[L0][CMDBUF] Optimize fence/event waits during update

- Track whether a command buffer was recently synchronized so we can skip unnecessary fence/event waits

Author: Bensuo

source/adapters/level_zero/command_buffer.cpp

@@ -1753,6 +1753,8 @@ ur_result_t urCommandBufferEnqueueExp(
                                  EventWaitList, OutEvent, ZeCommandListHelper,
                                  DoProfiling));
   }
+  // Mark that synchronization will be required for later updates
+  CommandBuffer->NeedsUpdateSynchronization = true;
 
   return UR_RESULT_SUCCESS;
 }
@@ -2117,6 +2119,13 @@ ur_result_t updateKernelCommand(
  */
 ur_result_t
 waitForOngoingExecution(ur_exp_command_buffer_handle_t CommandBuffer) {
+  // Calling function has taken a lock for the command buffer so we can safely
+  // check and modify this value here.
+  // If command buffer was recently synchronized we can return early.
+  if (!CommandBuffer->NeedsUpdateSynchronization) {
+    return UR_RESULT_SUCCESS;
+  }
+
   if (CommandBuffer->UseImmediateAppendPath) {
     if (ur_event_handle_t &CurrentSubmissionEvent =
             CommandBuffer->CurrentSubmissionEvent) {
@@ -2128,7 +2137,8 @@ waitForOngoingExecution(ur_exp_command_buffer_handle_t CommandBuffer) {
   } else if (ze_fence_handle_t &ZeFence = CommandBuffer->ZeActiveFence) {
     ZE2UR_CALL(zeFenceHostSynchronize, (ZeFence, UINT64_MAX));
   }
-
+  // Mark that command buffer was recently synchronized
+  CommandBuffer->NeedsUpdateSynchronization = false;
   return UR_RESULT_SUCCESS;
 }
 

source/adapters/level_zero/command_buffer.hpp

@@ -141,6 +141,10 @@ struct ur_exp_command_buffer_handle_t_ : public _ur_object {
   // This list is needed to release all kernels retained by the
   // command_buffer.
   std::vector<ur_kernel_handle_t> KernelsList;
+  // Track whether synchronization is required when updating the command buffer
+  // Set this value to true when a command buffer is enqueued, and false after
+  // any fence or event synchronization to avoid repeated calls to synchronize.
+  bool NeedsUpdateSynchronization = false;
 };
 
 struct ur_exp_command_buffer_command_handle_t_ : public _ur_object {