[SYCL] Add sync for host task after barrier (#15345)

PR includes the following fixes:
* When submitting a command to an out-of-order queue we don't need to
add the dependency from the last barrier to the scheduler if command is
enqueued via UR because that means that UR backend will take care of the
command from barrier.
  
* We used to update the last barrier only if barrier which is being
submitted to the sycl::queue is not enqueued to the backend which is not
correct, because there might be host tasks following that barrier which
have to depend on it.

Signed-off-by: Tikhomirova, Kseniya <kseniya.tikhomirova@intel.com>

sycl/source/detail/queue_impl.hpp

@@ -784,18 +784,19 @@ class queue_impl {
       if (Type == CGType::Barrier && !Deps.UnenqueuedCmdEvents.empty()) {
         Handler.depends_on(Deps.UnenqueuedCmdEvents);
       }
-      if (Deps.LastBarrier)
+      if (Deps.LastBarrier && (Type == CGType::CodeplayHostTask ||
+                               (!Deps.LastBarrier->isEnqueued())))
         Handler.depends_on(Deps.LastBarrier);
+
       EventRet = Handler.finalize();
       EventImplPtr EventRetImpl = getSyclObjImpl(EventRet);
       if (Type == CGType::CodeplayHostTask)
         Deps.UnenqueuedCmdEvents.push_back(EventRetImpl);
-      else if (!EventRetImpl->isEnqueued()) {
-        if (Type == CGType::Barrier || Type == CGType::BarrierWaitlist) {
-          Deps.LastBarrier = EventRetImpl;
-          Deps.UnenqueuedCmdEvents.clear();
-        } else
-          Deps.UnenqueuedCmdEvents.push_back(EventRetImpl);
+      else if (Type == CGType::Barrier || Type == CGType::BarrierWaitlist) {
+        Deps.LastBarrier = EventRetImpl;
+        Deps.UnenqueuedCmdEvents.clear();
+      } else if (!EventRetImpl->isEnqueued()) {
+        Deps.UnenqueuedCmdEvents.push_back(EventRetImpl);
       }
     }
   }

sycl/unittests/scheduler/HostTaskAndBarrier.cpp

@@ -196,7 +196,8 @@ TEST_F(BarrierHandlingWithHostTask, BarrierHostTaskKernel) {
   sycl::event HTEvent = AddTask(TestCGType::HOST_TASK);
   EventImplPtr HostTaskEventImpl = sycl::detail::getSyclObjImpl(HTEvent);
   auto HostTaskWaitList = HostTaskEventImpl->getWaitList();
-  ASSERT_EQ(HostTaskWaitList.size(), 0u);
+  ASSERT_EQ(HostTaskWaitList.size(), 1u);
+  EXPECT_EQ(HostTaskWaitList[0], BarrierEventImpl);
   EXPECT_EQ(HostTaskEventImpl->isEnqueued(), true);
 
   sycl::event KernelEvent = AddTask(TestCGType::KERNEL_TASK);
@@ -225,7 +226,8 @@ TEST_F(BarrierHandlingWithHostTask, BarrierKernelHostTask) {
   sycl::event HTEvent = AddTask(TestCGType::HOST_TASK);
   EventImplPtr HostTaskEventImpl = sycl::detail::getSyclObjImpl(HTEvent);
   auto HostTaskWaitList = HostTaskEventImpl->getWaitList();
-  ASSERT_EQ(HostTaskWaitList.size(), 0u);
+  ASSERT_EQ(HostTaskWaitList.size(), 1u);
+  EXPECT_EQ(HostTaskWaitList[0], BarrierEventImpl);
   EXPECT_EQ(HostTaskEventImpl->isEnqueued(), true);
 
   MainLock.unlock();
@@ -272,7 +274,8 @@ TEST_F(BarrierHandlingWithHostTask, KernelBarrierHostTask) {
   sycl::event HTEvent = AddTask(TestCGType::HOST_TASK);
   EventImplPtr HostTaskEventImpl = sycl::detail::getSyclObjImpl(HTEvent);
   auto HostTaskWaitList = HostTaskEventImpl->getWaitList();
-  ASSERT_EQ(HostTaskWaitList.size(), 0u);
+  ASSERT_EQ(HostTaskWaitList.size(), 1u);
+  EXPECT_EQ(HostTaskWaitList[0], BarrierEventImpl);
   EXPECT_EQ(HostTaskEventImpl->isEnqueued(), true);
 
   MainLock.unlock();