[SYCL] Make queue's non-USM event ownership temporary

sycl/source/detail/queue_impl.cpp

@@ -49,7 +49,7 @@ event queue_impl::memset(shared_ptr_class<detail::queue_impl> Impl, void *Ptr,
     return event();
 
   event ResEvent{pi::cast<cl_event>(Event), Context};
-  addEvent(ResEvent);
+  addUSMEvent(ResEvent);
   return ResEvent;
 }
 
@@ -63,7 +63,7 @@ event queue_impl::memcpy(shared_ptr_class<detail::queue_impl> Impl, void *Dest,
     return event();
 
   event ResEvent{pi::cast<cl_event>(Event), Context};
-  addEvent(ResEvent);
+  addUSMEvent(ResEvent);
   return ResEvent;
 }
 
@@ -81,13 +81,19 @@ event queue_impl::mem_advise(const void *Ptr, size_t Length,
                                                    Advice, &Event);
 
   event ResEvent{pi::cast<cl_event>(Event), Context};
-  addEvent(ResEvent);
+  addUSMEvent(ResEvent);
   return ResEvent;
 }
 
 void queue_impl::addEvent(event Event) {
+  std::weak_ptr<event_impl> EventWeakPtr{getSyclObjImpl(Event)};
   std::lock_guard<mutex_class> Guard(MMutex);
-  MEvents.push_back(std::move(Event));
+  MEvents.push_back(std::move(EventWeakPtr));
+}
+
+void queue_impl::addUSMEvent(event Event) {
+  std::lock_guard<mutex_class> Guard(MMutex);
+  MUSMEvents.push_back(std::move(Event));
 }
 
 void *queue_impl::instrumentationProlog(const detail::code_location &CodeLoc,
@@ -175,8 +181,13 @@ void queue_impl::wait(const detail::code_location &CodeLoc) {
 #endif
 
   std::lock_guard<mutex_class> Guard(MMutex);
-  for (auto &Event : MEvents)
+  for (std::weak_ptr<event_impl> &EventImplWeakPtr : MEvents) {
+    if (std::shared_ptr<event_impl> EventImplPtr = EventImplWeakPtr.lock())
+      EventImplPtr->wait(EventImplPtr);
+  }
+  for (event &Event : MUSMEvents) {
     Event.wait();
+  }
   MEvents.clear();
 
 #ifdef XPTI_ENABLE_INSTRUMENTATION

sycl/source/detail/queue_impl.hpp

@@ -382,12 +382,20 @@ class queue_impl {
   /// \param Event is the event to be stored
   void addEvent(event Event);
 
+  /// Stores a USM operation event that should be associated with the queue
+  ///
+  /// \param Event is the event to be stored
+  void addUSMEvent(event Event);
+
   /// Protects all the fields that can be changed by class' methods.
   mutex_class MMutex;
 
   DeviceImplPtr MDevice;
   const ContextImplPtr MContext;
-  vector_class<event> MEvents;
+  vector_class<std::weak_ptr<event_impl>> MEvents;
+  // USM operations are not added to the scheduler command graph,
+  // queue is the only owner on the runtime side.
+  vector_class<event> MUSMEvents;
   exception_list MExceptions;
   const async_handler MAsyncHandler;
   const property_list MPropList;

sycl/test/basic_tests/event_release.cpp

@@ -0,0 +1,38 @@
+// RUN: %clangxx -fsycl %s -o %t.out
+// RUN: env SYCL_PI_TRACE=1 %CPU_RUN_PLACEHOLDER %t.out 2>&1 %CPU_CHECK_PLACEHOLDER
+#include <CL/sycl.hpp>
+#include <cassert>
+#include <iostream>
+
+// The test checks that pi_events are released without queue destruction
+// or call to queue::wait, when the corresponding commands are cleaned up.
+
+using namespace cl::sycl;
+
+class Foo;
+
+int main() {
+  int Val = 0;
+  int Gold = 42;
+
+  queue Q;
+
+  {
+    buffer<int, 1> Buf{&Val, range<1>(1)};
+    Q.submit([&](handler &Cgh) {
+      auto Acc = Buf.get_access<access::mode::discard_write>(Cgh);
+      Cgh.single_task<Foo>([=]() {
+        Acc[0] = Gold;
+      });
+    });
+  }
+
+  // Buffer destruction triggers execution graph cleanup, check that both
+  // events (one for launching the kernel and one for memory transfer to host)
+  // are released.
+  // CHECK: piEventRelease
+  // CHECK: piEventRelease
+  assert(Val == Gold);
+  // CHECK: End of main scope
+  std::cout << "End of main scope" << std::endl;
+}