[SYCL][L0] Wait for all events completion to do full cleanup (#3985)

Signed-off-by: Sergey V Maslov <[email protected]>

Author: smaslov-intel

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -4371,8 +4371,21 @@ pi_result piEventRelease(pi_event Event) {
     die("piEventRelease: called on a destroyed event");
   }
 
+  // The event is no longer needed upstream, but we have to wait for its
+  // completion in order to do proper cleanup. Otherwise refcount may still be
+  // non-zero in the check below and we will get event leak.
+  //
+  // TODO: in case this potentially "early" wait causes performance problems,
+  // e.g. due to closing a batch too early, or blocking the host for no good
+  // reason, then we should look into moving the wait down to queue release
+  // (will need to remember all events in the queue for that).
+  //
+  if (!Event->CleanedUp)
+    PI_CALL(piEventsWait(1, &Event));
+
   if (--(Event->RefCount) == 0) {
-    cleanupAfterEvent(Event);
+    if (!Event->CleanedUp)
+      cleanupAfterEvent(Event);
 
     if (Event->CommandType == PI_COMMAND_TYPE_MEM_BUFFER_UNMAP &&
         Event->CommandData) {