Fix incorrect profiling timestamps when events are reused (#21086)

Fix issue where urEventGetProfilingInfo returns stale timestamps when
profiling events are reused from cache. zeEventHostReset() does not
clear profiling data, causing zeEventQueryKernelTimestamp() to return
old timestamps.

This patch includes these changes:
- Prevents reusing events with profiling enabled from the cache,
ensuring each profiling measurement uses a fresh event.
- Fix reset() to clear timestamp fields.
- COMMAND_START case for timestamped events to return correct start
timestamp.

---------

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

Author: winstonzhang-intel

source/adapters/level_zero/context.cpp

@@ -597,6 +597,13 @@ ur_result_t ur_context_handle_t_::getFreeSlotInExistingOrNewPool(
 ur_event_handle_t ur_context_handle_t_::getEventFromContextCache(
     bool HostVisible, bool WithProfiling, ur_device_handle_t Device,
     bool CounterBasedEventEnabled, bool InterruptBasedEventEnabled) {
+  // Don't reuse events with profiling enabled because zeEventHostReset
+  // does not clear the profiling timestamps, causing stale timestamp data
+  // to be returned by zeEventQueryKernelTimestamp after the event is reused.
+  if (WithProfiling) {
+    return nullptr;
+  }
+
   std::scoped_lock<ur_mutex> Lock(EventCacheMutex);
   auto Cache =
       getEventCache(HostVisible, WithProfiling, Device,

source/adapters/level_zero/event.cpp

@@ -1427,6 +1427,15 @@ EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
 }
 
 ur_result_t ur_event_handle_t_::reset() {
+  // Clean up timestamp recording entry from queue if this was a timestamped
+  // event
+  if (IsTimestamped && UrQueue) {
+    auto Entry = UrQueue->EndTimeRecordings.find(this);
+    if (Entry != UrQueue->EndTimeRecordings.end()) {
+      UrQueue->EndTimeRecordings.erase(Entry);
+    }
+  }
+
   UrQueue = nullptr;
   CleanedUp = false;
   Completed = false;
@@ -1439,6 +1448,8 @@ ur_result_t ur_event_handle_t_::reset() {
   completionBatch = std::nullopt;
   OriginAllocEvent = nullptr;
   IsTimestamped = false;
+  RecordEventStartTimestamp = 0;
+  RecordEventEndTimestamp = 0;
 
   if (!isHostVisible())
     HostVisibleEvent = nullptr;

test/adapters/level_zero/event_cache_tests.cpp

@@ -142,7 +142,11 @@ TEST_P(urEventCacheTest, eventsReuseWithVisibleEvent) {
     verifyData();
   }
 
-  ASSERT_LT(eventCreateCount, numIters * numEnqueues);
+  if (flags & UR_QUEUE_FLAG_PROFILING_ENABLE) {
+    ASSERT_GE(eventCreateCount, numIters * numEnqueues);
+  } else {
+    ASSERT_LT(eventCreateCount, numIters * numEnqueues);
+  }
 }
 
 TEST_P(urEventCacheTest, eventsReuseWithVisibleEventAndWait) {