Reduce the loop over exterior for GPU to indirection table

14TF/s on a Perlmutter node

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/allocator/MemoryManager.cc

@@ -40,7 +40,7 @@ void MemoryManager::PrintBytes(void)
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 8, 16, 8, 16 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils

Grid/allocator/MemoryManager.h

@@ -36,11 +36,6 @@ NAMESPACE_BEGIN(Grid);
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
 
-#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*/
 ////////////////////////////////////////////////////////////////////////////
 // Advise the LatticeAccelerator class
@@ -97,9 +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);
@@ -119,8 +113,6 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
-  static uint64_t     DeviceEvictions;
-  static uint64_t     DeviceDestroy;
  
  private:
 #ifndef GRID_UVM
@@ -178,7 +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 dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define MAXLINE 512
+#define dprintf(...)
-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(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -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(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\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
+  // Make CPU consistent, remove from Accelerator, remove entry
-  // Cannot be acclocked. If allocated must be in LRU pool.
+  // Cannot be locked. If allocated must be in LRU pool.
-  //
-  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
-  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
-  //                          but there is a weakness where CpuLock entries are attempted for erase
-  //                          Take these OUT LRU queue when CPU locked?
-  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
+  assert(AccCache.accLock==0);
-	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0); // Cannot evict so logic bomb
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   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;
+  uint64_t CpuPtr = AccCache.CpuPtr;
-  DeviceEvictions++;
+  EntryErase(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)bytes);
-		    (uint64_t)AccCache.accLock);
     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 remove from LRU
-  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
-    dprintf("AccCache entry removed from LRU \n");
     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) {
-  //  if (!AccCache.AccPtr) {
+     EvictVictims(bytes);
-  //    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 << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-}
-void  MemoryManager::PrintAll(void)
-{
-  Print();
-  std::cout << GridLogMessage << std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -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
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\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;
@@ -97,6 +98,7 @@ 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)    ;
 
@@ -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();
@@ -370,6 +372,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
+  if ( dor ) {
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
       ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
@@ -377,7 +380,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(rrq);
       off_node_bytes+=bytes;
     }
+  }
   
+  if (dox) {
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
       ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
@@ -385,16 +390,16 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       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);
     }
+  }
   
-  //  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
-  //    this->StencilSendToRecvFromComplete(list,dir);
+    this->StencilSendToRecvFromComplete(list,dir);
-  //  }
+    list.resize(0);
+  }
 
   return off_node_bytes;
 }

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
+  // 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 OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
-  static void OptimalCommunicatorSharedMemory(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,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

@@ -152,7 +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
@@ -165,8 +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);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
 static inline int divides(int a,int b)
 {
@@ -221,7 +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.
@@ -294,6 +294,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
+  SHM = ShmDims;
   
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
@@ -341,7 +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
@@ -353,6 +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/lattice/Lattice_base.h

@@ -129,7 +129,7 @@ public:
     
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWrite);
+    auto me  = View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -152,7 +152,7 @@ public:
 
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWrite);
+    auto me  = View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -174,7 +174,7 @@ public:
     this->checkerboard=cb;
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
-    auto me  = View(AcceleratorWrite);
+    auto me  = View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       auto tmp = eval(ss,exprCopy);
       coalescedWrite(me[ss],tmp);
@@ -245,7 +245,7 @@ public:
   ///////////////////////////////////////////
   // user defined constructor
   ///////////////////////////////////////////
-  Lattice(GridBase *grid,ViewMode mode=AcceleratorWrite) { 
+  Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
     assert((((uint64_t)&this->_odata[0])&0xF) ==0);
@@ -288,7 +288,7 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWrite);
+    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
@@ -303,7 +303,7 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWrite);
+    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));

Grid/lattice/Lattice_peekpoke.h

@@ -125,6 +125,12 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
+template<class vobj>
+typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
+  typename vobj::scalar_object s;
+  peekSite(s,l,site);
+  return s;
+}        
 template<class vobj,class sobj>
 void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         

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);
@@ -139,7 +136,7 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -148,7 +145,7 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sumD_gpu_large(arg,osites);
 #else
   return sumD_cpu(arg,osites);
@@ -158,13 +155,13 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-  Integer osites = arg.Grid()->oSites();
+#if defined(GRID_CUDA)||defined(GRID_HIP)
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  typename vobj::scalar_object ssum;
   autoView( arg_v, arg, AcceleratorRead);
-  ssum= sum_gpu(&arg_v[0],osites);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
+  Integer osites = arg.Grid()->oSites();
   auto ssum= sum_cpu(&arg_v[0],osites);
 #endif  
   arg.Grid()->GlobalSum(ssum);
@@ -174,7 +171,7 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 template<class vobj>
 inline typename vobj::scalar_object sum_large(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();
   auto ssum= sum_gpu_large(&arg_v[0],osites);
@@ -238,10 +235,11 @@ 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_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_slice_gpu.h

@@ -1,126 +0,0 @@
-NAMESPACE_BEGIN(Grid);
-
-// If NOT CUDA or HIP -- we should provide
-// -- atomicAdd(float *,float)
-// -- atomicAdd(double *,double)
-// 
-// Augment CUDA with complex atomics
-#if !defined(GRID_HIP) || !defined(GRID_CUDA)
-inline void atomicAdd(float *acc,float elem)
-{
-  *acc += elem;
-}
-inline void atomicAdd(double *acc,double elem)
-{
-  *acc += elem;
-}
-#endif
-inline void atomicAdd(ComplexD *accum,ComplexD & elem)
-{
-  double *a_p = (double *)accum;
-  double *e_p = (double *)&elem;
-  for(int w=0;w<2;w++){
-    atomicAdd(&a_p[w],e_p[w]);
-  }
-}
-inline void atomicAdd(ComplexF *accum,ComplexF & elem)
-{
-  float *a_p = (float *)accum;
-  float *e_p = (float *)&elem;
-  for(int w=0;w<2;w++){
-    atomicAdd(&a_p[w],e_p[w]);
-  }
-}
-// Augment CUDA with vobj atomics
-template<class vobj> accelerator_inline void atomicAdd(vobj *accum, vobj & elem)
-{
-  typedef typename vobj::scalar_type scalar_type;
-  scalar_type *a_p= (scalar_type *)accum;
-  scalar_type *e_p= (scalar_type *)& elem;
-  for(int w=0;w<vobj::Nsimd();w++){
-    atomicAdd(&a_p[w],e_p[w]);
-  }
-}
-// Atomics based slice sum
-template<class vobj> inline void sliceSumGpu(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
-{
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_object::scalar_type scalar_type;
-  GridBase  *grid = Data.Grid();
-  assert(grid!=NULL);
-
-  const int    Nd = grid->_ndimension;
-  const int Nsimd = grid->Nsimd();
-
-  assert(orthogdim >= 0);
-  assert(orthogdim < Nd);
-
-  int fd=grid->_fdimensions[orthogdim];
-  int ld=grid->_ldimensions[orthogdim];
-  int rd=grid->_rdimensions[orthogdim];
-
-  // Move to device memory and copy in / out
-  Vector<vobj> lvSum(rd); // will locally sum vectors first
-  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
-  ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
-
-  result.resize(fd); // And then global sum to return the same vector to every node 
-  for(int r=0;r<rd;r++){
-    lvSum[r]=Zero();
-  }
-
-  int e1=    grid->_slice_nblock[orthogdim];
-  int e2=    grid->_slice_block [orthogdim];
-  int stride=grid->_slice_stride[orthogdim];
-
-  // sum over reduced dimension planes, breaking out orthog dir
-  // Parallel over orthog direction
-  autoView( Data_v, Data, AcceleratorRead);
-  auto lvSum_p=&lvSum[0];
-  int ostride = grid->_ostride[orthogdim]; 
-  accelerator_for( ree,rd*e1*e2,1, {
-    int b = ree%e2;
-    int re= ree/e2;
-    int n=re%e1;
-    int r=re/e1;
-    int so=r*ostride;
-    int ss=so+n*stride+b;
-    atomicAdd(&lvSum_p[r],Data_v[ss]);
-  });
-
-  // Sum across simd lanes in the plane, breaking out orthog dir.
-  Coordinate icoor(Nd);
-
-  for(int rt=0;rt<rd;rt++){
-
-    extract(lvSum[rt],extracted);
-
-    for(int idx=0;idx<Nsimd;idx++){
-
-      grid->iCoorFromIindex(icoor,idx);
-
-      int ldx =rt+icoor[orthogdim]*rd;
-
-      lsSum[ldx]=lsSum[ldx]+extracted[idx];
-
-    }
-  }
-  
-  // sum over nodes.
-  for(int t=0;t<fd;t++){
-    int pt = t/ld; // processor plane
-    int lt = t%ld;
-    if ( pt == grid->_processor_coor[orthogdim] ) {
-      result[t]=lsSum[lt];
-    } else {
-      result[t]=Zero();
-    }
-
-  }
-  scalar_type * ptr = (scalar_type *) &result[0];
-  int words = fd*sizeof(sobj)/sizeof(scalar_type);
-  grid->GlobalSumVector(ptr, words);
-}
-
-
-NAMESPACE_END(Grid);

Grid/log/Log.cc

@@ -65,38 +65,30 @@ 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("Error"))       GridLogError.Active(1);
-    if (logstreams[i] == std::string("Memory"))      GridLogMemory.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("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 << log.colour() <<  std::left;
-      stream <<  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 << log.colour();
-      stream <<  std::right;
       stream.flags(f);
+
       return stream;
     } else { 
       return devnull;
@@ -181,12 +180,9 @@ 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/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;
+// Dress the output; use std::chrono
-typedef  std::chrono::high_resolution_clock   GridClock;
+// 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/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/ActionBase.h

@@ -40,6 +40,29 @@ class Action
 
 public:
   bool is_smeared = false;
+  RealD deriv_norm_sum;
+  RealD deriv_max_sum;
+  int   deriv_num;
+  RealD deriv_us;
+  RealD S_us;
+  RealD refresh_us;
+  void  reset_timer(void)        {
+    deriv_us = S_us = refresh_us = 0.0;
+    deriv_num=0;
+    deriv_norm_sum = deriv_max_sum=0.0;
+  }
+  void  deriv_log(RealD nrm, RealD max) { deriv_max_sum+=max; deriv_norm_sum+=nrm; deriv_num++;}
+  RealD deriv_max_average(void)         { return deriv_max_sum/deriv_num; };
+  RealD deriv_norm_average(void)        { return deriv_norm_sum/deriv_num; };
+  RealD deriv_timer(void)        { return deriv_us; };
+  RealD S_timer(void)            { return deriv_us; };
+  RealD refresh_timer(void)      { return deriv_us; };
+  void deriv_timer_start(void)   { deriv_us-=usecond(); }
+  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
+  void refresh_timer_start(void) { refresh_us-=usecond(); }
+  void refresh_timer_stop(void)  { refresh_us+=usecond(); }
+  void S_timer_start(void)       { S_us-=usecond(); }
+  void S_timer_stop(void)        { S_us+=usecond(); }
   // Heatbath?
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action

Grid/qcd/action/ActionCore.h

@@ -37,6 +37,10 @@ NAMESPACE_CHECK(ActionSet);
 #include <Grid/qcd/action/ActionParams.h>
 NAMESPACE_CHECK(ActionParams);
 
+#include <Grid/qcd/action/filters/MomentumFilter.h>
+#include <Grid/qcd/action/filters/DirichletFilter.h>
+#include <Grid/qcd/action/filters/DDHMCFilter.h>
+
 ////////////////////////////////////////////
 // Gauge Actions
 ////////////////////////////////////////////

Grid/qcd/action/ActionParams.h

@@ -37,24 +37,32 @@ NAMESPACE_BEGIN(Grid);
 // These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
   Coordinate twists;
-  GparityWilsonImplParams() : twists(Nd, 0) {};
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  GparityWilsonImplParams() : twists(Nd, 0), dirichlet(Nd, 0) {};
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
+    dirichlet.resize(Nd,0);
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
+    dirichlet.resize(Nd,0);
   }
 };
 
 struct StaggeredImplParams {
-  StaggeredImplParams()  {};
+  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  StaggeredImplParams()
+  {
+    dirichlet.resize(Nd,0);
+  };
 };
   
   struct OneFlavourRationalParams : Serializable {

Grid/qcd/action/fermion/CloverHelpers.h

@@ -204,18 +204,15 @@ public:
   typedef WilsonCloverHelpers<Impl> Helpers;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+  static void MassTerm(CloverField& Clover, RealD diag_mass) {
     Clover += diag_mass;
   }
 
-  static void InvertClover(CloverField& InvClover,
+  static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
-                            const CloverDiagonalField& diagonal,
+                          CloverTriangleField& Triangle,
-                            const CloverTriangleField& triangle,
+                          RealD csw_t, RealD diag_mass) {
-                            CloverDiagonalField&       diagonalInv,
-                            CloverTriangleField&       triangleInv,
-                            bool fixedBoundaries) {
 
-    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+    // Do nothing
   }
 
   // TODO: implement Cmunu for better performances with compact layout, but don't do it
@@ -240,17 +237,9 @@ public:
   template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
 
-  // Can this be avoided?
+  static void MassTerm(CloverField& Clover, RealD diag_mass) {
-  static void IdentityTimesC(const CloverField& in, RealD c) {
+    // do nothing!
-    int DimRep = Impl::Dimension;
+    // mass term is multiplied to exp(Clover) below
-
-    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) {
@@ -265,62 +254,175 @@ public:
     return NMAX;
   }
 
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+  static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
-  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
+  static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
 
-    GridBase* grid = Clover.Grid();
+  	  typedef iMatrix<ComplexD,6> mat;
-    CloverField ExpClover(grid);
 
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
+  	  RealD qn[6];
+  	  RealD qnold[6];
+  	  RealD p[5];
+  	  RealD trA2, trA3, trA4;
 
-    Clover *= (1.0/diag_mass);
+  	  mat A2, A3, A4, A5;
+  	  A2 = alpha * alpha * arg * arg;
+  	  A3 = alpha * arg * A2;
+  	  A4 = A2 * A2;
+  	  A5 = A2 * A3;
 
-    // Taylor expansion, slow but generic
+  	  trA2 = toReal( trace(A2) );
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
+  	  trA3 = toReal( trace(A3) );
-    // qN = cN
+  	  trA4 = toReal( trace(A4));
-    // qn = cn + qn+1 X
+
+  	  p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
+  	  p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
+  	  p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
+  	  p[3] = trA3 / 3.0;
+  	  p[4] = 0.5 * trA2;
+
+  	  qnold[0] = cN[Niter];
+  	  qnold[1] = 0.0;
+  	  qnold[2] = 0.0;
+  	  qnold[3] = 0.0;
+  	  qnold[4] = 0.0;
+  	  qnold[5] = 0.0;
+
+  	  for(int i = Niter-1; i >= 0; i--)
+  	  {
+  	   qn[0] = p[0] * qnold[5] + cN[i];
+  	   qn[1] = p[1] * qnold[5] + qnold[0];
+  	   qn[2] = p[2] * qnold[5] + qnold[1];
+  	   qn[3] = p[3] * qnold[5] + qnold[2];
+  	   qn[4] = p[4] * qnold[5] + qnold[3];
+  	   qn[5] = qnold[4];
+
+  	   qnold[0] = qn[0];
+  	   qnold[1] = qn[1];
+  	   qnold[2] = qn[2];
+  	   qnold[3] = qn[3];
+  	   qnold[4] = qn[4];
+  	   qnold[5] = qn[5];
+  	  }
+
+  	  mat unit(1.0);
+
+  	  dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
+
+    }
+
+  static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
+
+    GridBase* grid = Diagonal.Grid();
+    int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
+
+    //
+    // Implementation completely in Daniel's layout
+    //
+
+    // Taylor expansion with Cayley-Hamilton recursion
+    // underlying Horner scheme as above
     std::vector<RealD> cn(NMAX+1);
     cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
+    for (int i=1; i<=NMAX; i++){
       cn[i] = cn[i-1] / RealD(i);
-
-    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,
+      // Taken over from Daniel's implementation
-                            const CloverDiagonalField& diagonal,
+      conformable(Diagonal, Triangle);
-                            const CloverTriangleField& triangle,
-                            CloverDiagonalField&       diagonalInv,
-                            CloverTriangleField&       triangleInv,
-                            bool fixedBoundaries) {
 
-    if (fixedBoundaries)
+      long lsites = grid->lSites();
     {
-      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
+      typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
+      typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
+      typedef iMatrix<ComplexD,6> mat;
+
+      autoView(diagonal_v,  Diagonal,  CpuRead);
+      autoView(triangle_v,  Triangle,  CpuRead);
+      autoView(diagonalExp_v, Diagonal, CpuWrite);
+      autoView(triangleExp_v, Triangle, CpuWrite);
+
+      thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
+
+    	  mat srcCloverOpUL(0.0); // upper left block
+    	  mat srcCloverOpLR(0.0); // lower right block
+    	  mat ExpCloverOp;
+
+        scalar_object_diagonal diagonal_tmp     = Zero();
+        scalar_object_diagonal diagonal_exp_tmp = Zero();
+        scalar_object_triangle triangle_tmp     = Zero();
+        scalar_object_triangle triangle_exp_tmp = Zero();
+
+        Coordinate lcoor;
+        grid->LocalIndexToLocalCoor(site, lcoor);
+
+        peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
+        peekLocalSite(triangle_tmp, triangle_v, lcoor);
+
+        int block;
+        block = 0;
+        for(int i = 0; i < 6; i++){
+        	for(int j = 0; j < 6; j++){
+        		if (i == j){
+        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
         		}
-    else
+        		else{
-    {
+        			srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
-      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
         		}
         	}
+        }
+        block = 1;
+        for(int i = 0; i < 6; i++){
+          	for(int j = 0; j < 6; j++){
+           		if (i == j){
+           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
+           		}
+           		else{
+           			srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
+           		}
+            }
+        }
+
+        // exp(Clover)
+
+        ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
+
+        block = 0;
+        for(int i = 0; i < 6; i++){
+        	for(int j = 0; j < 6; j++){
+            	if (i == j){
+            		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
+            	}
+            	else if(i < j){
+            		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
+            	}
+           	}
+        }
+
+        ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
+
+        block = 1;
+        for(int i = 0; i < 6; i++){
+        	for(int j = 0; j < 6; j++){
+              	if (i == j){
+              		diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
+               	}
+               	else if(i < j){
+               		triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
+               	}
+            }
+        }
+
+        pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
+        pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
+      });
+    }
+
+    Diagonal *= diag_mass;
+    Triangle *= diag_mass;
+  }
+
 
   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
     assert(0);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

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

Grid/qcd/action/fermion/FermionOperator.h

@@ -49,6 +49,8 @@ public:
 
   virtual FermionField &tmp(void) = 0;
 
+  virtual void DirichletBlock(Coordinate & _Block) { assert(0); };
+  
   GridBase * Grid(void)   { return FermionGrid(); };   // this is all the linalg routines need to know
   GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };
 

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 *vp0 = &in[k];
-    const vobj *vp1 = &in[m];  // out1[j] = merge hi  bit of type from in[k] and in[m]
+    const vobj *vp1 = &in[m];
-    const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
+    const vobj *vp = (lane&mask) ? vp1:vp0;
-    auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
+    auto sa = coalescedRead(*vp,j0);
-    auto sb = coalescedRead(*vp,j1); // lane to read for out 1
+    auto sb = coalescedRead(*vp,j1);
     hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);  // spin project the result0
+    projector::Proj(psa,sa,mu,dag);
-    projector::Proj(psb,sb,mu,dag);  // spin project the result1
+    projector::Proj(psb,sb,mu,dag);
     coalescedWrite(out0[j],psa);
     coalescedWrite(out1[j],psb);
 #else
@@ -297,7 +297,7 @@ public:
   void ZeroCountersi(void)  {  }
   void Reporti(int calls)  {  }
 
-  std::vector<int> surface_list;
+  //  Vector<int> surface_list;
 
   WilsonStencil(GridBase *grid,
 		int npoints,
@@ -307,10 +307,11 @@ public:
     : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
   { 
     ZeroCountersi();
-    surface_list.resize(0);
+    //    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
 
+  /*
   void BuildSurfaceList(int Ls,int vol4){
 
     // find same node for SHM
@@ -331,6 +332,7 @@ public:
       }
     }
   }
+  */
   
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,6 +75,10 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  int Dirichlet;
+  Coordinate Block; 
+
+  /********** Deprecate timers **********/
   void Report(void);
   void ZeroCounters(void);
   double DhopCalls;
@@ -174,6 +178,17 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
 
+  virtual void DirichletBlock(Coordinate & block)
+  {
+    assert(block.size()==Nd+1);
+    if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
+      Dirichlet = 1;
+      Block = block;
+      Stencil.DirichletBlock(block); 
+      StencilEven.DirichletBlock(block); 
+      StencilOdd.DirichletBlock(block);
+    }
+  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

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

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

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

@@ -60,7 +60,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   UmuOdd (_FourDimRedBlackGrid),
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid),
-  _tmp(&FiveDimRedBlackGrid)
+  _tmp(&FiveDimRedBlackGrid),
+  Dirichlet(0)
 {
   // some assertions
   assert(FiveDimGrid._ndimension==5);
@@ -218,6 +219,14 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
 {
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
+  if ( Dirichlet ) {
+    std::cout << GridLogMessage << " Dirichlet BCs 5d " <<Block<<std::endl;
+    Coordinate GaugeBlock(Nd);
+    for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
+    std::cout << GridLogMessage << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
+    DirichletFilter<GaugeField> Filter(GaugeBlock);
+    Filter.applyFilter(HUmu);
+  }
   Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
   pickCheckerboard(Even,UmuEven,Umu);
   pickCheckerboard(Odd ,UmuOdd,Umu);

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

@@ -440,6 +440,17 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
+#define KERNEL_CALL_EXT(A)						\
+  const uint64_t    NN = Nsite*Ls;					\
+  const uint64_t    sz = st.surface_list.size();			\
+  auto ptr = &st.surface_list[0];					\
+  accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
+      int sF = ptr[ss];							\
+      int sU = ss/Ls;							\
+      WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
+    });									\
+  accelerator_barrier();
+
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
     int sU = ss;							\
@@ -498,7 +509,6 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
-     acceleratorFenceComputeStream();
    } else if( interior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
@@ -506,13 +516,11 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      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/filters/DDHMCFilter.h

@@ -0,0 +1,102 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/DirichletFilter.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+//--------------------------------------------------------------------
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+////////////////////////////////////////////////////
+// DDHMC filter with sub-block size B[mu]
+////////////////////////////////////////////////////
+
+template<typename GaugeField>
+struct DDHMCFilter: public MomentumFilterBase<GaugeField>
+{
+  Coordinate Block;
+  int Width;
+  
+  DDHMCFilter(const Coordinate &_Block,int _Width=2): Block(_Block) { Width=_Width; }
+
+  void applyFilter(GaugeField &U) const override
+  {
+    GridBase *grid = U.Grid();
+    Coordinate Global=grid->GlobalDimensions();
+    GaugeField zzz(grid); zzz = Zero();
+    LatticeInteger coor(grid); 
+    
+    auto zzz_mu = PeekIndex<LorentzIndex>(zzz,0);
+    ////////////////////////////////////////////////////
+    // Zero BDY layers
+    ////////////////////////////////////////////////////
+    std::cout<<GridLogMessage<<" DDHMC Force Filter Block "<<Block<<" width " <<Width<<std::endl;
+    for(int mu=0;mu<Nd;mu++) {
+
+      Integer B1 = Block[mu];
+      if ( B1 && (B1 <= Global[mu]) ) {
+	LatticeCoordinate(coor,mu);
+
+	////////////////////////////////
+	// OmegaBar - zero all links contained in slice B-1,0 and
+	// mu links connecting to Omega
+	////////////////////////////////
+	if ( Width==1) { 
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-2),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+	if ( Width==2) { 
+	  U    = where(mod(coor,B1)==Integer(B1-2),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-3),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+	if ( Width==3) { 
+	  U    = where(mod(coor,B1)==Integer(B1-3),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-2),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(B1-1),zzz,U);
+	  U    = where(mod(coor,B1)==Integer(0)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(1)   ,zzz,U); 
+	  U    = where(mod(coor,B1)==Integer(2)   ,zzz,U); 
+	  auto U_mu   = PeekIndex<LorentzIndex>(U,mu);
+	  U_mu = where(mod(coor,B1)==Integer(B1-4),zzz_mu,U_mu); 
+	  PokeIndex<LorentzIndex>(U, U_mu, mu);
+	}
+      }
+
+    }
+   
+  }
+};
+
+NAMESPACE_END(Grid);
+

Grid/qcd/action/filters/DirichletFilter.h

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/hmc/integrators/DirichletFilter.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+//--------------------------------------------------------------------
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<typename MomentaField>
+struct DirichletFilter: public MomentumFilterBase<MomentaField>
+{
+  typedef typename MomentaField::vector_type vector_type; //SIMD-vectorized complex type
+  typedef typename MomentaField::scalar_type scalar_type; //scalar complex type
+
+  typedef iScalar<iScalar<iScalar<vector_type> > >            ScalarType; //complex phase for each site
+  
+  Coordinate Block;
+  
+  DirichletFilter(const Coordinate &_Block): Block(_Block){}
+
+  void applyFilter(MomentaField &P) const override
+  {
+    GridBase *grid = P.Grid();
+    typedef decltype(PeekIndex<LorentzIndex>(P, 0)) LatCM;
+    ////////////////////////////////////////////////////
+    // Zero strictly links crossing between domains
+    ////////////////////////////////////////////////////
+    LatticeInteger coor(grid); 
+    LatCM zz(grid); zz = Zero();
+    for(int mu=0;mu<Nd;mu++) {
+      if ( (Block[mu]) && (Block[mu] < grid->GlobalDimensions()[mu] ) ) {
+	// If costly could provide Grid earlier and precompute masks
+	std::cout << GridLogMessage << " Dirichlet in mu="<<mu<<std::endl;
+	LatticeCoordinate(coor,mu);
+	auto P_mu = PeekIndex<LorentzIndex>(P, mu);
+	P_mu = where(mod(coor,Block[mu])==Integer(Block[mu]-1),zz,P_mu);
+	PokeIndex<LorentzIndex>(P, P_mu, mu);
+      }
+    }
+  }
+};
+
+
+
+NAMESPACE_END(Grid);
+

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/hmc/GenericHMCrunner.h

@@ -129,18 +129,10 @@ public:
     Runner(S);
   }
 
-  //////////////////////////////////////////////////////////////////
+  //Use the checkpointer to initialize the RNGs and the gauge field, writing the resulting gauge field into U.
-
+  //This is called automatically by Run but may be useful elsewhere, e.g. for integrator tuning experiments
-private:
+  void initializeGaugeFieldAndRNGs(Field &U){
-  template <class SmearingPolicy>
+    if(!Resources.haveRNGs()) Resources.AddRNGs();
-  void Runner(SmearingPolicy &Smearing) {
-    auto UGrid = Resources.GetCartesian();
-    Resources.AddRNGs();
-    Field U(UGrid);
-
-    // Can move this outside?
-    typedef IntegratorType<SmearingPolicy> TheIntegrator;
-    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
 
     if (Parameters.StartingType == "HotStart") {
       // Hot start
@@ -167,6 +159,25 @@ private:
 	<< "Valid [HotStart, ColdStart, TepidStart, CheckpointStart]\n";
       exit(1);
     }
+  }
+
+
+
+  //////////////////////////////////////////////////////////////////
+
+private:
+  template <class SmearingPolicy>
+  void Runner(SmearingPolicy &Smearing) {
+    auto UGrid = Resources.GetCartesian();
+    Field U(UGrid);
+
+    initializeGaugeFieldAndRNGs(U);
+
+    typedef IntegratorType<SmearingPolicy> TheIntegrator;
+    TheIntegrator MDynamics(UGrid, Parameters.MD, TheAction, Smearing);
+
+    // Sets the momentum filter
+    MDynamics.setMomentumFilter(*(Resources.GetMomentumFilter()));
 
     Smearing.set_Field(U);
 

Grid/qcd/hmc/HMC.h

@@ -34,6 +34,7 @@ directory
 			    * @brief Classes for Hybrid Monte Carlo update
 			    *
 			    * @author Guido Cossu
+			    * @author Peter Boyle
 			    */
 			   //--------------------------------------------------------------------
 #pragma once
@@ -115,22 +116,17 @@ private:
 
     random(sRNG, rn_test);
 
-    std::cout << GridLogMessage
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
-              << "--------------------------------------------------\n";
+    std::cout << GridLogHMC << "exp(-dH) = " << prob << "  Random = " << rn_test << "\n";
-    std::cout << GridLogMessage << "exp(-dH) = " << prob
+    std::cout << GridLogHMC << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
-              << "  Random = " << rn_test << "\n";
-    std::cout << GridLogMessage
-              << "Acc. Probability = " << ((prob < 1.0) ? prob : 1.0) << "\n";
 
     if ((prob > 1.0) || (rn_test <= prob)) {  // accepted
-      std::cout << GridLogMessage << "Metropolis_test -- ACCEPTED\n";
+      std::cout << GridLogHMC << "Metropolis_test -- ACCEPTED\n";
-      std::cout << GridLogMessage
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
-                << "--------------------------------------------------\n";
       return true;
     } else {  // rejected
-      std::cout << GridLogMessage << "Metropolis_test -- REJECTED\n";
+      std::cout << GridLogHMC << "Metropolis_test -- REJECTED\n";
-      std::cout << GridLogMessage
+      std::cout << GridLogHMC << "--------------------------------------------------\n";
-                << "--------------------------------------------------\n";
       return false;
     }
   }
@@ -139,18 +135,67 @@ private:
   // Evolution
   /////////////////////////////////////////////////////////
   RealD evolve_hmc_step(Field &U) {
-    TheIntegrator.refresh(U, sRNG, pRNG);  // set U and initialize P and phi's
 
-    RealD H0 = TheIntegrator.S(U);  // initial state action
+    GridBase *Grid = U.Grid();
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // Mainly for DDHMC perform a random translation of U modulo volume
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Random shifting gauge field by [";
+    for(int d=0;d<Grid->Nd();d++) {
+
+      int L = Grid->GlobalDimensions()[d];
+
+      RealD rn_uniform;  random(sRNG, rn_uniform);
+
+      int shift = (int) (rn_uniform*L);
+
+      std::cout << shift;
+      if(d<Grid->Nd()-1) std::cout <<",";
+      else               std::cout <<"]\n";
+      
+      U = Cshift(U,d,shift);
+    }
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    TheIntegrator.reset_timer();
+    
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // set U and initialize P and phi's
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Refresh momenta and pseudofermions";
+    TheIntegrator.refresh(U, sRNG, pRNG);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // initial state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Compute initial action";
+    RealD H0 = TheIntegrator.S(U);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
     std::streamsize current_precision = std::cout.precision();
     std::cout.precision(15);
-    std::cout << GridLogMessage << "Total H before trajectory = " << H0 << "\n";
+    std::cout << GridLogHMC << "Total H before trajectory = " << H0 << "\n";
     std::cout.precision(current_precision);
 
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << " Molecular Dynamics evolution ";
     TheIntegrator.integrate(U);
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // updated state action
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+    std::cout << GridLogMessage << "Compute final action";
+    RealD H1 = TheIntegrator.S(U);  
+    std::cout << GridLogMessage << "--------------------------------------------------\n";
+
 
-    RealD H1 = TheIntegrator.S(U);  // updated state action
     
     ///////////////////////////////////////////////////////////
     if(0){
@@ -163,18 +208,17 @@ private:
     }
     ///////////////////////////////////////////////////////////
 
-
     std::cout.precision(15);
-    std::cout << GridLogMessage << "Total H after trajectory  = " << H1
+
-	      << "  dH = " << H1 - H0 << "\n";
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+    std::cout << GridLogHMC << "Total H after trajectory  = " << H1 << "  dH = " << H1 - H0 << "\n";
+    std::cout << GridLogHMC << "--------------------------------------------------\n";
+
     std::cout.precision(current_precision);
     
     return (H1 - H0);
   }
 
-
-  
-
 public:
   /////////////////////////////////////////
   // Constructor
@@ -195,10 +239,13 @@ public:
 
     // Actual updates (evolve a copy Ucopy then copy back eventually)
     unsigned int FinalTrajectory = Params.Trajectories + Params.NoMetropolisUntil + Params.StartTrajectory;
+
     for (int traj = Params.StartTrajectory; traj < FinalTrajectory; ++traj) {
-      std::cout << GridLogMessage << "-- # Trajectory = " << traj << "\n";
+
+      std::cout << GridLogHMC << "-- # Trajectory = " << traj << "\n";
+
       if (traj < Params.StartTrajectory + Params.NoMetropolisUntil) {
-      	std::cout << GridLogMessage << "-- Thermalization" << std::endl;
+      	std::cout << GridLogHMC << "-- Thermalization" << std::endl;
       }
       
       double t0=usecond();
@@ -207,20 +254,19 @@ public:
       DeltaH = evolve_hmc_step(Ucopy);
       // Metropolis-Hastings test
       bool accept = true;
-      if (traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
+      if (Params.MetropolisTest && traj >= Params.StartTrajectory + Params.NoMetropolisUntil) {
         accept = metropolis_test(DeltaH);
       } else {
-      	std::cout << GridLogMessage << "Skipping Metropolis test" << std::endl;
+      	std::cout << GridLogHMC << "Skipping Metropolis test" << std::endl;
       }
 
       if (accept)
         Ucur = Ucopy; 
       
-     
-      
       double t1=usecond();
-      std::cout << GridLogMessage << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
+      std::cout << GridLogHMC << "Total time for trajectory (s): " << (t1-t0)/1e6 << std::endl;
 
+      TheIntegrator.print_timer();
 
       for (int obs = 0; obs < Observables.size(); obs++) {
       	std::cout << GridLogDebug << "Observables # " << obs << std::endl;
@@ -228,7 +274,7 @@ public:
       	std::cout << GridLogDebug << "Observables pointer " << Observables[obs] << std::endl;
         Observables[obs]->TrajectoryComplete(traj + 1, Ucur, sRNG, pRNG);
       }
-      std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
+      std::cout << GridLogHMC << ":::::::::::::::::::::::::::::::::::::::::::" << std::endl;
     }
   }
 

Grid/qcd/hmc/HMCResourceManager.h

@@ -72,6 +72,8 @@ class HMCResourceManager {
   typedef HMCModuleBase< BaseHmcCheckpointer<ImplementationPolicy> > CheckpointerBaseModule;
   typedef HMCModuleBase< HmcObservable<typename ImplementationPolicy::Field> > ObservableBaseModule;
   typedef ActionModuleBase< Action<typename ImplementationPolicy::Field>, GridModule > ActionBaseModule;
+  typedef typename ImplementationPolicy::Field MomentaField;
+  typedef typename ImplementationPolicy::Field Field;  
 
   // Named storage for grid pairs (std + red-black)
   std::unordered_map<std::string, GridModule> Grids;
@@ -80,6 +82,9 @@ class HMCResourceManager {
   // SmearingModule<ImplementationPolicy> Smearing;
   std::unique_ptr<CheckpointerBaseModule> CP;
 
+  // Momentum filter
+  std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> > Filter;
+  
   // A vector of HmcObservable modules
   std::vector<std::unique_ptr<ObservableBaseModule> > ObservablesList;
 
@@ -90,6 +95,7 @@ class HMCResourceManager {
 
   bool have_RNG;
   bool have_CheckPointer;
+  bool have_Filter;
 
   // NOTE: operator << is not overloaded for std::vector<string> 
   // so this function is necessary
@@ -101,7 +107,7 @@ class HMCResourceManager {
 
 
 public:
-  HMCResourceManager() : have_RNG(false), have_CheckPointer(false) {}
+  HMCResourceManager() : have_RNG(false), have_CheckPointer(false), have_Filter(false) {}
 
   template <class ReaderClass, class vector_type = vComplex >
   void initialize(ReaderClass &Read){
@@ -129,6 +135,7 @@ public:
     RNGModuleParameters RNGpar(Read);
     SetRNGSeeds(RNGpar);
 
+  
     // Observables
     auto &ObsFactory = HMC_ObservablesModuleFactory<observable_string, typename ImplementationPolicy::Field, ReaderClass>::getInstance(); 
     Read.push(observable_string);// here must check if existing...
@@ -208,6 +215,16 @@ public:
     AddGrid(s, Mod);
   }
 
+  void SetMomentumFilter( MomentumFilterBase<typename ImplementationPolicy::Field> * MomFilter) {
+    assert(have_Filter==false);
+    Filter = std::unique_ptr<MomentumFilterBase<typename ImplementationPolicy::Field> >(MomFilter);
+    have_Filter = true;
+  }
+  MomentumFilterBase<typename ImplementationPolicy::Field> *GetMomentumFilter(void) {
+    if ( !have_Filter)
+      SetMomentumFilter(new MomentumFilterNone<typename ImplementationPolicy::Field>());
+    return Filter.get();
+  }
 
   GridCartesian* GetCartesian(std::string s = "") {
     if (s.empty()) s = Grids.begin()->first;
@@ -227,6 +244,9 @@ public:
   // Random number generators
   //////////////////////////////////////////////////////
   
+  //Return true if the RNG objects have been instantiated
+  bool haveRNGs() const{ return have_RNG; }
+
   void AddRNGs(std::string s = "") {
     // Couple the RNGs to the GridModule tagged by s
     // the default is the first grid registered

Grid/qcd/hmc/integrators/Integrator.h

@@ -33,7 +33,6 @@ directory
 #define INTEGRATOR_INCLUDED
 
 #include <memory>
-#include "MomentumFilter.h"
 
 NAMESPACE_BEGIN(Grid);
 
@@ -67,6 +66,7 @@ public:
 template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
+
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 
@@ -119,36 +119,58 @@ protected:
     }
   } update_P_hireps{};
 
+ 
   void update_P(MomentaField& Mom, Field& U, int level, double ep) {
     // input U actually not used in the fundamental case
     // Fundamental updates, include smearing
 
     for (int a = 0; a < as[level].actions.size(); ++a) {
+
       double start_full = usecond();
       Field force(U.Grid());
       conformable(U.Grid(), Mom.Grid());
 
       Field& Us = Smearer.get_U(as[level].actions.at(a)->is_smeared);
       double start_force = usecond();
+      as[level].actions.at(a)->deriv_timer_start();
       as[level].actions.at(a)->deriv(Us, force);  // deriv should NOT include Ta
+      as[level].actions.at(a)->deriv_timer_stop();
 
       std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
+      auto name = as[level].actions.at(a)->action_name();
       if (as[level].actions.at(a)->is_smeared) Smearer.smeared_force(force);
+
       force = FieldImplementation::projectForce(force); // Ta for gauge fields
       double end_force = usecond();
-      Real force_abs = std::sqrt(norm2(force)/U.Grid()->gSites());
+
-      std::cout << GridLogIntegrator << "["<<level<<"]["<<a<<"] Force average: " << force_abs << std::endl;
+      MomFilter->applyFilter(force);
+      std::cout << GridLogIntegrator << " update_P : Level [" << level <<"]["<<a <<"] "<<name<< std::endl;
+      DumpSliceNorm("force ",force,Nd-1);
+      
+      Real force_abs   = std::sqrt(norm2(force)/U.Grid()->gSites()); //average per-site norm.  nb. norm2(latt) = \sum_x norm2(latt[x]) 
+      Real impulse_abs = force_abs * ep * HMC_MOMENTUM_DENOMINATOR;    
+
+      Real force_max   = std::sqrt(maxLocalNorm2(force));
+      Real impulse_max = force_max * ep * HMC_MOMENTUM_DENOMINATOR;    
+
+      as[level].actions.at(a)->deriv_log(force_abs,force_max);
+      
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force average: " << force_abs <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Force max    : " << force_max <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt average  : " << impulse_abs <<" "<<name<<std::endl;
+      std::cout << GridLogIntegrator<< "["<<level<<"]["<<a<<"] Fdt max      : " << impulse_max <<" "<<name<<std::endl;
+
       Mom -= force * ep* HMC_MOMENTUM_DENOMINATOR;; 
       double end_full = usecond();
       double time_full  = (end_full - start_full) / 1e3;
       double time_force = (end_force - start_force) / 1e3;
       std::cout << GridLogMessage << "["<<level<<"]["<<a<<"] P update elapsed time: " << time_full << " ms (force: " << time_force << " ms)"  << std::endl;
+
     }
 
     // Force from the other representations
     as[level].apply(update_P_hireps, Representations, Mom, U, ep);
 
-    MomFilter->applyFilter(Mom);
   }
 
   void update_U(Field& U, double ep) 
@@ -162,8 +184,12 @@ protected:
   
   void update_U(MomentaField& Mom, Field& U, double ep) 
   {
+    MomentaField MomFiltered(Mom.Grid());
+    MomFiltered = Mom;
+    MomFilter->applyFilter(MomFiltered);
+
     // exponential of Mom*U in the gauge fields case
-    FieldImplementation::update_field(Mom, U, ep);
+    FieldImplementation::update_field(MomFiltered, U, ep);
 
     // Update the smeared fields, can be implemented as observer
     Smearer.set_Field(U);
@@ -206,6 +232,66 @@ public:
   const MomentaField & getMomentum() const{ return P; }
   
 
+  void reset_timer(void)
+  {
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+        as[level].actions.at(actionID)->reset_timer();
+      }
+    }
+  }
+  void print_timer(void)
+  {
+    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::" << std::endl;
+    std::cout << GridLogMessage << " Refresh cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->refresh_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Action cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->S_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Force cumulative timings "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] "
+		  << as[level].actions.at(actionID)->deriv_us*1.0e-6<<" s"<< std::endl;
+      }
+    }
+    std::cout << GridLogMessage << "--------------------------- "<<std::endl;
+    std::cout << GridLogMessage << " Force average size "<<std::endl;
+    std::cout << GridLogMessage << "------------------------- "<<std::endl;
+    for (int level = 0; level < as.size(); ++level) {
+      for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
+	std::cout << GridLogMessage 
+		  << as[level].actions.at(actionID)->action_name()
+		  <<"["<<level<<"]["<< actionID<<"] : "
+		  <<" force max " << as[level].actions.at(actionID)->deriv_max_average()
+		  <<" norm "      << as[level].actions.at(actionID)->deriv_norm_average()
+		  <<" calls "     << as[level].actions.at(actionID)->deriv_num
+		  << std::endl;
+      }
+    }
+    std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
+  }
+  
   void print_parameters()
   {
     std::cout << GridLogMessage << "[Integrator] Name : "<< integrator_name() << std::endl;
@@ -224,7 +310,6 @@ public:
       }
     }
     std::cout << GridLogMessage << ":::::::::::::::::::::::::::::::::::::::::"<< std::endl;
-
   }
 
   void reverse_momenta()
@@ -267,15 +352,19 @@ public:
       for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
         // get gauge field from the SmearingPolicy and
         // based on the boolean is_smeared in actionID
+	auto name = as[level].actions.at(actionID)->action_name();
+        std::cout << GridLogMessage << "refresh [" << level << "][" << actionID << "] "<<name << std::endl;
+
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
+	as[level].actions.at(actionID)->refresh_timer_start();
         as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
+	as[level].actions.at(actionID)->refresh_timer_stop();
       }
 
       // Refresh the higher representation actions
       as[level].apply(refresh_hireps, Representations, sRNG, pRNG);
     }
 
-    MomFilter->applyFilter(P);
   }
 
   // to be used by the actionlevel class to iterate
@@ -310,7 +399,9 @@ public:
         // based on the boolean is_smeared in actionID
         Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
+	        as[level].actions.at(actionID)->S_timer_start();
         Hterm = as[level].actions.at(actionID)->S(Us);
+   	        as[level].actions.at(actionID)->S_timer_stop();
         std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
         H += Hterm;
       }

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/SimpleCompressor.h

@@ -52,6 +52,11 @@ public:
     return arg;
   }
 };
+class SimpleStencilParams{
+public:
+  Coordinate dirichlet;
+  SimpleStencilParams() {};
+};
 
 NAMESPACE_END(Grid);
 

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,
+				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
-				 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,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				 const Lattice<vobj> &rhs,
+				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
-				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -95,12 +92,11 @@ 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,
+      compress.CompressExchange(p0,p1, &rhs_v[0], j,
 			      so+tp[2*j  ].second,
 			      so+tp[2*j+1].second,
 			      type);
@@ -137,6 +133,8 @@ class CartesianStencilAccelerator {
   int           _osites;
   StencilVector _directions;
   StencilVector _distances;
+  StencilVector _comms_send;
+  StencilVector _comms_recv;
   StencilVector _comm_buf_size;
   StencilVector _permute_type;
   StencilVector same_node;
@@ -230,12 +228,14 @@ 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;
+    Vector<scalar_object *> rpointers;
-    std::vector<cobj *> vpointers;
+    Vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
   };
@@ -244,7 +244,20 @@ 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 +288,8 @@ 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 +302,6 @@ 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 +328,12 @@ 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
@@ -350,67 +342,30 @@ 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)
   {
     reqs.resize(Packets.size());
-    commtime-=usecond();
     for(int i=0;i<Packets.size();i++){
-      uint64_t bytes=_grid->StencilSendToRecvFromBegin(reqs[i],
+      _grid->StencilSendToRecvFromBegin(reqs[i],
 					Packets[i].send_buf,
-						     Packets[i].to_rank,
+					Packets[i].to_rank,Packets[i].do_send,
 					Packets[i].recv_buf,
-						     Packets[i].from_rank,
+					Packets[i].from_rank,Packets[i].do_recv,
 					Packets[i].bytes,i);
-      comms_bytes+=bytes;
-      shm_bytes  +=2*Packets[i].bytes-bytes;
     }
-    _grid->StencilBarrier();// Synch shared memory on a single nodes
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
@@ -418,36 +373,34 @@ public:
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
-    commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    if ( 0 ){
+    if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
-      thread_region {
+      /////////////////////////////////////////////////////////
-	// must be called in parallel region
+      // several way threaded on different communicators.
-	int mythread  = thread_num();
+      // Cannot combine with Dirichlet operators
-	int maxthreads= thread_max();
+      // This scheme is needed on Intel Omnipath for best performance
+      // Deprecate once there are very few omnipath clusters
+      /////////////////////////////////////////////////////////
       int nthreads = CartesianCommunicator::nCommThreads;
-	assert(nthreads <= maxthreads);
+      int old = GridThread::GetThreads();
-	if (nthreads == -1) nthreads = 1;
+      GridThread::SetThreads(nthreads);
-	if (mythread < nthreads) {
+      thread_for(i,Packets.size(),{
-	  for (int i = mythread; i < Packets.size(); i += nthreads) {
+	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-	    double start = usecond();
+				       Packets[i].to_rank,Packets[i].do_send,
-	    uint64_t bytes= _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-							 Packets[i].to_rank,
 				       Packets[i].recv_buf,
-							 Packets[i].from_rank,
+				       Packets[i].from_rank,Packets[i].do_recv,
 				       Packets[i].bytes,i);
-	    comm_bytes_thr[mythread] += bytes;
+      });
-	    shm_bytes_thr[mythread]  += Packets[i].bytes - bytes;
+      GridThread::SetThreads(old);
-	    comm_time_thr[mythread]  += usecond() - start;
+    } else { 
-	  }
+      /////////////////////////////////////////////////////////
-	}
+      // Concurrent and non-threaded asynch calls to MPI
-      }
+      /////////////////////////////////////////////////////////
-    } else { // Concurrent and non-threaded asynch calls to MPI
       std::vector<std::vector<CommsRequest_t> > reqs;
       this->CommunicateBegin(reqs);
       this->CommunicateComplete(reqs);
@@ -489,31 +442,23 @@ 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,point);
-	  auto tmp  = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
 	  is_same_node = is_same_node && tmp;
-	  splicetime+=usecond();
 	} else {
-	  nosplicetime-=usecond();
+	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx,point);
-	  auto tmp  = Gather(source,dimension,shift,0x3,compress,face_idx);
 	  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 tmp1 =  GatherSimd(source,dimension,shift,0x1,compress,face_idx,point);
-	  auto tmp2 =  GatherSimd(source,dimension,shift,0x2,compress,face_idx);
+	  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,point);
-	  auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
+	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx,point);
-	  auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
 	  is_same_node = is_same_node && tmp1 && tmp2;
-	  nosplicetime+=usecond();
 	}
       }
     }
@@ -523,13 +468,10 @@ 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;
 
@@ -543,7 +485,6 @@ public:
     assert(u_comm_offset==_unified_buffer_size);
 
     accelerator_barrier();
-    halogtime+=usecond();
   }
 
   /////////////////////////
@@ -556,14 +497,70 @@ 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)
+  {
+    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);
+
+      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)
+	     ){
+
+	AddCopy(CachedTransfers[i].recv_buf,recv_buf,bytes);
+	return 1;
+      }
+    }
+
+    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);
   }
@@ -574,7 +571,7 @@ 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;
@@ -583,23 +580,17 @@ public:
     mv.push_back(m);
   }
   template<class decompressor>  void CommsMerge(decompressor decompress)    {
+    CommsCopy();
     CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    mpi3synctime-=usecond();
-    accelerator_barrier();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime+=usecond();
-    shmmergetime-=usecond();
     CommsMerge(decompress,MergersSHM,DecompressionsSHM);
-    shmmergetime+=usecond();
   }
 
   template<class decompressor>
-  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
+  void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd)
-
+  {
-    
-    mergetime-=usecond();
     for(int i=0;i<mm.size();i++){
       auto mp = &mm[i].mpointer[0];
       auto vp0= &mm[i].vpointers[0][0];
@@ -609,9 +600,7 @@ public:
 	  decompress.Exchange(mp,vp0,vp1,type,o);
       });
     }
-    mergetime+=usecond();
 
-    decompresstime-=usecond();
     for(int i=0;i<dd.size();i++){
       auto kp = dd[i].kernel_p;
       auto mp = dd[i].mpi_p;
@@ -619,7 +608,6 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
-    decompresstime+=usecond();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -652,22 +640,60 @@ public:
 	}
       }
       if(local == 0) {
-	surface_list.push_back(site);
+	for(int s=0;s<Ls;s++){
+	  surface_list.push_back(site*Ls+s);
+	}
+      }
+    }
+  }
+  /// Introduce a block structure and switch off comms on boundaries
+  void DirichletBlock(const Coordinate &dirichlet_block)
+  {
+    for(int ii=0;ii<this->_npoints;ii++){
+      int dimension    = this->_directions[ii];
+      int displacement = this->_distances[ii];
+      int gd = _grid->_gdimensions[dimension];
+      int fd = _grid->_fdimensions[dimension];
+      int pd = _grid->_processors [dimension];
+      int pc = _grid->_processor_coor[dimension];
+      int ld = fd/pd;
+      ///////////////////////////////////////////
+      // 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 && comm_dim ) {
+	assert(abs(displacement) < ld );
+	// Quiesce communication across block boundaries
+	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)
   {
     face_table_computed=0;
     _grid    = grid;
@@ -681,8 +707,12 @@ 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);
 
+    if ( p.dirichlet.size() ) DirichletBlock(p.dirichlet); // comms send/recv set up
+
     _unified_buffer_size=0;
     surface_list.resize(0);
 
@@ -699,15 +729,16 @@ 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);
@@ -716,7 +747,6 @@ public:
       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
@@ -817,6 +847,7 @@ 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];
@@ -842,12 +873,14 @@ public:
     for(int x=0;x<rd;x++){
 
       int permute_type=grid->PermuteType(dimension);
+      int permute_slice;
 
       int sx        =  (x+sshift)%rd;
 
       int offnode = 0;
       if ( simd_layout > 1 ) {
 
+	permute_slice=1;
 	for(int i=0;i<Nsimd;i++){
 
 	  int inner_bit = (Nsimd>>(permute_type+1));
@@ -864,6 +897,7 @@ public:
       } else {
 	int comm_proc = ((x+sshift)/rd)%pd;
 	offnode = (comm_proc!= 0);
+	permute_slice=0;
       }
 
       int wraparound=0;
@@ -873,25 +907,31 @@ public:
       if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
 	wraparound = 1;
       }
-      if (!offnode) {
 
-	int permute_slice=0;
+      // Wrap locally dirichlet support case OR node local
+      if ( offnode==0 ) {
+
+	permute_slice=0;
 	CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
 	
       } else {
 
+	if ( comms_recv ) {
+
+	  ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
+
+	} else { 
+
+	  CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
+
+	}
+
+      }
+      
+      if ( offnode ) {
 	int words = buffer_size;
 	if (cbmask != 0x3) words=words>>1;
-
-	//	int rank           = grid->_processor;
-	//	int recv_from_rank;
-	//	int xmit_to_rank;
-
-	int unified_buffer_offset = _unified_buffer_size;
 	_unified_buffer_size    += words;
-
-	ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
-
       }
     }
   }
@@ -990,11 +1030,14 @@ 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());
 
@@ -1026,69 +1069,84 @@ public:
 	int bytes =  words * compress.CommDatumSize();
 
 	int so  = sx*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
+	int comm_off = u_comm_offset;
+
+	int recv_from_rank;
+	int xmit_to_rank;
+	cobj *recv_buf;
+	cobj *send_buf;
+	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+	assert (xmit_to_rank   != _grid->ThisRank());
+	assert (recv_from_rank != _grid->ThisRank());
+
+	if( comms_send ) {
+
 	  if ( !face_table_computed ) {
 	    face_table.resize(face_idx+1);
 	    std::vector<std::pair<int,int> >  face_table_host ;
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
+	    Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
 	    face_table[face_idx].resize(face_table_host.size());
 	    acceleratorCopyToDevice(&face_table_host[0],
 				    &face_table[face_idx][0],
 				    face_table[face_idx].size()*sizeof(face_table_host[0]));
 	  }
 
-	//      	int rank           = _grid->_processor;
-	int recv_from_rank;
-	int xmit_to_rank;
-	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
-	assert (xmit_to_rank   != _grid->ThisRank());
-	assert (recv_from_rank != _grid->ThisRank());
-
-	cobj *recv_buf;
 	  if ( compress.DecompressionStep() ) {
 	    recv_buf=u_simd_recv_buf[0];
 	  } else {
 	    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();
+	  Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,comm_off,so);
+	}
+
+	int duplicate = CheckForDuplicate(dimension,sx,comm_proc,(void *)&recv_buf[comm_off],0,bytes,cbmask);
+	if ( (!duplicate) ) { // Force comms for now
 
 	  ///////////////////////////////////////////////////////////
 	  // Build a list of things to do after we synchronise GPUs
 	  // Start comms now???
 	  ///////////////////////////////////////////////////////////
-	AddPacket((void *)&send_buf[u_comm_offset],
+	  AddPacket((void *)&send_buf[comm_off],
-		  (void *)&recv_buf[u_comm_offset],
+		    (void *)&recv_buf[comm_off],
-		  xmit_to_rank,
+		    xmit_to_rank, comms_send,
-		  recv_from_rank,
+		    recv_from_rank, comms_recv,
 		    bytes);
+	}
 	
-	if ( compress.DecompressionStep() ) {
+	if ( compress.DecompressionStep()  && comms_recv ) {
-	  AddDecompress(&this->u_recv_buf_p[u_comm_offset],
+	  AddDecompress(&this->u_recv_buf_p[comm_off],
-			&recv_buf[u_comm_offset],
+			&recv_buf[comm_off],
 			words,Decompressions);
 	}
+	
 	u_comm_offset+=words;
+	face_idx++;
+
       }
     }
     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];
@@ -1103,7 +1161,6 @@ public:
 
 
     int permute_type=_grid->PermuteType(dimension);
-    //    std::cout << "SimdNew permute type "<<permute_type<<std::endl;
 
     ///////////////////////////////////////////////
     // Simd direction uses an extract/merge pair
@@ -1120,8 +1177,8 @@ public:
     int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    std::vector<cobj *> rpointers(maxl);
+    Vector<cobj *> rpointers(maxl);
-    std::vector<cobj *> spointers(maxl);
+    Vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where
@@ -1137,8 +1194,9 @@ public:
 
       if ( any_offnode ) {
 
+	int comm_off = u_comm_offset;
 	for(int i=0;i<maxl;i++){
-	  spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
+	  spointers[i] = (cobj *) &u_simd_send_buf[i][comm_off];
 	}
 
 	int sx   = (x+sshift)%rd;
@@ -1147,18 +1205,17 @@ public:
 	  face_table.resize(face_idx+1);
 	  std::vector<std::pair<int,int> >  face_table_host ;
 				
-	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table_host);
+	  Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,comm_off,face_table_host);
 	  face_table[face_idx].resize(face_table_host.size());
 	  acceleratorCopyToDevice(&face_table_host[0],
 				  &face_table[face_idx][0],
 				  face_table[face_idx].size()*sizeof(face_table_host[0]));
 	}
-	gathermtime-=usecond();
 
+	if ( comms_send || comms_recv )
 	  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
 
@@ -1175,8 +1232,8 @@ public:
 	  int nbr_plane = nbr_ic;
 	  assert (sx == nbr_ox);
 
-	  auto rp = &u_simd_recv_buf[i        ][u_comm_offset];
+	  auto rp = &u_simd_recv_buf[i        ][comm_off];
-	  auto sp = &u_simd_send_buf[nbr_plane][u_comm_offset];
+	  auto sp = &u_simd_send_buf[nbr_plane][comm_off];
 
 	  if(nbr_proc){
 
@@ -1187,8 +1244,13 @@ 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  ) { 
+	      AddPacket((void *)sp,(void *)rp,
+			xmit_to_rank,comms_send,
+			recv_from_rank,comms_recv,
+			bytes);
+	    }
 
 	  } else {
 
@@ -1197,9 +1259,12 @@ 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[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
+	}
 
 	u_comm_offset     +=buffer_size;
+
       }
     }
     return 0;

Grid/tensors/Tensor_exp.h

@@ -55,7 +55,7 @@ template<class vtype, int N> accelerator_inline iVector<vtype, N> Exponentiate(c
 
 
 // Specialisation: Cayley-Hamilton exponential for SU(3)
-#ifndef GRID_CUDA
+#ifndef GRID_ACCELERATED
 template<class vtype, typename std::enable_if< GridTypeMapper<vtype>::TensorLevel == 0>::type * =nullptr> 
 accelerator_inline iMatrix<vtype,3> Exponentiate(const iMatrix<vtype,3> &arg, RealD alpha  , Integer Nexp = DEFAULT_MAT_EXP )
 {

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

@@ -247,6 +247,7 @@ inline int  acceleratorIsCommunicable(void *ptr)
 //////////////////////////////////////////////
 // SyCL acceleration
 //////////////////////////////////////////////
+
 #ifdef GRID_SYCL
 NAMESPACE_END(Grid);
 #include <CL/sycl.hpp>
@@ -261,7 +262,6 @@ NAMESPACE_END(Grid);
 NAMESPACE_BEGIN(Grid);
 
 extern cl::sycl::queue *theGridAccelerator;
-extern cl::sycl::queue *theCopyAccelerator;
 
 #ifdef __SYCL_DEVICE_ONLY__
 #define GRID_SIMT
@@ -289,7 +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);	      \
@@ -298,19 +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 acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {
-inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
+  theGridAccelerator->memcpy(to,from,bytes);
-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 acceleratorCopySynchronise(void) {  theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->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
@@ -441,7 +441,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
-  hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToDevice,copyStream);
+  hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);
 }
 inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
 
@@ -461,6 +461,8 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
   accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, { __VA_ARGS__ } ); \
   accelerator_barrier(dummy);
 
+#define GRID_ACCELERATED
+
 #endif
 
 //////////////////////////////////////////////
@@ -511,16 +513,7 @@ 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/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");
+    SHA256_CTX sha256;
-    EVP_Digest(data, bytes, &hash[0], NULL, digest, NULL);
+    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)

HMC/Mobius2p1f_DD_RHMC.cc

@@ -0,0 +1,232 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_hmc_EODWFRatio.cc
+
+Copyright (C) 2015-2016
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Guido Cossu <guido.cossu@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>
+
+int main(int argc, char **argv) {
+  using namespace Grid;
+
+  Grid_init(&argc, &argv);
+  int threads = GridThread::GetThreads();
+
+   // Typedefs to simplify notation
+  typedef WilsonImplR FermionImplPolicy;
+  typedef MobiusFermionR FermionAction;
+  typedef typename FermionAction::FermionField FermionField;
+
+  typedef Grid::XmlReader       Serialiser;
+
+  //::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+  IntegratorParameters MD;
+  //  typedef GenericHMCRunner<LeapFrog> HMCWrapper;
+  //  MD.name    = std::string("Leap Frog");
+  //  typedef GenericHMCRunner<ForceGradient> HMCWrapper;
+  //  MD.name    = std::string("Force Gradient");
+  typedef GenericHMCRunner<MinimumNorm2> HMCWrapper;
+  MD.name    = std::string("MinimumNorm2");
+  MD.MDsteps =  4;
+  MD.trajL   = 1.0;
+
+  HMCparameters HMCparams;
+  HMCparams.StartTrajectory  = 8;
+  HMCparams.Trajectories     = 200;
+  HMCparams.NoMetropolisUntil=  0;
+  // "[HotStart, ColdStart, TepidStart, CheckpointStart]\n";
+  //  HMCparams.StartingType     =std::string("ColdStart");
+  HMCparams.StartingType     =std::string("CheckpointStart");
+  HMCparams.MD = MD;
+  HMCWrapper TheHMC(HMCparams);
+
+  // Grid from the command line arguments --grid and --mpi
+  TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
+
+  CheckpointerParameters CPparams;
+  CPparams.config_prefix = "ckpoint_EODWF_lat";
+  CPparams.rng_prefix    = "ckpoint_EODWF_rng";
+  CPparams.saveInterval  = 1;
+  CPparams.format        = "IEEE64BIG";
+  TheHMC.Resources.LoadNerscCheckpointer(CPparams);
+
+  RNGModuleParameters RNGpar;
+  RNGpar.serial_seeds = "1 2 3 4 5";
+  RNGpar.parallel_seeds = "6 7 8 9 10";
+  TheHMC.Resources.SetRNGSeeds(RNGpar);
+
+  // Construct observables
+  // here there is too much indirection
+  typedef PlaquetteMod<HMCWrapper::ImplPolicy> PlaqObs;
+  TheHMC.Resources.AddObservable<PlaqObs>();
+  //////////////////////////////////////////////
+
+  const int Ls      = 16;
+  Real beta         = 2.13;
+  Real light_mass   = 0.01;
+  Real strange_mass = 0.04;
+  Real pv_mass      = 1.0;
+  RealD M5  = 1.8;
+  RealD b   = 1.0;
+  RealD c   = 0.0;
+
+  // FIXME:
+  // Same in MC and MD
+  // Need to mix precision too
+  OneFlavourRationalParams OFRp;
+  OFRp.lo       = 4.0e-3;
+  OFRp.hi       = 30.0;
+  OFRp.MaxIter  = 10000;
+  OFRp.tolerance= 1.0e-10;
+  OFRp.degree   = 16;
+  OFRp.precision= 50;
+
+  std::vector<Real> hasenbusch({ 0.01, 0.04, 0.2 , pv_mass });
+  std::vector<bool> dirichlet ({ true, true, true });
+
+  auto GridPtr   = TheHMC.Resources.GetCartesian();
+  auto GridRBPtr = TheHMC.Resources.GetRBCartesian();
+
+  ////////////////////////////////////////////////////////////////
+  // Domain decomposed
+  ////////////////////////////////////////////////////////////////
+  Coordinate latt4  = GridPtr->GlobalDimensions();
+  Coordinate mpi    = GridPtr->ProcessorGrid();
+  Coordinate shm;
+
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+  
+  Coordinate CommDim(Nd);
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+
+  Coordinate Dirichlet(Nd+1,0);
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Coordinate Block4(Nd);
+  Block4[0] = Dirichlet[1];
+  Block4[1] = Dirichlet[2];
+  Block4[2] = Dirichlet[3];
+  Block4[3] = Dirichlet[4];
+  TheHMC.Resources.SetMomentumFilter(new DDHMCFilter<WilsonImplR::Field>(Block4));
+
+  //////////////////////////
+  // Fermion Grid
+  //////////////////////////
+  auto FGrid     = SpaceTimeGrid::makeFiveDimGrid(Ls,GridPtr);
+  auto FrbGrid   = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,GridPtr);
+
+  IwasakiGaugeActionR GaugeAction(beta);
+
+  // temporarily need a gauge field
+  LatticeGaugeField U(GridPtr);
+
+  // These lines are unecessary if BC are all periodic
+  std::vector<Complex> boundary = {1,1,1,-1};
+  FermionAction::ImplParams Params(boundary);
+
+  double StoppingCondition = 1e-10;
+  double MaxCGIterations = 30000;
+  ConjugateGradient<FermionField>  CG(StoppingCondition,MaxCGIterations);
+
+  ////////////////////////////////////
+  // Collect actions
+  ////////////////////////////////////
+  ActionLevel<HMCWrapper::Field> Level1(1);
+  ActionLevel<HMCWrapper::Field> Level2(2);
+  ActionLevel<HMCWrapper::Field> Level3(8);
+
+  ////////////////////////////////////
+  // Strange action
+  ////////////////////////////////////
+  FermionAction StrangeOp (U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,strange_mass,M5,b,c, Params);
+  FermionAction StrangePauliVillarsOp(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,pv_mass,  M5,b,c, Params);
+
+  OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermion(StrangePauliVillarsOp,StrangeOp,OFRp);
+  //  Level1.push_back(&StrangePseudoFermion);
+
+  ////////////////////////////////////
+  // up down action
+  ////////////////////////////////////
+  std::vector<Real> light_den;
+  std::vector<Real> light_num;
+  std::vector<int> dirichlet_den;
+  std::vector<int> dirichlet_num;
+
+  int n_hasenbusch = hasenbusch.size();
+  light_den.push_back(light_mass);
+  dirichlet_den.push_back(0);
+  for(int h=0;h<n_hasenbusch;h++){
+    light_den.push_back(hasenbusch[h]);
+    light_num.push_back(hasenbusch[h]);
+    dirichlet_num.push_back(1);
+    dirichlet_den.push_back(1);
+  }
+  light_num.push_back(pv_mass);
+  dirichlet_num.push_back(0);
+
+  std::vector<FermionAction *> Numerators;
+  std::vector<FermionAction *> Denominators;
+  std::vector<TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy> *> Quotients;
+  
+  for(int h=0;h<n_hasenbusch+1;h++){
+    std::cout << GridLogMessage << " 2f quotient Action  "<< light_num[h]<< " (" << dirichlet_num[h]
+	      <<") / " << light_den[h]<< " (" << dirichlet_den[h]<<")"<< std::endl;
+    Numerators.push_back  (new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_num[h],M5,b,c, Params));
+    Denominators.push_back(new FermionAction(U,*FGrid,*FrbGrid,*GridPtr,*GridRBPtr,light_den[h],M5,b,c, Params));
+    Quotients.push_back   (new TwoFlavourEvenOddRatioPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],CG,CG));
+    if ( dirichlet_den[h]==1) Denominators[h]->DirichletBlock(Dirichlet);
+    if ( dirichlet_num[h]==1) Numerators[h]->DirichletBlock(Dirichlet);
+  }
+
+  int nquo=Quotients.size();
+  Level1.push_back(Quotients[0]);
+  Level1.push_back(Quotients[nquo-1]);
+  for(int h=1;h<nquo-1;h++){
+    Level2.push_back(Quotients[h]);
+  }
+
+  /////////////////////////////////////////////////////////////
+  // Gauge action
+  /////////////////////////////////////////////////////////////
+  Level3.push_back(&GaugeAction);
+  TheHMC.TheAction.push_back(Level1);
+  TheHMC.TheAction.push_back(Level2);
+  TheHMC.TheAction.push_back(Level3);
+  std::cout << GridLogMessage << " Action complete "<< std::endl;
+
+  /////////////////////////////////////////////////////////////
+
+  std::cout << GridLogMessage << " Running the HMC "<< std::endl;
+  TheHMC.Run();  // no smearing
+
+  Grid_finalize();
+} // main
+
+
+

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,
+	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,1,
-					       (void *)&rbuf[dir][0], recv_from_rank, bytes,tid);
+					       (void *)&rbuf[dir][0], recv_from_rank,1, bytes,tid);
 
 	    thread_critical { dbytes+=tbytes; }
 	  }

benchmarks/Benchmark_dwf_fp32.cc

@@ -32,10 +32,9 @@
 using namespace std;
 using namespace Grid;
 
-template<class d>
+////////////////////////
-struct scal {
+/// Move to domains ////
-  d internal;
+////////////////////////
-};
 
 Gamma::Algebra Gmu [] = {
 			 Gamma::Algebra::GammaX,
@@ -44,6 +43,7 @@ struct scal {
 			 Gamma::Algebra::GammaT
 };
 
+void Benchmark(int Ls, Coordinate Dirichlet);
 
 int main (int argc, char ** argv)
 {
@@ -52,24 +52,82 @@ 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;
     }
+  }
 
+  //////////////////
+  // With comms
+  //////////////////
+  Coordinate Dirichlet(Nd+1,0);
+
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing with full communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  
+  Benchmark(Ls,Dirichlet);
+
+  //////////////////
+  // Domain decomposed
+  //////////////////
+  Coordinate latt4  = GridDefaultLatt();
+  Coordinate mpi    = GridDefaultMpi();
+  Coordinate CommDim(Nd);
+  Coordinate shm;
+  GlobalSharedMemory::GetShmDims(mpi,shm);
+
+
+  //////////////////////
+  // Node level
+  //////////////////////
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing without internode communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+
+  for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
+  Dirichlet[0] = 0;
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
+
+  Benchmark(Ls,Dirichlet);
+
+  std::cout << "\n\n\n\n\n\n" <<std::endl;
+
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+  std::cout << GridLogMessage<< " Testing without intranode communication " <<std::endl;
+  std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
+
+  for(int d=0;d<Nd;d++) CommDim[d]= mpi[d]>1 ? 1 : 0;
+  Dirichlet[0] = 0;
+  Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0];
+  Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1];
+  Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2];
+  Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3];
+  
+  Benchmark(Ls,Dirichlet);
+
+  Grid_finalize();
+  exit(0);
+}
+void Benchmark(int Ls, Coordinate 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);
@@ -80,9 +138,9 @@ 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
@@ -100,7 +158,6 @@ int main (int argc, char ** argv)
   src = src*N2;
 #endif
 
-
   LatticeFermionF result(FGrid); result=Zero();
   LatticeFermionF    ref(FGrid);    ref=Zero();
   LatticeFermionF    tmp(FGrid);
@@ -108,29 +165,31 @@ int main (int argc, char ** argv)
 
   std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeFieldF Umu(UGrid);
+  LatticeGaugeFieldF UmuCopy(UGrid);
   SU<Nc>::HotConfiguration(RNG4,Umu);
+  UmuCopy=Umu;
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
-#if 0
+
-  Umu=1.0;
+  ////////////////////////////////////
-  for(int mu=0;mu<Nd;mu++){
+  // Apply BCs
-    LatticeColourMatrixF ttmp(UGrid);
+  ////////////////////////////////////
-    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
+  Coordinate Block(4);
-    //    if (mu !=2 ) ttmp = 0;
+  for(int d=0;d<4;d++)  Block[d]= Dirichlet[d+1];
-    //    ttmp = ttmp* pow(10.0,mu);
+
-    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
+  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block5 " << Dirichlet << std::endl;
-  }
+  std::cout << GridLogMessage << "Applying BCs for Dirichlet Block4 " << Block << std::endl;
-  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
+
-#endif
+  DirichletFilter<LatticeGaugeFieldF> Filter(Block);
+  Filter.applyFilter(Umu);
   
   ////////////////////////////////////
   // Naive wilson implementation
   ////////////////////////////////////
-  // replicate across fifth dimension
-  //  LatticeGaugeFieldF Umu5d(FGrid);
   std::vector<LatticeColourMatrixF> U(4,UGrid);
   for(int mu=0;mu<Nd;mu++){
     U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
   }
+
   std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
   if (1)
@@ -190,12 +249,15 @@ 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;
 
-  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  DomainWallFermionF::ImplParams p;
+  p.dirichlet=Dirichlet;
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,p);
+  Dw.ImportGauge(Umu);
+  
   int ncall =300;
 
   if (1) {
     FGrid->Barrier();
-    Dw.ZeroCounters();
     Dw.Dhop(src,result,0);
     std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
@@ -220,29 +282,20 @@ 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;
-    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
+    //    std::cout<<GridLogMessage << "RF  GiB/s (base 2) =   "<< 1000000. * data_rf/((t1-t0))<<std::endl;
-    std::cout<<GridLogMessage << "mem GiB/s (base 2) =   "<< 1000000. * data_mem/((t1-t0))<<std::endl;
+    //    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)
@@ -286,21 +339,20 @@ int main (int argc, char ** argv)
     }
     ref = -0.5*ref;
   }
-  //  dump=1;
+
-  Dw.Dhop(src,result,1);
+  Dw.Dhop(src,result,DaggerYes);
+
+  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
   std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
+  std::cout << GridLogMessage << "----------------------------------------------------------------" << std::endl;
+
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
   std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
   std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result;
   std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
-  if((norm2(err)>1.0e-4)){
+  assert((norm2(err)<1.0e-4));
-/*
+  
-	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
-	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
-	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
-*/
-  }
   LatticeFermionF src_e (FrbGrid);
   LatticeFermionF src_o (FrbGrid);
   LatticeFermionF r_e   (FrbGrid);
@@ -330,7 +382,6 @@ 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 +403,6 @@ 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 +417,7 @@ 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)){
+  assert(norm2(err)<1.0e-4);
-    /*
-	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
-	std::cout<< "Deo REF\n " <<result  << std::endl;
-	std::cout<< "Deo ERR   \n " << err <<std::endl;
-    */
-  }
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
@@ -382,6 +426,4 @@ 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);
-}

examples/Example_Laplacian.cc

@@ -93,14 +93,14 @@ template<class Field> class FreeLaplacianStencil : public SparseMatrixBase<Field
 {
 public:
   typedef typename Field::vector_object siteObject;
-  typedef CartesianStencil<siteObject, siteObject, int> StencilImpl;
+  typedef CartesianStencil<siteObject, siteObject, SimpleStencilParams> StencilImpl;
 
   GridBase *grid;
   StencilImpl Stencil;
   SimpleCompressor<siteObject> Compressor;
   
   FreeLaplacianStencil(GridBase *_grid)
-    : Stencil    (_grid,6,Even,directions,displacements,0), grid(_grid)
+    : Stencil    (_grid,6,Even,directions,displacements,SimpleStencilParams()), grid(_grid)
   {  };
   
   virtual GridBase *Grid(void) { return grid; };
@@ -168,7 +168,8 @@ public:
   typedef iImplDoubledGaugeField<Simd> SiteDoubledGaugeField;
   typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
 
-  typedef CartesianStencil<siteObject, siteObject, int> StencilImpl;
+  typedef CartesianStencil<siteObject, siteObject,SimpleStencilParams> StencilImpl;
+  SimpleStencilParams p;
 
   GridBase *grid;
   StencilImpl Stencil;
@@ -177,7 +178,7 @@ public:
   CovariantLaplacianStencil(GaugeField &Umu)
     :
       grid(Umu.Grid()),
-      Stencil    (grid,6,Even,directions,displacements,0),
+      Stencil    (grid,6,Even,directions,displacements,p),
       Uds(grid)
   {
     for (int mu = 0; mu < Nd; mu++) {

systems/Crusher/dwf.slurm

@@ -3,28 +3,28 @@
 #SBATCH -A LGT104
 #SBATCH -t 01:00:00
 ##SBATCH -U openmpThu
-##SBATCH -p ecp
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
 #SBATCH -N 1
-#SBATCH -n 1
+#SBATCH -n 8
 #SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
 
 DIR=.
 module list
-#export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-#export MPICH_SMP_SINGLE_COPY_MODE=NONE
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export OMP_NUM_THREADS=1
 
-AT=8
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
 
-PARAMS=" --accelerator-threads ${AT} --grid 24.24.24.24 --shm-mpi 0 --mpi 1.1.1.1"
+PARAMS=" --accelerator-threads 16 --grid 32.32.32.256 --mpi 1.1.1.8 --comms-overlap --shm 2048 --shm-mpi 0"
-
+echo $PARAMS
-srun --gpus-per-task 1 -n1 ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+srun --gpus-per-task 1 -n8  ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+
 
 

systems/Crusher/dwf8.slurm

@@ -6,22 +6,43 @@
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
 #SBATCH -e DWF.%J
-#SBATCH -N 1
+#SBATCH -N 8
-#SBATCH -n 8
+#SBATCH -n 64
 #SBATCH --exclusive
+#SBATCH --gpu-bind=map_gpu:0,1,2,3,7,6,5,4
 
 DIR=.
 module list
+export MPICH_OFI_NIC_POLICY=GPU
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
-#export MPICH_SMP_SINGLE_COPY_MODE=NONE
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
+export MPICH_SMP_SINGLE_COPY_MODE=NONE
 export OMP_NUM_THREADS=1
 
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads 8 --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
 
-srun --gpus-per-task 1 -n8 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.256 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.256.8node
+
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 1"
+echo $PARAMS
+srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 4.4.4.1 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_dwf_fp32 $PARAMS > dwf.64.64.64.32.8node.shm0
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 1"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node
+
+PARAMS=" --accelerator-threads 16 --grid 64.64.64.32 --mpi 2.2.2.8 --comms-overlap --shm 2048 --shm-mpi 0"
+echo $PARAMS
+#srun --gpus-per-task 1 -N8 -n64  ./benchmarks/Benchmark_ITT $PARAMS > itt.8node_shm0
+
 
 

systems/Crusher/mpiwrapper.sh

@@ -1,10 +1,11 @@
 #!/bin/bash
 
 lrank=$SLURM_LOCALID
+lgpu=(0 1 2 3 7 6 5 4)
 
-export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
+export ROCR_VISIBLE_DEVICES=${lgpu[$lrank]}
 
-echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
+echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES "
 
 $*
 

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

@@ -1,34 +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 
-do
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1  --device-mem 32768
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1  --device-mem 32768
-done
-#mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.1x2.log
-#mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.2x1.log
-
-

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/Perlmutter/config-command

@@ -1,9 +1,14 @@
+DIR=`pwd`
+PREFIX=$DIR/../Prequisites/install/
 ../../configure \
     --enable-comms=mpi \
     --enable-simd=GPU \
     --enable-shm=nvlink \
     --enable-gen-simd-width=64 \
     --enable-accelerator=cuda \
+    --enable-setdevice \
+    --disable-accelerator-cshift \
+    --with-gmp=$PREFIX \
     --disable-fermion-reps \
     --disable-unified \
     --disable-gparity \

systems/Perlmutter/dwf4.slurm

@@ -1,24 +1,27 @@
 #!/bin/bash
-#SBATCH -A mp13
+#SBATCH -A m3886_g
 #SBATCH -C gpu
-#SBATCH -q regular
+#SBATCH -q debug
 #SBATCH -t 0:20:00
-#SBATCH -n 16
-#SBATCH --ntasks-per-node=4
 #SBATCH -c 32
-#SBATCH --exclusive
+#SBATCH -N 1
+#SBATCH -n 4
+#SBATCH --ntasks-per-node=4
 #SBATCH --gpus-per-task=1
-#SBATCH --gpu-bind=map_gpu:0,1,2,3
+#SBATCH --exclusive
+#SBATCH --gpu-bind=none
 
 export SLURM_CPU_BIND="cores"
-export MPICH_RDMA_ENABLED_CUDA=1
 export MPICH_GPU_SUPPORT_ENABLED=1
-srun ./benchmarks/Benchmark_comms_host_device --mpi 2.2.2.2  --accelerator-threads 8 > comms.4node
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
 
-OPT="--comms-overlap --comms-concurrent --shm-mpi 0"
+OPT="--comms-sequential --shm-mpi 1"
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 64.64.64.64 --accelerator-threads 8 --shm 2048 $OPT > dwf.64.64.64.64.4node.opt0
+VOL=64.64.64.64
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 48.48.48.48 --accelerator-threads 8 --shm 2048 $OPT > dwf.48.48.48.48.4node.opt0
+srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
+#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.1.1.4 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
+#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.8 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
 
-OPT="--comms-overlap --comms-concurrent --shm-mpi 1"
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 64.64.64.64 --accelerator-threads 8 --shm 2048 $OPT > dwf.64.64.64.64.4node.opt1
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.2.2 --grid 48.48.48.48 --accelerator-threads 8 --shm 2048 $OPT > dwf.48.48.48.48.4node.opt1

systems/Perlmutter/sourceme.sh

@@ -1,4 +1,4 @@
 
 export CRAY_ACCEL_TARGET=nvidia80
 
-module load PrgEnv-gnu cpe-cuda cuda
+module load PrgEnv-gnu cpe-cuda cudatoolkit/11.4

systems/Spock/config-command

@@ -6,6 +6,8 @@
 --enable-simd=GPU \
 --disable-fermion-reps \
 --disable-gparity \
+--with-gmp=$OLCF_GMP_ROOT \
+--with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 CXX=hipcc MPICXX=mpicxx \
 CXXFLAGS="-fPIC -I/opt/rocm-4.3.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
 --prefix=/ccs/home/chulwoo/Grid \

systems/Spock/dwf8.slurm

@@ -1,8 +1,7 @@
 #!/bin/bash
 # Begin LSF Directives
 #SBATCH -A LGT104
-#SBATCH -t 01:00:00
+#SBATCH -t 3:00:00
-##SBATCH -U openmpThu
 #SBATCH -p ecp
 #SBATCH -J DWF
 #SBATCH -o DWF.%J
@@ -14,13 +13,12 @@ DIR=.
 module list
 export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
 export MPICH_GPU_SUPPORT_ENABLED=1
-#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
+export MPICH_SMP_SINGLE_COPY_MODE=CMA
-export MPICH_SMP_SINGLE_COPY_MODE=NONE
+
-#export MPICH_SMP_SINGLE_COPY_MODE=CMA
 export OMP_NUM_THREADS=8
 
 AT=8
 echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
-PARAMS=" --accelerator-threads ${AT} --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
+PARAMS=" --accelerator-threads ${AT} --grid 16.16.16.48 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
-srun -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS
+srun -N2 -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./HMC/Mobius2p1f_DD_RHMC $PARAMS
 

systems/Spock/sourceme.sh

@@ -1,5 +1,9 @@
+module load emacs
 module load PrgEnv-gnu
-module load rocm/4.3.0
+module load rocm/4.5.0
 module load gmp
 module load cray-fftw
 module load craype-accel-amd-gfx908
+export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
+export MPICH_GPU_SUPPORT_ENABLED=1
+export LD_LIBRARY_PATH=/opt/cray/pe/gcc/mpfr/3.1.4/lib/:$LD_LIBRARY_PATH

systems/Tursa/dwf.4node.perf

@@ -1,25 +1,25 @@
-tu-c0r0n00 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n00 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n00 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n00 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n09 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n00 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n00 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n00 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n06 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n06 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n06 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n06 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n09 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n06 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n09 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n06 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n03 - 1 device=1 binding=--interleave=2,3
+tu-c0r3n03 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n06 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n03 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n09 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n03 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n00 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n03 - 3 device=3 binding=--interleave=6,7
-tu-c0r0n03 - 0 device=0 binding=--interleave=0,1
+tu-c0r3n09 - 0 device=0 binding=--interleave=0,1
-tu-c0r0n03 - 2 device=2 binding=--interleave=4,5
+tu-c0r3n09 - 1 device=1 binding=--interleave=2,3
-tu-c0r0n06 - 3 device=3 binding=--interleave=6,7
+tu-c0r3n09 - 2 device=2 binding=--interleave=4,5
-tu-c0r0n03 - 3 device=3 binding=--interleave=6,7
+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: Configure options --enable-setdevice=no 
-AcceleratorCudaInit: ================================================
 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: Configure options --enable-setdevice=no 
-AcceleratorCudaInit: ================================================
 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
+local rank 0 device 0 bus id: 0000:03:00.0
-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: ================================================
-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 : 1.875883 s : Grid Layout
-Grid : Message : 1.198530 s : 	Global lattice size  : 64 64 64 64 
+Grid : Message : 1.875893 s : 	Global lattice size  : 64 64 64 64 
-Grid : Message : 1.198534 s : 	OpenMP threads       : 4
+Grid : Message : 1.875897 s : 	OpenMP threads       : 4
-Grid : Message : 1.198535 s : 	MPI tasks            : 2 2 2 2 
+Grid : Message : 1.875898 s : 	MPI tasks            : 2 2 2 2 
-Grid : Message : 1.397615 s : Making s innermost grids
+Grid : Message : 1.993571 s : Initialising 4d RNG
-Grid : Message : 1.441828 s : Initialising 4d RNG
+Grid : Message : 2.881990 s : Intialising parallel RNG with unique string 'The 4D RNG'
-Grid : Message : 1.547973 s : Intialising parallel RNG with unique string 'The 4D RNG'
+Grid : Message : 2.882370 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
-Grid : Message : 1.547998 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
+Grid : Message : 2.495044 s : Initialising 5d RNG
-Grid : Message : 1.954777 s : Initialising 5d RNG
+Grid : Message : 4.120900 s : Intialising parallel RNG with unique string 'The 5D RNG'
-Grid : Message : 3.633825 s : Intialising parallel RNG with unique string 'The 5D RNG'
+Grid : Message : 4.121350 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
-Grid : Message : 3.633869 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
+Grid : Message : 15.268010 s : Drawing gauge field
-Grid : Message : 12.162710 s : Initialised RNGs
+Grid : Message : 16.234025 s : Random gauge initialised 
-Grid : Message : 15.882520 s : Drawing gauge field
+Grid : Message : 16.234057 s : Applying BCs 
-Grid : Message : 15.816362 s : Random gauge initialised 
+Grid : Message : 16.365565 s : Setting up Cshift based reference 
-Grid : Message : 17.279671 s : Setting up Cshift based reference 
+Grid : Message : 44.512418 s : *****************************************************************
-Grid : Message : 26.331426 s : *****************************************************************
+Grid : Message : 44.512448 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
-Grid : Message : 26.331452 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
+Grid : Message : 44.512450 s : *****************************************************************
-Grid : Message : 26.331454 s : *****************************************************************
+Grid : Message : 44.512451 s : *****************************************************************
-Grid : Message : 26.331456 s : *****************************************************************
+Grid : Message : 44.512452 s : * Benchmarking DomainWallFermionR::Dhop                  
-Grid : Message : 26.331458 s : * Benchmarking DomainWallFermionR::Dhop                  
+Grid : Message : 44.512453 s : * Vectorising space-time by 8
-Grid : Message : 26.331459 s : * Vectorising space-time by 8
+Grid : Message : 44.512454 s : * VComplexF size is 64 B
-Grid : Message : 26.331463 s : * VComplexF size is 64 B
+Grid : Message : 44.512456 s : * SINGLE precision 
-Grid : Message : 26.331465 s : * SINGLE precision 
+Grid : Message : 44.512459 s : * Using Overlapped Comms/Compute
-Grid : Message : 26.331467 s : * Using Overlapped Comms/Compute
+Grid : Message : 44.512460 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 26.331468 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 44.512461 s : *****************************************************************
-Grid : Message : 26.331469 s : *****************************************************************
+Grid : Message : 46.389070 s : Called warmup
-Grid : Message : 28.413717 s : Called warmup
+Grid : Message : 49.211265 s : Called Dw 300 times in 2.82203e+06 us
-Grid : Message : 56.418423 s : Called Dw 3000 times in 2.80047e+07 us
+Grid : Message : 49.211295 s : mflop/s =   3.76681e+07
-Grid : Message : 56.418476 s : mflop/s =   3.79581e+07
+Grid : Message : 49.211297 s : mflop/s per rank =  2.35425e+06
-Grid : Message : 56.418479 s : mflop/s per rank =  2.37238e+06
+Grid : Message : 49.211299 s : mflop/s per node =  9.41702e+06
-Grid : Message : 56.418481 s : mflop/s per node =  9.48953e+06
+Grid : Message : 49.211301 s : RF  GiB/s (base 2) =   76540.6
-Grid : Message : 56.418483 s : RF  GiB/s (base 2) =   77130
+Grid : Message : 49.211308 s : mem GiB/s (base 2) =   47837.9
-Grid : Message : 56.418485 s : mem GiB/s (base 2) =   48206.3
+Grid : Message : 49.214868 s : norm diff   1.06409e-13
-Grid : Message : 56.422076 s : norm diff   1.03481e-13
+Grid : Message : 92.647781 s : Compare to naive wilson implementation Dag to verify correctness
-Grid : Message : 56.456894 s : #### Dhop calls report 
+Grid : Message : 92.647816 s : Called DwDag
-Grid : Message : 56.456899 s : WilsonFermion5D Number of DhopEO Calls   : 6002
+Grid : Message : 92.647817 s : norm dag result 12.0421
-Grid : Message : 56.456903 s : WilsonFermion5D TotalTime   /Calls        : 4710.93 us
+Grid : Message : 92.801806 s : norm dag ref    12.0421
-Grid : Message : 56.456905 s : WilsonFermion5D CommTime    /Calls        : 3196.15 us
+Grid : Message : 92.817724 s : norm dag diff   7.21921e-14
-Grid : Message : 56.456908 s : WilsonFermion5D FaceTime    /Calls        : 494.392 us
+Grid : Message : 92.858973 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
-Grid : Message : 56.456910 s : WilsonFermion5D ComputeTime1/Calls        : 44.4107 us
+Grid : Message : 93.210378 s : src_e0.499997
-Grid : Message : 56.456912 s : WilsonFermion5D ComputeTime2/Calls        : 1037.75 us
+Grid : Message : 93.583286 s : src_o0.500003
-Grid : Message : 56.456921 s : Average mflops/s per call                : 3.55691e+09
+Grid : Message : 93.682468 s : *********************************************************
-Grid : Message : 56.456925 s : Average mflops/s per call per rank       : 2.22307e+08
+Grid : Message : 93.682471 s : * Benchmarking DomainWallFermionF::DhopEO                
-Grid : Message : 56.456928 s : Average mflops/s per call per node       : 8.89228e+08
+Grid : Message : 93.682472 s : * Vectorising space-time by 8
-Grid : Message : 56.456930 s : Average mflops/s per call (full)         : 3.82915e+07
+Grid : Message : 93.682473 s : * SINGLE precision 
-Grid : Message : 56.456933 s : Average mflops/s per call per rank (full): 2.39322e+06
+Grid : Message : 93.682475 s : * Using Overlapped Comms/Compute
-Grid : Message : 56.456952 s : Average mflops/s per call per node (full): 9.57287e+06
+Grid : Message : 93.682476 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 56.456954 s : WilsonFermion5D Stencil
+Grid : Message : 93.682477 s : *********************************************************
-Grid : Message : 56.457016 s :  Stencil calls 3001
+Grid : Message : 95.162342 s : Deo mflop/s =   3.92487e+07
-Grid : Message : 56.457022 s :  Stencil halogtime 0
+Grid : Message : 95.162387 s : Deo mflop/s per rank   2.45305e+06
-Grid : Message : 56.457024 s :  Stencil gathertime 55.9154
+Grid : Message : 95.162389 s : Deo mflop/s per node   9.81219e+06
-Grid : Message : 56.457026 s :  Stencil gathermtime 20.1073
+Grid : Message : 95.232801 s : r_e6.02111
-Grid : Message : 56.457028 s :  Stencil mergetime 18.5585
+Grid : Message : 95.240061 s : r_o6.02102
-Grid : Message : 56.457030 s :  Stencil decompresstime 0.0639787
+Grid : Message : 95.245975 s : res12.0421
-Grid : Message : 56.457032 s :  Stencil comms_bytes 4.02653e+08
+Grid : Message : 95.833402 s : norm diff   0
-Grid : Message : 56.457034 s :  Stencil commtime 6379.93
+Grid : Message : 96.573829 s : norm diff even  0
-Grid : Message : 56.457036 s :  Stencil 63.1124 GB/s per rank
+Grid : Message : 96.868272 s : norm diff odd   0
-Grid : Message : 56.457038 s :  Stencil 252.45 GB/s per node
+ Dirichlet block [0 64 64 32 32]
-Grid : Message : 56.457040 s : WilsonFermion5D StencilEven
+Grid : Message : 97.756909 s : Grid Layout
-Grid : Message : 56.457048 s : WilsonFermion5D StencilOdd
+Grid : Message : 97.756911 s : 	Global lattice size  : 64 64 64 64 
-Grid : Message : 56.457062 s : WilsonFermion5D Stencil     Reporti()
+Grid : Message : 97.756921 s : 	OpenMP threads       : 4
-Grid : Message : 56.457065 s : WilsonFermion5D StencilEven Reporti()
+Grid : Message : 97.756922 s : 	MPI tasks            : 2 2 2 2 
-Grid : Message : 56.457066 s : WilsonFermion5D StencilOdd  Reporti()
+Grid : Message : 97.897085 s : Initialising 4d RNG
-Grid : Message : 79.259261 s : Compare to naive wilson implementation Dag to verify correctness
+Grid : Message : 97.965061 s : Intialising parallel RNG with unique string 'The 4D RNG'
-Grid : Message : 79.259287 s : Called DwDag
+Grid : Message : 97.965097 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
-Grid : Message : 79.259288 s : norm dag result 12.0421
+Grid : Message : 98.367431 s : Initialising 5d RNG
-Grid : Message : 79.271740 s : norm dag ref    12.0421
+Grid : Message : 99.752745 s : Intialising parallel RNG with unique string 'The 5D RNG'
-Grid : Message : 79.287759 s : norm dag diff   7.63236e-14
+Grid : Message : 99.752790 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
-Grid : Message : 79.328100 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
+Grid : Message : 111.290148 s : Drawing gauge field
-Grid : Message : 79.955951 s : src_e0.499997
+Grid : Message : 112.349289 s : Random gauge initialised 
-Grid : Message : 80.633620 s : src_o0.500003
+Grid : Message : 112.349320 s : Applying BCs 
-Grid : Message : 80.164163 s : *********************************************************
+Grid : Message : 113.948740 s : Setting up Cshift based reference 
-Grid : Message : 80.164168 s : * Benchmarking DomainWallFermionF::DhopEO                
+Grid : Message : 140.320415 s : *****************************************************************
-Grid : Message : 80.164170 s : * Vectorising space-time by 8
+Grid : Message : 140.320443 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
-Grid : Message : 80.164172 s : * SINGLE precision 
+Grid : Message : 140.320444 s : *****************************************************************
-Grid : Message : 80.164174 s : * Using Overlapped Comms/Compute
+Grid : Message : 140.320445 s : *****************************************************************
-Grid : Message : 80.164177 s : * Using GENERIC Nc WilsonKernels
+Grid : Message : 140.320446 s : * Benchmarking DomainWallFermionR::Dhop                  
-Grid : Message : 80.164178 s : *********************************************************
+Grid : Message : 140.320447 s : * Vectorising space-time by 8
-Grid : Message : 93.797635 s : Deo mflop/s =   3.93231e+07
+Grid : Message : 140.320448 s : * VComplexF size is 64 B
-Grid : Message : 93.797670 s : Deo mflop/s per rank   2.45769e+06
+Grid : Message : 140.320450 s : * SINGLE precision 
-Grid : Message : 93.797672 s : Deo mflop/s per node   9.83077e+06
+Grid : Message : 140.320451 s : * Using Overlapped Comms/Compute
-Grid : Message : 93.797674 s : #### Dhop calls report 
+Grid : Message : 140.320452 s : * Using GENERIC Nc WilsonKernels
-Grid : Message : 93.797675 s : WilsonFermion5D Number of DhopEO Calls   : 3001
+Grid : Message : 140.320453 s : *****************************************************************
-Grid : Message : 93.797677 s : WilsonFermion5D TotalTime   /Calls        : 4542.83 us
+Grid : Message : 142.296150 s : Called warmup
-Grid : Message : 93.797679 s : WilsonFermion5D CommTime    /Calls        : 2978.97 us
+Grid : Message : 144.397678 s : Called Dw 300 times in 2.36719e+06 us
-Grid : Message : 93.797681 s : WilsonFermion5D FaceTime    /Calls        : 602.287 us
+Grid : Message : 144.397700 s : mflop/s =   4.49058e+07
-Grid : Message : 93.797683 s : WilsonFermion5D ComputeTime1/Calls        : 67.1416 us
+Grid : Message : 144.397702 s : mflop/s per rank =  2.80661e+06
-Grid : Message : 93.797685 s : WilsonFermion5D ComputeTime2/Calls        : 1004.07 us
+Grid : Message : 144.397704 s : mflop/s per node =  1.12265e+07
-Grid : Message : 93.797713 s : Average mflops/s per call                : 3.30731e+09
+Grid : Message : 144.397706 s : RF  GiB/s (base 2) =   91247.6
-Grid : Message : 93.797717 s : Average mflops/s per call per rank       : 2.06707e+08
+Grid : Message : 144.397708 s : mem GiB/s (base 2) =   57029.7
-Grid : Message : 93.797719 s : Average mflops/s per call per node       : 8.26827e+08
+Grid : Message : 144.401269 s : norm diff   9.78944e-14
-Grid : Message : 93.797721 s : Average mflops/s per call (full)         : 3.97084e+07
+Grid : Message : 186.885460 s : Compare to naive wilson implementation Dag to verify correctness
-Grid : Message : 93.797727 s : Average mflops/s per call per rank (full): 2.48178e+06
+Grid : Message : 186.885492 s : Called DwDag
-Grid : Message : 93.797732 s : Average mflops/s per call per node (full): 9.92711e+06
+Grid : Message : 186.885493 s : norm dag result 10.4157
-Grid : Message : 93.797735 s : WilsonFermion5D Stencil
+Grid : Message : 186.897154 s : norm dag ref    11.2266
-Grid : Message : 93.797746 s : WilsonFermion5D StencilEven
+Grid : Message : 186.912538 s : norm dag diff   0.484633
-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

systems/Tursa/dwf4.slurm

@@ -1,14 +1,13 @@
 #!/bin/bash
 #SBATCH -J dslash
-#SBATCH -A tc002
+#SBATCH -A dp207
-#SBATCH -t 2:20:00
-#SBATCH --nodelist=tu-c0r0n[00,03,06,09]
 #SBATCH --exclusive
 #SBATCH --nodes=4
 #SBATCH --ntasks=16
+#SBATCH --qos=standard
 #SBATCH --ntasks-per-node=4
 #SBATCH --cpus-per-task=8
-#SBATCH --time=12:00:00
+#SBATCH --time=0:05:00
 #SBATCH --partition=gpu
 #SBATCH --gres=gpu:4
 #SBATCH --output=%x.%j.out

systems/mac-arm/config-command-mpi

@@ -1 +1 @@
-CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi --enable-unified=no
+CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi

tests/Test_simd.cc

@@ -793,7 +793,6 @@ int main (int argc, char ** argv)
     }
     std::cout <<" OK ! "<<std::endl;
 
-#ifdef USE_FP16
     // Double to Half
     std::cout << GridLogMessage<< "Double to half" ;
     precisionChange(&H[0],&D[0],Ndp);
@@ -823,7 +822,6 @@ int main (int argc, char ** argv)
       assert( tmp < 1.0e-3 );
     }
     std::cout <<" OK ! "<<std::endl;
-#endif
 
   }
   Grid_finalize();

tests/Test_stencil.cc

@@ -31,7 +31,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 using namespace std;
 using namespace Grid;
- ;
 
 int main(int argc, char ** argv) {
   Grid_init(&argc, &argv);
@@ -80,7 +79,8 @@ int main(int argc, char ** argv) {
     Foo=lex;
   }
 
-  typedef CartesianStencil<vobj,vobj,int> Stencil;
+  typedef CartesianStencil<vobj,vobj,SimpleStencilParams> Stencil;
+  SimpleStencilParams p;
     for(int dir=0;dir<4;dir++){
       for(int disp=0;disp<Fine._fdimensions[dir];disp++){
 
@@ -90,7 +90,7 @@ int main(int argc, char ** argv) {
 	std::vector<int> directions(npoint,dir);
 	std::vector<int> displacements(npoint,disp);
 
-	Stencil myStencil(&Fine,npoint,0,directions,displacements,0);
+	Stencil myStencil(&Fine,npoint,0,directions,displacements,p);
 	Coordinate ocoor(4);
 	for(int o=0;o<Fine.oSites();o++){
 	  Fine.oCoorFromOindex(ocoor,o);
@@ -183,8 +183,8 @@ int main(int argc, char ** argv) {
 	std::vector<int> directions(npoint,dir);
 	std::vector<int> displacements(npoint,disp);
 
-	Stencil EStencil(&rbFine,npoint,Even,directions,displacements,0);
+	Stencil EStencil(&rbFine,npoint,Even,directions,displacements,p);
-	Stencil OStencil(&rbFine,npoint,Odd,directions,displacements,0);
+	Stencil OStencil(&rbFine,npoint,Odd,directions,displacements,p);
 
 	Coordinate ocoor(4);
 	for(int o=0;o<Fine.oSites();o++){

tests/core/Test_fft_matt.cc

@@ -1,270 +0,0 @@
-    /*************************************************************************************
-    grid` physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./tests/Test_cshift.cc
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/Grid.h>
-
-using namespace Grid;
-
-Gamma::Algebra Gmu [] = {
-  Gamma::Algebra::GammaX,
-  Gamma::Algebra::GammaY,
-  Gamma::Algebra::GammaZ,
-  Gamma::Algebra::GammaT,
-  Gamma::Algebra::Gamma5
-};
-
-int main (int argc, char ** argv)
-{
-  Grid_init(&argc,&argv);
-
-  int threads = GridThread::GetThreads();
-  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
-
-  Coordinate latt_size   = GridDefaultLatt();
-  Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
-  Coordinate mpi_layout  = GridDefaultMpi();
-
-  int vol = 1;
-  for(int d=0;d<latt_size.size();d++){
-    vol = vol * latt_size[d];
-  }
-  GridCartesian         GRID(latt_size,simd_layout,mpi_layout);
-  GridRedBlackCartesian RBGRID(&GRID);
-
-  LatticeComplexD    coor(&GRID);
-  ComplexD ci(0.0,1.0);
-
-  std::vector<int> seeds({1,2,3,4});
-  GridSerialRNG          sRNG;  sRNG.SeedFixedIntegers(seeds); // naughty seeding
-  GridParallelRNG          pRNG(&GRID);
-  pRNG.SeedFixedIntegers(seeds);
-
-  LatticeGaugeFieldD Umu(&GRID);
-  SU<Nc>::ColdConfiguration(pRNG,Umu); // Unit gauge
-
-  ////////////////////////////////////////////////////
-  // Wilson test
-  ////////////////////////////////////////////////////
-  {
-    LatticeFermionD    src(&GRID); gaussian(pRNG,src);
-    LatticeFermionD    src_p(&GRID);
-    LatticeFermionD    tmp(&GRID);
-    LatticeFermionD    ref(&GRID);
-    LatticeFermionD    result(&GRID);
-    
-    RealD mass=0.1;
-    WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
-    
-    Dw.M(src,ref);
-    std::cout << "Norm src "<<norm2(src)<<std::endl;
-    std::cout << "Norm Dw x src "<<norm2(ref)<<std::endl;
-    {
-      FFT theFFT(&GRID);
-
-      ////////////////
-      // operator in Fourier space
-      ////////////////
-      tmp =ref;
-      theFFT.FFT_all_dim(result,tmp,FFT::forward);
-      std::cout<<"FFT[ Dw x src ]  "<< norm2(result)<<std::endl;    
-
-      tmp = src;
-      theFFT.FFT_all_dim(src_p,tmp,FFT::forward);
-      std::cout<<"FFT[ src      ]  "<< norm2(src_p)<<std::endl;
-      
-      /////////////////////////////////////////////////////////////////
-      // work out the predicted FT from Fourier
-      /////////////////////////////////////////////////////////////////
-      auto FGrid = &GRID;
-      LatticeFermionD    Kinetic(FGrid); Kinetic = Zero();
-      LatticeComplexD    kmu(FGrid); 
-      LatticeInteger     scoor(FGrid); 
-      LatticeComplexD    sk (FGrid); sk = Zero();
-      LatticeComplexD    sk2(FGrid); sk2= Zero();
-      LatticeComplexD    W(FGrid); W= Zero();
-      LatticeComplexD    one(FGrid); one =ComplexD(1.0,0.0);
-      ComplexD ci(0.0,1.0);
-    
-      for(int mu=0;mu<Nd;mu++) {
-	
-	RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
-
-	LatticeCoordinate(kmu,mu);
-
-	kmu = TwoPiL * kmu;
-      
-	sk2 = sk2 + 2.0*sin(kmu*0.5)*sin(kmu*0.5);
-	sk  = sk  +     sin(kmu)    *sin(kmu); 
-      
-	// -1/2 Dw ->  1/2 gmu (eip - emip) = i sinp gmu
-	Kinetic = Kinetic + sin(kmu)*ci*(Gamma(Gmu[mu])*src_p);
-	
-      }
-    
-      W = mass + sk2; 
-      Kinetic = Kinetic + W * src_p;
-    
-      std::cout<<"Momentum space src         "<< norm2(src_p)<<std::endl;
-      std::cout<<"Momentum space Dw x src    "<< norm2(Kinetic)<<std::endl;
-      std::cout<<"FT[Coordinate space Dw]    "<< norm2(result)<<std::endl;
-    
-      result = result - Kinetic;
-      std::cout<<"diff "<< norm2(result)<<std::endl;
-      
-    }
-
-    std::cout << " =======================================" <<std::endl;
-    std::cout << " Checking FourierFreePropagator x Dw = 1" <<std::endl;
-    std::cout << " =======================================" <<std::endl;
-    std::cout << "Dw src = " <<norm2(src)<<std::endl;
-    std::cout << "Dw tmp = " <<norm2(tmp)<<std::endl;
-    Dw.M(src,tmp);
-    Dw.FreePropagator(tmp,ref,mass);
-
-    std::cout << "Dw ref = " <<norm2(ref)<<std::endl;
-    
-    ref = ref - src;
-    
-    std::cout << "Dw ref-src = " <<norm2(ref)<<std::endl;
-  }
-
-
-  ////////////////////////////////////////////////////
-  // Wilson prop
-  ////////////////////////////////////////////////////
-  {
-    std::cout<<"****************************************"<<std::endl;
-    std::cout << "Wilson Mom space 4d propagator \n";
-    std::cout<<"****************************************"<<std::endl;
-
-    LatticeFermionD    src(&GRID); gaussian(pRNG,src);
-    LatticeFermionD    tmp(&GRID);
-    LatticeFermionD    ref(&GRID);
-    LatticeFermionD    diff(&GRID);
-
-    src=Zero();
-    Coordinate point(4,0); // 0,0,0,0
-    SpinColourVectorD ferm;
-    ferm=Zero();
-    ferm()(0)(0) = ComplexD(1.0);
-    pokeSite(ferm,src,point);
-
-    RealD mass=0.1;
-    WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
-
-    // Momentum space prop
-    std::cout << " Solving by FFT and Feynman rules" <<std::endl;
-    Dw.FreePropagator(src,ref,mass) ;
-
-    Gamma G5(Gamma::Algebra::Gamma5);
-
-    LatticeFermionD    result(&GRID); 
-    const int sdir=0;
-    
-    ////////////////////////////////////////////////////////////////////////
-    // Conjugate gradient on normal equations system
-    ////////////////////////////////////////////////////////////////////////
-    std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
-    Dw.Mdag(src,tmp);
-    src=tmp;
-    MdagMLinearOperator<WilsonFermionD,LatticeFermionD> HermOp(Dw);
-    ConjugateGradient<LatticeFermionD> CG(1.0e-10,10000);
-    CG(HermOp,src,result);
-    
-    ////////////////////////////////////////////////////////////////////////
-    std::cout << " Taking difference" <<std::endl;
-    std::cout << "Dw result "<<norm2(result)<<std::endl;
-    std::cout << "Dw ref     "<<norm2(ref)<<std::endl;
-    
-    diff = ref - result;
-    std::cout << "result - ref     "<<norm2(diff)<<std::endl;
-
-    DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
-  }
-
-  ////////////////////////////////////////////////////
-  //Gauge invariance test
-  ////////////////////////////////////////////////////
-  {
-    std::cout<<"****************************************"<<std::endl;
-    std::cout << "Gauge invariance test \n";
-    std::cout<<"****************************************"<<std::endl;
-    LatticeGaugeField     U_GT(&GRID); // Gauge transformed field
-    LatticeColourMatrix   g(&GRID);    // local Gauge xform matrix
-    U_GT = Umu;
-    // Make a random xform to teh gauge field
-    SU<Nc>::RandomGaugeTransform(pRNG,U_GT,g); // Unit gauge
-
-    LatticeFermionD    src(&GRID);
-    LatticeFermionD    tmp(&GRID);
-    LatticeFermionD    ref(&GRID);
-    LatticeFermionD    diff(&GRID);
-
-    // could loop over colors
-    src=Zero();
-    Coordinate point(4,0); // 0,0,0,0
-    SpinColourVectorD ferm;
-    ferm=Zero();
-    ferm()(0)(0) = ComplexD(1.0);
-    pokeSite(ferm,src,point);
-
-    RealD mass=0.1;
-    WilsonFermionD Dw(U_GT,GRID,RBGRID,mass);
-
-    // Momentum space prop
-    std::cout << " Solving by FFT and Feynman rules" <<std::endl;
-    Dw.FreePropagator(src,ref,mass) ;
-
-    Gamma G5(Gamma::Algebra::Gamma5);
-
-    LatticeFermionD    result(&GRID); 
-    const int sdir=0;
-    
-    ////////////////////////////////////////////////////////////////////////
-    // Conjugate gradient on normal equations system
-    ////////////////////////////////////////////////////////////////////////
-    std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
-    Dw.Mdag(src,tmp);
-    src=tmp;
-    MdagMLinearOperator<WilsonFermionD,LatticeFermionD> HermOp(Dw);
-    ConjugateGradient<LatticeFermionD> CG(1.0e-10,10000);
-    CG(HermOp,src,result);
-    
-    ////////////////////////////////////////////////////////////////////////
-    std::cout << " Taking difference" <<std::endl;
-    std::cout << "Dw result "<<norm2(result)<<std::endl;
-    std::cout << "Dw ref     "<<norm2(ref)<<std::endl;
-    
-    diff = ref - result;
-    std::cout << "result - ref     "<<norm2(diff)<<std::endl;
-
-    DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
-  }
-  
-  
-  Grid_finalize();
-}

tests/core/Test_memory_manager.cc

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

tests/core/Test_slicesum.cc

@@ -1,73 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./tests/Test_poisson_fft.cc
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/Grid.h>
-#include <Grid/lattice/Lattice_slice_gpu.h>
-
-using namespace Grid;
-
-int main (int argc, char ** argv)
-{
-  Grid_init(&argc,&argv);
-
-  int N=16;
-  std::vector<int> latt_size  ({N,N,N,N});
-  std::vector<int> simd_layout({vComplexD::Nsimd(),1,1,1});
-  std::vector<int> mpi_layout ({1,1,1,1});
-
-  GridCartesian         GRID(latt_size,simd_layout,mpi_layout);
-
-  LatticeComplexD  rn(&GRID);
-
-  GridParallelRNG RNG(&GRID);
-  RNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));  
-  gaussian(RNG,rn);
-
-  std::vector<TComplex> reduced_ref;
-  std::vector<TComplex> reduced_gpu;
-  for(int d=0;d<4;d++){
-    {
-      RealD t=-usecond();
-      sliceSum(rn,reduced_ref,d);
-      t+=usecond();
-      std::cout << " sliceSum took "<<t<<" usecs"<<std::endl;
-    }
-    {
-      RealD t=-usecond();
-      sliceSumGpu(rn,reduced_gpu,d);
-      t+=usecond();
-      std::cout << " sliceSumGpu took "<<t<<" usecs"<<std::endl;
-    }
-    for(int t=0;t<reduced_ref.size();t++){
-      std::cout << t<<" ref "<< reduced_ref[t] <<" opt " << reduced_gpu[t] << " diff "<<reduced_ref[t]-reduced_gpu[t]<<std::endl;
-      TComplex diff = reduced_ref[t]-reduced_gpu[t];
-      assert(abs(TensorRemove(diff)) < 1e-8 );
-    }
-  }
-  Grid_finalize();
-}