[L0] Add support for in-order lists using L0 driver

Signed-off-by: Raiyan Latif <raiyan.latif@intel.com>

source/adapters/level_zero/device.cpp

@@ -1054,6 +1054,19 @@ bool ur_device_handle_t_::useRelaxedAllocationLimits() {
   return EnableRelaxedAllocationLimits;
 }
 
+bool ur_device_handle_t_::useDriverInOrderLists() {
+  // Use in-order lists implementation from L0 driver instead
+  // of adapter's implementation.
+  static const bool UseDriverInOrderLists = [] {
+    const char *UrRet = std::getenv("UR_L0_USE_DRIVER_INORDER_LISTS");
+    if (!UrRet)
+      return false;
+    return std::atoi(UrRet) != 0;
+  }();
+
+  return UseDriverInOrderLists;
+}
+
 ur_result_t ur_device_handle_t_::initialize(int SubSubDeviceOrdinal,
                                             int SubSubDeviceIndex) {
   // Maintain various device properties cache.

source/adapters/level_zero/device.hpp

@@ -143,6 +143,9 @@ struct ur_device_handle_t_ : _ur_object {
   // Read env settings to select immediate commandlist mode.
   ImmCmdlistMode useImmediateCommandLists();
 
+  // Whether Adapter uses driver's implementation of in-order lists or not
+  bool useDriverInOrderLists();
+
   // Returns whether immediate command lists are used on this device.
   ImmCmdlistMode ImmCommandListUsed{};
 

source/adapters/level_zero/event.cpp

@@ -43,6 +43,20 @@ static const bool UseMultipleCmdlistBarriers = [] {
   return std::atoi(UseMultipleCmdlistBarriersFlag) > 0;
 }();
 
+bool WaitListEmptyOrAllEventsFromSameQueue(
+    ur_queue_handle_t Queue, uint32_t NumEventsInWaitList,
+    const ur_event_handle_t *EventWaitList) {
+  if (!NumEventsInWaitList)
+    return true;
+
+  for (uint32_t i = 0; i < NumEventsInWaitList; ++i) {
+    if (Queue != EventWaitList[i]->UrQueue)
+      return false;
+  }
+
+  return true;
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urEnqueueEventsWait(
     ur_queue_handle_t Queue,      ///< [in] handle of the queue object
     uint32_t NumEventsInWaitList, ///< [in] size of the event wait list
@@ -206,21 +220,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueEventsWaitWithBarrier(
   bool IsInternal = OutEvent == nullptr;
   ur_event_handle_t *Event = OutEvent ? OutEvent : &InternalEvent;
 
-  auto WaitListEmptyOrAllEventsFromSameQueue = [Queue, NumEventsInWaitList,
-                                                EventWaitList]() {
-    if (!NumEventsInWaitList)
-      return true;
-
-    for (uint32_t I = 0; I < NumEventsInWaitList; ++I)
-      if (Queue != EventWaitList[I]->UrQueue)
-        return false;
-
-    return true;
-  };
-
   // For in-order queue and wait-list which is empty or has events from
   // the same queue just use the last command event as the barrier event.
-  if (Queue->isInOrderQueue() && WaitListEmptyOrAllEventsFromSameQueue() &&
+  if (Queue->isInOrderQueue() &&
+      WaitListEmptyOrAllEventsFromSameQueue(Queue, NumEventsInWaitList,
+                                            EventWaitList) &&
       Queue->LastCommandEvent && !Queue->LastCommandEvent->IsDiscarded) {
     UR_CALL(urEventRetain(Queue->LastCommandEvent));
     *Event = Queue->LastCommandEvent;
@@ -1189,6 +1193,23 @@ ur_result_t _ur_ze_event_list_t::createAndRetainUrZeEventList(
       CurQueue->LastCommandEvent && CurQueue->LastCommandEvent->IsDiscarded)
     IncludeLastCommandEvent = false;
 
+  // If we are using L0 native implementation for handling in-order queues,
+  // then we don't need to add the last enqueued event into the waitlist, as
+  // the native driver implementation will already ensure in-order semantics.
+  // The only exception is when a different immediate command was last used on
+  // the same UR Queue.
+  if (CurQueue->Device->useDriverInOrderLists() && CurQueue->isInOrderQueue() &&
+      CurQueue->UsingImmCmdLists) {
+    auto QueueGroup = CurQueue->getQueueGroup(UseCopyEngine);
+    uint32_t QueueGroupOrdinal, QueueIndex;
+    auto NextIndex = QueueGroup.getQueueIndex(&QueueGroupOrdinal, &QueueIndex,
+                                              /*QueryOnly */ true);
+    auto NextImmCmdList = QueueGroup.ImmCmdLists[NextIndex];
+    IncludeLastCommandEvent &=
+        CurQueue->LastUsedCommandList != CurQueue->CommandListMap.end() &&
+        NextImmCmdList != CurQueue->LastUsedCommandList;
+  }
+
   try {
     uint32_t TmpListLength = 0;
 
@@ -1205,6 +1226,16 @@ ur_result_t _ur_ze_event_list_t::createAndRetainUrZeEventList(
       this->UrEventList = new ur_event_handle_t[EventListLength];
     }
 
+    // For in-order queue and wait-list which is empty or has events only from
+    // the same queue then we don't need to wait on any other additional events
+    if (CurQueue->Device->useDriverInOrderLists() &&
+        CurQueue->isInOrderQueue() &&
+        WaitListEmptyOrAllEventsFromSameQueue(CurQueue, EventListLength,
+                                              EventList)) {
+      this->Length = TmpListLength;
+      return UR_RESULT_SUCCESS;
+    }
+
     if (EventListLength > 0) {
       for (uint32_t I = 0; I < EventListLength; I++) {
         {

source/adapters/level_zero/kernel.cpp

@@ -214,8 +214,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   // the code can do a urKernelRelease on this kernel.
   (*Event)->CommandData = (void *)Kernel;
 
-  // Increment the reference count of the Kernel and indicate that the Kernel is
-  // in use. Once the event has been signalled, the code in
+  // Increment the reference count of the Kernel and indicate that the Kernel
+  // is in use. Once the event has been signalled, the code in
   // CleanupCompletedEvent(Event) will do a urKernelRelease to update the
   // reference count on the kernel, using the kernel saved in CommandData.
   UR_CALL(urKernelRetain(Kernel));

source/adapters/level_zero/queue.cpp

@@ -1870,6 +1870,10 @@ ur_result_t ur_queue_handle_t_::createCommandList(
   ZeStruct<ze_command_list_desc_t> ZeCommandListDesc;
   ZeCommandListDesc.commandQueueGroupOrdinal = QueueGroupOrdinal;
 
+  if (Device->useDriverInOrderLists() && isInOrderQueue()) {
+    ZeCommandListDesc.flags = ZE_COMMAND_LIST_FLAG_IN_ORDER;
+  }
+
   ZE2UR_CALL(zeCommandListCreate, (Context->ZeContext, Device->ZeDevice,
                                    &ZeCommandListDesc, &ZeCommandList));
 
@@ -1985,7 +1989,11 @@ ur_command_list_ptr_t &ur_queue_handle_t_::ur_queue_group_t::getImmCmdList() {
 
   // Evaluate performance of explicit usage for "0" index.
   if (QueueIndex != 0) {
-    ZeCommandQueueDesc.flags = ZE_COMMAND_QUEUE_FLAG_EXPLICIT_ONLY;
+    ZeCommandQueueDesc.flags |= ZE_COMMAND_QUEUE_FLAG_EXPLICIT_ONLY;
+  }
+
+  if (Queue->Device->useDriverInOrderLists() && Queue->isInOrderQueue()) {
+    ZeCommandQueueDesc.flags |= ZE_COMMAND_QUEUE_FLAG_IN_ORDER;
   }
 
   // Check if context's command list cache has an immediate command list with