[L0] fix Append Barrier to only occur when at least one inner event exists

Signed-off-by: Neil R. Spruit <[email protected]>

Author: nrspruit

source/adapters/level_zero/event.cpp

@@ -691,7 +691,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventWait(
 
           ze_event_handle_t ZeEvent = HostVisibleEvent->ZeEvent;
           logger::debug("ZeEvent = {}", ur_cast<std::uintptr_t>(ZeEvent));
-          // If this event was an inner batched event, then lock and sync with
+          // If this event was an inner batched event, then sync with
           // the Queue instead of waiting on the event.
           if (HostVisibleEvent->IsInnerBatchedEvent && Event->ZeBatchedQueue) {
             ZE2UR_CALL(zeHostSynchronize, (Event->ZeBatchedQueue));

source/adapters/level_zero/queue.cpp

@@ -1406,12 +1406,13 @@ ur_queue_handle_t_::executeCommandList(ur_command_list_ptr_t CommandList,
       this->signalEventFromCmdListIfLastEventDiscarded(CommandList);
     }
     // Append Signalling of the inner events at the end of the batch
-    if (CommandList->second.EventList.size() > 0 && AppendBarrierNeeded) {
-      ZE2UR_CALL(zeCommandListAppendBarrier,
-                 (CommandList->first, nullptr, 0, nullptr));
-    }
     for (auto &Event : CommandList->second.EventList) {
       if (Event->IsInnerBatchedEvent) {
+        if (AppendBarrierNeeded) {
+          ZE2UR_CALL(zeCommandListAppendBarrier,
+                     (CommandList->first, nullptr, 0, nullptr));
+          AppendBarrierNeeded = false;
+        }
         ZE2UR_CALL(zeCommandListAppendSignalEvent,
                    (CommandList->first, Event->ZeEvent));
       }