Modified spec wording and made creating an event branchless on CUDA and HIP

Author: konradkusiak97

unified-runtime/include/ur_api.h

@@ -310,7 +310,7 @@ returns:
     - $X_RESULT_ERROR_OUT_OF_HOST_MEMORY
 --- #--------------------------------------------------------------------------
 type: function
-desc: "Decrement the command-buffer object's reference count and delete the command-buffer object if the reference count becomes zero. It will try synchronizing the command-buffer, hence it is legal to call it while command-buffer is still executing."
+desc: "Decrement the command-buffer object's reference count and delete the command-buffer object if the reference count becomes zero. It is legal to call the entry-point while hCommandBuffer is still executing, which will block on completion if the reference count of hCommandBuffer becomes zero."
 class: $xCommandBuffer
 name: ReleaseExp
 params:

unified-runtime/scripts/core/exp-command-buffer.yml

@@ -1172,10 +1172,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCommandBufferExp(
   // Launch graph
   UR_CHECK_ERROR(cuGraphLaunch(hCommandBuffer->CudaGraphExec, CuStream));
 
+  UR_CHECK_ERROR(RetImplEvent->record());
+  hCommandBuffer->CurrentExecution = RetImplEvent.release();
   if (phEvent) {
-    UR_CHECK_ERROR(RetImplEvent->record());
-    *phEvent = RetImplEvent.release();
-    hCommandBuffer->CurrentExecution = *phEvent;
+    *phEvent = hCommandBuffer->CurrentExecution;
   }
   return UR_RESULT_SUCCESS;
 } catch (ur_result_t Err) {

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

@@ -192,7 +192,7 @@ struct ur_exp_command_buffer_handle_t_ : ur::cuda::handle_base {
   // Atomic variable counting the number of reference to this command_buffer
   // using std::atomic prevents data race when incrementing/decrementing.
   std::atomic_uint32_t RefCount;
-  // The event of current graph execution.
+  // Track the event of the current graph execution.
   ur_event_handle_t CurrentExecution = nullptr;
 
   // Ordered map of sync_points to ur_events, so that we can find the last

unified-runtime/source/adapters/cuda/command_buffer.hpp

@@ -811,10 +811,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCommandBufferExp(
     // Launch graph
     UR_CHECK_ERROR(hipGraphLaunch(hCommandBuffer->HIPGraphExec, HIPStream));
 
+    UR_CHECK_ERROR(RetImplEvent->record());
+    hCommandBuffer->CurrentExecution = RetImplEvent.release();
     if (phEvent) {
-      UR_CHECK_ERROR(RetImplEvent->record());
-      *phEvent = RetImplEvent.release();
-      hCommandBuffer->CurrentExecution = *phEvent;
+      *phEvent = hCommandBuffer->CurrentExecution;
     }
   } catch (ur_result_t Err) {
     return Err;

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

@@ -128,7 +128,7 @@ struct ur_exp_command_buffer_handle_t_ : ur::hip::handle_base {
   // Atomic variable counting the number of reference to this command_buffer
   // using std::atomic prevents data race when incrementing/decrementing.
   std::atomic_uint32_t RefCount;
-  // The event of current graph execution.
+  // Track the event of the current graph execution.
   ur_event_handle_t CurrentExecution = nullptr;
 
   // Ordered map of sync_points to ur_events