Dirichlet first cut - wrong answers on dagger multiply.

Struggling to get a compute node so changing systems

Grid/DisableWarnings.h

@@ -44,22 +44,14 @@ directory
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
-
-#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
- //disables nvcc specific warning in json.hpp
-#pragma nv_diag_suppress unsigned_compare_with_zero
-#pragma nv_diag_suppress cast_to_qualified_type
- //disables nvcc specific warning in many files
-#pragma nv_diag_suppress esa_on_defaulted_function_ignored
-#pragma nv_diag_suppress extra_semicolon
-#else
- //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
+
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon
-#endif
+
+//Eigen only
 #endif
 
 // Disable vectorisation in Eigen on the Power8/9 and PowerPC

Grid/GridStd.h

@@ -16,7 +16,6 @@
 #include <functional>
 #include <stdio.h>
 #include <stdlib.h>
-#include <strings.h>
 #include <stdio.h>
 #include <signal.h>
 #include <ctime>

Grid/Grid_Eigen_Dense.h

@@ -14,11 +14,7 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
-#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
-#pragma nv_diag_suppress code_is_unreachable
-#else
 #pragma diag_suppress code_is_unreachable
-#endif
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")

Grid/algorithms/Algorithms.h

@@ -54,7 +54,6 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.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/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
         });

Grid/algorithms/approx/Chebyshev.h

@@ -264,7 +264,7 @@ public:
       auto Tnp_v = Tnp->View();
       auto Tnm_v = Tnm->View();
       constexpr int Nsimd = vector_type::Nsimd();
-      accelerator_for(ss, in.Grid()->oSites(), Nsimd, {
+      accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
       });

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -1,213 +0,0 @@
-/*************************************************************************************
-
-    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/Deflation.h

@@ -113,43 +113,7 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
-
-  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
-    int Nevec = (int)evec_coarse.size();
-    int Nsrc = (int)src.size();
-    // make temp variables
-    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
-    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
-    //Preporcessing
-    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    guess_coarse[j] = Zero();
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockProject(src_coarse[j],src[j],subspace);
-    }
-    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
-    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
-    for (int i=0;i<Nevec;i++)
-    {
-      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
-      const CoarseField & tmp = evec_coarse[i];
-      for (int j=0;j<Nsrc;j++)
-      {
-        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
-      }
-    }
-    //postprocessing
-    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockPromote(guess_coarse[j],guess[j],subspace);
-    guess[j].Checkerboard() = src[j].Checkerboard();
-    }
-  };
-
-  };
+};
 
 
 

Grid/allocator/MemoryManager.cc

@@ -4,14 +4,11 @@ NAMESPACE_BEGIN(Grid);
 
 /*Allocation types, saying which pointer cache should be used*/
 #define Cpu      (0)
-#define CpuHuge  (1)
-#define CpuSmall (2)
-#define Acc      (3)
-#define AccHuge  (4)
-#define AccSmall (5)
-#define Shared   (6)
-#define SharedHuge  (7)
-#define SharedSmall (8)
+#define CpuSmall (1)
+#define Acc      (2)
+#define AccSmall (3)
+#define Shared   (4)
+#define SharedSmall (5)
 #undef GRID_MM_VERBOSE 
 uint64_t total_shared;
 uint64_t total_device;
@@ -38,15 +35,12 @@ 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, 0, 8, 8, 0, 16, 8, 0, 16 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils
@@ -176,16 +170,6 @@ 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);
@@ -206,9 +190,7 @@ 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 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;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
 #endif
   
 #ifdef GRID_UVM
@@ -240,11 +222,8 @@ void MemoryManager::InitMessage(void) {
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
 {
 #ifdef ALLOCATION_CACHE
-  int cache;
-  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
-  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
-  else                                     cache = type;
-
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type + small;
   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
@@ -253,12 +232,11 @@ 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;
 
@@ -293,11 +271,8 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 void *MemoryManager::Lookup(size_t bytes,int type)
 {
 #ifdef ALLOCATION_CACHE
-  int cache;
-  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
-  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
-  else                                     cache = type;
-
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache = type+small;
   return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
@@ -306,6 +281,7 @@ 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

@@ -35,12 +35,6 @@ 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)
-#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
-#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
 
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
@@ -71,21 +65,6 @@ 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:
 
@@ -99,7 +78,7 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=9;
+  static const int NallocType=6;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
@@ -113,9 +92,8 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
- public:
   static void PrintBytes(void);
-  static void Audit(std::string s);
+ public:
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -135,28 +113,7 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
-  static uint64_t     DeviceEvictions;
-  static uint64_t     DeviceDestroy;
-  
-  static uint64_t     DeviceCacheBytes();
-  static uint64_t     HostCacheBytes();
-
-  static MemoryStatus GetFootprint(void) {
-    MemoryStatus stat;
-    stat.DeviceBytes       = DeviceBytes;
-    stat.DeviceLRUBytes    = DeviceLRUBytes;
-    stat.DeviceMaxBytes    = DeviceMaxBytes;
-    stat.HostToDeviceBytes = HostToDeviceBytes;
-    stat.DeviceToHostBytes = DeviceToHostBytes;
-    stat.HostToDeviceXfer  = HostToDeviceXfer;
-    stat.DeviceToHostXfer  = DeviceToHostXfer;
-    stat.DeviceEvictions   = DeviceEvictions;
-    stat.DeviceDestroy     = DeviceDestroy;
-    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
-    stat.HostAllocCacheBytes   = HostCacheBytes();
-    return stat;
-  };
-  
+ 
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////
@@ -213,7 +170,6 @@ private:
 
  public:
   static void Print(void);
-  static void PrintAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -3,13 +3,8 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-
-#define MAXLINE 512
-static char print_buffer [ MAXLINE ];
-
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-//#define dprintf(...) 
+//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
+#define dprintf(...)
 
 
 ////////////////////////////////////////////////////////////
@@ -28,8 +23,6 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
-uint64_t  MemoryManager::DeviceEvictions;
-uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -111,17 +104,15 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    AccCache.AccPtr=(uint64_t) NULL;
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -130,36 +121,26 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
-  // Cannot be acclocked. If allocated must be in LRU pool.
-  //
-  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
-  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
-  //                          but there is a weakness where CpuLock entries are attempted for erase
-  //                          Take these OUT LRU queue when CPU locked?
-  //                          Cannot take out the table as cpuLock data is important.
+  // Make CPU consistent, remove from Accelerator, remove entry
+  // Cannot be locked. If allocated must be in LRU pool.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
-	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
-	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
-  assert(AccCache.accLock==0); // Cannot evict so logic bomb
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  assert(AccCache.accLock==0);
+  assert(AccCache.cpuLock==0);
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    LRUremove(AccCache);
-    AccCache.AccPtr=(uint64_t)NULL;
-    AccCache.state=CpuDirty; // CPU primary now
     DeviceBytes   -=AccCache.bytes;
-    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    LRUremove(AccCache);
+    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  //  uint64_t CpuPtr = AccCache.CpuPtr;
-  DeviceEvictions++;
-  //  EntryErase(CpuPtr);
+  uint64_t CpuPtr = AccCache.CpuPtr;
+  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -169,7 +150,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -184,7 +165,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -210,7 +191,6 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -222,7 +202,6 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -233,16 +212,13 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
-  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back(); // From the LRU
+      uint64_t victim = LRU.back();
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
-    } else {
-      return;
     }
   }
 }
@@ -265,12 +241,11 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
+    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-	            (uint64_t)bytes,
-		    (uint64_t)AccCache.accLock);
+		    (uint64_t)bytes);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -305,7 +280,6 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
-    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -318,30 +292,28 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
+    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  assert(AccCache.accLock>0);
-  // If view is opened on device must remove from LRU
+  // If view is opened on device remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
-    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -362,12 +334,10 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
+
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
-  } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -404,10 +374,9 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  // CPU doesn't need to free space
-  //  if (!AccCache.AccPtr) {
-  //    EvictVictims(bytes);
-  //  }
+  if (!AccCache.AccPtr) {
+     EvictVictims(bytes);
+  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -461,28 +430,20 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
-  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
-  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-}
-void  MemoryManager::PrintAll(void)
-{
-  Print();
-  std::cout << GridLogMessage << std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -492,13 +453,13 @@ void  MemoryManager::PrintAll(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -512,61 +473,6 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
-void MemoryManager::Audit(std::string s)
-{
-  uint64_t CpuBytes=0;
-  uint64_t AccBytes=0;
-  uint64_t LruBytes1=0;
-  uint64_t LruBytes2=0;
-  uint64_t LruCnt=0;
-  uint64_t LockedBytes=0;
-  
-  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
-  for(auto it=LRU.begin();it!=LRU.end();it++){
-    uint64_t cpuPtr = *it;
-    assert(EntryPresent(cpuPtr));
-    auto AccCacheIterator = EntryLookup(cpuPtr);
-    auto & AccCache = AccCacheIterator->second;
-    LruBytes2+=AccCache.bytes;
-    assert(AccCache.LRU_valid==1);
-    assert(AccCache.LRU_entry==it);
-  }
-  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
-  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
-    auto &AccCache = it->second;
-    
-    std::string str;
-    if ( AccCache.state==Empty    ) str = std::string("Empty");
-    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
-    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
-    if ( AccCache.state==Consistent)str = std::string("Consistent");
-
-    CpuBytes+=AccCache.bytes;
-    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
-    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
-    if( AccCache.LRU_valid ) LruCnt++;
-    
-    if ( AccCache.cpuLock || AccCache.accLock ) {
-      assert(AccCache.LRU_valid==0);
-      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
-		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
-		<< "\t cpuLock  " << AccCache.cpuLock
-		<< "\t accLock  " << AccCache.accLock
-		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
-    }
-
-    assert( AccCache.cpuLock== 0 ) ;
-    assert( AccCache.accLock== 0 ) ;
-  }
-  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
-  assert(LruBytes1==LruBytes2);
-  assert(LruBytes1==DeviceLRUBytes);
-  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
-  assert(AccBytes==DeviceBytes);
-  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
-  assert(LruCnt == LRU.size());
-  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
-}
 
 void MemoryManager::PrintState(void* _CpuPtr)
 {
@@ -583,8 +489,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
     if ( AccCache.state==EvictNext) str = std::string("EvictNext");
 
     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
     << "\t" << AccCache.cpuLock
     << "\t" << AccCache.accLock
     << "\t" << AccCache.LRU_valid<<std::endl;

Grid/allocator/MemoryManagerShared.cc

@@ -12,10 +12,7 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
-uint64_t  MemoryManager::DeviceEvictions;
-uint64_t  MemoryManager::DeviceDestroy;
 
-void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
@@ -24,7 +21,6 @@ void  MemoryManager::PrintState(void* CpuPtr)
 std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
 };
 void  MemoryManager::Print(void){};
-void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_base.h

@@ -53,10 +53,11 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
-  Coordinate _processor_coor;  // linear processor coordinate
   unsigned long    _ndimension;
+  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
+  Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  Coordinate _processor_coor;  // linear processor coordinate
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -97,8 +98,9 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
+  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                      ProcessorCount(void)    ;
+  int                ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -142,16 +144,16 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,
+			       int xmit_to_rank,int do_xmit,
 			       void *recv,
-			       int recv_from_rank,
+			       int recv_from_rank,int do_recv,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,
+				    int xmit_to_rank,int do_xmit,
 				    void *recv,
-				    int recv_from_rank,
+				    int recv_from_rank,int do_recv,
 				    int bytes,int dir);
   
   

Grid/communicator/Communicator_mpi3.cc

@@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
   // Remap using the shared memory optimising routine
   // The remap creates a comm which must be freed
   ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,12 +124,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
-
+  Coordinate shm_processors       (_ndimension,1);
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
   for(int d=0;d<parent_ndimension;d++){
     parent_processor_coor[pad+d]=parent._processor_coor[d];
     parent_processors    [pad+d]=parent._processors[d];
+    shm_processors       [pad+d]=parent._shm_processors[d];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -154,6 +155,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     ccoor[d] = parent_processor_coor[d] % processors[d];
     scoor[d] = parent_processor_coor[d] / processors[d];
     ssize[d] = parent_processors[d]     / processors[d];
+    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@@ -335,22 +337,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest,
+						     int dest, int dox,
 						     void *recv,
-						     int from,
+						     int from, int dor,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,
+							 int dest,int dox,
 							 void *recv,
-							 int from,
+							 int from,int dor,
 							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
@@ -370,37 +372,43 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+from*32;
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-    assert(ierr==0);
-    list.push_back(rrq);
-    off_node_bytes+=bytes;
+  if ( dox ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      list.push_back(rrq);
+      off_node_bytes+=bytes;
+    }
   }
-
-  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    tag= dir+_processor*32;
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-    assert(ierr==0);
-    list.push_back(xrq);
-    off_node_bytes+=bytes;
-  } else {
+  
+  if (dor) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+_processor*32;
+      ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      list.push_back(xrq);
+      off_node_bytes+=bytes;
+    } else {
     // TODO : make a OMP loop on CPU, call threaded bcopy
-    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-    assert(shm!=NULL);
-    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
-    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      assert(shm!=NULL);
+      //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+    }
+  }
+  
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+    this->StencilSendToRecvFromComplete(list,dir);
   }
-
-  //  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-  //    this->StencilSendToRecvFromComplete(list,dir);
-  //  }
 
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
   //   std::cout << "Copy Synchronised\n"<<std::endl;
+  acceleratorCopySynchronise();
+
   int nreq=list.size();
 
   if (nreq==0) return;

Grid/communicator/Communicator_none.cc

@@ -45,12 +45,14 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
+  _shm_processors = Coordinate(processors.size(),1);
   srank=0;
   SetCommunicator(communicator_world);
 }
 
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
+  _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
   _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
@@ -111,18 +113,18 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,
+						     int xmit_to_rank,int dox,
 						     void *recv,
-						     int recv_from_rank,
+						     int recv_from_rank,int dor,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,
+							 int xmit_to_rank,int dox,
 							 void *recv,
-							 int recv_from_rank,
+							 int recv_from_rank,int dor,
 							 int bytes, int dir)
 {
   return 2.0*bytes;

Grid/communicator/SharedMemory.h

@@ -93,9 +93,10 @@ public:
   // Create an optimal reordered communicator that makes MPI_Cart_create get it right
   //////////////////////////////////////////////////////////////////////////////////////
   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
   static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
   ///////////////////////////////////////////////////
   // Provide shared memory facilities off comm world

Grid/communicator/SharedMemoryMPI.cc

@@ -36,10 +36,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
 #endif
-#ifdef GRID_SYCL
-#define GRID_SYCL_LEVEL_ZERO_IPC
-#endif
+#ifdef GRID_SYCl
 
+#endif
 
 NAMESPACE_BEGIN(Grid); 
 #define header "SharedMemoryMpi: "
@@ -153,7 +152,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   //////////////////////////////////////////////////////////////////////////////
   // Look and see if it looks like an HPE 8600 based on hostname conventions
@@ -166,8 +165,8 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
 
-  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
 static inline int divides(int a,int b)
 {
@@ -222,7 +221,7 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
     dim=(dim+1) %ndimension;
   }
 }
-void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -295,7 +294,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-
+  SHM = ShmDims;
+  
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -342,7 +342,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
 }
-void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -354,6 +354,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
+  SHM=ShmDims;
+
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryNone.cc

@@ -48,9 +48,10 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   _ShmSetup=1;
 }
 
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   optimal_comm = WorldComm;
+  SHM = Coordinate(processors.size(),1);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////

Grid/cshift/Cshift_common.h

@@ -297,30 +297,6 @@ 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
 //////////////////////////////////////////////////////
@@ -345,20 +321,12 @@ 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++){
@@ -409,19 +377,11 @@ 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

@@ -46,4 +46,3 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_unary.h>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
-#include <Grid/lattice/Lattice_crc.h>

Grid/lattice/Lattice_base.h

@@ -288,8 +288,8 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto him= r.View(AcceleratorRead);
     auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
@@ -303,8 +303,8 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto him= r.View(AcceleratorRead);
     auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });

Grid/lattice/Lattice_crc.h

@@ -1,55 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/lattice/Lattice_crc.h
-
-    Copyright (C) 2021
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
-{
-  auto ff = localNorm2(f);
-  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
-  typedef typename vobj::tensor_reduced normtype;
-  typedef typename normtype::scalar_object scalar;
-  std::vector<scalar> sff;
-  sliceSum(ff,sff,mu);
-  for(int t=0;t<sff.size();t++){
-    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
-  }
-}
-
-template<class vobj> uint32_t crc(Lattice<vobj> & buf)
-{
-  autoView( buf_v , buf, CpuRead);
-  return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
-}
-
-#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
-
-NAMESPACE_END(Grid);
-
-

Grid/lattice/Lattice_reduction.h

@@ -28,9 +28,6 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
-#if defined(GRID_SYCL)
-#include <Grid/lattice/Lattice_reduction_sycl.h>
-#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -130,7 +127,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -139,61 +136,25 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
 #endif  
 }
-template<class vobj>
-inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  return sumD_gpu_large(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
-
-template<class vobj>
-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)
-  autoView( arg_v, arg, AcceleratorRead);
-  return sum_gpu(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  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 rankSumLarge(const Lattice<vobj> &arg)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  return sum_gpu_large(&arg_v[0],osites);
+  auto ssum= sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->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);
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif  
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -249,10 +210,11 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
+    
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
-    // This code could read coalesce
+
     // GPU - SIMT lane compliance...
     accelerator_for( ss, sites, 1,{
 	auto x_l = left_v[ss];

Grid/lattice/Lattice_reduction_gpu.h

@@ -23,7 +23,7 @@ unsigned int nextPow2(Iterator x) {
 }
 
 template <class Iterator>
-int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
   
   int device;
 #ifdef GRID_CUDA
@@ -37,13 +37,13 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  /*  
+  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  */  
+  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -53,12 +53,12 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   threads = warpSize;
   if ( threads*sizeofsobj > sharedMemPerBlock ) {
     std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
-    return 0;
+    exit(EXIT_FAILURE);
   }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  return 1;
+  
 }
 
 template <class sobj, class Iterator>
@@ -198,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -207,9 +207,7 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   Integer size = osites*nsimd;
 
   Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
-
+  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
   Integer smemSize = numThreads * sizeof(sobj);
 
   Vector<sobj> buffer(numBlocks);
@@ -220,54 +218,6 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   auto result = buffer_v[0];
   return result;
 }
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<vector> buffer(osites);
-  vector *dat = (vector *)lat;
-  vector *buf = &buffer[0];
-  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,1,{
-	buf[ss] = dat[ss*words+w];
-      });
-      
-    ret_p[w] = sumD_gpu_small(tbuf,osites);
-  }
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  
-  Integer nsimd= vobj::Nsimd();
-  Integer size = osites*nsimd;
-  Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  
-  if ( ok ) {
-    ret = sumD_gpu_small(lat,osites);
-  } else {
-    ret = sumD_gpu_large(lat,osites);
-  }
-  return ret;
-}
-
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Return as same precision as input performing reduction in double precision though
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -280,13 +230,6 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
+
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_sycl.h

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

Grid/lattice/Lattice_transfer.h

@@ -288,36 +288,7 @@ 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,
@@ -619,26 +590,6 @@ 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>

Grid/log/Log.cc

@@ -65,40 +65,29 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
-GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
-GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
-GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
-GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(1);
+  GridLogError.Active(0);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
-  GridLogMemory.Active(0); 
-  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
-  GridLogDslash.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
-  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
-    if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
+    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
-    if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
-    if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
+    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,8 +138,7 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      //      stream << log.colour() <<  std::left;
-      stream <<  std::left;
+      stream << log.colour() <<  std::left;
       if (log.chanWidth > 0)
       {
         stream << std::setw(log.chanWidth);
@@ -154,9 +153,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      //      stream << log.colour();
-      stream <<  std::right;
+      stream << log.colour();
       stream.flags(f);
+
       return stream;
     } else { 
       return devnull;
@@ -181,12 +180,8 @@ extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
 extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
-extern GridLogger GridLogDslash;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
-extern GridLogger GridLogHMC;
-extern GridLogger GridLogMemory;
-extern GridLogger GridLogTracing;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/parallelIO/IldgIO.h

@@ -31,7 +31,6 @@ directory
 #include <fstream>
 #include <iomanip>
 #include <iostream>
-#include <string>
 #include <map>
 
 #include <pwd.h>
@@ -655,8 +654,7 @@ class IldgWriter : public ScidacWriter {
     // Fill ILDG header data struct
     //////////////////////////////////////////////////////
     ildgFormat ildgfmt ;
-    const std::string stNC = std::to_string( Nc ) ;
-    ildgfmt.field          = std::string("su"+stNC+"gauge");
+    ildgfmt.field     = std::string("su3gauge");
 
     if ( format == std::string("IEEE32BIG") ) { 
       ildgfmt.precision = 32;
@@ -873,8 +871,7 @@ class IldgReader : public GridLimeReader {
     } else { 
 
       assert(found_ildgFormat);
-      const std::string stNC = std::to_string( Nc ) ;
-      assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
+      assert ( ildgFormat_.field == std::string("su3gauge") );
 
       ///////////////////////////////////////////////////////////////////////////////////////
       // Populate our Grid metadata as best we can
@@ -882,7 +879,7 @@ class IldgReader : public GridLimeReader {
 
       std::ostringstream vers; vers << ildgFormat_.version;
       FieldMetaData_.hdr_version = vers.str();
-      FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
+      FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
 
       FieldMetaData_.nd=4;
       FieldMetaData_.dimension.resize(4);

Grid/parallelIO/MetaData.h

@@ -6,8 +6,8 @@
 
     Copyright (C) 2015
 
+
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -182,8 +182,8 @@ class GaugeStatistics
 public:
   void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
   {
-    header.link_trace = WilsonLoops<Impl>::linkTrace(data);
-    header.plaquette  = WilsonLoops<Impl>::avgPlaquette(data);
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
   }
 };
 typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
@@ -203,24 +203,20 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
 //////////////////////////////////////////////////////////////////////
 inline void reconstruct3(LorentzColourMatrix & cm)
 {
-  assert( Nc < 4 && Nc > 1 ) ;
+  const int x=0;
+  const int y=1;
+  const int z=2;
   for(int mu=0;mu<Nd;mu++){
-    #if Nc == 2
-      cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ;
-      cm(mu)()(1,1) =  adj(cm(mu)()(0,x)) ;
-    #else
-      const int x=0 , y=1 , z=2 ; // a little disinenuous labelling
-      cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
-      cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
-      cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
-    #endif
+    cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
+    cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
+    cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
   }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Some data types for intermediate storage
 ////////////////////////////////////////////////////////////////////////////////
-template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, Nc-1>, Nd >;
+template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
 
 typedef iLorentzColour2x3<Complex>  LorentzColour2x3;
 typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
@@ -282,6 +278,7 @@ struct GaugeSimpleMunger{
 
 template <class fobj, class sobj>
 struct GaugeSimpleUnmunger {
+
   void operator()(sobj &in, fobj &out) {
     for (int mu = 0; mu < Nd; mu++) {
       for (int i = 0; i < Nc; i++) {
@@ -320,8 +317,8 @@ template<class fobj,class sobj>
 struct Gauge3x2munger{
   void operator() (fobj &in,sobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<Nc-1;i++){
-	for(int j=0;j<Nc;j++){
+      for(int i=0;i<2;i++){
+	for(int j=0;j<3;j++){
 	  out(mu)()(i,j) = in(mu)(i)(j);
 	}}
     }
@@ -333,8 +330,8 @@ template<class fobj,class sobj>
 struct Gauge3x2unmunger{
   void operator() (sobj &in,fobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<Nc-1;i++){
-	for(int j=0;j<Nc;j++){
+      for(int i=0;i<2;i++){
+	for(int j=0;j<3;j++){
 	  out(mu)(i)(j) = in(mu)()(i,j);
 	}}
     }

Grid/parallelIO/NerscIO.h

@@ -9,7 +9,6 @@
     Author: Matt Spraggs <matthew.spraggs@gmail.com>
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -31,8 +30,6 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
-#include <string>
-
 NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
@@ -148,17 +145,15 @@ public:
 
     std::string format(header.floating_point);
 
-    const int ieee32big = (format == std::string("IEEE32BIG"));
-    const int ieee32    = (format == std::string("IEEE32"));
-    const int ieee64big = (format == std::string("IEEE64BIG"));
-    const int ieee64    = (format == std::string("IEEE64") || \
-			   format == std::string("IEEE64LITTLE"));
+    int ieee32big = (format == std::string("IEEE32BIG"));
+    int ieee32    = (format == std::string("IEEE32"));
+    int ieee64big = (format == std::string("IEEE64BIG"));
+    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     // depending on datatype, set up munger;
     // munger is a function of <floating point, Real, data_type>
-    const std::string stNC = std::to_string( Nc ) ;
-    if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
+    if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@@ -169,7 +164,7 @@ public:
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
-    } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
+    } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@@ -214,29 +209,27 @@ public:
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
-					std::string ens_label = std::string("DWF"),
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
+					std::string ens_label = std::string("DWF"))
   {
-    writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
+    writeConfiguration(Umu,file,0,1,ens_label);
   }
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32,
-					std::string ens_label = std::string("DWF"),
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
+					std::string ens_label = std::string("DWF"))
   {
     typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
-    header.sequence_number = sequence_number;
-    header.ensemble_id     = ens_id;
+    ///////////////////////////////////////////
+    // Following should become arguments
+    ///////////////////////////////////////////
+    header.sequence_number = 1;
+    header.ensemble_id     = std::string("UKQCD");
     header.ensemble_label  = ens_label;
-    header.hdr_version     = "1.0" ;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -250,14 +243,10 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write NcxNc double
-    header.floating_point  = std::string("IEEE64BIG");
-    const std::string stNC = std::to_string( Nc ) ;
-    if( two_row ) {
-      header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
-    } else {
-      header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
-    }
+    // Sod it -- always write 3x3 double
+    header.floating_point = std::string("IEEE64BIG");
+    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
+    GaugeSimpleUnmunger<fobj3D,sobj> munge;
     if ( grid->IsBoss() ) { 
       truncate(file);
       offset = writeHeader(header,file);
@@ -265,15 +254,8 @@ public:
     grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    if( two_row ) {
-      Gauge3x2unmunger<fobj2D,sobj> munge;
-      BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
-						nersc_csum,scidac_csuma,scidac_csumb);
-    } else {
-      GaugeSimpleUnmunger<fobj3D,sobj> munge;
-      BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
-						nersc_csum,scidac_csuma,scidac_csumb);
-    }
+    BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
+					      nersc_csum,scidac_csuma,scidac_csumb);
     header.checksum = nersc_csum;
     if ( grid->IsBoss() ) { 
       writeHeader(header,file);
@@ -305,7 +287,8 @@ public:
     header.plaquette=0.0;
     MachineCharacteristics(header);
 
-    uint64_t offset;
+	uint64_t offset;
+  
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");
@@ -345,7 +328,7 @@ public:
 
     GridBase *grid = parallel.Grid();
 
-    uint64_t offset = readHeader(file,grid,header);
+	uint64_t offset = readHeader(file,grid,header);
 
     FieldMetaData clone(header);
 

Grid/perfmon/PerfCount.cc

@@ -27,12 +27,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
-
-#include <Grid/perfmon/Timer.h>
 #include <Grid/perfmon/PerfCount.h>
-NAMESPACE_BEGIN(Grid);
 
-GridTimePoint theProgramStart = GridClock::now();
+NAMESPACE_BEGIN(Grid);
 
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
 #define RawConfig(A,B) (A<<8|B)

Grid/perfmon/PerfCount.h

@@ -30,12 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
-
-#ifndef __SSC_START
-#define __SSC_START
-#define __SSC_STOP
-#endif
-
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>
@@ -78,9 +72,17 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
+#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
+#define __SSC_STOP  __SSC_MARK(0x110)
+#define __SSC_START __SSC_MARK(0x111)
+
 
 #else
 
+#define __SSC_MARK(mark) 
+#define __SSC_STOP  
+#define __SSC_START 
+
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

@@ -35,8 +35,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid)
 
-//typedef  std::chrono::system_clock          GridClock;
-typedef  std::chrono::high_resolution_clock   GridClock;
+// Dress the output; use std::chrono
+// C++11 time facilities better?
+inline double usecond(void) {
+  struct timeval tv;
+#ifdef TIMERS_ON
+  gettimeofday(&tv,NULL);
+#endif
+  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
+}
+
+typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
 
 typedef  std::chrono::seconds               GridSecs;
@@ -44,15 +53,6 @@ typedef  std::chrono::milliseconds          GridMillisecs;
 typedef  std::chrono::microseconds          GridUsecs;
 typedef  std::chrono::microseconds          GridTime;
 
-extern GridTimePoint theProgramStart;
-// Dress the output; use std::chrono
-// C++11 time facilities better?
-inline double usecond(void) {
-  auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart); 
-  return 1.0*usecs.count();
-}
-
-
 inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
   stream << time.count()<<" s";

Grid/perfmon/Tracing.h

@@ -1,70 +0,0 @@
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-#ifdef GRID_TRACING_NVTX
-#include <nvToolsExt.h>
-class GridTracer {
-public:
-  GridTracer(const char* name) {
-    nvtxRangePushA(name);
-  }
-  ~GridTracer() {
-    nvtxRangePop();
-  }
-};
-inline void tracePush(const char *name) { nvtxRangePushA(name); }
-inline void tracePop(const char *name) { nvtxRangePop(); }
-inline int  traceStart(const char *name) {  }
-inline void traceStop(int ID) {  }
-#endif
-
-#ifdef GRID_TRACING_ROCTX
-#include <roctracer/roctx.h>
-class GridTracer {
- public:
-  GridTracer(const char* name) {
-    roctxRangePushA(name);
-    std::cout << "roctxRangePush "<<name<<std::endl;
-  }
-  ~GridTracer() {
-    roctxRangePop();
-    std::cout << "roctxRangePop "<<std::endl;
-  }
-};
-inline void tracePush(const char *name) { roctxRangePushA(name); }
-inline void tracePop(const char *name) { roctxRangePop(); }
-inline int  traceStart(const char *name) { roctxRangeStart(name); }
-inline void traceStop(int ID) { roctxRangeStop(ID); }
-#endif
-
-#ifdef GRID_TRACING_TIMER
-class GridTracer {
- public:
-  const char *name;
-  double elapsed;
-  GridTracer(const char* _name) {
-    name = _name;
-    elapsed=-usecond();
-  }
-  ~GridTracer() {
-    elapsed+=usecond();
-    std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl;
-  }
-};
-inline void tracePush(const char *name) {  }
-inline void tracePop(const char *name) {  }
-inline int  traceStart(const char *name) { return 0; }
-inline void traceStop(int ID) {  }
-#endif
-
-#ifdef GRID_TRACING_NONE
-#define GRID_TRACE(name) 
-inline void tracePush(const char *name) {  }
-inline void tracePop(const char *name) {  }
-inline int  traceStart(const char *name) { return 0;  }
-inline void traceStop(int ID) {  }
-#else
-#define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
-#endif
-NAMESPACE_END(Grid);

Grid/pugixml/pugixml.cc

@@ -16,12 +16,8 @@
 
 #ifdef __NVCC__
 #pragma push
-#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
-#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
-#else
 #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
 #endif
-#endif
 
 #include "pugixml.h"
 

Grid/qcd/QCD.h

@@ -451,20 +451,9 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
-#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
-#ifdef FAST_FERM_TO_PROP
-  autoView(p_v,p,CpuWrite);
-  autoView(f_v,f,CpuRead);
-  thread_for(idx,p_v.oSites(),{
-      for(int ss = 0; ss < Ns; ++ss) {
-      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
-	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
-      }}
-    });
-#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -476,23 +465,12 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
-#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
-#ifdef FAST_FERM_TO_PROP
-  autoView(p_v,p,CpuRead);
-  autoView(f_v,f,CpuWrite);
-  thread_for(idx,p_v.oSites(),{
-      for(int ss = 0; ss < Ns; ++ss) {
-      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
-	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
-      }}
-    });
-#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -504,7 +482,6 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
-#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -68,16 +68,9 @@ public:
   ///////////////////////////////////////////////////////////////
   // Support for MADWF tricks
   ///////////////////////////////////////////////////////////////
-  RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
-  RealD MassPlus(void) { return mass_plus; };
-  RealD MassMinus(void) { return mass_minus; };
+  RealD Mass(void) { return mass; };
   void  SetMass(RealD _mass) { 
-    mass_plus=mass_minus=_mass; 
-    SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
-  } ;
-  void  SetMass(RealD _mass_plus, RealD _mass_minus) { 
-    mass_plus=_mass_plus;
-    mass_minus=_mass_minus;
+    mass=_mass; 
     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
   } ;
   void  P(const FermionField &psi, FermionField &chi);
@@ -115,7 +108,7 @@ public:
   void   MeooeDag5D    (const FermionField &in, FermionField &out);
 
   //    protected:
-  RealD mass_plus, mass_minus;
+  RealD mass;
 
   // Save arguments to SetCoefficientsInternal
   Vector<Coeff_t> _gamma;

Grid/qcd/action/fermion/CloverHelpers.h

@@ -1,333 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
-
-    Copyright (C) 2017 - 2022
-
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
-
-    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/Grid.h>
-#include <Grid/qcd/spin/Dirac.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-
-////////////////////////////////////////////
-// Standard Clover
-//   (4+m0) + csw * clover_term
-// Exp Clover
-//   (4+m0) * exp(csw/(4+m0) clover_term)
-//   = (4+m0) + csw * clover_term + ...
-////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-
-//////////////////////////////////
-// Generic Standard Clover
-//////////////////////////////////
-
-template<class Impl>
-class CloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
-    GridBase *grid = CloverTerm.Grid();
-    CloverTerm += diag_mass;
-
-    int lvol = grid->lSites();
-    int DimRep = Impl::Dimension;
-    {
-      autoView(CTv,CloverTerm,CpuRead);
-      autoView(CTIv,CloverTermInv,CpuWrite);
-      thread_for(site, lvol, {
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-        Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
-        peekLocalSite(Qx, CTv, lcoor);
-
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++){
-                auto zz =  Qx()(j, k)(a, b);
-                EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
-              }
-
-        EigenInvCloverOp = EigenCloverOp.inverse();
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++)
-                Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-               pokeLocalSite(Qxinv, CTIv, lcoor);
-      });
-    }
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-
-};
-
-
-//////////////////////////////////
-// Generic Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  // Can this be avoided?
-  static void IdentityTimesC(const CloverField& in, RealD c) {
-    int DimRep = Impl::Dimension;
-
-    autoView(in_v, in, AcceleratorWrite);
-
-    accelerator_for(ss, in.Grid()->oSites(), 1, {
-      for (int sa=0; sa<Ns; sa++)
-        for (int ca=0; ca<DimRep; ca++)
-          in_v[ss]()(sa,sa)(ca,ca) = c;
-    });
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-    GridBase* grid = Clover.Grid();
-    CloverField ExpClover(grid);
-
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-
-    Clover *= (1.0/diag_mass);
-
-    // Taylor expansion, slow but generic
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-    // qN = cN
-    // qn = cn + qn+1 X
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
-      cn[i] = cn[i-1] / RealD(i);
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * Clover + cn[i];
-
-    // prepare inverse
-    CloverInv = (-1.0)*Clover;
-
-    Clover = ExpClover * diag_mass;
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * CloverInv + cn[i];
-
-    CloverInv = ExpClover * (1.0/diag_mass);
-
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-//////////////////////////////////
-// Compact Standard Clover
-//////////////////////////////////
-
-
-template<class Impl>
-class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
-                            public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-    Clover += diag_mass;
-  }
-
-  static void InvertClover(CloverField& InvClover,
-                            const CloverDiagonalField& diagonal,
-                            const CloverTriangleField& triangle,
-                            CloverDiagonalField&       diagonalInv,
-                            CloverTriangleField&       triangleInv,
-                            bool fixedBoundaries) {
-
-    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
-  }
-
-  // TODO: implement Cmunu for better performances with compact layout, but don't do it
-  // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-};
-
-//////////////////////////////////
-// Compact Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  // Can this be avoided?
-  static void IdentityTimesC(const CloverField& in, RealD c) {
-    int DimRep = Impl::Dimension;
-
-    autoView(in_v, in, AcceleratorWrite);
-
-    accelerator_for(ss, in.Grid()->oSites(), 1, {
-      for (int sa=0; sa<Ns; sa++)
-        for (int ca=0; ca<DimRep; ca++)
-          in_v[ss]()(sa,sa)(ca,ca) = c;
-    });
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Clover.Grid();
-    CloverField ExpClover(grid);
-
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-
-    Clover *= (1.0/diag_mass);
-
-    // Taylor expansion, slow but generic
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-    // qN = cN
-    // qn = cn + qn+1 X
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
-      cn[i] = cn[i-1] / RealD(i);
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * Clover + cn[i];
-
-    // prepare inverse
-    CloverInv = (-1.0)*Clover;
-
-    Clover = ExpClover * diag_mass;
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * CloverInv + cn[i];
-
-    CloverInv = ExpClover * (1.0/diag_mass);
-
-  }
-
-  static void InvertClover(CloverField& InvClover,
-                            const CloverDiagonalField& diagonal,
-                            const CloverTriangleField& triangle,
-                            CloverDiagonalField&       diagonalInv,
-                            CloverTriangleField&       triangleInv,
-                            bool fixedBoundaries) {
-
-    if (fixedBoundaries)
-    {
-      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
-    }
-    else
-    {
-      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
-    }
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -31,7 +31,6 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -86,7 +85,7 @@ NAMESPACE_BEGIN(Grid);
 //           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
 //                                                                 =  84 complex words per site
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
 class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
                                    public WilsonCloverHelpers<Impl>,
                                    public CompactWilsonCloverHelpers<Impl> {
@@ -225,7 +224,7 @@ public:
   RealD csw_t;
   RealD cF;
 
-  bool fixedBoundaries;
+  bool open_boundaries;
 
   CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
   CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;

Grid/qcd/action/fermion/Fermion.h

@@ -138,52 +138,38 @@ typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
 // Clover fermions
-template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
-template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
+typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
+typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
-typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
-typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
+typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
+typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
 
-typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
-typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
-typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
 
-typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
-typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
-typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
-
-typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
-
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Compact Clover fermions
-template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
-template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
+typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
+typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
 
-typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
-typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
-typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
+typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
+typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
 
-typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
-typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
-typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
 
-typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
-typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
-typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
-
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
-
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Domain Wall fermions
 typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -32,7 +32,6 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -52,7 +51,7 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
 class WilsonCloverFermion : public WilsonFermion<Impl>,
                             public WilsonCloverHelpers<Impl>
 {

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -209,8 +209,6 @@ public:
 };
 
 
-////////////////////////////////////////////////////////
-
 template<class Impl> class CompactWilsonCloverHelpers {
 public:
 
@@ -728,8 +726,8 @@ public:
   static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
     conformable(f, m);
     auto grid  = f.Grid();
-    const uint32_t Nsite = grid->oSites();
-    const uint32_t Nsimd = grid->Nsimd();
+    const int Nsite = grid->oSites();
+    const int Nsimd = grid->Nsimd();
     autoView(f_v, f, AcceleratorWrite);
     autoView(m_v, m, AcceleratorRead);
     // NOTE: this function cannot be 'private' since nvcc forbids this for kernels

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -47,6 +47,8 @@ class CompactWilsonCloverTypes {
 public:
   INHERIT_IMPL_TYPES(Impl);
 
+  static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
+
   static constexpr int Nred      = Nc * Nhs;        // 6
   static constexpr int Nblock    = Nhs;             // 2
   static constexpr int Ndiagonal = Nred;            // 6

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -117,19 +117,19 @@ public:
     typedef decltype(coalescedRead(*in))    sobj;
     typedef decltype(coalescedRead(*out0)) hsobj;
 
-    constexpr unsigned int Nsimd = vobj::Nsimd();
+    unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];  // out0[j] = merge low bit of type from in[k] and in[m] 
-    const vobj *vp1 = &in[m];  // out1[j] = merge hi  bit of type from in[k] and in[m]
-    const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
-    auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
-    auto sb = coalescedRead(*vp,j1); // lane to read for out 1
+    const vobj *vp0 = &in[k];
+    const vobj *vp1 = &in[m];
+    const vobj *vp = (lane&mask) ? vp1:vp0;
+    auto sa = coalescedRead(*vp,j0);
+    auto sb = coalescedRead(*vp,j1);
     hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);  // spin project the result0
-    projector::Proj(psb,sb,mu,dag);  // spin project the result1
+    projector::Proj(psa,sa,mu,dag);
+    projector::Proj(psb,sb,mu,dag);
     coalescedWrite(out0[j],psa);
     coalescedWrite(out1[j],psb);
 #else

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -173,7 +173,12 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-    
+
+  void DirichletBlock(std::vector<int> & block){
+    Stencil.DirichletBlock(block); 
+    StencilEven.DirichletBlock(block); 
+    StencilOdd.DirichletBlock(block); 
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h

@@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FiveDimRedBlackGrid,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
-  mass_plus(_mass), mass_minus(_mass)
+  mass(_mass)
 { 
 }
 
@@ -209,8 +209,8 @@ void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
   Vector<Coeff_t> diag (Ls,1.0);
-  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
-  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -220,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
   Vector<Coeff_t> diag = bs;
   Vector<Coeff_t> upper= cs;
   Vector<Coeff_t> lower= cs; 
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
 }
 // FIXME Redunant with the above routine; check this and eliminate
@@ -235,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -250,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &
     upper[i]=-cee[i];
     lower[i]=-cee[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -266,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &
     // Assemble the 5d matrix
     if ( s==0 ) {
       upper[s] = -cee[s+1] ;
-      lower[s] = mass_minus*cee[Ls-1];
+      lower[s] = mass*cee[Ls-1];
     } else if ( s==(Ls-1)) { 
-      upper[s] = mass_plus*cee[0];
+      upper[s] = mass*cee[0];
       lower[s] = -cee[s-1];
     } else {
       upper[s]=-cee[s+1];
@@ -291,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
   Vector<Coeff_t> diag(Ls,1.0);
   Vector<Coeff_t> upper(Ls,-1.0);
   Vector<Coeff_t> lower(Ls,-1.0);
-  upper[Ls-1]=-mass_plus*upper[Ls-1];
-  lower[0]   =-mass_minus*lower[0];
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
 }
 
@@ -307,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
       upper[s] = cs[s+1];
-      lower[s] =-mass_minus*cs[Ls-1];
+      lower[s] =-mass*cs[Ls-1];
     } else if ( s==(Ls-1) ) { 
-      upper[s] =-mass_plus*cs[0];
+      upper[s] =-mass*cs[0];
       lower[s] = cs[s-1];
     } else { 
       upper[s] = cs[s+1];
@@ -552,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
-      leem[i]=mass_minus*cee[Ls-1]/bee[0];
+      leem[i]=mass*cee[Ls-1]/bee[0];
       for(int j=0;j<i;j++) {
 	assert(bee[j+1]!=Coeff_t(0.0));
 	leem[i]*= aee[j]/bee[j+1];
@@ -560,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
-      ueem[i]=mass_plus;
+      ueem[i]=mass;
       for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
       ueem[i]*= aee[0]/bee[0];
       
@@ -573,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
   }
 	
   { 
-    Coeff_t delta_d=mass_minus*cee[Ls-1];
+    Coeff_t delta_d=mass*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
       assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
@@ -642,10 +642,6 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
-  
 #if (!defined(GRID_HIP))
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
@@ -781,8 +777,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   assert(mu>=0);
   assert(mu<Nd);
 
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
 
 #if 0
   int tshift = (mu == Nd-1) ? 1 : 0;

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

@@ -32,23 +32,22 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 
-
 NAMESPACE_BEGIN(Grid);
-template<class Impl, class CloverHelpers>
-CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
-                                                                            GridCartesian& Fgrid,
-                                                                            GridRedBlackCartesian& Hgrid,
-                                                                            const RealD _mass,
-                                                                            const RealD _csw_r,
-                                                                            const RealD _csw_t,
-                                                                            const RealD _cF,
-                                                                            const WilsonAnisotropyCoefficients& clover_anisotropy,
-                                                                            const ImplParams& impl_p)
+template<class Impl>
+CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
+                                                             GridCartesian& Fgrid,
+                                                             GridRedBlackCartesian& Hgrid,
+                                                             const RealD _mass,
+                                                             const RealD _csw_r,
+                                                             const RealD _csw_t,
+                                                             const RealD _cF,
+                                                             const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                                             const ImplParams& impl_p)
   : WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
   , csw_r(_csw_r)
   , csw_t(_csw_t)
   , cF(_cF)
-  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
   , Diagonal(&Fgrid),        Triangle(&Fgrid)
   , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
   , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
@@ -59,85 +58,80 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , BoundaryMask(&Fgrid)
   , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
 {
-  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
-
   csw_r *= 0.5;
   csw_t *= 0.5;
   if (clover_anisotropy.isAnisotropic)
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (fixedBoundaries) {
-    this->BoundaryMaskEven.Checkerboard() = Even;
-    this->BoundaryMaskOdd.Checkerboard() = Odd;
+  if (open_boundaries)
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
-  }
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->fixedBoundaries) ApplyBoundaryMask(out);
+  if(this->open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
-  if(this->fixedBoundaries) {
+  if(this->open_boundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
@@ -147,16 +141,16 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
   Mooee(in, out); // blocks are hermitian
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
@@ -166,27 +160,27 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
   MooeeInv(in, out); // blocks are hermitian
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
   DhopDir(in, out, dir, disp);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
   DhopDirAll(in, out);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!fixedBoundaries); // TODO check for changes required for open bc
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+  assert(!open_boundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -257,7 +251,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force,
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);   // checked
       count++;
     }
 
@@ -267,18 +261,18 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force,
   force += clover_force;
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&        in,
                     FermionField&              out,
                     const CloverDiagonalField& diagonal,
                     const CloverTriangleField& triangle) {
@@ -291,8 +285,8 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const Fermio
   CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
 }
 
-template<class Impl, class CloverHelpers>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
+template<class Impl>
+void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   // NOTE: parts copied from original implementation
 
   // Import gauge into base class
@@ -305,7 +299,6 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   GridBase* grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
   CloverField TmpOriginal(grid);
-  CloverField TmpInverse(grid);
 
   // Compute the field strength terms mu>nu
   double t2 = usecond();
@@ -325,30 +318,22 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-
-  // Instantiate the clover term
-  // - In case of the standard clover the mass term is added
-  // - In case of the exponential clover the clover term is exponentiated
-  double t4 = usecond();
-  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
+  TmpOriginal += this->diag_mass;
 
   // Convert the data layout of the clover term
-  double t5 = usecond();
+  double t4 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Modify the clover term at the temporal boundaries in case of open boundary conditions
-  double t6 = usecond();
-  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+  // Possible modify the boundary values
+  double t5 = usecond();
+  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the Clover term
-  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
-  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
-  // TODO: For now this inversion is explictly done on the CPU
-  double t7 = usecond();
-  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
+  // Invert the clover term in the improved layout
+  double t6 = usecond();
+  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
 
   // Fill the remaining clover fields
-  double t8 = usecond();
+  double t7 = usecond();
   pickCheckerboard(Even, DiagonalEven,    Diagonal);
   pickCheckerboard(Even, TriangleEven,    Triangle);
   pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
@@ -359,19 +344,20 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
 
   // Report timings
-  double t9 = usecond();
-
-  std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
+  double t8 = usecond();
+#if 0
+  std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
+            << ", allocations = "               << (t2 - t1) / 1e6
+            << ", field strength = "            << (t3 - t2) / 1e6
+            << ", fill clover = "               << (t4 - t3) / 1e6
+            << ", convert = "                   << (t5 - t4) / 1e6
+            << ", boundaries = "                << (t6 - t5) / 1e6
+            << ", inversions = "                << (t7 - t6) / 1e6
+            << ", pick cbs = "                  << (t8 - t7) / 1e6
+            << ", total = "                     << (t8 - t0) / 1e6
+            << std::endl;
+#endif
 }
 
 NAMESPACE_END(Grid);

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

@@ -34,8 +34,8 @@
 
 NAMESPACE_BEGIN(Grid);
 
-template<class Impl, class CloverHelpers>
-WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField&                         _Umu,
+template<class Impl>
+WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField&                         _Umu,
                                                GridCartesian&                      Fgrid,
                                                GridRedBlackCartesian&              Hgrid,
                                                const RealD                         _mass,
@@ -74,8 +74,8 @@ WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField&
 }
 
 // *NOT* EO
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -89,8 +89,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, Fermion
   out += temp;
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
 {
   FermionField temp(out.Grid());
 
@@ -104,8 +104,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, Ferm
   out += temp;
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
+template <class Impl>
+void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   double t0 = usecond();
   WilsonFermion<Impl>::ImportGauge(_Umu);
@@ -131,11 +131,47 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
   CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
   CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
   CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
-   
+  CloverTerm += diag_mass;
+
   double t4 = usecond();
-  CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
+  int lvol = _Umu.Grid()->lSites();
+  int DimRep = Impl::Dimension;
 
   double t5 = usecond();
+  {
+    autoView(CTv,CloverTerm,CpuRead);
+    autoView(CTIv,CloverTermInv,CpuWrite);
+    thread_for(site, lvol, {
+      Coordinate lcoor;
+      grid->LocalIndexToLocalCoor(site, lcoor);
+      Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
+      typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
+      peekLocalSite(Qx, CTv, lcoor);
+      //if (csw!=0){
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++){
+	      auto zz =  Qx()(j, k)(a, b);
+	      EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
+	    }
+      //   if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
+      
+      EigenInvCloverOp = EigenCloverOp.inverse();
+      //std::cout << EigenInvCloverOp << std::endl;
+      for (int j = 0; j < Ns; j++)
+	for (int k = 0; k < Ns; k++)
+	  for (int a = 0; a < DimRep; a++)
+	    for (int b = 0; b < DimRep; b++)
+	      Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
+      //    if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
+      //  }
+      pokeLocalSite(Qxinv, CTIv, lcoor);
+    });
+  }
+
+  double t6 = usecond();
   // Separate the even and odd parts
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@@ -148,44 +184,48 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
 
   pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
   pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
-  double t6 = usecond();
+  double t7 = usecond();
 
-  std::cout << GridLogDebug << "WilsonCloverFermion::ImportGauge timings:" << std::endl;
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "instantiation =              " << (t5 - t4) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "pick cbs =                   " << (t6 - t5) / 1e6 << std::endl;
-  std::cout << GridLogDebug << "total =                      " << (t6 - t0) / 1e6 << std::endl;
+#if 0
+  std::cout << GridLogMessage << "WilsonCloverFermion::ImportGauge timings:"
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
+            << ", allocations = "               << (t2 - t1) / 1e6
+            << ", field strength = "            << (t3 - t2) / 1e6
+            << ", fill clover = "               << (t4 - t3) / 1e6
+            << ", misc = "                      << (t5 - t4) / 1e6
+            << ", inversions = "                << (t6 - t5) / 1e6
+            << ", pick cbs = "                  << (t7 - t6) / 1e6
+            << ", total = "                     << (t7 - t0) / 1e6
+            << std::endl;
+#endif
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseNo);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseNo);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerNo, InverseYes);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
 {
   this->MooeeInternal(in, out, DaggerYes, InverseYes);
 }
 
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
 {
   out.Checkerboard() = in.Checkerboard();
   CloverField *Clover;
@@ -238,8 +278,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField
 } // MooeeInternal
 
 // Derivative parts unpreconditioned pseudofermions
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
 {
   conformable(X.Grid(), Y.Grid());
   conformable(X.Grid(), force.Grid());
@@ -309,7 +349,7 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const F
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);                   // checked
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);                   // checked
       count++;
     }
 
@@ -320,15 +360,15 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const F
 }
 
 // Derivative parts
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
 {
   assert(0);
 }
 
 // Derivative parts
-template<class Impl, class CloverHelpers>
-void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
+template <class Impl>
+void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
 {
   assert(0); // not implemented yet
 }

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

@@ -4,13 +4,12 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
 
-Copyright (C) 2022
+Copyright (C) 2015
 
 Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Fabian Joswig <fabian.joswig@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
@@ -600,47 +599,11 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
                                                    Current curr_type,
                                                    unsigned int mu)
 {
-  if(curr_type != Current::Vector)
-  {
-    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
-    exit(1);
-  }
-
   Gamma g5(Gamma::Algebra::Gamma5);
   conformable(_grid, q_in_1.Grid());
   conformable(_grid, q_in_2.Grid());
   conformable(_grid, q_out.Grid());
-  auto UGrid= this->GaugeGrid();
-
-  PropagatorField tmp_shifted(UGrid);
-  PropagatorField g5Lg5(UGrid);
-  PropagatorField R(UGrid);
-  PropagatorField gmuR(UGrid);
-
-    Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT,
-  };
-  Gamma gmu=Gamma(Gmu[mu]);
-
-  g5Lg5=g5*q_in_1*g5;
-  tmp_shifted=Cshift(q_in_2,mu,1);
-  Impl::multLinkField(R,this->Umu,tmp_shifted,mu);
-  gmuR=gmu*R;
-
-  q_out=adj(g5Lg5)*R;
-  q_out-=adj(g5Lg5)*gmuR;
-
-  tmp_shifted=Cshift(q_in_1,mu,1);
-  Impl::multLinkField(g5Lg5,this->Umu,tmp_shifted,mu);
-  g5Lg5=g5*g5Lg5*g5;
-  R=q_in_2;
-  gmuR=gmu*R;
-
-  q_out-=adj(g5Lg5)*R;
-  q_out-=adj(g5Lg5)*gmuR;
+  assert(0);
 }
 
 
@@ -654,51 +617,9 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
                                               unsigned int tmax,
 					      ComplexField &lattice_cmplx)
 {
-  if(curr_type != Current::Vector)
-  {
-    std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
-    exit(1);
-  }
-
-  int tshift = (mu == Nd-1) ? 1 : 0;
-  unsigned int LLt    = GridDefaultLatt()[Tp];
   conformable(_grid, q_in.Grid());
   conformable(_grid, q_out.Grid());
-  auto UGrid= this->GaugeGrid();
-
-  PropagatorField tmp(UGrid);
-  PropagatorField Utmp(UGrid);
-  PropagatorField L(UGrid);
-  PropagatorField zz (UGrid);
-  zz=Zero();
-  LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
-
-    Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT,
-  };
-  Gamma gmu=Gamma(Gmu[mu]);
-
-  tmp = Cshift(q_in,mu,1);
-  Impl::multLinkField(Utmp,this->Umu,tmp,mu);
-  tmp = ( Utmp*lattice_cmplx - gmu*Utmp*lattice_cmplx ); // Forward hop
-  tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
-  q_out = where((lcoor<=tmax),tmp,zz); // Position of current complicated
-
-  tmp = q_in *lattice_cmplx;
-  tmp = Cshift(tmp,mu,-1);
-  Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
-  tmp = -( Utmp + gmu*Utmp );
-  // Mask the time
-  if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
-    unsigned int t0 = 0;
-    tmp = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
-  } else {
-    tmp = where((lcoor>=tmin+tshift),tmp,zz);
-  }
-  q_out+= where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
+  assert(0);
 }
 
 NAMESPACE_END(Grid);

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

@@ -459,7 +459,11 @@ 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_TMP(HandDhopSiteSycl);    return; }
+#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
+#endif     
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
@@ -470,7 +474,6 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
-     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
@@ -496,19 +499,17 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
    } else if( interior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteDagInt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteDagInt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
-     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(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/CompactWilsonCloverFermionInstantiation.cc.master

@@ -9,7 +9,6 @@
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
     Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
 
     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
@@ -33,12 +32,10 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>; 
-template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>; 
+template class CompactWilsonCloverFermion<IMPLEMENTATION>; 
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

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

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

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

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonCloverFermionInstantiation.cc.master

@@ -8,8 +8,7 @@
 
     Author: paboyle <paboyle@ph.ed.ac.uk>
     Author: Guido Cossu <guido.cossu@ed.ac.uk>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
-    
+
     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
@@ -32,12 +31,10 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
 #include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
 #include "impl.h"
-template class WilsonCloverFermion<IMPLEMENTATION, CloverHelpers<IMPLEMENTATION>>; 
-template class WilsonCloverFermion<IMPLEMENTATION, ExpCloverHelpers<IMPLEMENTATION>>; 
+template class WilsonCloverFermion<IMPLEMENTATION>; 
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

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

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

@@ -0,0 +1,51 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
+
+Copyright (C) 2015, 2020
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
+
+#ifndef AVX512
+#ifndef QPX
+#ifndef A64FX
+#ifndef A64FXFIXEDSIZE
+#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
+#endif
+#endif
+#endif
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+#include "impl.h"
+template class WilsonKernels<IMPLEMENTATION>;
+
+NAMESPACE_END(Grid);

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

@@ -18,10 +18,6 @@ WILSON_IMPL_LIST=" \
 	   GparityWilsonImplF \
 	   GparityWilsonImplD "
 
-COMPACT_WILSON_IMPL_LIST=" \
-	   WilsonImplF \
-	   WilsonImplD "
-
 DWF_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
@@ -44,23 +40,13 @@ EOF
 
 done
 
-CC_LIST="WilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
+CC_LIST="WilsonCloverFermionInstantiation CompactWilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
 
 for impl in $WILSON_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
-done
-done
-
-CC_LIST="CompactWilsonCloverFermionInstantiation"
-
-for impl in $COMPACT_WILSON_IMPL_LIST
-do
-for f in $CC_LIST
-do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 
@@ -77,14 +63,14 @@ for impl in $DWF_IMPL_LIST $GDWF_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 
 # overwrite the .cc file in Gparity directories
 for impl in $GDWF_IMPL_LIST
 do
-  ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
+  ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc 
 done
 
 
@@ -98,7 +84,7 @@ for impl in $STAG_IMPL_LIST
 do
 for f in $CC_LIST
 do
-  ln -f -s ../$f.cc.master $impl/$f$impl.cc
+  ln -f -s ../$f.cc.master $impl/$f$impl.cc 
 done
 done
 

Grid/qcd/action/gauge/Photon.h

@@ -49,7 +49,7 @@ NAMESPACE_BEGIN(Grid);
     
     typedef Lattice<SiteLink>  LinkField;
     typedef Lattice<SiteField> Field;
-    typedef LinkField          ComplexField;
+    typedef Field              ComplexField;
   };
   
   typedef QedGImpl<vComplex> QedGImplR;

Grid/qcd/utils/GaugeFix.h

@@ -55,12 +55,12 @@ public:
     }
   }  
 
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
-    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
+    SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog);
   }
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
 
     GridBase *grid = Umu.Grid();
 
@@ -122,8 +122,6 @@ public:
 
       }
     }
-    std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
-    if (err_on_no_converge) assert(0);
   };
   static Real SteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform,Real & alpha, GaugeMat & dmuAmu,int orthog) {
     GridBase *grid = U[0].Grid();

Grid/qcd/utils/WilsonLoops.h

@@ -125,6 +125,7 @@ public:
     return sumplaq / vol / faces / Nc; // Nd , Nc dependent... FIXME
   }
 
+
   //////////////////////////////////////////////////
   // average over all x,y,z the temporal loop
   //////////////////////////////////////////////////
@@ -164,7 +165,7 @@ public:
 
     double vol = Umu.Grid()->gSites();
 
-    return p.real() / vol / (4.0 * Nc ) ;
+    return p.real() / vol / 4.0 / 3.0;
   };
 
   //////////////////////////////////////////////////

Grid/serialisation/JSON_IO.cc

@@ -26,7 +26,7 @@
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/Grid.h>
-#ifndef GRID_HIP
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
 
 NAMESPACE_BEGIN(Grid);
 
@@ -82,7 +82,7 @@ void JSONWriter::writeDefault(const std::string &s,	const std::string &x)
   if (s.size())
     ss_ << "\""<< s << "\" : \"" << os.str() << "\" ," ;
   else
-    ss_ << "\""<< os.str() << "\" ," ;
+    ss_ << os.str() << " ," ;
 }
 
 // Reader implementation ///////////////////////////////////////////////////////

Grid/serialisation/JSON_IO.h

@@ -54,7 +54,7 @@ namespace Grid
     void pop(void);
     template <typename U>
     void writeDefault(const std::string &s, const U &x);
-#if defined(GRID_CUDA) || defined(GRID_HIP)
+#ifdef __NVCC__
     void writeDefault(const std::string &s, const Grid::ComplexD &x) 
     { 
       std::complex<double> z(real(x),imag(x));
@@ -101,7 +101,7 @@ namespace Grid
     void readDefault(const std::string &s, std::vector<U> &output);
     template <typename U, typename P>
     void readDefault(const std::string &s, std::pair<U,P> &output);
-#if defined(GRID_CUDA) || defined(GRID_HIP)
+#ifdef __NVCC__
     void readDefault(const std::string &s, ComplexD &output)
     { 
       std::complex<double> z;

Grid/serialisation/Serialisation.h

@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include "BinaryIO.h"
 #include "TextIO.h"
 #include "XmlIO.h"
-#ifndef GRID_HIP
+#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
 #include "JSON_IO.h"
 #endif
 

Grid/stencil/Stencil.h

@@ -80,14 +80,11 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
 ///////////////////////////////////////////////////////////////////
 template<class cobj,class vobj,class compressor>
 void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
-				 commVector<cobj *> pointers,
-				 int dimension,int plane,
-				 int cbmask,compressor &compress,int type) __attribute__((noinline));
+				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
 
 template<class cobj,class vobj,class compressor>
-void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
-				 const Lattice<vobj> &rhs,
-				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -95,15 +92,14 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
   int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
   auto rhs_v = rhs.View(AcceleratorRead);
-  auto rhs_p = &rhs_v[0];
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
   accelerator_forNB(j, num, vobj::Nsimd(), {
-      compress.CompressExchange(p0,p1, rhs_p, j,
-				so+tp[2*j  ].second,
-				so+tp[2*j+1].second,
-				type);
+      compress.CompressExchange(p0,p1, &rhs_v[0], j,
+			      so+tp[2*j  ].second,
+			      so+tp[2*j+1].second,
+			      type);
   });
   rhs_v.ViewClose();
 }
@@ -135,8 +131,11 @@ class CartesianStencilAccelerator {
   int           _checkerboard;
   int           _npoints; // Move to template param?
   int           _osites;
+  int           _dirichlet;
   StencilVector _directions;
   StencilVector _distances;
+  StencilVector _comms_send;
+  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -230,12 +229,14 @@ public:
     void * recv_buf;
     Integer to_rank;
     Integer from_rank;
+    Integer do_send;
+    Integer do_recv;
     Integer bytes;
   };
   struct Merge {
     cobj * mpointer;
-    //    std::vector<scalar_object *> rpointers;
-    std::vector<cobj *> vpointers;
+    Vector<scalar_object *> rpointers;
+    Vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
   };
@@ -244,7 +245,20 @@ public:
     cobj * mpi_p;
     Integer buffer_size;
   };
-
+  struct CopyReceiveBuffer {
+    void * from_p;
+    void * to_p;
+    Integer bytes;
+  };
+  struct CachedTransfer {
+    Integer direction;
+    Integer OrthogPlane;
+    Integer DestProc;
+    Integer bytes;
+    Integer lane;
+    Integer cb;
+    void *recv_buf;
+  };
 
 protected:
   GridBase *                        _grid;
@@ -275,7 +289,8 @@ public:
   std::vector<Merge> MergersSHM;
   std::vector<Decompress> Decompressions;
   std::vector<Decompress> DecompressionsSHM;
-
+  std::vector<CopyReceiveBuffer> CopyReceiveBuffers ;
+  std::vector<CachedTransfer> CachedTransfers;
   ///////////////////////////////////////////////////////////
   // Unified Comms buffers for all directions
   ///////////////////////////////////////////////////////////
@@ -288,29 +303,6 @@ public:
   int u_comm_offset;
   int _unified_buffer_size;
 
-  /////////////////////////////////////////
-  // Timing info; ugly; possibly temporary
-  /////////////////////////////////////////
-  double commtime;
-  double mpi3synctime;
-  double mpi3synctime_g;
-  double shmmergetime;
-  double gathertime;
-  double gathermtime;
-  double halogtime;
-  double mergetime;
-  double decompresstime;
-  double comms_bytes;
-  double shm_bytes;
-  double splicetime;
-  double nosplicetime;
-  double calls;
-  std::vector<double> comm_bytes_thr;
-  std::vector<double> shm_bytes_thr;
-  std::vector<double> comm_time_thr;
-  std::vector<double> comm_enter_thr;
-  std::vector<double> comm_leave_thr;
-
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
@@ -337,11 +329,12 @@ public:
   //////////////////////////////////////////
   // Comms packet queue for asynch thread
   // Use OpenMP Tasks for cleaner ???
+  // must be called *inside* parallel region
   //////////////////////////////////////////
+  /*
   void CommunicateThreaded()
   {
 #ifdef GRID_OMP
-    // must be called in parallel region
     int mythread = omp_get_thread_num();
     int nthreads = CartesianCommunicator::nCommThreads;
 #else
@@ -350,90 +343,69 @@ public:
 #endif
     if (nthreads == -1) nthreads = 1;
     if (mythread < nthreads) {
-      comm_enter_thr[mythread] = usecond();
       for (int i = mythread; i < Packets.size(); i += nthreads) {
 	uint64_t bytes = _grid->StencilSendToRecvFrom(Packets[i].send_buf,
 						      Packets[i].to_rank,
 						      Packets[i].recv_buf,
 						      Packets[i].from_rank,
 						      Packets[i].bytes,i);
-	comm_bytes_thr[mythread] += bytes;
-	shm_bytes_thr[mythread] += 2*Packets[i].bytes-bytes; // Send + Recv.
-
       }
-      comm_leave_thr[mythread]= usecond();
-      comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
     }
   }
-
-  void CollateThreads(void)
-  {
-    int nthreads = CartesianCommunicator::nCommThreads;
-    double first=0.0;
-    double last =0.0;
-
-    for(int t=0;t<nthreads;t++) {
-
-      double t0 = comm_enter_thr[t];
-      double t1 = comm_leave_thr[t];
-      comms_bytes+=comm_bytes_thr[t];
-      shm_bytes  +=shm_bytes_thr[t];
-
-      comm_enter_thr[t] = 0.0;
-      comm_leave_thr[t] = 0.0;
-      comm_time_thr[t]   = 0.0;
-      comm_bytes_thr[t]=0;
-      shm_bytes_thr[t]=0;
-
-      if ( first == 0.0 ) first = t0;                   // first is t0
-      if ( (t0 > 0.0) && ( t0 < first ) ) first = t0;   // min time seen
-
-      if ( t1 > last ) last = t1;                       // max time seen
-
-    }
-    commtime+= last-first;
-  }
+  */
   ////////////////////////////////////////////////////////////////////////
   // Non blocking send and receive. Necessarily parallel.
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
-    // Buffers are gathered AND synchronised
-    // Copies are MPI ISend OR asynch copy on copy stream
     reqs.resize(Packets.size());
-    commtime-=usecond();
     for(int i=0;i<Packets.size();i++){
-      uint64_t bytes=_grid->StencilSendToRecvFromBegin(reqs[i],
-						     Packets[i].send_buf,
-						     Packets[i].to_rank,
-						     Packets[i].recv_buf,
-						     Packets[i].from_rank,
-						     Packets[i].bytes,i);
-      comms_bytes+=bytes;
-      shm_bytes  +=2*Packets[i].bytes-bytes;
+      _grid->StencilSendToRecvFromBegin(reqs[i],
+					Packets[i].send_buf,
+					Packets[i].to_rank,Packets[i].do_send,
+					Packets[i].recv_buf,
+					Packets[i].from_rank,Packets[i].do_recv,
+					Packets[i].bytes,i);
     }
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
-    // complete intranode
-    acceleratorCopySynchronise();
-    // complete MPI
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
-    // Everyone agrees we are all done
-    _grid->StencilBarrier(); 
-    commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    std::vector<std::vector<CommsRequest_t> > reqs;
-    this->CommunicateBegin(reqs);
-    this->CommunicateComplete(reqs);
+    if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
+      /////////////////////////////////////////////////////////
+      // several way threaded on different communicators.
+      // Cannot combine with Dirichlet operators
+      // This scheme is needed on Intel Omnipath for best performance
+      // Deprecate once there are very few omnipath clusters
+      /////////////////////////////////////////////////////////
+      int nthreads = CartesianCommunicator::nCommThreads;
+      int old = GridThread::GetThreads();
+      GridThread::SetThreads(nthreads);
+      thread_for(i,Packets.size(),{
+	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
+				       Packets[i].to_rank,Packets[i].do_send,
+				       Packets[i].recv_buf,
+				       Packets[i].from_rank,Packets[i].do_recv,
+				       Packets[i].bytes,i);
+      });
+      GridThread::SetThreads(old);
+    } else { 
+      /////////////////////////////////////////////////////////
+      // Concurrent and non-threaded asynch calls to MPI
+      /////////////////////////////////////////////////////////
+      std::vector<std::vector<CommsRequest_t> > reqs;
+      this->CommunicateBegin(reqs);
+      this->CommunicateComplete(reqs);
+    }
   }
 
   template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
@@ -471,31 +443,23 @@ public:
       sshift[1] = _grid->CheckerBoardShiftForCB(this->_checkerboard,dimension,shift,Odd);
       if ( sshift[0] == sshift[1] ) {
 	if (splice_dim) {
-	  splicetime-=usecond();
-	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
+	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
-	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
+	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp;
-	  nosplicetime+=usecond();
 	}
       } else {
 	if(splice_dim){
-	  splicetime-=usecond();
 	  // if checkerboard is unfavourable take two passes
 	  // both with block stride loop iteration
-	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx);
-	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx);
+	  auto tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx,point);
+	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
-	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
-	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
+	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx,point);
+	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx,point);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  nosplicetime+=usecond();
 	}
       }
     }
@@ -505,12 +469,10 @@ public:
   template<class compressor>
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
-    mpi3synctime_g-=usecond();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime_g+=usecond();
 
+    // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
-    halogtime-=usecond();
 
     u_comm_offset=0;
 
@@ -524,7 +486,6 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
-    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -537,14 +498,72 @@ public:
     Mergers.resize(0);
     MergersSHM.resize(0);
     Packets.resize(0);
-    calls++;
+    CopyReceiveBuffers.resize(0);
+    CachedTransfers.resize(0);
   }
-  void AddPacket(void *xmit,void * rcv, Integer to,Integer from,Integer bytes){
+  void AddCopy(void *from,void * to, Integer bytes)
+  {
+    //    std::cout << "Adding CopyReceiveBuffer "<<std::hex<<from<<" "<<to<<std::dec<<" "<<bytes<<std::endl;
+    CopyReceiveBuffer obj;
+    obj.from_p = from;
+    obj.to_p = to;
+    obj.bytes= bytes;
+    CopyReceiveBuffers.push_back(obj);
+  }
+  void CommsCopy()
+  {
+    //    These are device resident MPI buffers.
+    for(int i=0;i<CopyReceiveBuffers.size();i++){
+      cobj *from=(cobj *)CopyReceiveBuffers[i].from_p;
+      cobj *to  =(cobj *)CopyReceiveBuffers[i].to_p;
+      Integer words = CopyReceiveBuffers[i].bytes/sizeof(cobj);
+      //    std::cout << "CopyReceiveBuffer "<<std::hex<<from<<" "<<to<<std::dec<<" "<<words*sizeof(cobj)<<std::endl;
+      accelerator_forNB(j, words, cobj::Nsimd(), {
+	  coalescedWrite(to[j] ,coalescedRead(from [j]));
+      });
+    }
+  }
+  
+  Integer CheckForDuplicate(Integer direction, Integer OrthogPlane, Integer DestProc, void *recv_buf,Integer lane,Integer bytes,Integer cb)
+  {
+    CachedTransfer obj;
+    obj.direction   = direction;
+    obj.OrthogPlane = OrthogPlane;
+    obj.DestProc    = DestProc;
+    obj.recv_buf    = recv_buf;
+    obj.lane        = lane;
+    obj.bytes       = bytes;
+    obj.cb          = cb;
+
+    for(int i=0;i<CachedTransfers.size();i++){
+      if (   (CachedTransfers[i].direction  ==direction)
+	   &&(CachedTransfers[i].OrthogPlane==OrthogPlane)
+	   &&(CachedTransfers[i].DestProc   ==DestProc)
+	   &&(CachedTransfers[i].bytes      ==bytes)
+	   &&(CachedTransfers[i].lane       ==lane)
+	   &&(CachedTransfers[i].cb         ==cb)
+	     ){
+	//	std::cout << "Found duplicate plane dir "<<direction<<" plane "<< OrthogPlane<< " simd "<<lane << " relproc "<<DestProc<< " bytes "<<bytes <<std::endl;
+	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
+	return 1;
+      }
+    }
+
+    //    std::cout << "No duplicate plane dir "<<direction<<" plane "<< OrthogPlane<< " simd "<<lane << " relproc "<<DestProc<<"  bytes "<<bytes<<std::endl;
+    CachedTransfers.push_back(obj);
+    return 0;
+  }
+  void AddPacket(void *xmit,void * rcv,
+		 Integer to, Integer do_send,
+		 Integer from, Integer do_recv,
+		 Integer bytes){
     Packet p;
     p.send_buf = xmit;
     p.recv_buf = rcv;
     p.to_rank  = to;
     p.from_rank= from;
+    p.do_send  = do_send;
+    p.do_recv  = do_recv;
     p.bytes    = bytes;
     Packets.push_back(p);
   }
@@ -555,7 +574,7 @@ public:
     d.buffer_size = buffer_size;
     dv.push_back(d);
   }
-  void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
+  void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
     Merge m;
     m.type     = type;
     m.mpointer = merge_p;
@@ -564,18 +583,17 @@ public:
     mv.push_back(m);
   }
   template<class decompressor>  void CommsMerge(decompressor decompress)    {
+    CommsCopy();
     CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    assert(MergersSHM.size()==0);
-    assert(DecompressionsSHM.size()==0);
+    _grid->StencilBarrier();// Synch shared memory on a single nodes
+    CommsMerge(decompress,MergersSHM,DecompressionsSHM);
   }
 
   template<class decompressor>
-  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
-
-    
-    mergetime-=usecond();
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd)
+  {
     for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
@@ -585,10 +603,7 @@ public:
 	  decompress.Exchange(mp,vp0,vp1,type,o);
       });
     }
-    if ( mm.size() )    acceleratorFenceComputeStream();
-    mergetime+=usecond();
 
-    decompresstime-=usecond();
     for(int i=0;i<dd.size();i++){
       auto kp = dd[i].kernel_p;
       auto mp = dd[i].mpi_p;
@@ -596,9 +611,6 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
-    if ( dd.size() )    acceleratorFenceComputeStream();
-    decompresstime+=usecond();
-    
   }
   ////////////////////////////////////////
   // Set up routines
@@ -635,19 +647,58 @@ public:
       }
     }
   }
-
+  /// Introduce a block structure and switch off comms on boundaries
+  void DirichletBlock(const std::vector<int> &dirichlet_block)
+  {
+    this->_dirichlet = 1;
+    for(int ii=0;ii<this->_npoints;ii++){
+      int dimension    = this->_directions[ii];
+      int displacement = this->_distances[ii];
+      int shift = displacement;
+      int gd = _grid->_gdimensions[dimension];
+      int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      int ld = gd/pd;
+      int pc = _grid->_processor_coor[dimension];
+      ///////////////////////////////////////////
+      // Figure out dirichlet send and receive
+      // on this leg of stencil.
+      ///////////////////////////////////////////
+      int comm_dim        = _grid->_processors[dimension] >1 ;
+      int block = dirichlet_block[dimension];
+      this->_comms_send[ii] = comm_dim;
+      this->_comms_recv[ii] = comm_dim;
+      if ( block ) {
+	assert(abs(displacement) < ld );
+      
+	if( displacement > 0 ) {
+	  // High side, low side
+	  // | <--B--->|
+	  // |    |    |
+	  //           noR
+	  // noS
+	  if ( (ld*(pc+1) ) % block == 0 ) this->_comms_recv[ii] = 0;
+	  if ( ( ld*pc ) % block == 0    ) this->_comms_send[ii] = 0;
+	} else {
+	  // High side, low side
+	  // | <--B--->|
+	  // |    |    |
+	  //           noS
+	  // noR
+	  if ( (ld*(pc+1) ) % block == 0 ) this->_comms_send[ii] = 0;
+	  if ( ( ld*pc ) % block    == 0 ) this->_comms_recv[ii] = 0;
+	}
+      }
+    }
+  }
   CartesianStencil(GridBase *grid,
 		   int npoints,
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
 		   Parameters p)
-    : shm_bytes_thr(npoints),
-      comm_bytes_thr(npoints),
-      comm_enter_thr(npoints),
-      comm_leave_thr(npoints),
-      comm_time_thr(npoints)
   {
+    this->_dirichlet = 0;
     face_table_computed=0;
     _grid    = grid;
     this->parameters=p;
@@ -660,6 +711,8 @@ public:
     this->_simd_layout = _grid->_simd_layout; // copy simd_layout to give access to Accelerator Kernels
     this->_directions = StencilVector(directions);
     this->_distances  = StencilVector(distances);
+    this->_comms_send.resize(npoints); 
+    this->_comms_recv.resize(npoints); 
     this->same_node.resize(npoints);
 
     _unified_buffer_size=0;
@@ -678,24 +731,27 @@ public:
       int displacement = distances[i];
       int shift = displacement;
 
+      int gd = _grid->_gdimensions[dimension];
       int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      int ld = gd/pd;
       int rd = _grid->_rdimensions[dimension];
+      int pc = _grid->_processor_coor[dimension];
       this->_permute_type[point]=_grid->PermuteType(dimension);
 
       this->_checkerboard = checkerboard;
 
-      //////////////////////////
-      // the permute type
-      //////////////////////////
       int simd_layout     = _grid->_simd_layout[dimension];
       int comm_dim        = _grid->_processors[dimension] >1 ;
       int splice_dim      = _grid->_simd_layout[dimension]>1 && (comm_dim);
       int rotate_dim      = _grid->_simd_layout[dimension]>2;
 
+      this->_comms_send[ii] = comm_dim;
+      this->_comms_recv[ii] = comm_dim;
+
       assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
 
       int sshift[2];
-
       //////////////////////////
       // Underlying approach. For each local site build
       // up a table containing the npoint "neighbours" and whether they
@@ -796,6 +852,7 @@ public:
     GridBase *grid=_grid;
     const int Nsimd = grid->Nsimd();
 
+    int comms_recv      = this->_comms_recv[point];
     int fd              = _grid->_fdimensions[dimension];
     int ld              = _grid->_ldimensions[dimension];
     int rd              = _grid->_rdimensions[dimension];
@@ -852,7 +909,9 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-      if (!offnode) {
+
+      // Wrap locally dirichlet support case OR node local
+      if ( (offnode==0) || (comms_recv==0)  ) {
 
 	int permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
@@ -969,11 +1028,14 @@ public:
   }
 
   template<class compressor>
-  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
+  int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx, int point)
   {
     typedef typename cobj::vector_type vector_type;
     typedef typename cobj::scalar_type scalar_type;
 
+    int comms_send   = this->_comms_send[point] ;
+    int comms_recv   = this->_comms_recv[point] ;
+
     assert(rhs.Grid()==_grid);
     //	  conformable(_grid,rhs.Grid());
 
@@ -996,9 +1058,11 @@ public:
 
       int sx        = (x+sshift)%rd;
       int comm_proc = ((x+sshift)/rd)%pd;
-
+      
       if (comm_proc) {
 
+	
+	
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
 
@@ -1030,44 +1094,53 @@ public:
 	  recv_buf=this->u_recv_buf_p;
 	}
 
+
 	cobj *send_buf;
 	send_buf = this->u_send_buf_p; // Gather locally, must send
-
+	
 	////////////////////////////////////////////////////////
 	// Gather locally
 	////////////////////////////////////////////////////////
-	gathertime-=usecond();
 	assert(send_buf!=NULL);
-	Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so); face_idx++;
-	gathertime+=usecond();
+	if ( comms_send ) 
+	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so);
+	face_idx++;
 
-	///////////////////////////////////////////////////////////
-	// Build a list of things to do after we synchronise GPUs
-	// Start comms now???
-	///////////////////////////////////////////////////////////
-	AddPacket((void *)&send_buf[u_comm_offset],
-		  (void *)&recv_buf[u_comm_offset],
-		  xmit_to_rank,
-		  recv_from_rank,
-		  bytes);
+	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[u_comm_offset],0,bytes,cbmask);
+	if ( (!duplicate) ) { // Force comms for now
 
-	if ( compress.DecompressionStep() ) {
+	  ///////////////////////////////////////////////////////////
+	  // Build a list of things to do after we synchronise GPUs
+	  // Start comms now???
+	  ///////////////////////////////////////////////////////////
+	  AddPacket((void *)&send_buf[u_comm_offset],
+		    (void *)&recv_buf[u_comm_offset],
+		    xmit_to_rank, comms_send,
+		    recv_from_rank, comms_recv,
+		    bytes);
+	}
+	
+	if ( compress.DecompressionStep() && comms_recv ) {
 	  AddDecompress(&this->u_recv_buf_p[u_comm_offset],
 			&recv_buf[u_comm_offset],
 			words,Decompressions);
 	}
 	u_comm_offset+=words;
-      }
+	}
     }
     return 0;
   }
 
   template<class compressor>
-  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
+  int  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx,int point)
   {
     const int Nsimd = _grid->Nsimd();
 
     const int maxl =2;// max layout in a direction
+
+    int comms_send   = this->_comms_send[point] ;
+    int comms_recv   = this->_comms_recv[point] ;
+
     int fd = _grid->_fdimensions[dimension];
     int rd = _grid->_rdimensions[dimension];
     int ld = _grid->_ldimensions[dimension];
@@ -1099,8 +1172,8 @@ public:
     int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    std::vector<cobj *> rpointers(maxl);
-    std::vector<cobj *> spointers(maxl);
+    Vector<cobj *> rpointers(maxl);
+    Vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where
@@ -1132,12 +1205,11 @@ public:
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
 	}
-	gathermtime-=usecond();
 
-	Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
+	if ( comms_send )
+	  Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
 	face_idx++;
 
-	gathermtime+=usecond();
 	//spointers[0] -- low
 	//spointers[1] -- high
 
@@ -1166,8 +1238,13 @@ public:
 
 	    rpointers[i] = rp;
 
-	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
-
+	    int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,bytes,cbmask);
+	    if ( (!duplicate) ) { // Force comms for now
+	      AddPacket((void *)sp,(void *)rp,
+			xmit_to_rank,comms_send,
+			recv_from_rank,comms_recv,
+			bytes);
+	    }
 
 	  } else {
 
@@ -1176,7 +1253,9 @@ public:
 	  }
 	}
 
-	AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
+	if ( comms_recv ) {
+	  AddMerge(&this->u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
+	}
 
 	u_comm_offset     +=buffer_size;
       }

Grid/threads/Accelerator.cc

@@ -195,15 +195,12 @@ void acceleratorInit(void)
 #ifdef GRID_SYCL
 
 cl::sycl::queue *theGridAccelerator;
-cl::sycl::queue *theCopyAccelerator;
 void acceleratorInit(void)
 {
   int nDevices = 1;
   cl::sycl::gpu_selector selector;
   cl::sycl::device selectedDevice { selector };
   theGridAccelerator = new sycl::queue (selectedDevice);
-  //  theCopyAccelerator = new sycl::queue (selectedDevice);
-  theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
 
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
   zeInit(0);

Grid/threads/Accelerator.h

@@ -247,23 +247,21 @@ inline int  acceleratorIsCommunicable(void *ptr)
 //////////////////////////////////////////////
 // SyCL acceleration
 //////////////////////////////////////////////
+
 #ifdef GRID_SYCL
 NAMESPACE_END(Grid);
-#if 0
 #include <CL/sycl.hpp>
 #include <CL/sycl/usm.hpp>
+
+#define GRID_SYCL_LEVEL_ZERO_IPC
+
+#ifdef GRID_SYCL_LEVEL_ZERO_IPC
 #include <level_zero/ze_api.h>
 #include <CL/sycl/backend/level_zero.hpp>
-#else
-#include <sycl/CL/sycl.hpp>
-#include <sycl/usm.hpp>
-#include <level_zero/ze_api.h>
-#include <sycl/ext/oneapi/backend/level_zero.hpp>
 #endif
 NAMESPACE_BEGIN(Grid);
 
 extern cl::sycl::queue *theGridAccelerator;
-extern cl::sycl::queue *theCopyAccelerator;
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define GRID_SIMT
@@ -291,7 +289,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
       cgh.parallel_for(					\
       cl::sycl::nd_range<3>(global,local), \
       [=] (cl::sycl::nd_item<3> item) /*mutable*/     \
-      [[intel::reqd_sub_group_size(16)]]	      \
+      [[intel::reqd_sub_group_size(8)]]	      \
       {						      \
       auto iter1    = item.get_global_id(0);	      \
       auto iter2    = item.get_global_id(1);	      \
@@ -300,19 +298,19 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
      });	   			              \
     });
 
-#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
+#define accelerator_barrier(dummy) theGridAccelerator->wait();
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
-
-inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {  theCopyAccelerator->memcpy(to,from,bytes);}
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
-
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {
+  theGridAccelerator->memcpy(to,from,bytes);
+}
+inline void acceleratorCopySynchronise(void) {  theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
+inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
 inline int  acceleratorIsCommunicable(void *ptr)
 {
 #if 0
@@ -483,10 +481,9 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
 #define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) thread_for2d(iter1,num1,iter2,num2,{ __VA_ARGS__ });
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
-
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes);}
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
 inline void acceleratorCopySynchronise(void) {};
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }
@@ -513,16 +510,7 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
 inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 #endif
 
-//////////////////////////////////////////////
-// Fencing needed ONLY for SYCL
-//////////////////////////////////////////////
 
-#ifdef GRID_SYCL
-inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
-#else
-// Ordering within a stream guaranteed on Nvidia & AMD
-inline void acceleratorFenceComputeStream(void){ };
-#endif
 
 ///////////////////////////////////////////////////
 // Synchronise across local threads for divergence resynch

Grid/threads/Threads.h

@@ -72,20 +72,3 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define thread_region                                       DO_PRAGMA(omp parallel)
 #define thread_critical                                     DO_PRAGMA(omp critical)
 
-#ifdef GRID_OMP
-inline void thread_bcopy(void *from, void *to,size_t bytes)
-{
-  uint64_t *ufrom = (uint64_t *)from;
-  uint64_t *uto   = (uint64_t *)to;
-  assert(bytes%8==0);
-  uint64_t words=bytes/8;
-  thread_for(w,words,{
-      uto[w] = ufrom[w];
-  });
-}
-#else
-inline void thread_bcopy(void *from, void *to,size_t bytes)
-{
-  bcopy(from,to,bytes);
-}
-#endif

Grid/util/Init.cc

@@ -534,7 +534,6 @@ void Grid_init(int *argc,char ***argv)
 void Grid_finalize(void)
 {
 #if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
-  MPI_Barrier(MPI_COMM_WORLD);
   MPI_Finalize();
   Grid_unquiesce_nodes();
 #endif

Grid/util/Sha.h

@@ -27,7 +27,6 @@
     /*  END LEGAL */
 extern "C" {
 #include <openssl/sha.h>
-#include <openssl/evp.h>
 }
 #ifdef USE_IPP
 #include "ipp.h"
@@ -71,8 +70,10 @@ public:
   static inline std::vector<unsigned char> sha256(const void *data,size_t bytes)
   {
     std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
-    auto digest = EVP_get_digestbyname("SHA256");
-    EVP_Digest(data, bytes, &hash[0], NULL, digest, NULL);
+    SHA256_CTX sha256;
+    SHA256_Init  (&sha256);
+    SHA256_Update(&sha256, data,bytes);
+    SHA256_Final (&hash[0], &sha256);
     return hash;
   }
   static inline std::vector<int> sha256_seeds(const std::string &s)

README.md

@@ -148,7 +148,7 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-mkl[=<path>]`: use Intel MKL for FFT (and LAPACK if enabled) routines. A UNIX prefix containing the library can be specified (optional).
 - `--enable-numa`: enable NUMA first touch optimisation
 - `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
-- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 64 bytes).
+- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
 - `--enable-comms=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
 - `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.

benchmarks/Benchmark_comms.cc

@@ -217,9 +217,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu);
 	
 	    comm_proc = mpi_layout[mu]-1;
@@ -228,9 +228,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu+4][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu+4);
 	  
 	  }
@@ -309,9 +309,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu);
 	    Grid.StencilSendToRecvFromComplete(requests,mu);
 	    requests.resize(0);
@@ -322,9 +322,9 @@ int main (int argc, char ** argv)
 	    dbytes+=
 	      Grid.StencilSendToRecvFromBegin(requests,
 					      (void *)&xbuf[mu+4][0],
-					      xmit_to_rank,
+					      xmit_to_rank,1,
 					      (void *)&rbuf[mu+4][0],
-					      recv_from_rank,
+					      recv_from_rank,1,
 					      bytes,mu+4);
 	    Grid.StencilSendToRecvFromComplete(requests,mu+4);
 	    requests.resize(0);
@@ -411,8 +411,8 @@ int main (int argc, char ** argv)
 	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	    }
             int tid = omp_get_thread_num();
-	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
-					       (void *)&rbuf[dir][0], recv_from_rank, bytes,tid);
+	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,1,
+					       (void *)&rbuf[dir][0], recv_from_rank,1, bytes,tid);
 
 	    thread_critical { dbytes+=tbytes; }
 	  }

benchmarks/Benchmark_dwf_fp32.cc

@@ -32,18 +32,112 @@
 using namespace std;
 using namespace Grid;
 
-template<class d>
-struct scal {
-  d internal;
+////////////////////////
+/// Move to domains ////
+////////////////////////
+
+struct DomainDecomposition
+{
+  Coordinate Block;
+
+  DomainDecomposition(const Coordinate &_Block): Block(_Block){ assert(Block.size()==Nd);};
+  
+  template<class Field>
+  void ProjectDomain(Field &f,Integer domain)
+  {
+    GridBase *grid = f.Grid();
+    int dims = grid->Nd();
+    int isDWF= (dims==Nd+1);
+    assert((dims==Nd)||(dims==Nd+1));
+
+    Field   zz(grid);  zz = Zero();
+    LatticeInteger coor(grid);
+    LatticeInteger domaincoor(grid);
+    LatticeInteger mask(grid); mask = Integer(1);
+    LatticeInteger zi(grid);     zi = Integer(0);
+    for(int d=0;d<Nd;d++){
+      Integer B= Block[d];
+      if ( B ) {
+	LatticeCoordinate(coor,d+isDWF);
+	domaincoor = mod(coor,B);
+	mask = where(domaincoor==Integer(0),zi,mask);
+	mask = where(domaincoor==Integer(B-1),zi,mask);
+      }
+    }
+    if ( !domain )
+      f = where(mask==Integer(1),f,zz);
+    else 
+      f = where(mask==Integer(0),f,zz);
+  };
 };
 
-  Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT
-  };
+template<typename MomentaField>
+struct DirichletFilter: public MomentumFilterBase<MomentaField>
+{
+  Coordinate Block;
+  
+  DirichletFilter(const Coordinate &_Block): Block(_Block) {}
 
+  // Edge detect using domain projectors
+  void applyFilter (MomentaField &U) const override
+  {
+    DomainDecomposition Domains(Block);
+    GridBase *grid = U.Grid();
+    LatticeInteger  coor(grid);
+    LatticeInteger  face(grid);
+    LatticeInteger  one(grid);   one = 1;
+    LatticeInteger  zero(grid); zero = 0;
+    LatticeInteger  omega(grid);
+    LatticeInteger  omegabar(grid);
+    LatticeInteger  tmp(grid);
+
+    omega=one;    Domains.ProjectDomain(omega,0);
+    omegabar=one; Domains.ProjectDomain(omegabar,1);
+    
+    LatticeInteger nface(grid); nface=Zero();
+    
+    MomentaField projected(grid); projected=Zero();
+    typedef decltype(PeekIndex<LorentzIndex>(U,0)) MomentaLinkField;
+    MomentaLinkField  Umu(grid);
+    MomentaLinkField   zz(grid); zz=Zero();
+
+    int dims = grid->Nd();
+    Coordinate Global=grid->GlobalDimensions();
+    assert(dims==Nd);
+
+    for(int mu=0;mu<Nd;mu++){
+
+      if ( Block[mu]!=0 ) {
+
+	Umu = PeekIndex<LorentzIndex>(U,mu);
+
+	// Upper face 
+ 	tmp = Cshift(omegabar,mu,1);
+	tmp = tmp + omega;
+	face = where(tmp == Integer(2),one,zero );
+
+ 	tmp = Cshift(omega,mu,1);
+	tmp = tmp + omegabar;
+	face = where(tmp == Integer(2),one,face );
+
+	Umu = where(face,zz,Umu);
+
+	PokeIndex<LorentzIndex>(U, Umu, mu);
+      }
+    }
+  }
+};
+
+
+
+Gamma::Algebra Gmu [] = {
+			 Gamma::Algebra::GammaX,
+			 Gamma::Algebra::GammaY,
+			 Gamma::Algebra::GammaZ,
+			 Gamma::Algebra::GammaT
+};
+
+void Benchmark(int Ls, std::vector<int> Dirichlet);
 
 int main (int argc, char ** argv)
 {
@@ -52,24 +146,48 @@ int main (int argc, char ** argv)
 
   int threads = GridThread::GetThreads();
 
-  Coordinate latt4 = GridDefaultLatt();
   int Ls=16;
-  for(int i=0;i<argc;i++)
+  for(int i=0;i<argc;i++) {
     if(std::string(argv[i]) == "-Ls"){
       std::stringstream ss(argv[i+1]); ss >> Ls;
     }
-
+  }
+  std::vector<int> Dirichlet(5,0);
+  Benchmark(Ls,Dirichlet);
+  Coordinate latt4  = GridDefaultLatt();
+  Coordinate mpi    = GridDefaultMpi();
+  Coordinate shm;
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  /*
+  Dirichlet = std::vector<int>({0,
+				latt4[0]/mpi[0] * shm[0],
+				latt4[1]/mpi[1] * shm[1],
+				latt4[2]/mpi[2] * shm[2],
+				latt4[3]/mpi[3] * shm[3]});
+  */
+  Dirichlet = std::vector<int>({0,
+				latt4[0]/mpi[0] ,
+				latt4[1]/mpi[1] ,
+				latt4[2]/mpi[2] ,
+				latt4[3]/mpi[3] });
+  
+  std::cout << " Dirichlet block "<< Dirichlet<< std::endl;
+  Benchmark(Ls,Dirichlet);
+  Grid_finalize();
+  exit(0);
+}
+void Benchmark(int Ls, std::vector<int> Dirichlet)
+{
+  Coordinate latt4 = GridDefaultLatt();
   GridLogLayout();
 
   long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
 
-
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
-  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
   GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
   GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
   GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
@@ -80,26 +198,13 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
   GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
+
   std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
   GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
-  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
 
   LatticeFermionF src   (FGrid); random(RNG5,src);
-#if 0
-  src = Zero();
-  {
-    Coordinate origin({0,0,0,latt4[2]-1,0});
-    SpinColourVectorF tmp;
-    tmp=Zero();
-    tmp()(0)(0)=Complex(-2.0,0.0);
-    std::cout << " source site 0 " << tmp<<std::endl;
-    pokeSite(tmp,src,origin);
-  }
-#else
   RealD N2 = 1.0/::sqrt(norm2(src));
   src = src*N2;
-#endif
-
 
   LatticeFermionF result(FGrid); result=Zero();
   LatticeFermionF    ref(FGrid);    ref=Zero();
@@ -110,18 +215,18 @@ int main (int argc, char ** argv)
   LatticeGaugeFieldF Umu(UGrid);
   SU<Nc>::HotConfiguration(RNG4,Umu);
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
-#if 0
-  Umu=1.0;
-  for(int mu=0;mu<Nd;mu++){
-    LatticeColourMatrixF ttmp(UGrid);
-    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
-    //    if (mu !=2 ) ttmp = 0;
-    //    ttmp = ttmp* pow(10.0,mu);
-    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
-  }
-  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
-#endif
 
+  ////////////////////////////////////
+  // Apply BCs
+  ////////////////////////////////////
+  std::cout << GridLogMessage << "Applying BCs " << std::endl;
+  Coordinate Block(4);
+  for(int d=0;d<4;d++)  Block[d]= Dirichlet[d+1];
+
+  std::cout << GridLogMessage << "Dirichlet Block " << Block<< std::endl;
+  DirichletFilter<LatticeGaugeFieldF> Filter(Block);
+  Filter.applyFilter(Umu);
+  
   ////////////////////////////////////
   // Naive wilson implementation
   ////////////////////////////////////
@@ -191,11 +296,11 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
   DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  Dw.DirichletBlock(Dirichlet);
   int ncall =300;
 
   if (1) {
     FGrid->Barrier();
-    Dw.ZeroCounters();
     Dw.Dhop(src,result,0);
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
@@ -220,8 +325,6 @@ int main (int argc, char ** argv)
     double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
 
     std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
@@ -229,20 +332,13 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
     err = ref-result;
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-    //exit(0);
 
     if(( norm2(err)>1.0e-4) ) {
-      /*
-      std::cout << "RESULT\n " << result<<std::endl;
-      std::cout << "REF   \n " << ref   <<std::endl;
-      std::cout << "ERR   \n " << err   <<std::endl;
-      */
       std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
       FGrid->Barrier();
       exit(-1);
     }
     assert (norm2(err)< 1.0e-4 );
-    Dw.Report();
   }
 
   if (1)
@@ -294,13 +390,14 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result;
   std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-/*
-	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
-	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
-	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
-*/
+
+  if (  norm2(err) > 1.0e-4 ) {
+    std::cout << "Error vector is\n" <<err << std::endl;
+    std::cout << "Ref   vector is\n" <<ref << std::endl;
+    std::cout << "Result  vector is\n" <<result << std::endl;
   }
+  assert((norm2(err)<1.0e-4));
+  
   LatticeFermionF src_e (FrbGrid);
   LatticeFermionF src_o (FrbGrid);
   LatticeFermionF r_e   (FrbGrid);
@@ -330,7 +427,6 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
   {
-    Dw.ZeroCounters();
     FGrid->Barrier();
     Dw.DhopEO(src_o,r_e,DaggerNo);
     double t0=usecond();
@@ -352,7 +448,6 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
-    Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
   Dw.DhopOE(src_e,r_o,DaggerNo);
@@ -367,13 +462,7 @@ int main (int argc, char ** argv)
 
   err = r_eo-result;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
-    /*
-	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
-	std::cout<< "Deo REF\n " <<result  << std::endl;
-	std::cout<< "Deo ERR   \n " << err <<std::endl;
-    */
-  }
+  assert(norm2(err)<1.0e-4);
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
@@ -382,6 +471,4 @@ int main (int argc, char ** argv)
 
   assert(norm2(src_e)<1.0e-4);
   assert(norm2(src_o)<1.0e-4);
-  Grid_finalize();
-  exit(0);
 }

benchmarks/Benchmark_halo.cc

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

benchmarks/Benchmark_mooee.cc

@@ -81,8 +81,8 @@ int main (int argc, char ** argv)
     Vector<Coeff_t> diag = Dw.bs;
     Vector<Coeff_t> upper= Dw.cs;
     Vector<Coeff_t> lower= Dw.cs;
-    upper[Ls-1]=-Dw.mass_minus*upper[Ls-1];
-    lower[0]   =-Dw.mass_plus*lower[0];
+    upper[Ls-1]=-Dw.mass*upper[Ls-1];
+    lower[0]   =-Dw.mass*lower[0];
     
     LatticeFermion r_eo(FGrid);
     LatticeFermion src_e (FrbGrid);

benchmarks/Benchmark_wilson_sweep.cc

@@ -44,13 +44,6 @@ void bench_wilson (
 		   double const     volume,
 		   int const           dag );
 
-void bench_wilson_eo (
-       LatticeFermion &    src,
-       LatticeFermion & result,
-       WilsonFermionR &     Dw,
-       double const     volume,
-       int const           dag );
-
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
@@ -117,8 +110,8 @@ int main (int argc, char ** argv)
 	  bench_wilson(src,result,Dw,volume,DaggerYes);
 	  std::cout << "\t";
     // EO
-    bench_wilson_eo(src_o,result_e,Dw,volume,DaggerNo);
-    bench_wilson_eo(src_o,result_e,Dw,volume,DaggerYes);
+	  bench_wilson(src,result,Dw,volume,DaggerNo);
+	  bench_wilson(src,result,Dw,volume,DaggerYes);
 	  std::cout << std::endl;
 	}
     }

configure.ac

@@ -159,7 +159,7 @@ case ${ac_ZMOBIUS} in
 esac
 ############### Nc
 AC_ARG_ENABLE([Nc],
-    [AC_HELP_STRING([--enable-Nc=2|3|4|5], [enable number of colours])],
+    [AC_HELP_STRING([--enable-Nc=2|3|4], [enable number of colours])],
     [ac_Nc=${enable_Nc}], [ac_Nc=3])
 
 case ${ac_Nc} in

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

@@ -1,62 +0,0 @@
-#!/bin/sh
-##SBATCH -p PVC-SPR-QZEH 
-##SBATCH -p PVC-ICX-QZNW
-#SBATCH -p QZ1J-ICX-PVC
-##SBATCH -p QZ1J-SPR-PVC-2C
-
-#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
-
-export NT=8
-
-export I_MPI_OFFLOAD=1
-export I_MPI_OFFLOAD_TOPOLIB=level_zero
-export I_MPI_OFFLOAD_DOMAIN_SIZE=-1
-
-# export IGC_EnableLSCFenceUGMBeforeEOT=0
-# export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file=False"
-export SYCL_DEVICE_FILTER=gpu,level_zero
-#export IGC_ShaderDumpEnable=1 
-#export IGC_DumpToCurrentDir=1
-export I_MPI_OFFLOAD_CELL=tile
-export EnableImplicitScaling=0
-export EnableWalkerPartition=0
-export ZE_AFFINITY_MASK=0.0
-mpiexec -launcher ssh -n 1 -host localhost  ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 --device-mem 32768
-
-export ZE_AFFINITY_MASK=0
-export I_MPI_OFFLOAD_CELL=device
-export EnableImplicitScaling=1
-export EnableWalkerPartition=1
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-#mpiexec -launcher ssh -n 2 -host localhost  vtune -collect gpu-hotspots -knob gpu-sampling-interval=1 -data-limit=0 -r ./vtune_run4 -- ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-overlap --shm-mpi 1
-
-#mpiexec  -launcher ssh -n 1 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-overlap --shm-mpi 1
-
-#mpiexec  -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1
-
-#mpiexec  -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-overlap --shm-mpi 1
-
-#mpiexec  -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0
-
-#mpirun -np 2 ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 16.32.32.64 --accelerator-threads $NT --comms-sequential --shm-mpi 0
-#mpirun -np 2 ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT --comms-sequential --shm-mpi 1
-

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

@@ -1,32 +0,0 @@
-#!/bin/bash
-##SBATCH -p PVC-SPR-QZEH 
-##SBATCH -p PVC-ICX-QZNW
-
-#SBATCH -p QZ1J-ICX-PVC
-
-#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
-
-export NT=16
-
-# export IGC_EnableLSCFenceUGMBeforeEOT=0
-# export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file=False"
-#export IGC_ShaderDumpEnable=1 
-#export IGC_DumpToCurrentDir=1
-export I_MPI_OFFLOAD=1
-export I_MPI_OFFLOAD_TOPOLIB=level_zero
-export I_MPI_OFFLOAD_DOMAIN_SIZE=-1
-export SYCL_DEVICE_FILTER=gpu,level_zero
-export I_MPI_OFFLOAD_CELL=tile
-export EnableImplicitScaling=0
-export EnableWalkerPartition=0
-#export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
-#export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
-export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
-
-for i in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
-do
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 0  --device-mem 32768 > 1.1.1.2.log$i
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 0  --device-mem 32768 > 2.1.1.1.log$i 
-done
-
-

systems/PVC/benchmarks/wrap.sh

@@ -1,14 +0,0 @@
-#!/bin/sh
-
-export ZE_AFFINITY_MASK=0.$MPI_LOCALRANKID
-
-echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
-
-
-#if [ $MPI_LOCALRANKID = "0" ] 
-#then
-#  ~psteinbr/build_pti/ze_tracer -h $@
-#  onetrace --chrome-device-timeline $@
-#else
-  $@
-#fi

systems/PVC/config-command

@@ -1,16 +0,0 @@
-INSTALL=/nfs/site/home/azusayax/install
-../../configure \
-	--enable-simd=GPU \
-	--enable-gen-simd-width=64 \
-	--enable-comms=mpi-auto \
-	--disable-accelerator-cshift \
-	--disable-gparity \
-	--disable-fermion-reps \
-	--enable-shm=nvlink \
-	--enable-accelerator=sycl \
-	--enable-unified=no \
-	MPICXX=mpicxx \
-	CXX=dpcpp \
-	LDFLAGS="-fsycl-device-code-split=per_kernel -fsycl-device-lib=all -lze_loader -L$INSTALL/lib" \
-	CXXFLAGS="-fsycl-unnamed-lambda -fsycl -no-fma -I$INSTALL/include -Wno-tautological-compare"
-

systems/PVC/setup.sh

@@ -1,11 +0,0 @@
-export https_proxy=http://proxy-chain.intel.com:911
-export LD_LIBRARY_PATH=/nfs/site/home/azusayax/install/lib:$LD_LIBRARY_PATH
-
-module load intel-release
-source /opt/intel/oneapi/PVC_setup.sh
-#source /opt/intel/oneapi/ATS_setup.sh
-module load intel/mpich/pvc45.3
-export PATH=~/ATS/pti-gpu/tools/onetrace/:$PATH
-
-#clsh embargo-ci-neo-022845
-#source /opt/intel/vtune_amplifier/amplxe-vars.sh

systems/Tursa/dwf.4node.perf

@@ -1,25 +1,25 @@
-tu-c0r0n00 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n00 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n09 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n00 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n06 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n06 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n09 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n09 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n03 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n06 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n09 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n00 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n03 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n03 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n06 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n03 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n00 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n00 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n00 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n00 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n06 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n06 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n06 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n06 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n03 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n03 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n03 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n03 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n09 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n09 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n09 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n09 - 3 device=3 binding=--interleave=6,7
 OPENMPI detected
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
 AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
-AcceleratorCudaInit: ================================================
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
 OPENMPI detected
 AcceleratorCudaInit[0]: ========================
 AcceleratorCudaInit[0]: Device Number    : 0
@@ -33,11 +33,41 @@ AcceleratorCudaInit[0]:   pciBusID: 3
 AcceleratorCudaInit[0]:   pciDeviceID: 0 
 AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
 AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
-AcceleratorCudaInit: ================================================
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
 OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+OPENMPI detected
+AcceleratorCudaInit: using default device 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
+AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
 AcceleratorCudaInit[0]: ========================
 AcceleratorCudaInit[0]: Device Number    : 0
 AcceleratorCudaInit[0]: ========================
@@ -50,43 +80,25 @@ AcceleratorCudaInit[0]:   pciBusID: 3
 AcceleratorCudaInit[0]:   pciDeviceID: 0 
 AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
 AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
+AcceleratorCudaInit: assume user either uses
+AcceleratorCudaInit: a) IBM jsrun, or 
 AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
+AcceleratorCudaInit: Configure options --enable-setdevice=no 
+local rank 1 device 0 bus id: 0000:44:00.0
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
+local rank 0 device 0 bus id: 0000:03:00.0
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
-OPENMPI detected
-AcceleratorCudaInit: using default device 
-AcceleratorCudaInit: assume user either uses a) IBM jsrun, or 
-AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding 
-AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no 
 AcceleratorCudaInit: ================================================
+local rank 0 device 0 bus id: 0000:03:00.0
+AcceleratorCudaInit: ================================================
+AcceleratorCudaInit: ================================================
+local rank 2 device 0 bus id: 0000:84:00.0
 SharedMemoryMpi:  World communicator of size 16
 SharedMemoryMpi:  Node  communicator of size 4
-0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x7fcd80000000 for comms buffers 
+0SharedMemoryMpi:  SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x153960000000 for comms buffers 
 Setting up IPC
 
 __|__|__|__|__|__|__|__|__|__|__|__|__|__|__
@@ -116,7 +128,7 @@ 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.
-Current Grid git commit hash=9d2238148c56e3fbadfa95dcabf2b83d4bde14cd: (HEAD -> develop) uncommited changes
+Current Grid git commit hash=da06d15f73184ceb15d66d4e7e702b02fed7b940: (HEAD -> feature/dirichlet, develop) uncommited changes
 
 Grid : Message : ================================================ 
 Grid : Message : MPI is initialised and logging filters activated 
@@ -124,122 +136,102 @@ Grid : Message : ================================================
 Grid : Message : Requested 2147483648 byte stencil comms buffers 
 Grid : Message : MemoryManager Cache 34004218675 bytes 
 Grid : Message : MemoryManager::Init() setting up
-Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 32 LARGE 8
+Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
 Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
 Grid : Message : MemoryManager::Init() Using cudaMalloc
-Grid : Message : 1.198523 s : Grid Layout
-Grid : Message : 1.198530 s : 	Global lattice size  : 64 64 64 64 
-Grid : Message : 1.198534 s : 	OpenMP threads       : 4
-Grid : Message : 1.198535 s : 	MPI tasks            : 2 2 2 2 
-Grid : Message : 1.397615 s : Making s innermost grids
-Grid : Message : 1.441828 s : Initialising 4d RNG
-Grid : Message : 1.547973 s : Intialising parallel RNG with unique string 'The 4D RNG'
-Grid : Message : 1.547998 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
-Grid : Message : 1.954777 s : Initialising 5d RNG
-Grid : Message : 3.633825 s : Intialising parallel RNG with unique string 'The 5D RNG'
-Grid : Message : 3.633869 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
-Grid : Message : 12.162710 s : Initialised RNGs
-Grid : Message : 15.882520 s : Drawing gauge field
-Grid : Message : 15.816362 s : Random gauge initialised 
-Grid : Message : 17.279671 s : Setting up Cshift based reference 
-Grid : Message : 26.331426 s : *****************************************************************
-Grid : Message : 26.331452 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
-Grid : Message : 26.331454 s : *****************************************************************
-Grid : Message : 26.331456 s : *****************************************************************
-Grid : Message : 26.331458 s : * Benchmarking DomainWallFermionR::Dhop                  
-Grid : Message : 26.331459 s : * Vectorising space-time by 8
-Grid : Message : 26.331463 s : * VComplexF size is 64 B
-Grid : Message : 26.331465 s : * SINGLE precision 
-Grid : Message : 26.331467 s : * Using Overlapped Comms/Compute
-Grid : Message : 26.331468 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 26.331469 s : *****************************************************************
-Grid : Message : 28.413717 s : Called warmup
-Grid : Message : 56.418423 s : Called Dw 3000 times in 2.80047e+07 us
-Grid : Message : 56.418476 s : mflop/s =   3.79581e+07
-Grid : Message : 56.418479 s : mflop/s per rank =  2.37238e+06
-Grid : Message : 56.418481 s : mflop/s per node =  9.48953e+06
-Grid : Message : 56.418483 s : RF  GiB/s (base 2) =   77130
-Grid : Message : 56.418485 s : mem GiB/s (base 2) =   48206.3
-Grid : Message : 56.422076 s : norm diff   1.03481e-13
-Grid : Message : 56.456894 s : #### Dhop calls report 
-Grid : Message : 56.456899 s : WilsonFermion5D Number of DhopEO Calls   : 6002
-Grid : Message : 56.456903 s : WilsonFermion5D TotalTime   /Calls        : 4710.93 us
-Grid : Message : 56.456905 s : WilsonFermion5D CommTime    /Calls        : 3196.15 us
-Grid : Message : 56.456908 s : WilsonFermion5D FaceTime    /Calls        : 494.392 us
-Grid : Message : 56.456910 s : WilsonFermion5D ComputeTime1/Calls        : 44.4107 us
-Grid : Message : 56.456912 s : WilsonFermion5D ComputeTime2/Calls        : 1037.75 us
-Grid : Message : 56.456921 s : Average mflops/s per call                : 3.55691e+09
-Grid : Message : 56.456925 s : Average mflops/s per call per rank       : 2.22307e+08
-Grid : Message : 56.456928 s : Average mflops/s per call per node       : 8.89228e+08
-Grid : Message : 56.456930 s : Average mflops/s per call (full)         : 3.82915e+07
-Grid : Message : 56.456933 s : Average mflops/s per call per rank (full): 2.39322e+06
-Grid : Message : 56.456952 s : Average mflops/s per call per node (full): 9.57287e+06
-Grid : Message : 56.456954 s : WilsonFermion5D Stencil
-Grid : Message : 56.457016 s :  Stencil calls 3001
-Grid : Message : 56.457022 s :  Stencil halogtime 0
-Grid : Message : 56.457024 s :  Stencil gathertime 55.9154
-Grid : Message : 56.457026 s :  Stencil gathermtime 20.1073
-Grid : Message : 56.457028 s :  Stencil mergetime 18.5585
-Grid : Message : 56.457030 s :  Stencil decompresstime 0.0639787
-Grid : Message : 56.457032 s :  Stencil comms_bytes 4.02653e+08
-Grid : Message : 56.457034 s :  Stencil commtime 6379.93
-Grid : Message : 56.457036 s :  Stencil 63.1124 GB/s per rank
-Grid : Message : 56.457038 s :  Stencil 252.45 GB/s per node
-Grid : Message : 56.457040 s : WilsonFermion5D StencilEven
-Grid : Message : 56.457048 s : WilsonFermion5D StencilOdd
-Grid : Message : 56.457062 s : WilsonFermion5D Stencil     Reporti()
-Grid : Message : 56.457065 s : WilsonFermion5D StencilEven Reporti()
-Grid : Message : 56.457066 s : WilsonFermion5D StencilOdd  Reporti()
-Grid : Message : 79.259261 s : Compare to naive wilson implementation Dag to verify correctness
-Grid : Message : 79.259287 s : Called DwDag
-Grid : Message : 79.259288 s : norm dag result 12.0421
-Grid : Message : 79.271740 s : norm dag ref    12.0421
-Grid : Message : 79.287759 s : norm dag diff   7.63236e-14
-Grid : Message : 79.328100 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
-Grid : Message : 79.955951 s : src_e0.499997
-Grid : Message : 80.633620 s : src_o0.500003
-Grid : Message : 80.164163 s : *********************************************************
-Grid : Message : 80.164168 s : * Benchmarking DomainWallFermionF::DhopEO                
-Grid : Message : 80.164170 s : * Vectorising space-time by 8
-Grid : Message : 80.164172 s : * SINGLE precision 
-Grid : Message : 80.164174 s : * Using Overlapped Comms/Compute
-Grid : Message : 80.164177 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 80.164178 s : *********************************************************
-Grid : Message : 93.797635 s : Deo mflop/s =   3.93231e+07
-Grid : Message : 93.797670 s : Deo mflop/s per rank   2.45769e+06
-Grid : Message : 93.797672 s : Deo mflop/s per node   9.83077e+06
-Grid : Message : 93.797674 s : #### Dhop calls report 
-Grid : Message : 93.797675 s : WilsonFermion5D Number of DhopEO Calls   : 3001
-Grid : Message : 93.797677 s : WilsonFermion5D TotalTime   /Calls        : 4542.83 us
-Grid : Message : 93.797679 s : WilsonFermion5D CommTime    /Calls        : 2978.97 us
-Grid : Message : 93.797681 s : WilsonFermion5D FaceTime    /Calls        : 602.287 us
-Grid : Message : 93.797683 s : WilsonFermion5D ComputeTime1/Calls        : 67.1416 us
-Grid : Message : 93.797685 s : WilsonFermion5D ComputeTime2/Calls        : 1004.07 us
-Grid : Message : 93.797713 s : Average mflops/s per call                : 3.30731e+09
-Grid : Message : 93.797717 s : Average mflops/s per call per rank       : 2.06707e+08
-Grid : Message : 93.797719 s : Average mflops/s per call per node       : 8.26827e+08
-Grid : Message : 93.797721 s : Average mflops/s per call (full)         : 3.97084e+07
-Grid : Message : 93.797727 s : Average mflops/s per call per rank (full): 2.48178e+06
-Grid : Message : 93.797732 s : Average mflops/s per call per node (full): 9.92711e+06
-Grid : Message : 93.797735 s : WilsonFermion5D Stencil
-Grid : Message : 93.797746 s : WilsonFermion5D StencilEven
-Grid : Message : 93.797758 s : WilsonFermion5D StencilOdd
-Grid : Message : 93.797769 s :  Stencil calls 3001
-Grid : Message : 93.797773 s :  Stencil halogtime 0
-Grid : Message : 93.797776 s :  Stencil gathertime 56.7458
-Grid : Message : 93.797780 s :  Stencil gathermtime 22.6504
-Grid : Message : 93.797782 s :  Stencil mergetime 21.1913
-Grid : Message : 93.797786 s :  Stencil decompresstime 0.0556481
-Grid : Message : 93.797788 s :  Stencil comms_bytes 2.01327e+08
-Grid : Message : 93.797791 s :  Stencil commtime 2989.33
-Grid : Message : 93.797795 s :  Stencil 67.3484 GB/s per rank
-Grid : Message : 93.797798 s :  Stencil 269.394 GB/s per node
-Grid : Message : 93.797801 s : WilsonFermion5D Stencil     Reporti()
-Grid : Message : 93.797803 s : WilsonFermion5D StencilEven Reporti()
-Grid : Message : 93.797805 s : WilsonFermion5D StencilOdd  Reporti()
-Grid : Message : 93.873429 s : r_e6.02111
-Grid : Message : 93.879931 s : r_o6.02102
-Grid : Message : 93.885912 s : res12.0421
-Grid : Message : 94.876555 s : norm diff   0
-Grid : Message : 95.485643 s : norm diff even  0
-Grid : Message : 95.581236 s : norm diff odd   0
+Grid : Message : 1.875883 s : Grid Layout
+Grid : Message : 1.875893 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 1.875897 s : 	OpenMP threads       : 4
+Grid : Message : 1.875898 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 1.993571 s : Initialising 4d RNG
+Grid : Message : 2.881990 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 2.882370 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 2.495044 s : Initialising 5d RNG
+Grid : Message : 4.120900 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 4.121350 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 15.268010 s : Drawing gauge field
+Grid : Message : 16.234025 s : Random gauge initialised 
+Grid : Message : 16.234057 s : Applying BCs 
+Grid : Message : 16.365565 s : Setting up Cshift based reference 
+Grid : Message : 44.512418 s : *****************************************************************
+Grid : Message : 44.512448 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 44.512450 s : *****************************************************************
+Grid : Message : 44.512451 s : *****************************************************************
+Grid : Message : 44.512452 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 44.512453 s : * Vectorising space-time by 8
+Grid : Message : 44.512454 s : * VComplexF size is 64 B
+Grid : Message : 44.512456 s : * SINGLE precision 
+Grid : Message : 44.512459 s : * Using Overlapped Comms/Compute
+Grid : Message : 44.512460 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 44.512461 s : *****************************************************************
+Grid : Message : 46.389070 s : Called warmup
+Grid : Message : 49.211265 s : Called Dw 300 times in 2.82203e+06 us
+Grid : Message : 49.211295 s : mflop/s =   3.76681e+07
+Grid : Message : 49.211297 s : mflop/s per rank =  2.35425e+06
+Grid : Message : 49.211299 s : mflop/s per node =  9.41702e+06
+Grid : Message : 49.211301 s : RF  GiB/s (base 2) =   76540.6
+Grid : Message : 49.211308 s : mem GiB/s (base 2) =   47837.9
+Grid : Message : 49.214868 s : norm diff   1.06409e-13
+Grid : Message : 92.647781 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 92.647816 s : Called DwDag
+Grid : Message : 92.647817 s : norm dag result 12.0421
+Grid : Message : 92.801806 s : norm dag ref    12.0421
+Grid : Message : 92.817724 s : norm dag diff   7.21921e-14
+Grid : Message : 92.858973 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 93.210378 s : src_e0.499997
+Grid : Message : 93.583286 s : src_o0.500003
+Grid : Message : 93.682468 s : *********************************************************
+Grid : Message : 93.682471 s : * Benchmarking DomainWallFermionF::DhopEO                
+Grid : Message : 93.682472 s : * Vectorising space-time by 8
+Grid : Message : 93.682473 s : * SINGLE precision 
+Grid : Message : 93.682475 s : * Using Overlapped Comms/Compute
+Grid : Message : 93.682476 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 93.682477 s : *********************************************************
+Grid : Message : 95.162342 s : Deo mflop/s =   3.92487e+07
+Grid : Message : 95.162387 s : Deo mflop/s per rank   2.45305e+06
+Grid : Message : 95.162389 s : Deo mflop/s per node   9.81219e+06
+Grid : Message : 95.232801 s : r_e6.02111
+Grid : Message : 95.240061 s : r_o6.02102
+Grid : Message : 95.245975 s : res12.0421
+Grid : Message : 95.833402 s : norm diff   0
+Grid : Message : 96.573829 s : norm diff even  0
+Grid : Message : 96.868272 s : norm diff odd   0
+ Dirichlet block [0 64 64 32 32]
+Grid : Message : 97.756909 s : Grid Layout
+Grid : Message : 97.756911 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 97.756921 s : 	OpenMP threads       : 4
+Grid : Message : 97.756922 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 97.897085 s : Initialising 4d RNG
+Grid : Message : 97.965061 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 97.965097 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 98.367431 s : Initialising 5d RNG
+Grid : Message : 99.752745 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 99.752790 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 111.290148 s : Drawing gauge field
+Grid : Message : 112.349289 s : Random gauge initialised 
+Grid : Message : 112.349320 s : Applying BCs 
+Grid : Message : 113.948740 s : Setting up Cshift based reference 
+Grid : Message : 140.320415 s : *****************************************************************
+Grid : Message : 140.320443 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 140.320444 s : *****************************************************************
+Grid : Message : 140.320445 s : *****************************************************************
+Grid : Message : 140.320446 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 140.320447 s : * Vectorising space-time by 8
+Grid : Message : 140.320448 s : * VComplexF size is 64 B
+Grid : Message : 140.320450 s : * SINGLE precision 
+Grid : Message : 140.320451 s : * Using Overlapped Comms/Compute
+Grid : Message : 140.320452 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 140.320453 s : *****************************************************************
+Grid : Message : 142.296150 s : Called warmup
+Grid : Message : 144.397678 s : Called Dw 300 times in 2.36719e+06 us
+Grid : Message : 144.397700 s : mflop/s =   4.49058e+07
+Grid : Message : 144.397702 s : mflop/s per rank =  2.80661e+06
+Grid : Message : 144.397704 s : mflop/s per node =  1.12265e+07
+Grid : Message : 144.397706 s : RF  GiB/s (base 2) =   91247.6
+Grid : Message : 144.397708 s : mem GiB/s (base 2) =   57029.7
+Grid : Message : 144.401269 s : norm diff   9.78944e-14
+Grid : Message : 186.885460 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 186.885492 s : Called DwDag
+Grid : Message : 186.885493 s : norm dag result 10.4157
+Grid : Message : 186.897154 s : norm dag ref    11.2266
+Grid : Message : 186.912538 s : norm dag diff   0.484633