Updated Driver In order lists check and required version (#17620)

@nrspruit's [patch ](#17106
and posted here as I work on fixing this patch.

- Refactor's the check for driver in order lists into one place, but
still disables driver in order lists by default.

---------

Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Zhang, Winston <[email protected]>
Co-authored-by: Wu, Yingcong <[email protected]>

Author: winstonzhang-intel

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

@@ -76,7 +76,9 @@ bool checkCounterBasedEventsSupport(ur_device_handle_t Device) {
     return std::atoi(UrRet) != 0;
   }();
 
-  return Device->ImmCommandListUsed && Device->useDriverInOrderLists() &&
+  return Device->ImmCommandListUsed &&
+         Device->Platform->allowDriverInOrderLists(
+             true /*Only Allow Driver In Order List if requested*/) &&
          useDriverCounterBasedEvents &&
          Device->Platform->ZeDriverEventPoolCountingEventsExtensionFound;
 }
@@ -647,12 +649,9 @@ ur_result_t createMainCommandList(ur_context_handle_t Context,
  */
 bool canBeInOrder(ur_context_handle_t Context,
                   const ur_exp_command_buffer_desc_t *CommandBufferDesc) {
-  const char *UrRet = std::getenv("UR_L0_USE_DRIVER_INORDER_LISTS");
-  // In-order command-lists are not available in old driver version.
-  bool DriverInOrderRequested = UrRet ? std::atoi(UrRet) != 0 : false;
-  bool CompatibleDriver = Context->getPlatform()->isDriverVersionNewerOrSimilar(
-      1, 3, L0_DRIVER_INORDER_MIN_VERSION);
-  bool CanUseDriverInOrderLists = CompatibleDriver && DriverInOrderRequested;
+  bool CanUseDriverInOrderLists =
+      Context->getPlatform()->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/);
   return CanUseDriverInOrderLists ? CommandBufferDesc->isInOrder : false;
 }
 

unified-runtime/source/adapters/level_zero/common.hpp

@@ -432,6 +432,4 @@ extern thread_local int32_t ErrorAdapterNativeCode;
 // Utility function for setting a message and warning
 [[maybe_unused]] void setErrorMessage(const char *pMessage,
                                       ur_result_t ErrorCode,
-                                      int32_t AdapterErrorCode);
-
-#define L0_DRIVER_INORDER_MIN_VERSION 29534
+                                      int32_t AdapterErrorCode);

unified-runtime/source/adapters/level_zero/context.cpp

@@ -207,7 +207,8 @@ ur_result_t ur_context_handle_t_::initialize() {
 
   ZeCommandQueueDesc.index = 0;
   ZeCommandQueueDesc.mode = ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS;
-  if (Device->useDriverInOrderLists() &&
+  if (Device->Platform->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/) &&
       Device->useDriverCounterBasedEvents()) {
     logger::debug(
         "L0 Synchronous Immediate Command List needed with In Order property.");
@@ -673,8 +674,9 @@ ur_result_t ur_context_handle_t_::getAvailableCommandList(
     for (auto ZeCommandListIt = ZeCommandListCache.begin();
          ZeCommandListIt != ZeCommandListCache.end(); ++ZeCommandListIt) {
       // If this is an InOrder Queue, then only allow lists which are in order.
-      if (Queue->Device->useDriverInOrderLists() && Queue->isInOrderQueue() &&
-          !(ZeCommandListIt->second.InOrderList)) {
+      if (Queue->Device->Platform->allowDriverInOrderLists(
+              true /*Only Allow Driver In Order List if requested*/) &&
+          Queue->isInOrderQueue() && !(ZeCommandListIt->second.InOrderList)) {
         continue;
       }
       // Only allow to reuse Regular Command Lists
@@ -740,8 +742,9 @@ ur_result_t ur_context_handle_t_::getAvailableCommandList(
       continue;
 
     // If this is an InOrder Queue, then only allow lists which are in order.
-    if (Queue->Device->useDriverInOrderLists() && Queue->isInOrderQueue() &&
-        !(it->second.IsInOrderList)) {
+    if (Queue->Device->Platform->allowDriverInOrderLists(
+            true /*Only Allow Driver In Order List if requested*/) &&
+        Queue->isInOrderQueue() && !(it->second.IsInOrderList)) {
       continue;
     }
 

unified-runtime/source/adapters/level_zero/device.cpp

@@ -1676,22 +1676,6 @@ 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");
-    // bool CompatibleDriver = this->Platform->isDriverVersionNewerOrSimilar(
-    //     1, 3, L0_DRIVER_INORDER_MIN_VERSION);
-    if (!UrRet)
-      return false;
-    return std::atoi(UrRet) != 0;
-  }();
-
-  return UseDriverInOrderLists;
-}
-
 bool ur_device_handle_t_::useDriverCounterBasedEvents() {
   // Use counter-based events implementation from L0 driver.
 

unified-runtime/source/adapters/level_zero/device.hpp

@@ -155,9 +155,6 @@ 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();
-
   // Whether Adapter uses driver's implementation of counter-based events or not
   bool useDriverCounterBasedEvents();
 

unified-runtime/source/adapters/level_zero/event.cpp

@@ -223,22 +223,6 @@ ur_result_t urEnqueueEventsWaitWithBarrierExt(
         if (Queue->isInOrderQueue() && InOrderBarrierBySignal &&
             !Queue->isProfilingEnabled()) {
           if (EventWaitList.Length) {
-            if (CmdList->second.IsInOrderList) {
-              for (unsigned i = EventWaitList.Length; i-- > 0;) {
-                // If the event is a multidevice event, then given driver in
-                // order lists, we cannot include this into the wait event list
-                // due to driver limitations.
-                if (EventWaitList.UrEventList[i]->IsMultiDevice) {
-                  EventWaitList.Length--;
-                  if (EventWaitList.Length != i) {
-                    std::swap(EventWaitList.UrEventList[i],
-                              EventWaitList.UrEventList[EventWaitList.Length]);
-                    std::swap(EventWaitList.ZeEventList[i],
-                              EventWaitList.ZeEventList[EventWaitList.Length]);
-                  }
-                }
-              }
-            }
             ZE2UR_CALL(zeCommandListAppendWaitOnEvents,
                        (CmdList->first, EventWaitList.Length,
                         EventWaitList.ZeEventList));
@@ -1505,8 +1489,9 @@ ur_result_t _ur_ze_event_list_t::createAndRetainUrZeEventList(
   // 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) {
+  if (CurQueue->Device->Platform->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/) &&
+      CurQueue->isInOrderQueue() && CurQueue->UsingImmCmdLists) {
     auto QueueGroup = CurQueue->getQueueGroup(UseCopyEngine);
     uint32_t QueueGroupOrdinal, QueueIndex;
     auto NextIndex = QueueGroup.getQueueIndex(&QueueGroupOrdinal, &QueueIndex,
@@ -1535,7 +1520,8 @@ ur_result_t _ur_ze_event_list_t::createAndRetainUrZeEventList(
 
     // 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() &&
+    if (CurQueue->Device->Platform->allowDriverInOrderLists(
+            true /*Only Allow Driver In Order List if requested*/) &&
         CurQueue->isInOrderQueue() &&
         WaitListEmptyOrAllEventsFromSameQueue(CurQueue, EventListLength,
                                               EventList)) {

unified-runtime/source/adapters/level_zero/platform.cpp

@@ -549,6 +549,28 @@ ur_result_t ur_platform_handle_t_::initialize() {
   return UR_RESULT_SUCCESS;
 }
 
+bool ur_platform_handle_t_::allowDriverInOrderLists(bool OnlyIfRequested) {
+  // Use in-order lists implementation from L0 driver instead
+  // of adapter's implementation.
+
+  // The following driver version is known to be passing and only this or newer
+  // drivers should be allowed by default for in order lists.
+#define L0_DRIVER_INORDER_MINOR_VERSION 6
+#define L0_DRIVER_INORDER_PATCH_VERSION 32149
+
+  static const bool UseEnvVarDriverInOrderLists = [&] {
+    const char *UrRet = std::getenv("UR_L0_USE_DRIVER_INORDER_LISTS");
+    return UrRet ? std::atoi(UrRet) != 0 : false;
+  }();
+  static const bool UseDriverInOrderLists = [this] {
+    bool CompatibleDriver = this->isDriverVersionNewerOrSimilar(
+        1, L0_DRIVER_INORDER_MINOR_VERSION, L0_DRIVER_INORDER_PATCH_VERSION);
+    return CompatibleDriver || UseEnvVarDriverInOrderLists;
+  }();
+
+  return OnlyIfRequested ? UseEnvVarDriverInOrderLists : UseDriverInOrderLists;
+}
+
 /// Checks the version of the level-zero driver.
 /// @param VersionMajor Major verion number to compare to.
 /// @param VersionMinor Minor verion number to compare to.

unified-runtime/source/adapters/level_zero/platform.hpp

@@ -48,6 +48,10 @@ struct ur_platform_handle_t_ : public _ur_platform {
   // Zero.
   ZeDriverVersionStringExtension ZeDriverVersionString;
 
+  // Helper function to check if the driver supports Driver In Order Lists or
+  // the User has Requested this support.
+  bool allowDriverInOrderLists(bool OnlyIfRequested = false);
+
   // Cache versions info from zeDriverGetProperties.
   std::string ZeDriverVersion;
   std::string ZeDriverApiVersion;

unified-runtime/source/adapters/level_zero/queue.cpp

@@ -1198,11 +1198,15 @@ ur_queue_handle_t_::ur_queue_handle_t_(
   CopyCommandBatch.QueueBatchSize = ZeCommandListBatchCopyConfig.startSize();
 
   this->CounterBasedEventsEnabled =
-      UsingImmCmdLists && isInOrderQueue() && Device->useDriverInOrderLists() &&
+      UsingImmCmdLists && isInOrderQueue() &&
+      Device->Platform->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/) &&
       Device->useDriverCounterBasedEvents() &&
       Device->Platform->ZeDriverEventPoolCountingEventsExtensionFound;
   this->InterruptBasedEventsEnabled =
-      isLowPowerEvents() && isInOrderQueue() && Device->useDriverInOrderLists();
+      isLowPowerEvents() && isInOrderQueue() &&
+      Device->Platform->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/);
 }
 
 void ur_queue_handle_t_::adjustBatchSizeForFullBatch(bool IsCopy) {
@@ -2297,7 +2301,9 @@ ur_result_t ur_queue_handle_t_::createCommandList(
   ZeCommandListDesc.commandQueueGroupOrdinal = QueueGroupOrdinal;
 
   bool IsInOrderList = false;
-  if (Device->useDriverInOrderLists() && isInOrderQueue()) {
+  if (Device->Platform->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/) &&
+      isInOrderQueue()) {
     ZeCommandListDesc.flags = ZE_COMMAND_LIST_FLAG_IN_ORDER;
     IsInOrderList = true;
   }
@@ -2434,7 +2440,9 @@ ur_command_list_ptr_t &ur_queue_handle_t_::ur_queue_group_t::getImmCmdList() {
     ZeCommandQueueDesc.flags |= ZE_COMMAND_QUEUE_FLAG_EXPLICIT_ONLY;
   }
 
-  if (Queue->Device->useDriverInOrderLists() && Queue->isInOrderQueue()) {
+  if (Queue->Device->Platform->allowDriverInOrderLists(
+          true /*Only Allow Driver In Order List if requested*/) &&
+      Queue->isInOrderQueue()) {
     isInOrderList = true;
     ZeCommandQueueDesc.flags |= ZE_COMMAND_QUEUE_FLAG_IN_ORDER;
   }