src: trace end dependencies

src/common/fxt.h

@@ -242,6 +242,7 @@
 #define _STARPU_FUT_PAPI_TASK_EVENT_VALUE   0x5186
 #define _STARPU_FUT_TASK_EXCLUDE_FROM_DAG   0x5187
 
+#define _STARPU_FUT_TASK_END_DEP	0x5188
 
 /* Predefined FUT key masks */
 #define _STARPU_FUT_KEYMASK_META           FUT_KEYMASK0
@@ -856,6 +857,9 @@ do {									\
 #define _STARPU_TRACE_TASK_DEPS(job_prev, job_succ)	\
 	_STARPU_FUT_FULL_PROBE4STR(_STARPU_FUT_KEYMASK_TASK, _STARPU_FUT_TASK_DEPS, (job_prev)->job_id, (job_succ)->job_id, (job_succ)->task->type, 1, "task")
 
+#define _STARPU_TRACE_TASK_END_DEP(job_prev, job_succ) \
+	FUT_DO_PROBE2(_STARPU_FUT_TASK_END_DEP, (job_prev)->job_id, (job_succ)->job_id)
+
 #define _STARPU_TRACE_GHOST_TASK_DEPS(ghost_prev_id, job_succ)		\
 	_STARPU_FUT_FULL_PROBE4STR(_STARPU_FUT_KEYMASK_TASK_VERBOSE, _STARPU_FUT_TASK_DEPS, (ghost_prev_id), (job_succ)->job_id, (job_succ)->task->type, 1, "ghost")
 
@@ -1364,6 +1368,7 @@ do {										\
 #define _STARPU_TRACE_TAG(tag, job)		do {(void)(tag); (void)(job);} while(0)
 #define _STARPU_TRACE_TAG_DEPS(a, b)		do {(void)(a); (void)(b);} while(0)
 #define _STARPU_TRACE_TASK_DEPS(a, b)		do {(void)(a); (void)(b);} while(0)
+#define _STARPU_TRACE_TASK_END_DEP(a, b)	do {(void)(a); (void)(b);} while(0)
 #define _STARPU_TRACE_GHOST_TASK_DEPS(a, b)	do {(void)(a); (void)(b);} while(0)
 #define _STARPU_TRACE_TASK_EXCLUDE_FROM_DAG(a)	do {(void)(a);} while(0)
 #define _STARPU_TRACE_TASK_NAME(a)		do {(void)(a);} while(0)

src/core/jobs.c

@@ -303,6 +303,16 @@ void _starpu_handle_job_submission(struct _starpu_job *j)
 void starpu_task_end_dep_release(struct starpu_task *t)
 {
 	struct _starpu_job *j = _starpu_get_job_associated_to_task(t);
+
+#ifdef STARPU_USE_FXT
+	struct starpu_task *current = starpu_task_get_current();
+	if (current)
+	{
+		struct _starpu_job *jcurrent = _starpu_get_job_associated_to_task(current);
+		_STARPU_TRACE_TASK_END_DEP(jcurrent, j);
+	}
+#endif
+
 	_starpu_handle_job_termination(j);
 }
 

src/debug/traces/starpu_fxt.c

@@ -114,6 +114,8 @@ struct task_info
 	char *parameters;
 	unsigned int ndeps;
 	unsigned long *dependencies;
+	unsigned int nend_deps;
+	unsigned long *end_dependencies;
 	char **dep_labels;
 	unsigned long ndata;
 	struct data_parameter_info *data;
@@ -155,6 +157,8 @@ static struct task_info *get_task(unsigned long job_id, int mpi_rank)
 		task->parameters = NULL;
 		task->ndeps = 0;
 		task->dependencies = NULL;
+		task->nend_deps = 0;
+		task->end_dependencies = NULL;
 		task->dep_labels = NULL;
 		task->ndata = 0;
 		task->data = NULL;
@@ -265,6 +269,14 @@ static void task_dump(struct task_info *task, struct starpu_fxt_options *options
 		fprintf(tasks_file, "\n");
 	}
 	fprintf(tasks_file, "MPIRank: %d\n", task->mpi_rank);
+	if (task->nend_deps)
+	{
+		fprintf(tasks_file, "EndDependencies: ");
+		unsigned int j=0;
+		for(j=0 ; j<task->nend_deps-1 ; j++)
+			fprintf(tasks_file, "%lu, ", task->end_dependencies[j]);
+		fprintf(tasks_file, "%lu ", task->end_dependencies[task->nend_deps-1]);
+	}
 	fprintf(tasks_file, "\n");
 
 out:
@@ -2736,6 +2748,36 @@ static void handle_task_deps(struct fxt_ev_64 *ev, struct starpu_fxt_options *op
 	}
 }
 
+static void handle_task_end_dep(struct fxt_ev_64 *ev, struct starpu_fxt_options *options)
+{
+	unsigned long dep_prev = ev->param[0];
+	unsigned long dep_succ = ev->param[1];
+
+	struct task_info *task = get_task(dep_succ, options->file_rank);
+	unsigned alloc = 0;
+
+	if (task->nend_deps == 0)
+		/* Start with 8=2^3, should be plenty in most cases */
+		alloc = 8;
+	else if (task->nend_deps >= 8)
+	{
+		/* Allocate dependencies array by powers of two */
+		if (! ((task->nend_deps - 1) & task->nend_deps)) /* Is task->ndeps a power of two? */
+		{
+			/* We have filled the previous power of two, get another one */
+			alloc = task->nend_deps * 2;
+		}
+	}
+	if (alloc)
+	{
+		_STARPU_REALLOC(task->end_dependencies, sizeof(*task->end_dependencies) * alloc);
+	}
+	task->end_dependencies[task->nend_deps++] = dep_prev;
+
+	if (!task->exclude_from_dag && show_task(task, options))
+		_starpu_fxt_dag_add_task_end_dep(options->file_prefix, dep_succ, dep_prev);
+}
+
 static void handle_task_submit(struct fxt_ev_64 *ev, struct starpu_fxt_options *options)
 {
 	unsigned long job_id = ev->param[0];
@@ -3514,6 +3556,10 @@ void _starpu_fxt_parse_new_file(char *filename_in, struct starpu_fxt_options *op
 				handle_task_deps(&ev, options);
 				break;
 
+			case _STARPU_FUT_TASK_END_DEP:
+				handle_task_end_dep(&ev, options);
+				break;
+
 			case _STARPU_FUT_TASK_SUBMIT:
 				handle_task_submit(&ev, options);
 				break;

src/debug/traces/starpu_fxt.h

@@ -51,6 +51,7 @@ void _starpu_fxt_dag_add_tag(const char *prefix, uint64_t tag, unsigned long job
 void _starpu_fxt_dag_add_tag_deps(const char *prefix, uint64_t child, uint64_t father, const char *label);
 void _starpu_fxt_dag_set_tag_done(const char *prefix, uint64_t tag, const char *color);
 void _starpu_fxt_dag_add_task_deps(const char *prefix, unsigned long dep_prev, unsigned long dep_succ, const char *label);
+void _starpu_fxt_dag_add_task_end_dep(const char *prefix, unsigned long prev, unsigned long succ);
 void _starpu_fxt_dag_set_task_name(const char *prefix, unsigned long job_id, const char *label, const char *color);
 void _starpu_fxt_dag_set_task_line(const char *prefix, unsigned long job_id, const char *file, int line);
 void _starpu_fxt_dag_add_send(int src, unsigned long dep_prev, unsigned long tag, unsigned long id);

src/debug/traces/starpu_fxt_dag.c

@@ -104,6 +104,12 @@ void _starpu_fxt_dag_set_tag_done(const char *prefix, uint64_t tag, const char *
 			prefix, (unsigned long long)tag, color);
 }
 
+void _starpu_fxt_dag_add_task_end_dep(const char *prefix, unsigned long prev, unsigned long succ)
+{
+	if (out_file)
+		fprintf(out_file, "\t \"task_%s%lu\" [ end_dep=\"%lu\"]\n", prefix, prev, succ);
+}
+
 void _starpu_fxt_dag_set_task_name(const char *prefix, unsigned long job_id, const char *label, const char *color)
 {
 	if (out_file)