when we verify if there is at least a worker able to execute a certain task check out the list of workers of a context not all of the workers of the machine

(there is this weird verification in starpu_worker_exists "if (!task->cl->can_execute)" that I am not sure what is it doing over here, sam & nat could you please check this out? )

examples/sched_ctx/sched_ctx.c

@@ -13,7 +13,7 @@
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  *
  * See the GNU Lesser General Public License in COPYING.LGPL for more details.
- */
+OB */
 
 #include <starpu.h>
 
@@ -26,16 +26,21 @@
 int tasks_executed = 0;
 starpu_pthread_mutex_t mut;
 
-static void sched_ctx_func(void *descr[] STARPU_ATTRIBUTE_UNUSED, void *arg STARPU_ATTRIBUTE_UNUSED)
+static void sched_ctx_cpu_func(void *descr[] STARPU_ATTRIBUTE_UNUSED, void *arg STARPU_ATTRIBUTE_UNUSED)
 {
 	starpu_pthread_mutex_lock(&mut);
 	tasks_executed++;
 	starpu_pthread_mutex_unlock(&mut);
 }
 
-static struct starpu_codelet sched_ctx_codelet =
+static void sched_ctx_cuda_func(void *descr[] STARPU_ATTRIBUTE_UNUSED, void *arg STARPU_ATTRIBUTE_UNUSED)
 {
-	.cpu_funcs = {sched_ctx_func, NULL},
+	
+}
+
+static struct starpu_codelet sched_ctx_codelet1 =
+{
+	.cpu_funcs = {sched_ctx_cpu_func, NULL},
 	.cuda_funcs = {NULL},
 	.opencl_funcs = {NULL},
 	.model = NULL,
@@ -43,6 +48,16 @@ static struct starpu_codelet sched_ctx_codelet =
 	.name = "sched_ctx"
 };
 
+static struct starpu_codelet sched_ctx_codelet2 =
+{
+	.cpu_funcs = {sched_ctx_cpu_func, NULL},
+	.cuda_funcs = {sched_ctx_cuda_func, NULL},
+	.opencl_funcs = {NULL},
+	.model = NULL,
+	.nbuffers = 0,
+	.name = "sched_ctx"
+};
+
 
 int main(int argc, char **argv)
 {
@@ -87,12 +102,11 @@ int main(int argc, char **argv)
 	{
 		struct starpu_task *task = starpu_task_create();
 
-		task->cl = &sched_ctx_codelet;
+		task->cl = &sched_ctx_codelet1;
 		task->cl_arg = NULL;
 
 		/*submit tasks to context*/
 		ret = starpu_task_submit_to_ctx(task,sched_ctx1);
-
 		STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
 	}
 
@@ -102,11 +116,22 @@ int main(int argc, char **argv)
 
 	starpu_sched_ctx_finished_submit(sched_ctx1);
 
+	/* task with no cuda impl submitted to a ctx with gpus only */
+	struct starpu_task *task2 = starpu_task_create();
+	
+	task2->cl = &sched_ctx_codelet1;
+	task2->cl_arg = NULL;
+	
+	/*submit tasks to context*/
+	ret = starpu_task_submit_to_ctx(task2,sched_ctx2);
+	STARPU_ASSERT_MSG(ret == -ENODEV, "submit task should ret enodev when the ctx does not have the PUs needed by the task");	
+//	STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
+
 	for (i = 0; i < ntasks/2; i++)
 	{
 		struct starpu_task *task = starpu_task_create();
 
-		task->cl = &sched_ctx_codelet;
+		task->cl = &sched_ctx_codelet2;
 		task->cl_arg = NULL;
 
 		ret = starpu_task_submit_to_ctx(task,sched_ctx2);
@@ -121,7 +146,7 @@ int main(int argc, char **argv)
 
 	starpu_sched_ctx_delete(sched_ctx1);
 	starpu_sched_ctx_delete(sched_ctx2);
-	printf("tasks executed %d out of %d\n", tasks_executed, ntasks);
+	printf("tasks executed %d out of %d\n", tasks_executed, ntasks/2);
 	starpu_shutdown();
 
 	return 0;

src/core/workers.c

@@ -95,12 +95,17 @@ static uint32_t _starpu_worker_exists_and_can_execute(struct starpu_task *task,
 						      enum starpu_worker_archtype arch)
 {
 	int i;
-	int nworkers = starpu_worker_get_count();
-
 	_starpu_codelet_check_deprecated_fields(task->cl);
+	struct _starpu_sched_ctx *sched_ctx = _starpu_get_sched_ctx_struct(task->sched_ctx);
+	struct starpu_worker_collection *workers = sched_ctx->workers;
 
-	for (i = 0; i < nworkers; i++)
-	{
+        struct starpu_sched_ctx_iterator it;
+        if(workers->init_iterator)
+                workers->init_iterator(workers, &it);
+
+        while(workers->has_next(workers, &it))
+        {
+                i = workers->get_next(workers, &it);
 		if (starpu_worker_get_type(i) != arch)
 			continue;
 
@@ -141,7 +146,10 @@ static uint32_t _starpu_worker_exists_and_can_execute(struct starpu_task *task,
 			if (!test_implementation)
 				break;
 
-			if (task->cl->can_execute(i, task, impl))
+			if (task->cl->can_execute && task->cl->can_execute(i, task, impl))
+				return 1;
+
+			if(test_implementation)
 				return 1;
 		}
 	}
@@ -155,11 +163,18 @@ uint32_t _starpu_worker_exists(struct starpu_task *task)
 {
 	_starpu_codelet_check_deprecated_fields(task->cl);
 
-	if (!(task->cl->where & config.worker_mask))
-		return 0;
-
-	if (!task->cl->can_execute)
-		return 1;
+	/* if the task belongs to the init context we can
+	   check out all the worker mask of the machine
+	   if not we should iterate on the workers of the ctx
+	   and verify if it exists a worker able to exec the task */
+	if(task->sched_ctx == 0)
+	{
+		if (!(task->cl->where & config.worker_mask))
+			return 0;
+		
+		if (!task->cl->can_execute)
+			return 1;
+	}
 
 #if defined(STARPU_USE_CPU) || defined(STARPU_SIMGRID)
 	if ((task->cl->where & STARPU_CPU) &&
@@ -186,6 +201,7 @@ uint32_t _starpu_worker_exists(struct starpu_task *task)
 	    _starpu_worker_exists_and_can_execute(task, STARPU_SCC_WORKER))
 		return 1;
 #endif
+
 	return 0;
 }
 

src/sched_policies/deque_modeling_policy_data_aware.c

@@ -547,10 +547,6 @@ static void compute_all_performance_predictions(struct starpu_task *task,
 	{
 		worker = workers->get_next(workers, &it);
 
-		if (worker >= nworkers)
-			/* This is a just-added worker, discard it */
-			continue;
-
 		struct _starpu_fifo_taskq *fifo = dt->queue_array[worker];
 		struct starpu_perfmodel_arch* perf_arch = starpu_worker_get_perf_archtype(worker);
 		unsigned memory_node = starpu_worker_get_memory_node(worker);