[Openmp-commits] [openmp] 95e11a9 - [OpenMP][FIX] Unbreak a fencing issue

[Johannes Doerfert via Openmp-commits]

Author: Johannes Doerfert
Date: 2023-11-02T15:04:10-07:00
New Revision: 95e11a97f6666400e89c66bdf9a61bb87ede5708

URL: https://github.com/llvm/llvm-project/commit/95e11a97f6666400e89c66bdf9a61bb87ede5708
DIFF: https://github.com/llvm/llvm-project/commit/95e11a97f6666400e89c66bdf9a61bb87ede5708.diff

LOG: [OpenMP][FIX] Unbreak a fencing issue

A recent update caused the fences to be team only while we always need
kernel fences. Broke OpenMC on NVIDIA A100.

Added: 
    

Modified: 
    openmp/libomptarget/DeviceRTL/src/Reduction.cpp

Removed: 
    


################################################################################
diff  --git a/openmp/libomptarget/DeviceRTL/src/Reduction.cpp b/openmp/libomptarget/DeviceRTL/src/Reduction.cpp
index efa09cafa879ec1..bae4237194590d6 100644
--- a/openmp/libomptarget/DeviceRTL/src/Reduction.cpp
+++ b/openmp/libomptarget/DeviceRTL/src/Reduction.cpp
@@ -220,17 +220,20 @@ int32_t __kmpc_nvptx_teams_reduce_nowait_v2(
     } else
       lgredFct(GlobalBuffer, ModBockId, reduce_data);
 
+    // Propagate the memory writes above to the world.
+    fence::kernel(atomic::release);
+
     // Increment team counter.
     // This counter is incremented by all teams in the current
-    // BUFFER_SIZE chunk.
+    // num_of_records chunk.
     ChunkTeamCount = atomic::inc(&Cnt, num_of_records - 1u, atomic::seq_cst,
                                  atomic::MemScopeTy::device);
   }
-  // Synchronize
+
+  // Synchronize in SPMD mode as in generic mode all but 1 threads are in the
+  // state machine.
   if (mapping::isSPMDMode())
     synchronize::threadsAligned(atomic::acq_rel);
-  else
-    fence::kernel(atomic::acq_rel);
 
   // reduce_data is global or shared so before being reduced within the
   // warp we need to bring it in local memory:
@@ -257,6 +260,9 @@ int32_t __kmpc_nvptx_teams_reduce_nowait_v2(
   // Check if this is the very last team.
   unsigned NumRecs = kmpcMin(NumTeams, uint32_t(num_of_records));
   if (ChunkTeamCount == NumTeams - Bound - 1) {
+    // Ensure we see the global memory writes by other teams
+    fence::kernel(atomic::aquire);
+
     //
     // Last team processing.
     //