[L0] Interrupt-based event implementation

To expose this functionality in UR, we want two ways of enabling low-power events:

Queue-wide enabling so all events created on the queue are low-powered.
As a property passed to urEnqueueEventsWaitWithBarrier making the resulting event a low-power event. This will require the existing interface to be extended with properties, potentially through a new experimental function.

Signed-off-by: Zhang, Winston <winston.zhang@intel.com>

Author: winstonzhang-intel

source/adapters/level_zero/command_buffer.cpp

@@ -215,7 +215,7 @@ ur_result_t createSyncPointAndGetZeEvents(
   UR_CALL(EventCreate(CommandBuffer->Context, nullptr /*Queue*/,
                       false /*IsMultiDevice*/, HostVisible, &LaunchEvent,
                       false /*CounterBasedEventEnabled*/,
-                      !CommandBuffer->IsProfilingEnabled));
+                      !CommandBuffer->IsProfilingEnabled, false));
   LaunchEvent->CommandType = CommandType;
   ZeLaunchEvent = LaunchEvent->ZeEvent;
 

source/adapters/level_zero/context.cpp

@@ -478,7 +478,8 @@ static const uint32_t MaxNumEventsPerPool = [] {
 ur_result_t ur_context_handle_t_::getFreeSlotInExistingOrNewPool(
     ze_event_pool_handle_t &Pool, size_t &Index, bool HostVisible,
     bool ProfilingEnabled, ur_device_handle_t Device,
-    bool CounterBasedEventEnabled, bool UsingImmCmdList) {
+    bool CounterBasedEventEnabled, bool UsingImmCmdList,
+    bool InterruptBasedEventEnabled) {
   // Lock while updating event pool machinery.
   std::scoped_lock<ur_mutex> Lock(ZeEventPoolCacheMutex);
 
@@ -537,6 +538,14 @@ ur_result_t ur_context_handle_t_::getFreeSlotInExistingOrNewPool(
                     counterBasedExt.flags);
       ZeEventPoolDesc.pNext = &counterBasedExt;
     }
+    if (InterruptBasedEventEnabled) {
+      ze_intel_event_sync_mode_exp_desc_t eventSyncMode = {
+          ZE_INTEL_STRUCTURE_TYPE_EVENT_SYNC_MODE_EXP_DESC, nullptr, 0};
+      eventSyncMode.syncModeFlags =
+          ZE_INTEL_EVENT_SYNC_MODE_EXP_FLAG_LOW_POWER_WAIT |
+          ZE_INTEL_EVENT_SYNC_MODE_EXP_FLAG_SIGNAL_INTERRUPT;
+      ZeEventPoolDesc.pNext = &eventSyncMode;
+    }
 
     std::vector<ze_device_handle_t> ZeDevices;
     if (ZeDevice) {

source/adapters/level_zero/context.hpp

@@ -33,6 +33,24 @@ struct l0_command_list_cache_info {
   bool IsImmediate = false;
 };
 
+typedef uint32_t ze_intel_event_sync_mode_exp_flags_t;
+typedef enum _ze_intel_event_sync_mode_exp_flag_t {
+  ZE_INTEL_EVENT_SYNC_MODE_EXP_FLAG_LOW_POWER_WAIT = ZE_BIT(0),
+  ZE_INTEL_EVENT_SYNC_MODE_EXP_FLAG_SIGNAL_INTERRUPT = ZE_BIT(1),
+  ZE_INTEL_EVENT_SYNC_MODE_EXP_EXP_FLAG_FORCE_UINT32 = 0x7fffffff
+
+} ze_intel_event_sync_mode_exp_flag_t;
+
+#define ZE_INTEL_STRUCTURE_TYPE_EVENT_SYNC_MODE_EXP_DESC                       \
+  (ze_structure_type_t)0x00030016
+
+typedef struct _ze_intel_event_sync_mode_exp_desc_t {
+  ze_structure_type_t stype;
+  const void *pNext;
+
+  ze_intel_event_sync_mode_exp_flags_t syncModeFlags;
+} ze_intel_event_sync_mode_exp_desc_t;
+
 struct ur_context_handle_t_ : _ur_object {
   ur_context_handle_t_(ze_context_handle_t ZeContext, uint32_t NumDevices,
                        const ur_device_handle_t *Devs, bool OwnZeContext)
@@ -199,7 +217,8 @@ struct ur_context_handle_t_ : _ur_object {
                                              bool ProfilingEnabled,
                                              ur_device_handle_t Device,
                                              bool CounterBasedEventEnabled,
-                                             bool UsingImmCmdList);
+                                             bool UsingImmCmdList,
+                                             bool InterruptBasedEventEnabled);
 
   // Get ur_event_handle_t from cache.
   ur_event_handle_t getEventFromContextCache(bool HostVisible,

source/adapters/level_zero/device.cpp

@@ -485,6 +485,8 @@ ur_result_t urDeviceGetInfo(
   case UR_DEVICE_INFO_BUILT_IN_KERNELS:
     // TODO: To find out correct value
     return ReturnValue("");
+  case UR_DEVICE_INFO_LOW_POWER_EVENTS_EXP:
+    return ReturnValue(UR_DEVICE_INFO_LOW_POWER_EVENTS_EXP);
   case UR_DEVICE_INFO_QUEUE_PROPERTIES:
     return ReturnValue(
         ur_queue_flag_t(UR_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE |

source/adapters/level_zero/event.cpp

@@ -422,7 +422,8 @@ ur_result_t urEnqueueEventsWaitWithBarrier(
 ur_result_t urEnqueueEventsWaitWithBarrierExt(
     ur_queue_handle_t Queue, ///< [in] handle of the queue object
     const ur_exp_enqueue_ext_properties_t
-        *, ///< [in][optional] pointer to the extended enqueue properties
+        *EnqueueExtProp, ///< [in][optional] pointer to the extended enqueue
+                         ///< properties
     uint32_t NumEventsInWaitList, ///< [in] size of the event wait list
     const ur_event_handle_t
         *EventWaitList, ///< [in][optional][range(0, numEventsInWaitList)]
@@ -913,7 +914,7 @@ ur_result_t urExtEventCreate(
   UR_CALL(EventCreate(Context, nullptr /*Queue*/, false /*IsMultiDevice*/,
                       true /*HostVisible*/, Event,
                       false /*CounterBasedEventEnabled*/,
-                      false /*ForceDisableProfiling*/));
+                      false /*ForceDisableProfiling*/, false));
 
   (*Event)->RefCountExternal++;
   if (!(*Event)->CounterBasedEventsEnabled)
@@ -935,7 +936,7 @@ ur_result_t urEventCreateWithNativeHandle(
     UR_CALL(EventCreate(Context, nullptr /*Queue*/, false /*IsMultiDevice*/,
                         true /*HostVisible*/, Event,
                         false /*CounterBasedEventEnabled*/,
-                        false /*ForceDisableProfiling*/));
+                        false /*ForceDisableProfiling*/, false));
 
     (*Event)->RefCountExternal++;
     if (!(*Event)->CounterBasedEventsEnabled)
@@ -1293,7 +1294,8 @@ ur_result_t EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
                         bool IsMultiDevice, bool HostVisible,
                         ur_event_handle_t *RetEvent,
                         bool CounterBasedEventEnabled,
-                        bool ForceDisableProfiling) {
+                        bool ForceDisableProfiling,
+                        bool InterruptBasedEventEnabled) {
   bool ProfilingEnabled =
       ForceDisableProfiling ? false : (!Queue || Queue->isProfilingEnabled());
   bool UsingImmediateCommandlists = !Queue || Queue->UsingImmCmdLists;
@@ -1317,14 +1319,15 @@ ur_result_t EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
 
   if (auto Res = Context->getFreeSlotInExistingOrNewPool(
           ZeEventPool, Index, HostVisible, ProfilingEnabled, Device,
-          CounterBasedEventEnabled, UsingImmediateCommandlists))
+          CounterBasedEventEnabled, UsingImmediateCommandlists,
+          Queue->interruptBasedEventsEnabled()))
     return Res;
 
   ZeStruct<ze_event_desc_t> ZeEventDesc;
   ZeEventDesc.index = Index;
   ZeEventDesc.wait = 0;
 
-  if (HostVisible || CounterBasedEventEnabled) {
+  if (HostVisible || CounterBasedEventEnabled || InterruptBasedEventEnabled) {
     ZeEventDesc.signal = ZE_EVENT_SCOPE_FLAG_HOST;
   } else {
     //
@@ -1350,6 +1353,7 @@ ur_result_t EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
     return UR_RESULT_ERROR_UNKNOWN;
   }
   (*RetEvent)->CounterBasedEventsEnabled = CounterBasedEventEnabled;
+  (*RetEvent)->InterruptBasedEventsEnabled = InterruptBasedEventEnabled;
   if (HostVisible)
     (*RetEvent)->HostVisibleEvent =
         reinterpret_cast<ur_event_handle_t>(*RetEvent);

source/adapters/level_zero/event.hpp

@@ -33,7 +33,8 @@ ur_result_t EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
                         bool IsMultiDevice, bool HostVisible,
                         ur_event_handle_t *RetEvent,
                         bool CounterBasedEventEnabled,
-                        bool ForceDisableProfiling);
+                        bool ForceDisableProfiling,
+                        bool InterruptBasedEventEnabled);
 } // extern "C"
 
 // This is an experimental option that allows to disable caching of events in
@@ -251,6 +252,8 @@ struct ur_event_handle_t_ : _ur_object {
   std::optional<ur_completion_batch_it> completionBatch;
   // Keeps track of whether we are using Counter-based Events.
   bool CounterBasedEventsEnabled = false;
+  // Keeps track of whether we are using Interrupt-based Events.
+  bool InterruptBasedEventsEnabled = false;
 };
 
 // Helper function to implement zeHostSynchronize.

source/adapters/level_zero/queue.cpp

@@ -104,10 +104,10 @@ ur_result_t ur_completion_batch::seal(ur_queue_handle_t queue,
   assert(st == ACCUMULATING);
 
   if (!barrierEvent) {
-    UR_CALL(EventCreate(queue->Context, queue, false /*IsMultiDevice*/,
-                        true /*HostVisible*/, &barrierEvent,
-                        false /*CounterBasedEventEnabled*/,
-                        false /*ForceDisableProfiling*/));
+    UR_CALL(EventCreate(
+        queue->Context, queue, false /*IsMultiDevice*/, true /*HostVisible*/,
+        &barrierEvent, false /*CounterBasedEventEnabled*/,
+        false /*ForceDisableProfiling*/, false /*InterruptBasedEventEnabled*/));
   }
 
   // Instead of collecting all the batched events, we simply issue a global
@@ -1494,6 +1494,11 @@ bool ur_queue_handle_t_::doReuseDiscardedEvents() {
   return ReuseDiscardedEvents && isInOrderQueue() && isDiscardEvents();
 }
 
+bool ur_queue_handle_t_::interruptBasedEventsEnabled() {
+  return isInOrderQueue() && Device->useDriverInOrderLists() &&
+         isLowPowerEvents();
+}
+
 ur_result_t
 ur_queue_handle_t_::resetDiscardedEvent(ur_command_list_ptr_t CommandList) {
   if (LastCommandEvent && LastCommandEvent->IsDiscarded) {
@@ -1654,6 +1659,10 @@ bool ur_queue_handle_t_::isInOrderQueue() const {
           0);
 }
 
+bool ur_queue_handle_t_::isLowPowerEvents() const {
+  return ((this->Properties & UR_QUEUE_FLAG_LOW_POWER_EVENTS_EXP) != 0);
+}
+
 // Helper function to perform the necessary cleanup of the events from reset cmd
 // list.
 ur_result_t CleanupEventListFromResetCmdList(
@@ -1868,12 +1877,10 @@ ur_result_t setSignalEvent(ur_queue_handle_t Queue, bool UseCopyEngine,
 //        visible pool.
 // \param HostVisible tells if the event must be created in the
 //        host-visible pool. If not set then this function will decide.
-ur_result_t createEventAndAssociateQueue(ur_queue_handle_t Queue,
-                                         ur_event_handle_t *Event,
-                                         ur_command_t CommandType,
-                                         ur_command_list_ptr_t CommandList,
-                                         bool IsInternal, bool IsMultiDevice,
-                                         std::optional<bool> HostVisible) {
+ur_result_t createEventAndAssociateQueue(
+    ur_queue_handle_t Queue, ur_event_handle_t *Event, ur_command_t CommandType,
+    ur_command_list_ptr_t CommandList, bool IsInternal, bool IsMultiDevice,
+    std::optional<bool> HostVisible, std::optional<bool> InterruptBasedEvents) {
 
   if (!HostVisible.has_value()) {
     // Internal/discarded events do not need host-scope visibility.
@@ -1888,7 +1895,8 @@ ur_result_t createEventAndAssociateQueue(ur_queue_handle_t Queue,
   if (*Event == nullptr)
     UR_CALL(EventCreate(
         Queue->Context, Queue, IsMultiDevice, HostVisible.value(), Event,
-        Queue->CounterBasedEventsEnabled, false /*ForceDisableProfiling*/));
+        Queue->CounterBasedEventsEnabled, false /*ForceDisableProfiling*/,
+        HostVisible.has_value() ? true : Queue->interruptBasedEventsEnabled()));
 
   (*Event)->UrQueue = Queue;
   (*Event)->CommandType = CommandType;

source/adapters/level_zero/queue.hpp

@@ -533,6 +533,8 @@ struct ur_queue_handle_t_ : _ur_object {
   // queue.
   bool doReuseDiscardedEvents();
 
+  bool interruptBasedEventsEnabled();
+
   // Append command to provided command list to wait and reset the last event if
   // it is discarded and create new ur_event_handle_t wrapper using the same
   // native event and put it to the cache. We call this method after each
@@ -557,6 +559,9 @@ struct ur_queue_handle_t_ : _ur_object {
   // Returns true if the queue has discard events property.
   bool isDiscardEvents() const;
 
+  // Returns true if the queue has low power events property.
+  bool isLowPowerEvents() const;
+
   // Returns true if the queue has explicit priority set by user.
   bool isPriorityLow() const;
   bool isPriorityHigh() const;
@@ -708,7 +713,7 @@ struct ur_queue_handle_t_ : _ur_object {
 ur_result_t createEventAndAssociateQueue(
     ur_queue_handle_t Queue, ur_event_handle_t *Event, ur_command_t CommandType,
     ur_command_list_ptr_t CommandList, bool IsInternal, bool IsMultiDevice,
-    std::optional<bool> HostVisible = std::nullopt);
+    std::optional<bool> HostVisible = std::nullopt, std::optional<bool> InterruptBasedEvents = std::nullopt);
 
 // This helper function checks to see if an event for a command can be included
 // at the end of a command list batch. This will only be true if the event does