Merge branch 'develop' of https://github.com/paboyle/Grid into develop

Could place before comms, or in Gather, but decided to insist Gather means Gather is done

Grid/qcd/action/fermion/WilsonCompressor.h

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

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

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

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

@@ -428,9 +428,10 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
   auto ptr = &st.surface_list[0];					\
   accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
       int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
+      int sU = sF/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
+    });									\
+  accelerator_barrier();
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
@@ -474,9 +475,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
+     // dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
@@ -506,9 +508,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     // Dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteDagExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif

Grid/stencil/Stencil.h

@@ -348,7 +348,7 @@ public:
   ////////////////////////////////////////
   // Stencil query
   ////////////////////////////////////////
-#ifdef SHM_FAST_PATH
+#if 1
   inline int SameNode(int point) {
 
     int dimension    = this->_directions[point];
@@ -434,7 +434,6 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
-    accelerator_barrier();
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromBegin(MpiReqs,
 					Packets[i].send_buf,
@@ -666,11 +665,9 @@ public:
     for(int i=0;i<mm.size();i++){
       decompressor::MergeFace(decompress,mm[i]);
     }
-    if ( mm.size() )    acceleratorFenceComputeStream();
     for(int i=0;i<dd.size();i++){
       decompressor::DecompressFace(decompress,dd[i]);
     }
-    if ( dd.size() )    acceleratorFenceComputeStream();
   }
   ////////////////////////////////////////
   // Set up routines
@@ -708,6 +705,7 @@ public:
 	}
       }
     }
+    std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
   }
   /// Introduce a block structure and switch off comms on boundaries
   void DirichletBlock(const Coordinate &dirichlet_block)

Grid/threads/Accelerator.h

@@ -526,7 +526,7 @@ inline void acceleratorFreeCpu  (void *ptr){free(ptr);};
 //////////////////////////////////////////////
 
 #ifdef GRID_SYCL
-inline void acceleratorFenceComputeStream(void){ theGridAccelerator->submit_barrier();};
+inline void acceleratorFenceComputeStream(void){ theGridAccelerator->ext_oneapi_submit_barrier(); };
 #else
 // Ordering within a stream guaranteed on Nvidia & AMD
 inline void acceleratorFenceComputeStream(void){ };

benchmarks/Benchmark_dwf_fp32.cc

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

systems/PVC/setup.sh

@@ -3,8 +3,14 @@ export https_proxy=http://proxy-chain.intel.com:911
 export LD_LIBRARY_PATH=$HOME/prereqs/lib/:$LD_LIBRARY_PATH
 
 module load intel-release
-source /opt/intel/oneapi/PVC_setup.sh
+module load intel-comp-rt/embargo-ci-neo
+
+#source /opt/intel/oneapi/PVC_setup.sh
 #source /opt/intel/oneapi/ATS_setup.sh
+#module load intel-nightly/20230331
+#module load intel-comp-rt/ci-neo-master/026093
+
+#module load intel/mpich
 module load intel/mpich/pvc45.3
 export PATH=~/ATS/pti-gpu/tools/onetrace/:$PATH