diff --git a/runtime/src/include/30/ompt.h.var b/runtime/src/include/30/ompt.h.var
index 5e17697a986f72887f7aa951d522830ba4eeba6f..b1b4fc380e22f52959280ab6392e90e2a61071af 100644
--- a/runtime/src/include/30/ompt.h.var
+++ b/runtime/src/include/30/ompt.h.var
@@ -177,8 +177,8 @@
\
macro (ompt_event_thread_state_begin, ompt_threadstate_callback_t, 72) /* begin state */ \
macro (ompt_event_thread_state_end, ompt_threadstate_callback_t, 73) /* end state */ \
- macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */
-
+ macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */ \
+ macro (ompt_event_thread_schedinfo, ompt_threadstate_callback_t, 75) /* set sched info */
/*****************************************************************************
diff --git a/runtime/src/include/40/ompt.h.var b/runtime/src/include/40/ompt.h.var
index f4d09d93de9a59a35069bec91b3d7b16de555be7..bb5492b099f8120443351fa3d86e2e1b27d92138 100644
--- a/runtime/src/include/40/ompt.h.var
+++ b/runtime/src/include/40/ompt.h.var
@@ -180,9 +180,8 @@
\
macro (ompt_event_thread_state_begin, ompt_threadstate_callback_t, 72) /* begin state */ \
macro (ompt_event_thread_state_end, ompt_threadstate_callback_t, 73) /* end state */ \
- macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */
-
-
+ macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */ \
+ macro (ompt_event_thread_schedinfo, ompt_threadstate_callback_t, 75) /* set sched info */
/*****************************************************************************
* data types
*****************************************************************************/
diff --git a/runtime/src/include/45/ompt.h.var b/runtime/src/include/45/ompt.h.var
index 59635af27cbb38f3e5b7fc39c22dfbe60d4b34a1..488f0888d72fdce99783672a83eef27f81bff89d 100644
--- a/runtime/src/include/45/ompt.h.var
+++ b/runtime/src/include/45/ompt.h.var
@@ -180,7 +180,8 @@
\
macro (ompt_event_thread_state_begin, ompt_threadstate_callback_t, 72) /* begin state */ \
macro (ompt_event_thread_state_end, ompt_threadstate_callback_t, 73) /* end state */ \
- macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */
+ macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */ \
+ macro (ompt_event_thread_schedinfo, ompt_threadstate_callback_t, 75) /* set sched info */
/*****************************************************************************
* data types
diff --git a/runtime/src/include/50/ompt.h.var b/runtime/src/include/50/ompt.h.var
index dfc55fe4b3db6b39609b53d9425740b4b53b5560..f26ec5cac669053f22087ad437c2b3ec037b1913 100644
--- a/runtime/src/include/50/ompt.h.var
+++ b/runtime/src/include/50/ompt.h.var
@@ -180,7 +180,8 @@
\
macro (ompt_event_thread_state_begin, ompt_threadstate_callback_t, 72) /* begin state */ \
macro (ompt_event_thread_state_end, ompt_threadstate_callback_t, 73) /* end state */ \
- macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */
+ macro (ompt_event_thread_steal, ompt_thread_steal_callback_t, 74) /* steal event */ \
+ macro (ompt_event_thread_schedinfo, ompt_threadstate_callback_t, 75) /* set sched info */
/*****************************************************************************
* data types
diff --git a/runtime/src/kaapi_ompt.c b/runtime/src/kaapi_ompt.c
index 1311e19418de1b95e677435b2b6b022a8d407376..e31df12044e8d2e5d1756ada532bd5a5c795a3da 100644
--- a/runtime/src/kaapi_ompt.c
+++ b/runtime/src/kaapi_ompt.c
@@ -827,6 +827,23 @@ on_ompt_event_task_dependence_pair(
#endif
}
+static void
+on_ompt_event_thread_schedinfo(
+ ompt_thread_id_t thread_id,
+ uint64_t state,
+ uint32_t cpu,
+ uint32_t node
+)
+{
+#if LOG
+ printf("%" PRIu64 ": ompt_event_thread_schedinfo: state id=%" PRIu64 ", cpu=%i, node=%i\n", thread_id, state, cpu, node);
+#endif
+#if USE_KAAPI
+ kaapi_ompt_thread_info_t* koti = &__kaapi_oth_info[thread_id];
+ kaapi_tracelib_thread_state( koti->kproc, 0, cpu, node, state );
+#endif
+}
+
static void
on_ompt_event_thread_state_begin(
ompt_thread_id_t thread_id,
@@ -939,6 +956,7 @@ void kaapi_ompt_initialize(
//ompt_set_callback(ompt_event_idle_begin, (ompt_callback_t) &on_ompt_event_idle_begin);
//ompt_set_callback(ompt_event_idle_end, (ompt_callback_t) &on_ompt_event_idle_end);
+ ompt_set_callback(ompt_event_thread_schedinfo, (ompt_callback_t) &on_ompt_event_thread_schedinfo);
ompt_set_callback(ompt_event_thread_state_begin, (ompt_callback_t) &on_ompt_event_thread_state_begin);
ompt_set_callback(ompt_event_thread_state_end, (ompt_callback_t) &on_ompt_event_thread_state_end);
ompt_set_callback(ompt_event_thread_steal, (ompt_callback_t) &on_ompt_event_thread_steal);
diff --git a/runtime/src/kaapi_trace_lib.c b/runtime/src/kaapi_trace_lib.c
index cc41bf2bba2068303e2172e5bbb117df2c67459d..d429d8d3ce6902db3cc902438164b8f4d1894923 100644
--- a/runtime/src/kaapi_trace_lib.c
+++ b/runtime/src/kaapi_trace_lib.c
@@ -650,7 +650,8 @@ int kaapi_tracelib_init(
if (getenv("KAAPI_RECORD_MASK") !=0)
{
/* actual grammar:
- eventno[,eventno]*
+ descr[,descr]*
+ descr -> groupname |Â eventno
eventno is an integer less than 2^sizeof(kaapi_event_mask_type_t)
grammar must be more complex using predefined set
*/
@@ -1234,7 +1235,7 @@ void kaapi_tracelib_thread_fini(
*/
void kaapi_tracelib_thread_state (
kaapi_tracelib_thread_t* kproc,
- uint8_t begend, /* 0==begin, 1==end */
+ uint8_t begend, /* 0==begin, 1==end, 2==running */
uint32_t cpu,
uint32_t node,
uint64_t state
@@ -1247,7 +1248,7 @@ void kaapi_tracelib_thread_state (
evt->u.s.d1.i32[0] = cpu;
evt->u.s.d1.i32[1] = node;
evt->u.s.d2.i64[0] = state;
- KAAPI_EVENT_PUSH(kproc,0, KAAPI_EVT_THREAD_STATE);
+ KAAPI_EVENT_PUSH(kproc, 0, KAAPI_EVT_THREAD_STATE);
}
}
diff --git a/runtime/src/kmp_runtime.cpp b/runtime/src/kmp_runtime.cpp
index 35f8c77b95d6f53de7b1894e94b1b00ee95770d2..9ca49e128a34c84ec3f99b73e93cca8c7c4d81ef 100644
--- a/runtime/src/kmp_runtime.cpp
+++ b/runtime/src/kmp_runtime.cpp
@@ -1376,6 +1376,19 @@ int __kmp_fork_call(ident_t *loc, int gtid,
#if KMP_NESTED_HOT_TEAMS
kmp_hot_team_ptr_t **p_hot_teams;
#endif
+#if OMPT_SUPPORT
+ ompt_thread_id_t tid = GTID_TO_OMPT_THREAD_ID(gtid);
+#if defined(__linux__)
+ kmp_int32 cpu = sched_getcpu();
+#else
+ kmp_int32 cpu = tid;
+#endif
+#if LIBOMP_USE_AFFINITY
+ kmp_int32 node = __kmp_cpu2node(cpu);
+#else
+ kmp_int32 node = 0;
+#endif
+#endif // OMPT_SUPPORT
{ // KMP_TIME_BLOCK
KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
@@ -1502,9 +1515,15 @@ int __kmp_fork_call(ident_t *loc, int gtid,
/* OMPT implicit task begin */
my_task_id = lw_taskteam.ompt_task_info.task_id;
my_parallel_id = parent_team->t.ompt_team_info.parallel_id;
+ ompt_thread_id_t tid = GTID_TO_OMPT_THREAD_ID(gtid);
if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
- GTID_TO_OMPT_THREAD_ID(gtid), my_parallel_id, my_task_id);
+ tid, my_parallel_id, my_task_id);
+ }
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
}
#endif
@@ -1530,11 +1549,16 @@ int __kmp_fork_call(ident_t *loc, int gtid,
*exit_runtime_p = NULL;
if (ompt_enabled) {
#if OMPT_TRACE
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
- GTID_TO_OMPT_THREAD_ID(gtid), ompt_parallel_id, ompt_task_id);
+ tid, ompt_parallel_id, ompt_task_id);
}
__ompt_lw_taskteam_unlink(master_th);
@@ -1544,7 +1568,7 @@ int __kmp_fork_call(ident_t *loc, int gtid,
if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
- GTID_TO_OMPT_THREAD_ID(gtid), ompt_parallel_id,
+ tid, ompt_parallel_id,
ompt_task_id, OMPT_INVOKER(call_context));
}
master_th->th.ompt_thread_info.state = ompt_state_overhead;
@@ -1724,7 +1748,12 @@ int __kmp_fork_call(ident_t *loc, int gtid,
my_task_id = lw_taskteam.ompt_task_info.task_id;
if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
- GTID_TO_OMPT_THREAD_ID(gtid), ompt_parallel_id, my_task_id);
+ tid, ompt_parallel_id, my_task_id);
+ }
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
}
#endif
@@ -1753,6 +1782,11 @@ int __kmp_fork_call(ident_t *loc, int gtid,
lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
#if OMPT_TRACE
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
GTID_TO_OMPT_THREAD_ID(gtid), ompt_parallel_id, ompt_task_id);
@@ -1837,6 +1871,11 @@ int __kmp_fork_call(ident_t *loc, int gtid,
ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
GTID_TO_OMPT_THREAD_ID(gtid), my_parallel_id, my_task_id);
}
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
#endif
/* OMPT state */
@@ -1863,6 +1902,11 @@ int __kmp_fork_call(ident_t *loc, int gtid,
#if OMPT_TRACE
lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
GTID_TO_OMPT_THREAD_ID(gtid), my_parallel_id, my_task_id);
@@ -5505,6 +5549,18 @@ void *__kmp_launch_thread(kmp_info_t *this_thr) {
}
#if OMPT_SUPPORT
+ ompt_thread_id_t tid = GTID_TO_OMPT_THREAD_ID(gtid);
+#if defined(__linux__)
+ kmp_int32 cpu = sched_getcpu();
+#else
+ kmp_int32 cpu = GTID_TO_OMPT_THREAD_ID(gtid);
+#endif
+#if LIBOMP_USE_AFFINITY
+ kmp_int32 node = __kmp_cpu2node(cpu);
+#else
+ kmp_int32 node = 0;
+#endif
+
if (ompt_enabled) {
this_thr->th.ompt_thread_info.state = ompt_state_overhead;
this_thr->th.ompt_thread_info.wait_id = 0;
@@ -5514,6 +5570,11 @@ void *__kmp_launch_thread(kmp_info_t *this_thr) {
if (ompt_callbacks.ompt_callback(ompt_event_thread_begin)) {
__ompt_thread_begin(ompt_thread_worker, gtid);
}
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
}
#endif
@@ -5567,6 +5628,11 @@ void *__kmp_launch_thread(kmp_info_t *this_thr) {
0 /* loc is missing */
);
}
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
}
#endif
/* we were just woken up, so run our new task */
@@ -5587,7 +5653,6 @@ void *__kmp_launch_thread(kmp_info_t *this_thr) {
task_info->task_id = __ompt_task_id_new(tid);
}
#endif
-
{
KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
@@ -5612,6 +5677,11 @@ void *__kmp_launch_thread(kmp_info_t *this_thr) {
__kmp_join_barrier(gtid);
#if OMPT_SUPPORT && OMPT_TRACE
if (ompt_enabled) {
+ if (ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)) {
+ ompt_callbacks.ompt_callback(ompt_event_thread_schedinfo)(
+ tid, (uint64_t)3, cpu, node
+ );
+ }
if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
// don't access *pteam here: it may have already been freed
// by the master thread behind the barrier (possible race)
diff --git a/runtime/src/kmp_tasking.cpp b/runtime/src/kmp_tasking.cpp
index 54a0b2fae68b684b3e345796322d1f83b26baa9b..823bf4415bb6692a4bfe813dbc753a0a06a74993 100644
--- a/runtime/src/kmp_tasking.cpp
+++ b/runtime/src/kmp_tasking.cpp
@@ -2622,7 +2622,7 @@ static inline int __kmp_execute_tasks_template(
#else
kmp_int32 node = 0;
#endif
- enum { INIT_STATE, IDLE_STATE, ACTIVE_STATE };
+ enum { INIT_STATE=0, IDLE_STATE=1, ACTIVE_STATE=2, UNKNOWN_STATE=3 };
kmp_int32 state = INIT_STATE;
kmp_base_queue_data_t *own_queue = &task_team->tt.tt_task_queues[KMP_LEVEL_THREAD][tid].qd;
diff --git a/runtime/src/ompt-event-specific.h b/runtime/src/ompt-event-specific.h
index 848072145ba782037e6e1ef641faa3b5f8afa189..be168b5b6c84f7170fb8681d57b0fc264666ea97 100644
--- a/runtime/src/ompt-event-specific.h
+++ b/runtime/src/ompt-event-specific.h
@@ -146,6 +146,7 @@
#define ompt_event_thread_state_begin_implemented ompt_event_MAY_ALWAYS_TRACE
#define ompt_event_thread_state_end_implemented ompt_event_MAY_ALWAYS_TRACE
#define ompt_event_thread_steal_implemented ompt_event_MAY_ALWAYS_TRACE
+#define ompt_event_thread_schedinfo_implemented ompt_event_MAY_ALWAYS_TRACE
#if OMP_40_ENABLED
#define ompt_event_task_dependences_implemented ompt_event_MAY_ALWAYS_TRACE