diff --git a/runtime/src/kaapi_trace.h b/runtime/src/kaapi_trace.h
index 263afe37a7f49e2b09bcaa63970e62a47ef1b56a..f2db61a3ac8d08ea479ba41d2e62553917d316c7 100644
--- a/runtime/src/kaapi_trace.h
+++ b/runtime/src/kaapi_trace.h
@@ -242,8 +242,8 @@ typedef struct kaapi_named_perfctr {
#define KAAPI_EVT_TASK_END 3 /* end execution of tasks, d0: task, d1: numaid */
#define KAAPI_EVT_TASK_SUCC 4 /* T0 has successor T1 */
#define KAAPI_EVT_TASK_ACCESS 5 /* d0: task, d1: mode, d2: pointer */
-#define KAAPI_EVT_COMP_DAG_BEG 6 /* begin of computing the dag */
-#define KAAPI_EVT_COMP_DAG_END 7 /* end of computing the dag */
+#define KAAPI_EVT_COMP_DAG 6 /* computing the dag, i0[0]=1 iff beg, =0 iff else. d1: key */
+/*#define KAAPI_EVT_FREE0 7*/
#define KAAPI_EVT_UNDEFINED_0 8
#define KAAPI_EVT_UNDEFINED_1 9
#define KAAPI_EVT_SCHED_IDLE_BEG 10 /* begin when k-processor begins to be idle and try to steal */
@@ -268,21 +268,21 @@ typedef struct kaapi_named_perfctr {
#define KAAPI_EVT_OFFLOAD_DTOD_END 29
#define KAAPI_EVT_OFFLOAD_KERNEL_BEG 30
#define KAAPI_EVT_OFFLOAD_KERNEL_END 31
-#define KAAPI_EVT_PARALLEL_BEG 32
-#define KAAPI_EVT_PARALLEL_END 33
-#define KAAPI_EVT_TASKWAIT_BEG 34
-#define KAAPI_EVT_TASKWAIT_END 35
-#define KAAPI_EVT_TASKGROUP_BEG 36
-#define KAAPI_EVT_TASKGROUP_END 37
+#define KAAPI_EVT_PARALLEL 32 /* i0[] = 1 iff beg, = 0 iff end, d1: parallel_id */
+/*#define KAAPI_EVT_FREE0 33*/
+#define KAAPI_EVT_TASKWAIT 34 /* i0[] = 1 iff beg, = 0 iff end, d1: task_id */
+/*#define KAAPI_EVT_FREE0 35*/
+#define KAAPI_EVT_TASKGROUP 36
+/*#define KAAPI_EVT_FREE0 37*/
#define KAAPI_EVT_PERFCOUNTER 38 /* format <perf id (0, 1, 2..)>, <value> */
#define KAAPI_EVT_TASK_PERFCOUNTER 39 /* d0=task; d1.i8[0..2]: perf counter id; d2, d3: values */
/* several KAAPI_EVT_TASK_PERFCOUNTER may follow KAAPI_EVT_TASK_END */
#define KAAPI_EVT_LOCK_BEG 40
#define KAAPI_EVT_LOCK_END 41
-#define KAAPI_EVT_YIELD_BEG 42
-#define KAAPI_EVT_YIELD_END 43
-#define KAAPI_EVT_BARRIER_BEG 44 /* d0: running task */
-#define KAAPI_EVT_BARRIER_END 45 /* d0: running task */
+#define KAAPI_EVT_YIELD 42 /* i0[] = 1 iff beg, = 0 iff end */
+/*#define KAAPI_EVT_FREE0 43*/
+#define KAAPI_EVT_BARRIER 44 /* d0: running task */
+/*#define KAAPI_EVT_FREE0 45*/
#define KAAPI_EVT_TASK_STEAL 46 /* d0: executed task, d1: original task */
#define KAAPI_EVT_LOOP_BEGIN 47 /* d0: workshare id, d1: sched type, d2: iteration count */
#define KAAPI_EVT_LOOP_END 48 /* d0: workshare id */
@@ -321,28 +321,23 @@ typedef uint64_t kaapi_event_mask_type_t;
| KAAPI_EVT_MASK(KAAPI_EVT_TASK_ACCESS) \
| KAAPI_EVT_MASK(KAAPI_EVT_TASK_END) \
| KAAPI_EVT_MASK(KAAPI_EVT_TASK_STEAL) \
- | KAAPI_EVT_MASK(KAAPI_EVT_COMP_DAG_BEG) \
- | KAAPI_EVT_MASK(KAAPI_EVT_COMP_DAG_END) \
+ | KAAPI_EVT_MASK(KAAPI_EVT_COMP_DAG) \
| KAAPI_EVT_MASK(KAAPI_EVT_OFFLOAD_KERNEL_BEG) \
| KAAPI_EVT_MASK(KAAPI_EVT_OFFLOAD_KERNEL_END) \
)
#define KAAPI_EVT_MASK_OMP \
- ( KAAPI_EVT_MASK(KAAPI_EVT_PARALLEL_BEG) \
- | KAAPI_EVT_MASK(KAAPI_EVT_PARALLEL_END) \
- | KAAPI_EVT_MASK(KAAPI_EVT_BARRIER_BEG) \
- | KAAPI_EVT_MASK(KAAPI_EVT_BARRIER_END) \
- | KAAPI_EVT_MASK(KAAPI_EVT_YIELD_BEG) \
- | KAAPI_EVT_MASK(KAAPI_EVT_YIELD_END) \
- | KAAPI_EVT_MASK(KAAPI_EVT_TASKWAIT_BEG) \
- | KAAPI_EVT_MASK(KAAPI_EVT_TASKWAIT_END) \
- | KAAPI_EVT_MASK(KAAPI_EVT_TASKGROUP_BEG) \
- | KAAPI_EVT_MASK(KAAPI_EVT_TASKGROUP_END) \
+ ( KAAPI_EVT_MASK(KAAPI_EVT_PARALLEL) \
+ | KAAPI_EVT_MASK(KAAPI_EVT_BARRIER) \
+ | KAAPI_EVT_MASK(KAAPI_EVT_YIELD) \
+ | KAAPI_EVT_MASK(KAAPI_EVT_TASKWAIT) \
+ | KAAPI_EVT_MASK(KAAPI_EVT_TASKGROUP) \
| KAAPI_EVT_MASK(KAAPI_EVT_LOOP_BEGIN) \
| KAAPI_EVT_MASK(KAAPI_EVT_LOOP_END) \
| KAAPI_EVT_MASK(KAAPI_EVT_LOOP_NEXT) \
| KAAPI_EVT_MASK(KAAPI_EVT_LOOP_MDATA) \
| KAAPI_EVT_MASK(KAAPI_EVT_TASK_ATTR) \
+ | KAAPI_EVT_MASK(KAAPI_EVT_COMP_DAG) \
)
#define KAAPI_EVT_MASK_SCHED \
diff --git a/runtime/src/kaapi_trace_lib.c b/runtime/src/kaapi_trace_lib.c
index 5a1af882c970bf7ccb094dc00422af1a89c2036a..5c5945c4b19fb6cda907e8913375da6747e240c7 100644
--- a/runtime/src/kaapi_trace_lib.c
+++ b/runtime/src/kaapi_trace_lib.c
@@ -1443,7 +1443,7 @@ void kaapi_tracelib_team_start(
int parallel_id
)
{
- KAAPI_EVENT_PUSH1( kproc, 0, KAAPI_EVT_PARALLEL_BEG, parallel_id );
+ KAAPI_EVENT_PUSH2( kproc, 0, KAAPI_EVT_PARALLEL, 1, parallel_id );
//Here information is correct
//printf("%i/%i:: parallel on numa:%i <-> %i\n", kproc->kid, sched_getcpu(), kproc->numaid, __kmp_cpu2node(sched_getcpu()));
kaapi_tracelib_thread_read(kproc, kproc->perfset, &team->perf0[0]);
@@ -1461,7 +1461,7 @@ void kaapi_tracelib_team_stop(
int parallel_id
)
{
- KAAPI_EVENT_PUSH1( kproc, 0, KAAPI_EVT_PARALLEL_END, parallel_id );
+ KAAPI_EVENT_PUSH2( kproc, 0, KAAPI_EVT_PARALLEL, 0, parallel_id );
kaapi_perf_counter_t tmp[ kaapi_tracelib_thread_idsetsize(kproc) ];
kaapi_perf_counter_t* readbuff;
if (parent)
@@ -1832,7 +1832,7 @@ void kaapi_tracelib_barrier_begin(
uint64_t task
)
{
- KAAPI_EVENT_PUSH1(kproc, 0, KAAPI_EVT_BARRIER_BEG, task );
+ KAAPI_EVENT_PUSH2(kproc, 0, KAAPI_EVT_BARRIER, 1, task );
}
@@ -1843,7 +1843,7 @@ void kaapi_tracelib_barrier_end(
uint64_t task
)
{
- KAAPI_EVENT_PUSH1(kproc, 0, KAAPI_EVT_BARRIER_END, task );
+ KAAPI_EVENT_PUSH2(kproc, 0, KAAPI_EVT_BARRIER, 0, task );
}
diff --git a/runtime/src/kaapi_trace_simulator.cpp b/runtime/src/kaapi_trace_simulator.cpp
index f5b94d16f427cd58a67d20e75e7f092a220347fb..b7c9563eb681e556ed0118a2af35eff321e74f26 100644
--- a/runtime/src/kaapi_trace_simulator.cpp
+++ b/runtime/src/kaapi_trace_simulator.cpp
@@ -724,8 +724,7 @@ static void processor_simulate_event(
++proc->active;
break;
- case KAAPI_EVT_COMP_DAG_BEG:
- case KAAPI_EVT_COMP_DAG_END:
+ case KAAPI_EVT_COMP_DAG:
break;
/* processing request */
@@ -743,16 +742,12 @@ static void processor_simulate_event(
proc->perfctr[KAAPI_EVENT_DATA(event,0,i)] = event->u.data[1].u;
break;
- case KAAPI_EVT_PARALLEL_BEG :
- case KAAPI_EVT_PARALLEL_END :
- case KAAPI_EVT_TASKWAIT_BEG :
- case KAAPI_EVT_TASKWAIT_END :
- case KAAPI_EVT_TASKGROUP_BEG:
- case KAAPI_EVT_TASKGROUP_END:
+ case KAAPI_EVT_PARALLEL :
+ case KAAPI_EVT_TASKWAIT :
+ case KAAPI_EVT_TASKGROUP:
break;
- case KAAPI_EVT_BARRIER_BEG:
- case KAAPI_EVT_BARRIER_END:
+ case KAAPI_EVT_BARRIER:
case KAAPI_EVT_LOOP_BEGIN:
case KAAPI_EVT_LOOP_END :
case KAAPI_EVT_LOOP_NEXT :
diff --git a/runtime/src/katracereader.cpp b/runtime/src/katracereader.cpp
index 57bf96dd2a8a99288287361cb1639a3a06b9cc4b..967dc9569f8c1fdc3db548ae47cbeadd69b09d33 100644
--- a/runtime/src/katracereader.cpp
+++ b/runtime/src/katracereader.cpp
@@ -323,11 +323,9 @@ static void callback_print_event(
break;
/* unroll graph for static schedule */
- case KAAPI_EVT_COMP_DAG_BEG:
- break;
-
- case KAAPI_EVT_COMP_DAG_END:
- std::cout << ", delay:" << KAAPI_EVENT_DATA(event,0,i);
+ case KAAPI_EVT_COMP_DAG:
+ std::cout << (KAAPI_EVENT_DATA(event,0,i) == 1 ? "BEGIN":"END")
+ << ", key:" << KAAPI_EVENT_DATA(event,1,u);
break;
/* idle = steal state */
@@ -389,25 +387,21 @@ static void callback_print_event(
<< ">, value: " << KAAPI_EVENT_DATA(event,1,u);
break;
- case KAAPI_EVT_PARALLEL_BEG :
- std::cout << " parallel_id:" << KAAPI_EVENT_DATA(event,0,i);
- break;
- case KAAPI_EVT_PARALLEL_END :
- std::cout << " parallel_id:" << KAAPI_EVENT_DATA(event,0,i);
+ case KAAPI_EVT_PARALLEL :
+ std::cout << (KAAPI_EVENT_DATA(event,0,i) == 1 ? "BEGIN":"END")
+ << ", parallel_id:" << KAAPI_EVENT_DATA(event,1,u);
break;
- case KAAPI_EVT_TASKWAIT_BEG :
- case KAAPI_EVT_TASKWAIT_END :
- case KAAPI_EVT_TASKGROUP_BEG:
- case KAAPI_EVT_TASKGROUP_END:
+ case KAAPI_EVT_TASKWAIT :
+ case KAAPI_EVT_TASKGROUP:
case KAAPI_EVT_LOCK_BEG:
case KAAPI_EVT_LOCK_END:
break;
- case KAAPI_EVT_BARRIER_BEG:
- case KAAPI_EVT_BARRIER_END:
- std::cout << " task: " << KAAPI_EVENT_DATA(event,0,u);
+ case KAAPI_EVT_BARRIER:
+ std::cout << (KAAPI_EVENT_DATA(event,0,i) == 1 ? " BEGIN" : "END")
+ << ", task: " << KAAPI_EVENT_DATA(event,1,u);
break;
case KAAPI_EVT_LOOP_BEGIN:
@@ -1036,32 +1030,32 @@ static void callback_display_paje_event(
break;
/* unroll graph for static schedule */
- case KAAPI_EVT_COMP_DAG_BEG:
- kaapi_trace_poti_PushState(d0, name, "STATE", "db");
- break;
- case KAAPI_EVT_COMP_DAG_END:
- kaapi_trace_poti_PopState (d0, name, "STATE");
+ case KAAPI_EVT_COMP_DAG:
+ if (KAAPI_EVENT_DATA(event,0,i) == 1)
+ kaapi_trace_poti_PushState(d0, name, "STATE", "db");
+ else
+ kaapi_trace_poti_PopState (d0, name, "STATE");
break;
- case KAAPI_EVT_PARALLEL_BEG :
- kaapi_trace_poti_PushState(d0, name, "STATE", "pi");
- break;
- case KAAPI_EVT_PARALLEL_END :
- kaapi_trace_poti_PopState (d0, name, "STATE");
+ case KAAPI_EVT_PARALLEL:
+ if (KAAPI_EVENT_DATA(event,0,i) == 1)
+ kaapi_trace_poti_PushState(d0, name, "STATE", "pi");
+ else
+ kaapi_trace_poti_PopState (d0, name, "STATE");
break;
- case KAAPI_EVT_TASKWAIT_BEG:
- kaapi_trace_poti_PushState(d0, name, "STATE", "ps");
- break;
- case KAAPI_EVT_TASKWAIT_END:
- kaapi_trace_poti_PopState (d0, name, "STATE");
+ case KAAPI_EVT_TASKWAIT:
+ if (KAAPI_EVENT_DATA(event,0,i) == 1)
+ kaapi_trace_poti_PushState(d0, name, "STATE", "ps");
+ else
+ kaapi_trace_poti_PopState (d0, name, "STATE");
break;
- case KAAPI_EVT_TASKGROUP_BEG :
- kaapi_trace_poti_PushState(d0, name, "STATE", "pf");
- break;
- case KAAPI_EVT_TASKGROUP_END :
- kaapi_trace_poti_PopState (d0, name, "STATE");
+ case KAAPI_EVT_TASKGROUP :
+ if (KAAPI_EVENT_DATA(event,0,i) == 1)
+ kaapi_trace_poti_PushState(d0, name, "STATE", "pf");
+ else
+ kaapi_trace_poti_PopState (d0, name, "STATE");
break;
case KAAPI_EVT_LOCK_BEG :
@@ -1982,60 +1976,57 @@ printf("pred[id:%i/%p] = {id:%i}/%p\n", succ->id, (void*)succ->addr, task->id, (
break;
/* unroll graph for static schedule */
- case KAAPI_EVT_COMP_DAG_BEG:
- case KAAPI_EVT_COMP_DAG_END:
+ case KAAPI_EVT_COMP_DAG:
break;
- case KAAPI_EVT_TASKWAIT_BEG :
- case KAAPI_EVT_TASKWAIT_END :
- case KAAPI_EVT_TASKGROUP_BEG:
- case KAAPI_EVT_TASKGROUP_END:
+ case KAAPI_EVT_TASKWAIT :
+ case KAAPI_EVT_TASKGROUP:
break;
- case KAAPI_EVT_PARALLEL_BEG :
+ case KAAPI_EVT_PARALLEL :
{
- uintptr_t parallel_id = KAAPI_EVENT_DATA(event,0,i);
- if (rastello_parallel_region.size() <= (size_t)parallel_id)
- {
- rastello_parallel_region.reserve( parallel_id * 2);
- rastello_parallel_region.resize( parallel_id +1 );
+ int begend = KAAPI_EVENT_DATA(event,0,i);
+ uintptr_t parallel_id = KAAPI_EVENT_DATA(event,1,u);
+ if (begend ==1)
+ { /* begin */
+ if (rastello_parallel_region.size() <= (size_t)parallel_id)
+ {
+ rastello_parallel_region.reserve( parallel_id * 2);
+ rastello_parallel_region.resize( parallel_id +1 );
+ }
+ rastello_parallel_region_t* rpr = (rastello_parallel_region_t*)rastello_parallel_region[parallel_id];
+ if (rpr == 0)
+ {
+ /* allocation */
+ rpr = rprf->create(parallel_id);
+ rastello_parallel_region[parallel_id] = rpr;
+ }
+ /* start is the min over all kprocs */
+ if (rpr->start ==0)
+ rpr->start = event->date;
+ else if (rpr->start > event->date)
+ rpr->start = event->date;
+ ++rpr->nproc;
+
+ if (rastello_front_parallel_region.size() <= kid)
+ rastello_front_parallel_region.resize(kid+1);
+ rastello_front_parallel_region[kid] = rpr;
}
- rastello_parallel_region_t* rpr = (rastello_parallel_region_t*)rastello_parallel_region[parallel_id];
- if (rpr == 0)
- {
- /* allocation */
- rpr = rprf->create(parallel_id);
- rastello_parallel_region[parallel_id] = rpr;
+ else
+ { /* end */
+ rastello_parallel_region_t* rpr = (rastello_parallel_region_t*)rastello_parallel_region[parallel_id];
+ /* stop is the max over all kprocs */
+ if (rpr->stop < event->date)
+ rpr->stop = event->date;
+ kaapi_assert( rpr->nproc >= 0);
}
- /* start is the min over all kprocs */
- if (rpr->start ==0)
- rpr->start = event->date;
- else if (rpr->start > event->date)
- rpr->start = event->date;
- ++rpr->nproc;
-
- if (rastello_front_parallel_region.size() <= kid)
- rastello_front_parallel_region.resize(kid+1);
- rastello_front_parallel_region[kid] = rpr;
- } break;
-
- case KAAPI_EVT_PARALLEL_END :
- {
- uintptr_t parallel_id = KAAPI_EVENT_DATA(event,0,i);
- rastello_parallel_region_t* rpr = (rastello_parallel_region_t*)rastello_parallel_region[parallel_id];
- /* stop is the max over all kprocs */
- if (rpr->stop < event->date)
- rpr->stop = event->date;
- kaapi_assert( rpr->nproc >= 0);
} break;
case KAAPI_EVT_LOCK_BEG :
case KAAPI_EVT_LOCK_END :
break;
- case KAAPI_EVT_BARRIER_BEG:
- break;
- case KAAPI_EVT_BARRIER_END:
+ case KAAPI_EVT_BARRIER:
break;
case KAAPI_EVT_LOOP_BEGIN: /* both: a state creation and an event */
diff --git a/runtime/src/kmp.h b/runtime/src/kmp.h
index 259bfe68d616f0bd27be99e54c242441db8b9f8f..628cad49230a60f860ff370a1089e27d002eaf72 100644
--- a/runtime/src/kmp.h
+++ b/runtime/src/kmp.h
@@ -2353,7 +2353,7 @@ struct kmp_taskdata { /* aligned during dynamic allocation */
kmp_int32 td_untied_count; /* untied task active parts counter */
ident_t *td_ident; /* task identifier */
#if OMPT_SUPPORT
- intptr_t td_keys[2];
+ intptr_t td_keys[2]; /* user level values */
#endif
#if LIBOMP_USE_AFFINITY
kmp_uint8 td_aff_kind; // Affinity kind & tag & strict for the task