[UR] [V2] Add wait before enqueue in command buffer

unified-runtime/source/adapters/level_zero/v2/command_buffer.cpp

@@ -47,7 +47,38 @@ ur_result_t ur_exp_command_buffer_handle_t_::finalizeCommandBuffer() {
   isFinalized = true;
   return UR_RESULT_SUCCESS;
 }
+ur_event_handle_t ur_exp_command_buffer_handle_t_::getCurrentExecutionEvent(
+    [[maybe_unused]] locked<ur_command_list_manager> &commandList) {
+  assert(
+      commandList->getZeCommandList() ==
+          commandListManager.get_no_lock()->getZeCommandList() &&
+      "Provided command list is not the same as the one in the command buffer");
+  return currentExecution;
+}
+
+ur_result_t ur_exp_command_buffer_handle_t_::registerExecutionEvent(
+    [[maybe_unused]] locked<ur_command_list_manager> &commandList,
+    ur_event_handle_t nextExecutionEvent) {
+  assert(
+      commandList->getZeCommandList() ==
+          commandListManager.get_no_lock()->getZeCommandList() &&
+      "Provided command list is not the same as the one in the command buffer");
+  if (currentExecution) {
+    UR_CALL(currentExecution->release());
+    currentExecution = nullptr;
+  }
+  if (nextExecutionEvent) {
+    currentExecution = nextExecutionEvent;
+    UR_CALL(nextExecutionEvent->retain());
+  }
+  return UR_RESULT_SUCCESS;
+}
 
+ur_exp_command_buffer_handle_t_::~ur_exp_command_buffer_handle_t_() {
+  if (currentExecution) {
+    currentExecution->release();
+  }
+}
 namespace ur::level_zero {
 
 ur_result_t

unified-runtime/source/adapters/level_zero/v2/command_buffer.hpp

@@ -23,7 +23,13 @@ struct ur_exp_command_buffer_handle_t_ : public _ur_object {
       v2::raii::command_list_unique_handle &&commandList,
       const ur_exp_command_buffer_desc_t *desc);
 
-  ~ur_exp_command_buffer_handle_t_() = default;
+  ~ur_exp_command_buffer_handle_t_();
+
+  ur_event_handle_t
+  getCurrentExecutionEvent(locked<ur_command_list_manager> &commandList);
+  ur_result_t
+  registerExecutionEvent(locked<ur_command_list_manager> &commandList,
+                         ur_event_handle_t nextExecutionEvent);
 
   lockable<ur_command_list_manager> commandListManager;
 
@@ -36,6 +42,8 @@ struct ur_exp_command_buffer_handle_t_ : public _ur_object {
 private:
   // Indicates if command-buffer was finalized.
   bool isFinalized = false;
+
+  ur_event_handle_t currentExecution = nullptr;
 };
 
 struct ur_exp_command_buffer_command_handle_t_ : public _ur_object {

unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp

@@ -928,9 +928,29 @@ ur_result_t ur_queue_immediate_in_order_t::enqueueCommandBufferExp(
   auto commandListLocked = hCommandBuffer->commandListManager.lock();
   ze_command_list_handle_t commandBufferCommandList =
       commandListLocked->getZeCommandList();
-  return enqueueGenericCommandListsExp(1, &commandBufferCommandList, phEvent,
-                                       numEventsInWaitList, phEventWaitList,
-                                       UR_COMMAND_ENQUEUE_COMMAND_BUFFER_EXP);
+  ur_event_handle_t internalEvent = nullptr;
+  if (phEvent == nullptr) {
+    phEvent = &internalEvent;
+  }
+  ur_event_handle_t executionEvent =
+      hCommandBuffer->getCurrentExecutionEvent(commandListLocked);
+  std::vector<ur_event_handle_t> extendedWaitList;
+  if (executionEvent != nullptr) {
+    extendedWaitList.resize(numEventsInWaitList + 1);
+    std::copy(phEventWaitList, phEventWaitList + numEventsInWaitList,
+              extendedWaitList.begin());
+    extendedWaitList[numEventsInWaitList] = executionEvent;
+    phEventWaitList = extendedWaitList.data();
+    numEventsInWaitList++;
+  }
+  UR_CALL(enqueueGenericCommandListsExp(1, &commandBufferCommandList, phEvent,
+                                        numEventsInWaitList, phEventWaitList,
+                                        UR_COMMAND_ENQUEUE_COMMAND_BUFFER_EXP));
+  UR_CALL(hCommandBuffer->registerExecutionEvent(commandListLocked, *phEvent));
+  if (internalEvent != nullptr) {
+    internalEvent->release();
+  }
+  return UR_RESULT_SUCCESS;
 }
 
 ur_result_t ur_queue_immediate_in_order_t::enqueueKernelLaunchCustomExp(

unified-runtime/test/conformance/exp_command_buffer/fixtures.h

@@ -45,7 +45,6 @@ static void checkCommandBufferUpdateSupport(
 
 struct urCommandBufferExpTest : uur::urContextTest {
   void SetUp() override {
-    UUR_KNOWN_FAILURE_ON(uur::LevelZeroV2{});
 
     UUR_RETURN_ON_FATAL_FAILURE(uur::urContextTest::SetUp());
 
@@ -72,7 +71,6 @@ struct urCommandBufferExpTest : uur::urContextTest {
 template <class T>
 struct urCommandBufferExpTestWithParam : urQueueTestWithParam<T> {
   void SetUp() override {
-    UUR_KNOWN_FAILURE_ON(uur::LevelZeroV2{});
 
     UUR_RETURN_ON_FATAL_FAILURE(uur::urQueueTestWithParam<T>::SetUp());
 
@@ -97,7 +95,6 @@ struct urCommandBufferExpTestWithParam : urQueueTestWithParam<T> {
 
 struct urCommandBufferExpExecutionTest : uur::urKernelExecutionTest {
   void SetUp() override {
-    UUR_KNOWN_FAILURE_ON(uur::LevelZeroV2{});
 
     UUR_RETURN_ON_FATAL_FAILURE(uur::urKernelExecutionTest::SetUp());
 

unified-runtime/test/conformance/exp_command_buffer/update/invalid_update.cpp

@@ -309,6 +309,8 @@ TEST_P(InvalidUpdateTest, CommandBufferMismatch) {
 // that isn't supported.
 struct InvalidUpdateCommandBufferExpExecutionTest : uur::urKernelExecutionTest {
   void SetUp() override {
+    UUR_KNOWN_FAILURE_ON(uur::LevelZeroV2{});
+
     program_name = "fill_usm";
     UUR_RETURN_ON_FATAL_FAILURE(uur::urKernelExecutionTest::SetUp());