[OMPT] Initialize task fields only if needed

[Hahnfeld]

Summary

Because __kmp_task_init_ompt is called for every initial task in each thread and always generated task ids, this was a big performance issue on bigger systems even without any tool attached.

After changing the initialization interface to ompt_tool, we can now rely on already knowing whether a tool is attached and OMPT is enabled at this point.

[jmellorcrummey]

LGTM. Thanks for finding and correcting this performance problem.

[jlpeyton]

LGTM as well.

openmp/trunk/runtime/src/kmp.h

 void ompc_set_num_threads( int arg );
 
 extern void __kmp_push_current_task_to_thread( kmp_info_t *this_thr,
                   kmp_team_t *team, int tid );
 extern void __kmp_pop_current_task_from_thread( kmp_info_t *this_thr );
 extern kmp_task_t* __kmp_task_alloc( ident_t *loc_ref, kmp_int32 gtid,
   kmp_tasking_flags_t *flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds,
   kmp_routine_entry_t task_entry );
-#if OMPT_SUPPORT
-extern void __kmp_task_init_ompt( kmp_taskdata_t * task, int tid );
-#endif
 extern void __kmp_init_implicit_task( ident_t *loc_ref, kmp_info_t *this_thr,
                   kmp_team_t *team, int tid, int set_curr_task );
 
 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32 *flag, int final_spin,
                            int *thread_finished,
 #if USE_ITT_BUILD
                            void * itt_sync_obj,
 #endif /* USE_ITT_BUILD */

openmp/trunk/runtime/src/kmp_tasking.c

__kmpc_omp_task_complete( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task )
     return;
 }
 #endif // TASK_UNUSED
 
 
 #if OMPT_SUPPORT
 //----------------------------------------------------------------------------------------------------
 // __kmp_task_init_ompt:
-//   Initialize OMPT fields maintained by a task. Since the serial task is initialized before
-//   ompt_initialize is called, at the point the serial task is initialized we don't know whether
-//   OMPT will be used or not when the serial task is initialized. This function provides the support
-//   needed to initialize OMPT for the serial task after the fact.
+//   Initialize OMPT fields maintained by a task. This will only be called after
+//   ompt_tool, so we already know whether ompt is enabled or not.
 
-void
-__kmp_task_init_ompt( kmp_taskdata_t * task, int tid )
+static inline void
+__kmp_task_init_ompt( kmp_taskdata_t * task, int tid, void * function )
 {
+    if (ompt_enabled) {
         task->ompt_task_info.task_id = __ompt_task_id_new(tid);
-    task->ompt_task_info.function = NULL;
+        task->ompt_task_info.function = function;
         task->ompt_task_info.frame.exit_runtime_frame = NULL;
         task->ompt_task_info.frame.reenter_runtime_frame = NULL;
     }
+}
 #endif
 
 
 //----------------------------------------------------------------------------------------------------
 // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit task for a given thread
 //
 // loc_ref:  reference to source location of parallel region
 // this_thr:  thread data structure corresponding to implicit task

 #endif
         __kmp_push_current_task_to_thread( this_thr, team, tid );
     } else {
         KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0);
         KMP_DEBUG_ASSERT(task->td_allocated_child_tasks  == 0);
     }
 
 #if OMPT_SUPPORT
-    __kmp_task_init_ompt(task, tid);
+    __kmp_task_init_ompt(task, tid, NULL);
 #endif
 
     KF_TRACE(10, ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n",
                   tid, team, task ) );
 }
 
 // Round up a size to a power of two specified by val
 // Used to insert padding between structures co-allocated using a single malloc() call
#endif
             KMP_TEST_THEN_INC32( (kmp_int32 *)(& taskdata->td_parent->td_allocated_child_tasks) );
         }
     }
 
     KA_TRACE(20, ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n",
                   gtid, taskdata, taskdata->td_parent) );
 
 #if OMPT_SUPPORT
-    if (ompt_enabled) {
-        taskdata->ompt_task_info.task_id = __ompt_task_id_new(gtid);
-        taskdata->ompt_task_info.function = (void*) task_entry;
-        taskdata->ompt_task_info.frame.exit_runtime_frame = NULL; 
-        taskdata->ompt_task_info.frame.reenter_runtime_frame = NULL;
-    }
+    __kmp_task_init_ompt(taskdata, gtid, (void*) task_entry);
 #endif
 
     return task;
 }
 
 
 kmp_task_t *
 __kmpc_omp_task_alloc( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags,