From a00ae981e0e6e83caa70f5bed346242ef1eb0bf9 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Wed, 29 Mar 2023 15:00:40 -0400
Subject: [PATCH] Fence propagation from SYCL

---
 Grid/communicator/SharedMemoryMPI.cc                   |  3 +--
 .../implementation/WilsonKernelsImplementation.h       | 10 +++-------
 Grid/stencil/Stencil.h                                 |  3 +++
 3 files changed, 7 insertions(+), 9 deletions(-)

diff --git a/Grid/communicator/SharedMemoryMPI.cc b/Grid/communicator/SharedMemoryMPI.cc
index 792f8405..a4f5731e 100644
--- a/Grid/communicator/SharedMemoryMPI.cc
+++ b/Grid/communicator/SharedMemoryMPI.cc
@@ -37,9 +37,8 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #include <hip/hip_runtime_api.h>
 #endif
 #ifdef GRID_SYCL
-
-#endif
 #define GRID_SYCL_LEVEL_ZERO_IPC
+#endif
 
 
 NAMESPACE_BEGIN(Grid); 
diff --git a/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h b/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
index 9f6960af..bbf9937c 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
@@ -459,11 +459,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
-#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
-#endif     
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
@@ -474,6 +470,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
@@ -498,10 +495,9 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
-     acceleratorFenceComputeStream();
    } else if( interior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteDagInt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteDagInt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
diff --git a/Grid/stencil/Stencil.h b/Grid/stencil/Stencil.h
index 6296df4e..0a1214a1 100644
--- a/Grid/stencil/Stencil.h
+++ b/Grid/stencil/Stencil.h
@@ -585,6 +585,7 @@ public:
 	  decompress.Exchange(mp,vp0,vp1,type,o);
       });
     }
+    if ( mm.size() )    acceleratorFenceComputeStream();
     mergetime+=usecond();
 
     decompresstime-=usecond();
@@ -595,7 +596,9 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
+    if ( dd.size() )    acceleratorFenceComputeStream();
     decompresstime+=usecond();
+    
   }
   ////////////////////////////////////////
   // Set up routines