Apply comments

Author: sergey-semenov

sycl/source/detail/queue_impl.cpp

@@ -122,6 +122,7 @@ void queue_impl::addEvent(const event &Event) {
 /// but some events have no other owner. In this case,
 /// addSharedEvent will have the queue track the events via a shared pointer.
 void queue_impl::addSharedEvent(const event &Event) {
+  assert(is_host() || !MSupportOOO);
   std::lock_guard<std::mutex> Lock(MMutex);
   // Events stored in MEventsShared are not released anywhere else aside from
   // calls to queue::wait/wait_and_throw, which a user application might not
@@ -254,26 +255,27 @@ void queue_impl::wait(const detail::code_location &CodeLoc) {
   // directly. Otherwise, only wait for unenqueued or host task events, starting
   // from the latest submitted task in order to minimize total amount of calls,
   // then handle the rest with piQueueFinish.
+  bool SupportsPiFinish = !is_host() && MSupportOOO;
   for (auto EventImplWeakPtrIt = WeakEvents.rbegin();
        EventImplWeakPtrIt != WeakEvents.rend(); ++EventImplWeakPtrIt) {
     if (std::shared_ptr<event_impl> EventImplSharedPtr =
             EventImplWeakPtrIt->lock()) {
       // A nullptr PI event indicates that piQueueFinish will not cover it,
       // either because it's a host task event or an unenqueued one.
-      if (is_host() || !MSupportOOO ||
-          nullptr == EventImplSharedPtr->getHandleRef()) {
+      if (!SupportsPiFinish || nullptr == EventImplSharedPtr->getHandleRef()) {
         EventImplSharedPtr->wait(EventImplSharedPtr);
       }
     }
   }
-  if (!is_host() && MSupportOOO) {
+  if (SupportsPiFinish) {
     const detail::plugin &Plugin = getPlugin();
     Plugin.call<detail::PiApiKind::piQueueFinish>(getHandleRef());
     for (std::weak_ptr<event_impl> &EventImplWeakPtr : WeakEvents)
       if (std::shared_ptr<event_impl> EventImplSharedPtr =
               EventImplWeakPtr.lock())
         EventImplSharedPtr->cleanupCommand(EventImplSharedPtr);
-    assert(SharedEvents.empty());
+    assert(SharedEvents.empty() && "Queues that support calling piQueueFinish "
+                                   "shouldn't have shared events");
   } else {
     for (event &Event : SharedEvents)
       Event.wait();

sycl/unittests/queue/Wait.cpp

@@ -91,12 +91,6 @@ bool preparePiMock(platform &Plt) {
               << std::endl;
     return false;
   }
-  // TODO: Skip tests for CUDA temporarily
-  if (detail::getSyclObjImpl(Plt)->getPlugin().getBackend() == backend::cuda) {
-    std::cout << "Not run on CUDA - usm is not supported for CUDA backend yet"
-              << std::endl;
-    return false;
-  }
 
   unittest::PiMock Mock{Plt};
   Mock.redefine<detail::PiApiKind::piQueueCreate>(redefinedQueueCreate);