[SYCL] "No last event mode" support for handler-less kernel submission

The handler-based kernel submission uses a "no last event mode" atomic
variable to mark, if the last event is recorded for the in order queue.
The handler-less submission needs to update that variable, in order
to support mixed APIs usage (handler-based and handler-less APIs,
which might require synchronization).

Author: slawekptak

sycl/source/detail/queue_impl.cpp

@@ -621,10 +621,6 @@ queue_impl::submit_direct(bool CallerNeedsEvent,
   detail::CG::StorageInitHelper CGData;
   std::unique_lock<std::mutex> Lock(MMutex);
 
-  // Set the No Last Event Mode to false, since the no-handler path
-  // does not support it yet.
-  MNoLastEventMode.store(false, std::memory_order_relaxed);
-
   // Used by queue_empty() and getLastEvent()
   MEmpty.store(false, std::memory_order_release);
 
@@ -662,6 +658,14 @@ queue_impl::submit_direct(bool CallerNeedsEvent,
            : true) &&
       !hasCommandGraph();
 
+  // Synchronize with the "no last event mode", used by the handler-based
+  // kernel submit path
+  if (SchedulerBypass) {
+    MNoLastEventMode.store(true, std::memory_order_relaxed);
+  } else {
+    MNoLastEventMode.store(false, std::memory_order_relaxed);
+  }
+
   EventImplPtr EventImpl = SubmitCommandFunc(CGData, SchedulerBypass);
 
   // Sync with the last event for in order queue. For scheduler-bypass flow,