[OpenMP][OMPT] Add missing callbacks for asynchronous target tasks (llvm#93472)

- The first hidden-helper-thread did not trigger thread-begin
- The "detaching" from a target-task when waiting for completion missed
to call task-switch
- Target tasks identified themself as explicit task

Co-authored-by: Kaloyan Ignatov <[email protected]>

Author: jprotze

openmp/runtime/src/kmp.h

@@ -2675,11 +2675,12 @@ typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
   /* Same fields as in the #else branch, but in reverse order */
 #if OMPX_TASKGRAPH
-  unsigned reserved31 : 6;
+  unsigned reserved31 : 5;
   unsigned onced : 1;
 #else
-  unsigned reserved31 : 7;
+  unsigned reserved31 : 6;
 #endif
+  unsigned target : 1;
   unsigned native : 1;
   unsigned freed : 1;
   unsigned complete : 1;
@@ -2728,11 +2729,12 @@ typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
   unsigned complete : 1; /* 1==complete, 0==not complete   */
   unsigned freed : 1; /* 1==freed, 0==allocated        */
   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
+  unsigned target : 1;
 #if OMPX_TASKGRAPH
   unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
-  unsigned reserved31 : 6; /* reserved for library use */
+  unsigned reserved31 : 5; /* reserved for library use */
 #else
-  unsigned reserved31 : 7; /* reserved for library use */
+  unsigned reserved31 : 6; /* reserved for library use */
 #endif
 #endif
 } kmp_tasking_flags_t;

openmp/runtime/src/kmp_runtime.cpp

@@ -3972,7 +3972,7 @@ int __kmp_register_root(int initial_thread) {
   __kmp_root_counter++;
 
 #if OMPT_SUPPORT
-  if (!initial_thread && ompt_enabled.enabled) {
+  if (ompt_enabled.enabled) {
 
     kmp_info_t *root_thread = ompt_get_thread();
 

openmp/runtime/src/kmp_taskdeps.cpp

@@ -739,7 +739,7 @@ kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
           &(current_task->ompt_task_info.task_data),
           &(current_task->ompt_task_info.frame),
           &(new_taskdata->ompt_task_info.task_data),
-          ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 1,
+          TASK_TYPE_DETAILS_FORMAT(new_taskdata), 1,
           OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
     }
 

openmp/runtime/src/kmp_tasking.cpp

@@ -811,8 +811,7 @@ static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid,
       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
           &(parent_info->task_data), &(parent_info->frame),
           &(taskdata->ompt_task_info.task_data),
-          ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(taskdata), 0,
-          return_address);
+          TASK_TYPE_DETAILS_FORMAT(taskdata), 0, return_address);
     }
     __ompt_task_start(task, current_task, gtid);
   }
@@ -1156,6 +1155,11 @@ static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task,
     // Note: no need to translate gtid to its shadow. If the current thread is a
     // hidden helper one, then the gtid is already correct. Otherwise, hidden
     // helper threads are disabled, and gtid refers to a OpenMP thread.
+#if OMPT_SUPPORT
+    if (ompt) {
+      __ompt_task_finish(task, resumed_task, ompt_task_switch);
+    }
+#endif
     __kmpc_give_task(task, __kmp_tid_from_gtid(gtid));
     if (KMP_HIDDEN_HELPER_THREAD(gtid))
       __kmp_hidden_helper_worker_thread_signal();
@@ -1711,6 +1715,7 @@ kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
   auto &input_flags = reinterpret_cast<kmp_tasking_flags_t &>(flags);
   // target task is untied defined in the specification
   input_flags.tiedness = TASK_UNTIED;
+  input_flags.target = 1;
 
   if (__kmp_enable_hidden_helper)
     input_flags.hidden_helper = TRUE;
@@ -1942,6 +1947,11 @@ __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task,
 #endif
       __kmp_task_finish<false>(gtid, task, current_task);
   }
+#if OMPT_SUPPORT
+  else if (UNLIKELY(ompt_enabled.enabled && taskdata->td_flags.target)) {
+    __ompt_task_finish(task, current_task, ompt_task_switch);
+  }
+#endif
 
   KA_TRACE(
       30,
@@ -1974,7 +1984,8 @@ kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
     if (ompt_enabled.ompt_callback_task_create) {
       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
           &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame),
-          &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0,
+          &(new_taskdata->ompt_task_info.task_data),
+          TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
           OMPT_GET_RETURN_ADDRESS(0));
     }
   }
@@ -2132,7 +2143,7 @@ kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
             &(parent->ompt_task_info.task_data),
             &(parent->ompt_task_info.frame),
             &(new_taskdata->ompt_task_info.task_data),
-            ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
+            TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
             OMPT_LOAD_RETURN_ADDRESS(gtid));
       }
     } else {
@@ -2193,8 +2204,7 @@ kmp_int32 __kmp_omp_taskloop_task(ident_t *loc_ref, kmp_int32 gtid,
       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
           &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame),
           &(new_taskdata->ompt_task_info.task_data),
-          ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
-          codeptr_ra);
+          TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0, codeptr_ra);
     }
   }
 #endif

openmp/runtime/src/ompt-internal.h

@@ -50,6 +50,10 @@ typedef struct ompt_callbacks_active_s {
        : 0x0) |                                                                \
       ((!(info->td_flags.tiedness)) ? ompt_task_untied : 0x0) |                \
       (info->td_flags.final ? ompt_task_final : 0x0) |                         \
+      (info->td_flags.target                                                   \
+           ? ompt_task_target                                                  \
+           : (info->td_flags.tasktype ? ompt_task_explicit                     \
+                                      : ompt_task_implicit)) |                 \
       (info->td_flags.merged_if0 ? ompt_task_mergeable : 0x0)
 
 typedef struct {

openmp/runtime/src/ompt-specific.cpp

@@ -421,9 +421,7 @@ int __ompt_get_task_info_internal(int ancestor_level, int *type,
       team_info = &team->t.ompt_team_info;
       if (type) {
         if (taskdata->td_parent) {
-          *type = (taskdata->td_flags.tasktype ? ompt_task_explicit
-                                               : ompt_task_implicit) |
-                  TASK_TYPE_DETAILS_FORMAT(taskdata);
+          *type = TASK_TYPE_DETAILS_FORMAT(taskdata);
         } else {
           *type = ompt_task_initial;
         }