From 1392dd29977fe2ae6eddd64d6d80609ced56de9f Mon Sep 17 00:00:00 2001
From: Gilbert Lee <gilbert.lee@amd.com>
Date: Fri, 11 Oct 2019 09:16:19 -0700
Subject: [PATCH] Performing __threadfence_system() with only first thread

---
 src/collectives/device/primitives.h | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/collectives/device/primitives.h b/src/collectives/device/primitives.h
index 8c63a6a4f1..b0d4a2938d 100644
--- a/src/collectives/device/primitives.h
+++ b/src/collectives/device/primitives.h
@@ -189,10 +189,13 @@ class ncclPrimitives {
         }
       }
       exitIfAbortBarrier(abort, abortCount);
-      if (tid == 0) FOR_SEND(postSendSize, realSize*sizeof(T));
-      if (SEND) __threadfence_system();
-      if (tid == 0) FOR_SEND(postSend);
-      if (tid == 0) FOR_RECV(postRecv);
+      if (tid == 0)
+      {
+        FOR_SEND(postSendSize, realSize*sizeof(T));
+        __threadfence_system();
+        FOR_SEND(postSend);
+        FOR_RECV(postRecv);
+      }
 
       for (int i=0; i<RECV*NRECV+SRC; i++) srcs[i] += sliceSize;
       for (int i=0; i<SEND*NSEND+DST; i++) dsts[i] += sliceSize;