[EXP][CMDBUF] Add extra event to get CommandBuffer start time

Adds an extra event in the first command list associated to the CommandBuffer execution to obtain the start time of the graph execution.

Author: Maxime France-Pillois

source/adapters/level_zero/command_buffer.cpp

@@ -941,6 +941,25 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferEnqueueExp(
     ZE2UR_CALL(zeCommandListAppendBarrier,
                (SignalCommandList->first, RetEvent->ZeEvent, 1,
                 &(CommandBuffer->SignalEvent->ZeEvent)));
+
+    if ((Queue->Properties & UR_QUEUE_FLAG_PROFILING_ENABLE)) {
+      // We create an additional signal specific to the current execution of the
+      // CommandBuffer. This signal is needed for profiling the execution time
+      // of the CommandBuffer. It waits for the WaitEvent to be signaled
+      // which indicates the start of the CommandBuffer actual execution.
+      // This event is embedded into the Event return to the user to allow
+      // the profiling engine to retrieve it.
+      ur_event_handle_t StartEvent{};
+      UR_CALL(createEventAndAssociateQueue(
+          Queue, &StartEvent, UR_COMMAND_COMMAND_BUFFER_ENQUEUE_EXP,
+          WaitCommandList, false));
+
+      ZE2UR_CALL(zeCommandListAppendBarrier,
+                 (WaitCommandList->first, StartEvent->ZeEvent, 1,
+                  &(CommandBuffer->WaitEvent->ZeEvent)));
+
+      RetEvent->CommandData = StartEvent;
+    }
   }
 
   // Execution our command-lists asynchronously

source/adapters/level_zero/event.cpp

@@ -13,6 +13,7 @@
 #include <mutex>
 #include <string.h>
 
+#include "command_buffer.hpp"
 #include "common.hpp"
 #include "event.hpp"
 #include "ur_level_zero.hpp"
@@ -454,6 +455,17 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
                              ///< bytes returned in propValue
 ) {
   std::shared_lock<ur_shared_mutex> EventLock(Event->Mutex);
+
+  // A Command-buffer consists of three command-lists.
+  // The start time should therefore be taken from an event associated
+  // to the first command-list.
+  if ((Event->CommandType == UR_COMMAND_COMMAND_BUFFER_ENQUEUE_EXP) &&
+      (PropName == UR_PROFILING_INFO_COMMAND_START) && (Event->CommandData)) {
+    auto StartEvent = static_cast<ur_event_handle_t>(Event->CommandData);
+    return urEventGetProfilingInfo(StartEvent, UR_PROFILING_INFO_COMMAND_END,
+                                   PropValueSize, PropValue, PropValueSizeRet);
+  }
+
   if (Event->UrQueue &&
       (Event->UrQueue->Properties & UR_QUEUE_FLAG_PROFILING_ENABLE) == 0) {
     return UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE;
@@ -763,6 +775,13 @@ ur_result_t urEventReleaseInternal(ur_event_handle_t Event) {
       return Res;
     Event->CommandData = nullptr;
   }
+  if (Event->CommandType == UR_COMMAND_COMMAND_BUFFER_ENQUEUE_EXP &&
+      Event->CommandData) {
+    // Free the memory extra event allocated for profiling purposed.
+    auto AssociateEvent = static_cast<ur_event_handle_t>(Event->CommandData);
+    urEventRelease(AssociateEvent);
+    Event->CommandData = nullptr;
+  }
   if (Event->OwnNativeHandle) {
     if (DisableEventsCaching) {
       auto ZeResult = ZE_CALL_NOCHECK(zeEventDestroy, (Event->ZeEvent));