Performance improve for Tesseract

Grid/qcd/action/fermion/WilsonCloverHelpers.h

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

Grid/threads/Accelerator.h

@@ -481,9 +481,10 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
 #define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) thread_for2d(iter1,num1,iter2,num2,{ __VA_ARGS__ });
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
-inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { memcpy(to,from,bytes);}
+
+inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes); }
+inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
+inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { thread_bcopy(from,to,bytes);}
 inline void acceleratorCopySynchronise(void) {};
 
 inline int  acceleratorIsCommunicable(void *ptr){ return 1; }

Grid/threads/Threads.h

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

Grid/util/Init.cc

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

systems/mac-arm/config-command-mpi

@@ -0,0 +1 @@
+CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi