqrt(maxLocalNorm2(force));
-      Real impulse_max = force_max * ep * HMC_MOMENTUM_DENOMINATOR;    
-
-      as[level].actions.at(a)->deriv_log(force_abs,force_max,impulse_abs,impulse_max);
-      
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] dt           : " << ep <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force average: " << force_abs <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force max    : " << force_max <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt average  : " << impulse_abs <<" "<<name<<std::endl;
-      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt max      : " << impulse_max <<" "<<name<<std::endl;
-
+      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
+      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << std::endl;
       Mom -= force * ep* HMC_MOMENTUM_DENOMINATOR;; 
       double end_full = usecond();
       double time_full  = (end_full - start_full) / 1e3;
       double time_force = (end_force - start_force) / 1e3;
       std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)"  << std::endl;
-
     }
 
     // Force from the other representations
     as[level].apply(update_P_hireps, Representations, Mom, U, ep);
+  }
 
+  void update_P2(MomentaField& Mom, Field& U, int level, double ep) {
+    // input U actually not used in the fundamental case
+    // Fundamental updates, include smearing
+
+    std::cout << GridLogIntegrator << "U before update_P2: " << std::sqrt(norm2(U)) << std::endl;
+    // Generalised momenta  
+    // Derivative of the kinetic term must be computed before
+    // Mom is the momenta and gets updated by the 
+    // actions derivatives
+    MomentaField MomDer(P.Mom.Grid());
+    P.M.ImportGauge(U);
+    P.DerivativeU(P.Mom, MomDer);
+    std::cout << GridLogIntegrator << "MomDer update_P2: " << std::sqrt(norm2(MomDer)) << std::endl;
+//    Mom -= MomDer * ep;
+    Mom -= MomDer * ep * HMC_MOMENTUM_DENOMINATOR;
+    std::cout << GridLogIntegrator << "Mom update_P2: " << std::sqrt(norm2(Mom)) << std::endl;
+
+    // Auxiliary fields
+    P.update_auxiliary_momenta(ep*0.5 );
+    P.AuxiliaryFieldsDerivative(MomDer);
+    std::cout << GridLogIntegrator << "MomDer(Aux) update_P2: " << std::sqrt(norm2(Mom)) << std::endl;
+//    Mom -= MomDer * ep;
+    Mom -= MomDer * ep * HMC_MOMENTUM_DENOMINATOR;
+    P.update_auxiliary_momenta(ep*0.5 );
+
+    for (int a = 0; a < as[level].actions.size(); ++a) {
+      double start_full = usecond();
+      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();
+      as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
+
+      std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
+      if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
+      force = FieldImplementation::projectForce(force); // Ta for gauge fields
+      double end_force = usecond();
+      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
+      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << std::endl;
+      Mom -= force * ep* HMC_MOMENTUM_DENOMINATOR;; 
+      double end_full = usecond();
+      double time_full  = (end_full - start_full) / 1e3;
+      double time_force = (end_force - start_force) / 1e3;
+      std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)"  << std::endl;
+    }
+
+    // Force from the other representations
+    as[level].apply(update_P_hireps, Representations, Mom, U, ep);
+  }
+
+  void implicit_update_P(Field& U, int level, double ep, double ep1, bool intermediate = false) {
+    t_P[level] += ep;
+
+    double ep2= ep-ep1;
+
+    std::cout << GridLogIntegrator << "[" << level << "] P "
+              << " dt " << ep << " : t_P " << t_P[level] << std::endl;
+    std::cout << GridLogIntegrator << "U before implicit_update_P: " << std::sqrt(norm2(U)) << std::endl;
+    // Fundamental updates, include smearing
+    MomentaField Msum(P.Mom.Grid());
+    Msum = Zero();
+    for (int a = 0; a < as[level].actions.size(); ++a) {
+      // Compute the force terms for the lagrangian part
+      // We need to compute the derivative of the actions
+      // only once
+      Field force(U.Grid());
+      conformable(U.Grid(), P.Mom.Grid());
+      Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
+      as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
+
+      std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
+      if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
+      force = FieldImplementation::projectForce(force);  // Ta for gauge fields
+      Real force_abs = std::sqrt(norm2(force) / U.Grid()->gSites());
+      std::cout << GridLogIntegrator << "|Force| site average: " << force_abs
+                << std::endl;
+      Msum += force;
+    }
+
+    MomentaField NewMom = P.Mom;
+    MomentaField OldMom = P.Mom;
+    double threshold = Params.RMHMCTol;
+    P.M.ImportGauge(U);
+    MomentaField MomDer(P.Mom.Grid());
+    MomentaField MomDer1(P.Mom.Grid());
+    MomentaField AuxDer(P.Mom.Grid());
+    MomDer1 = Zero();
+    MomentaField diff(P.Mom.Grid());
+    double factor = 2.0;
+    if (intermediate){
+      P.DerivativeU(P.Mom, MomDer1);
+      factor = 1.0;
+    }
+//    std::cout << GridLogIntegrator << "MomDer1 implicit_update_P: " << std::sqrt(norm2(MomDer1)) << std::endl;
+
+    // Auxiliary fields
+    P.update_auxiliary_momenta(ep1);
+    P.AuxiliaryFieldsDerivative(AuxDer);
+    Msum += AuxDer;
+    
+
+    // Here run recursively
+    int counter = 1;
+    RealD RelativeError;
+    do {
+      std::cout << GridLogIntegrator << "UpdateP implicit step "<< counter << std::endl;
+
+      // Compute the derivative of the kinetic term
+      // with respect to the gauge field
+      P.DerivativeU(NewMom, MomDer);
+      Real force_abs = std::sqrt(norm2(MomDer) / U.Grid()->gSites());
+      std::cout << GridLogIntegrator << "|Force| laplacian site average: " << force_abs
+                << std::endl;
+
+//      NewMom = P.Mom - ep* 0.5 * HMC_MOMENTUM_DENOMINATOR * (2.0*Msum + factor*MomDer + MomDer1);// simplify
+      NewMom = P.Mom -  HMC_MOMENTUM_DENOMINATOR * (ep*Msum + ep1* factor*MomDer + ep2* MomDer1);// simplify
+      diff = NewMom - OldMom;
+      counter++;
+      RelativeError = std::sqrt(norm2(diff))/std::sqrt(norm2(NewMom));
+      std::cout << GridLogIntegrator << "UpdateP RelativeError: " << RelativeError << std::endl;
+      OldMom = NewMom;
+    } while (RelativeError > threshold);
+
+    P.Mom = NewMom;
+    std::cout << GridLogIntegrator << "NewMom implicit_update_P: " << std::sqrt(norm2(NewMom)) << std::endl;
+
+    // update the auxiliary fields momenta    
+    P.update_auxiliary_momenta(ep2);
+  }
+
+  void implicit_update_P(Field& U, int level, double ep, bool intermediate = false) {
+      implicit_update_P( U, level, ep, ep*0.5, intermediate ); 
   }
 
   void update_U(Field& U, double ep) 
   {
-    update_U(P, U, ep);
+    update_U(P.Mom, U, ep);
 
     t_U += ep;
     int fl = levels - 1;
@@ -193,12 +318,8 @@ protected:
   
   void update_U(MomentaField& Mom, Field& U, double ep) 
   {
-    MomentaField MomFiltered(Mom.Grid());
-    MomFiltered = Mom;
-    MomFilter->applyFilter(MomFiltered);
-
     // exponential of Mom*U in the gauge fields case
-    FieldImplementation::update_field(MomFiltered, U, ep);
+    FieldImplementation::update_field(Mom, U, ep);
 
     // Update the smeared fields, can be implemented as observer
     Smearer.set_Field(U);
@@ -207,18 +328,74 @@ protected:
     Representations.update(U);  // void functions if fundamental representation
   }
 
+  void implicit_update_U(Field&U, double ep, double ep1 ){
+    double ep2=ep-ep1;
+    t_U += ep;
+    int fl = levels - 1;
+    std::cout << GridLogIntegrator << "   " << "[" << fl << "] U " << " dt " << ep << " : t_U " << t_U << std::endl;
+    std::cout << GridLogIntegrator << "U before implicit_update_U: " << std::sqrt(norm2(U)) << std::endl;
+
+    MomentaField Mom1(P.Mom.Grid());
+    MomentaField Mom2(P.Mom.Grid());
+    RealD RelativeError;
+    Field diff(U.Grid());
+    Real threshold =  Params.RMHMCTol;
+    int counter = 1;
+    int MaxCounter = 100;
+
+    Field OldU = U;
+    Field NewU = U;
+
+    P.M.ImportGauge(U);
+    P.DerivativeP(Mom1); // first term in the derivative 
+    std::cout << GridLogIntegrator << "implicit_update_U: Mom1: " << std::sqrt(norm2(Mom1)) << std::endl;
+
+    P.update_auxiliary_fields(ep1);
+
+
+    MomentaField sum=Mom1;
+    do {
+      std::cout << GridLogIntegrator << "UpdateU implicit step "<< counter << std::endl;
+      
+      P.DerivativeP(Mom2); // second term in the derivative, on the updated U
+      std::cout << GridLogIntegrator << "implicit_update_U: Mom1: " << std::sqrt(norm2(Mom1)) << std::endl;
+      sum = (Mom1*ep1 + Mom2*ep2);
+
+      for (int mu = 0; mu < Nd; mu++) {
+        auto Umu = PeekIndex<LorentzIndex>(U, mu);
+        auto Pmu = PeekIndex<LorentzIndex>(sum, mu);
+        Umu = expMat(Pmu, 1, 12) * Umu;
+        PokeIndex<LorentzIndex>(NewU, ProjectOnGroup(Umu), mu);
+      }
+
+      diff = NewU - OldU;
+      RelativeError = std::sqrt(norm2(diff))/std::sqrt(norm2(NewU));
+      std::cout << GridLogIntegrator << "UpdateU RelativeError: " << RelativeError << std::endl;
+      
+      P.M.ImportGauge(NewU);
+      OldU = NewU; // some redundancy to be eliminated
+      counter++;
+    } while (RelativeError > threshold && counter < MaxCounter);
+
+    U = NewU;
+    std::cout << GridLogIntegrator << "NewU implicit_update_U: " << std::sqrt(norm2(U)) << std::endl;
+    P.update_auxiliary_fields(ep2);
+  }
+
+
   virtual void step(Field& U, int level, int first, int last) = 0;
 
 public:
   Integrator(GridBase* grid, IntegratorParameters Par,
              ActionSet<Field, RepresentationPolicy>& Aset,
-             SmearingPolicy& Sm)
+             SmearingPolicy& Sm, Metric<MomentaField>& M)
     : Params(Par),
       as(Aset),
-      P(grid),
+      P(grid, M),
       levels(Aset.size()),
       Smearer(Sm),
-      Representations(grid) 
+      Representations(grid),
+      Saux(0.),Smom(0.),Sg(0.)
   {
     t_P.resize(levels, 0.0);
     t_U = 0.0;
@@ -334,7 +511,8 @@ public:
 
   void reverse_momenta()
   {
-    P *= -1.0;
+    P.Mom *= -1.0;
+    P.AuxMom *= -1.0;
   }
 
   // to be used by the actionlevel class to iterate
@@ -353,11 +531,14 @@ public:
   // Initialization of momenta and actions
   void refresh(Field& U,  GridSerialRNG & sRNG, GridParallelRNG& pRNG) 
   {
-    assert(P.Grid() == U.Grid());
+    assert(P.Mom.Grid() == U.Grid());
     std::cout << GridLogIntegrator << "Integrator refresh" << std::endl;
 
     std::cout << GridLogIntegrator << "Generating momentum" << std::endl;
-    FieldImplementation::generate_momenta(P, sRNG, pRNG);
+//    FieldImplementation::generate_momenta(P.Mom, sRNG, pRNG);
+    P.M.ImportGauge(U);
+    P.MomentaDistribution(sRNG,pRNG);
+
 
     // Update the smeared fields, can be implemented as observer
     // necessary to keep the fields updated even after a reject
@@ -379,15 +560,10 @@ 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
@@ -417,24 +593,36 @@ public:
 
     std::cout << GridLogIntegrator << "Integrator action\n";
 
-    RealD H = - FieldImplementation::FieldSquareNorm(P)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
+//    RealD H = - FieldImplementation::FieldSquareNorm(P.Mom)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
+//    RealD Hterm;
+
+//    static RealD Saux=0.,Smom=0.,Sg=0.;
+
+    RealD H = - FieldImplementation::FieldSquareNorm(P.Mom)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
+    std::cout << GridLogMessage << "S:FieldSquareNorm H_p = " << H << "\n";
+    std::cout << GridLogMessage << "S:dSField = " << H-Smom << "\n";
+    Smom=H;
+    P.M.ImportGauge(U);
+    RealD Hterm = - P.MomentaAction();
+    std::cout << GridLogMessage << "S:Momentum action H_p = " << Hterm << "\n";
+    std::cout << GridLogMessage << "S:dSMom = " << Hterm-Saux << "\n";
+    Saux=Hterm;
+    H = Hterm;
 
-    RealD Hterm;
 
     // 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);
     }
@@ -447,9 +635,9 @@ 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;
 
@@ -462,22 +650,30 @@ public:
 
     std::cout << GridLogIntegrator << "Integrator initial action\n";
 
-    RealD H = - FieldImplementation::FieldSquareNorm(P)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
-
-    RealD Hterm;
+//    RealD H = - FieldImplementation::FieldSquareNorm(P.Mom)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
+//    RealD Hterm;
+    RealD H = - FieldImplementation::FieldSquareNorm(P.Mom)/HMC_MOMENTUM_DENOMINATOR; // - trace (P*P)/denom
+    std::cout << GridLogMessage << "S:FieldSquareNorm H_p = " << H << "\n";
+    std::cout << GridLogMessage << "S:dSField = " << H-Smom << "\n";
+    Smom=H;
+    P.M.ImportGauge(U);
+    RealD Hterm = - P.MomentaAction();
+    std::cout << GridLogMessage << "S:Momentum action H_p = " << Hterm << "\n";
+    std::cout << GridLogMessage << "S:dSMom = " << Hterm-Saux << "\n";
+    Saux=Hterm;
+    H = Hterm;
 
     // Actions
     for (int level = 0; level < as.size(); ++level) {
       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;
       }
@@ -488,9 +684,8 @@ public:
   }
 
   
-  void integrate(Field& U) 
+  void integrate(Field& U, int traj=-1 ) 
   {
-    AUDIT();
     // reset the clocks
     t_U = 0;
     for (int level = 0; level < as.size(); ++level) {
@@ -501,6 +696,12 @@ public:
       int first_step = (stp == 0);
       int last_step = (stp == Params.MDsteps - 1);
       this->step(U, 0, first_step, last_step);
+      if (traj>=0){
+        std::string file("./config."+std::to_string(traj)+"_"+std::to_string(stp+1) );
+        int precision32 = 0;
+        int tworow      = 0;
+        NerscIO::writeConfiguration(U,file,tworow,precision32);
+      }
     }
 
     // Check the clocks all match on all levels
@@ -508,11 +709,8 @@ 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/hmc/integrators/Integrator_algorithm.h

@@ -102,8 +102,8 @@ public:
 
   std::string integrator_name(){return "LeapFrog";}
 
-  LeapFrog(GridBase* grid, IntegratorParameters Par, ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm)
-    : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(grid, Par, Aset, Sm){};
+  LeapFrog(GridBase* grid, IntegratorParameters Par, ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm, Metric<Field>& M)
+    : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(grid, Par, Aset, Sm,M){};
 
   void step(Field& U, int level, int _first, int _last) {
     int fl = this->as.size() - 1;
@@ -140,14 +140,14 @@ template <class FieldImplementation_, class SmearingPolicy, class Representation
 class MinimumNorm2 : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
 {
 private:
-  const RealD lambda = 0.1931833275037836;
+//  const RealD lambda = 0.1931833275037836;
 
 public:
   typedef FieldImplementation_ FieldImplementation;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
-  MinimumNorm2(GridBase* grid, IntegratorParameters Par, ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm)
-    : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(grid, Par, Aset, Sm){};
+  MinimumNorm2(GridBase* grid, IntegratorParameters Par, ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm, Metric<Field>& M)
+    : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(grid, Par, Aset, Sm,M){};
 
   std::string integrator_name(){return "MininumNorm2";}
 
@@ -155,6 +155,11 @@ public:
     // level  : current level
     // fl     : final level
     // eps    : current step size
+    assert(level<3);
+    RealD lambda= this->Params.lambda0;
+    if (level>0) lambda= this->Params.lambda1;
+    if (level>1) lambda= this->Params.lambda2;
+    std::cout << GridLogMessage << "level: "<<level<< "lambda: "<<lambda<<std::endl;
 
     int fl = this->as.size() - 1;
 
@@ -210,9 +215,9 @@ public:
   // Looks like dH scales as dt^4. tested wilson/wilson 2 level.
   ForceGradient(GridBase* grid, IntegratorParameters Par,
                 ActionSet<Field, RepresentationPolicy>& Aset,
-                SmearingPolicy& Sm)
+                SmearingPolicy& Sm, Metric<Field>& M)
     : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(
-									    grid, Par, Aset, Sm){};
+									    grid, Par, Aset, Sm,M){};
 
   std::string integrator_name(){return "ForceGradient";}
   
@@ -275,6 +280,255 @@ public:
   }
 };
 
+////////////////////////////////
+// Riemannian Manifold HMC
+// Girolami et al
+////////////////////////////////
+
+
+
+// correct
+template <class FieldImplementation, class SmearingPolicy,
+          class RepresentationPolicy =
+              Representations<FundamentalRepresentation> >
+class ImplicitLeapFrog : public Integrator<FieldImplementation, SmearingPolicy,
+                                           RepresentationPolicy> {
+ public:
+  typedef ImplicitLeapFrog<FieldImplementation, SmearingPolicy, RepresentationPolicy>
+      Algorithm;
+  INHERIT_FIELD_TYPES(FieldImplementation);
+
+  // Riemannian manifold metric operator
+  // Hermitian operator Fisher
+
+  std::string integrator_name(){return "ImplicitLeapFrog";}
+
+  ImplicitLeapFrog(GridBase* grid, IntegratorParameters Par,
+           ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm, Metric<Field>& M)
+      : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(
+            grid, Par, Aset, Sm, M){};
+
+  void step(Field& U, int level, int _first, int _last) {
+    int fl = this->as.size() - 1;
+    // level  : current level
+    // fl     : final level
+    // eps    : current step size
+
+    // Get current level step size
+    RealD eps = this->Params.trajL/this->Params.MDsteps;
+    for (int l = 0; l <= level; ++l) eps /= this->as[l].multiplier;
+
+    int multiplier = this->as[level].multiplier;
+    for (int e = 0; e < multiplier; ++e) {
+      int first_step = _first && (e == 0);
+      int last_step = _last && (e == multiplier - 1);
+
+      if (first_step) {  // initial half step
+       this->implicit_update_P(U, level, eps / 2.0);
+      }
+
+      if (level == fl) {  // lowest level
+        this->implicit_update_U(U, eps,eps/2.);
+      } else {  // recursive function call
+        this->step(U, level + 1, first_step, last_step);
+      }
+
+      //int mm = last_step ? 1 : 2;
+      if (last_step){
+        this->update_P2(U, level, eps / 2.0);
+      } else {
+      this->implicit_update_P(U, level, eps, true);// works intermediate step
+      }
+    }
+  }
+};
+
+
+template <class FieldImplementation, class SmearingPolicy,
+          class RepresentationPolicy =
+              Representations<FundamentalRepresentation> >
+class ImplicitMinimumNorm2 : public Integrator<FieldImplementation, SmearingPolicy,
+                                       RepresentationPolicy> {
+ private:
+//  const RealD lambda = 0.1931833275037836;
+
+ public:
+  INHERIT_FIELD_TYPES(FieldImplementation);
+
+  ImplicitMinimumNorm2(GridBase* grid, IntegratorParameters Par,
+               ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm, Metric<Field>& M)
+      : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(
+            grid, Par, Aset, Sm, M){};
+
+  std::string integrator_name(){return "ImplicitMininumNorm2";}
+
+  void step(Field& U, int level, int _first, int _last) {
+    // level  : current level
+    // fl     : final level
+    // eps    : current step size
+
+    int fl = this->as.size() - 1;
+//    assert(Params.lambda.size()>level);
+//    RealD lambda= Params.lambda[level];
+    assert(level<3);
+    RealD lambda= this->Params.lambda0;
+    if (level>0) lambda= this->Params.lambda1;
+    if (level>1) lambda= this->Params.lambda2;
+    std::cout << GridLogMessage << "level: "<<level<< "lambda: "<<lambda<<std::endl;
+
+  if(level<fl){
+
+    RealD eps = this->Params.trajL/this->Params.MDsteps * 2.0;
+    for (int l = 0; l <= level; ++l) eps /= 2.0 * this->as[l].multiplier;
+
+    // Nesting:  2xupdate_U of size eps/2
+    // Next level is eps/2/multiplier
+
+    int multiplier = this->as[level].multiplier;
+    for (int e = 0; e < multiplier; ++e) {  // steps per step
+
+      int first_step = _first && (e == 0);
+      int last_step = _last && (e == multiplier - 1);
+
+      if (first_step) {  // initial half step
+        this->update_P(U, level, lambda * eps);
+      }
+
+        this->step(U, level + 1, first_step, 0);
+
+      this->update_P(U, level, (1.0 - 2.0 * lambda) * eps);
+
+        this->step(U, level + 1, 0, last_step);
+
+      int mm = (last_step) ? 1 : 2;
+      this->update_P(U, level, lambda * eps * mm);
+    }
+  } 
+  else 
+  { // last level
+    RealD eps = this->Params.trajL/this->Params.MDsteps * 2.0;
+    for (int l = 0; l <= level; ++l) eps /= 2.0 * this->as[l].multiplier;
+
+    // Nesting:  2xupdate_U of size eps/2
+    // Next level is eps/2/multiplier
+
+    int multiplier = this->as[level].multiplier;
+    for (int e = 0; e < multiplier; ++e) {  // steps per step
+
+      int first_step = _first && (e == 0);
+      int last_step = _last && (e == multiplier - 1);
+
+      if (first_step) {  // initial half step
+        this->implicit_update_P(U, level, lambda * eps);
+      }
+
+      this->implicit_update_U(U, 0.5 * eps,lambda*eps);
+
+      this->implicit_update_P(U, level, (1.0 - 2.0 * lambda) * eps, true);
+
+      this->implicit_update_U(U, 0.5 * eps, (0.5-lambda)*eps);
+
+      if (last_step) {
+        this->update_P2(U, level, eps * lambda);
+      } else {
+        this->implicit_update_P(U, level, lambda * eps*2.0, true);
+      }
+    }
+  }
+
+  }
+};
+
+template <class FieldImplementation, class SmearingPolicy,
+          class RepresentationPolicy =
+              Representations<FundamentalRepresentation> >
+class ImplicitCampostrini : public Integrator<FieldImplementation, SmearingPolicy,
+                                       RepresentationPolicy> {
+ private:
+//  const RealD lambda = 0.1931833275037836;
+
+ public:
+  INHERIT_FIELD_TYPES(FieldImplementation);
+
+  ImplicitCampostrini(GridBase* grid, IntegratorParameters Par,
+               ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm, Metric<Field>& M)
+      : Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy>(
+            grid, Par, Aset, Sm, M){};
+
+  std::string integrator_name(){return "ImplicitCampostrini";}
+
+  void step(Field& U, int level, int _first, int _last) {
+    // level  : current level
+    // fl     : final level
+    // eps    : current step size
+
+    int fl = this->as.size() - 1;
+//    assert(Params.lambda.size()>level);
+//    RealD lambda= Params.lambda[level];
+    assert(level<3);
+    RealD lambda= this->Params.lambda0;
+    if (level>0) lambda= this->Params.lambda1;
+    if (level>1) lambda= this->Params.lambda2;
+    std::cout << GridLogMessage << "level: "<<level<< "lambda: "<<lambda<<std::endl;
+    
+    RealD sigma=pow(2.0,1./3.);
+
+  if(level<fl){
+//Still Omelyan. Needs to change step() to accept variable stepsize
+    RealD eps = this->Params.trajL/this->Params.MDsteps * 2.0;
+    for (int l = 0; l <= level; ++l) eps /= 2.0 * this->as[l].multiplier;
+
+    // Nesting:  2xupdate_U of size eps/2
+    // Next level is eps/2/multiplier
+
+    int multiplier = this->as[level].multiplier;
+    for (int e = 0; e < multiplier; ++e) {  // steps per step
+
+      int first_step = _first && (e == 0);
+      int last_step = _last && (e == multiplier - 1);
+
+      if (first_step) {  // initial half step
+        this->update_P(U, level, lambda * eps);
+      }
+
+        this->step(U, level + 1, first_step, 0);
+
+      this->update_P(U, level, (1.0 - 2.0 * lambda) * eps);
+
+        this->step(U, level + 1, 0, last_step);
+
+      int mm = (last_step) ? 1 : 2;
+      this->update_P(U, level, lambda * eps * mm);
+    }
+  } 
+  else 
+  { // last level
+    RealD dt = this->Params.trajL/this->Params.MDsteps * 2.0;
+    for (int l = 0; l <= level; ++l) dt /= 2.0 * this->as[l].multiplier;
+
+    RealD epsilon = dt/(2.0 - sigma);
+
+    int multiplier = this->as[level].multiplier;
+    for (int e = 0; e < multiplier; ++e) {  // steps per step
+
+      int first_step = _first && (e == 0);
+      int last_step = _last && (e == multiplier - 1);
+      // initial half step
+      if (first_step) {  this->implicit_update_P(U, level, epsilon*0.5); }
+      this->implicit_update_U(U, epsilon,epsilon*0.5);
+      this->implicit_update_P(U, level, (1.0 - sigma) * epsilon *0.5, epsilon*0.5, true);
+      this->implicit_update_U(U, -epsilon*sigma, -epsilon*sigma*0.5);
+      this->implicit_update_P(U, level, (1.0 - sigma) * epsilon *0.5, -epsilon*sigma*0.5, true);
+      this->implicit_update_U(U, epsilon,epsilon*0.5);
+      if (last_step) { this->update_P2(U, level, epsilon*0.5 ); } 
+      else
+      this->implicit_update_P(U, level, epsilon,epsilon*0.5);
+    }
+  }
+
+  }
+};
+
 NAMESPACE_END(Grid);
 
 #endif  // INTEGRATOR_INCLUDED

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,