static int ws_push_task(struct starpu_task *task) { unsigned sched_ctx_id = task->sched_ctx; struct _starpu_work_stealing_data *ws = (struct _starpu_work_stealing_data*)starpu_sched_ctx_get_policy_data(sched_ctx_id); struct _starpu_deque_jobq *deque_queue; struct _starpu_job *j = _starpu_get_job_associated_to_task(task); int workerid = starpu_worker_get_id(); unsigned worker = 0; struct starpu_worker_collection *workers = starpu_sched_ctx_get_worker_collection(sched_ctx_id); struct starpu_sched_ctx_iterator it; workers->init_iterator(workers, &it); /* !! C'est ballot de tout locker! */ while(workers->has_next(workers, &it)) { worker = workers->get_next(workers, &it); starpu_pthread_mutex_t *sched_mutex; starpu_pthread_cond_t *sched_cond; starpu_worker_get_sched_condition(worker, &sched_mutex, &sched_cond); STARPU_PTHREAD_MUTEX_LOCK(sched_mutex); } /* If the current thread is not a worker but * the main thread (-1), we find the better one to * put task on its queue */ if (workerid == -1) workerid = select_worker(sched_ctx_id); deque_queue = ws->queue_array[workerid]; #ifdef HAVE_AYUDAME_H if (AYU_event) { intptr_t id = workerid; AYU_event(AYU_ADDTASKTOQUEUE, j->job_id, &id); } #endif _starpu_job_list_push_back(&deque_queue->jobq, j); deque_queue->njobs++; starpu_push_task_end(task); while(workers->has_next(workers, &it)) { worker = workers->get_next(workers, &it); starpu_pthread_mutex_t *sched_mutex; starpu_pthread_cond_t *sched_cond; starpu_worker_get_sched_condition(worker, &sched_mutex, &sched_cond); #ifndef STARPU_NON_BLOCKING_DRIVERS STARPU_PTHREAD_COND_SIGNAL(sched_cond); #endif STARPU_PTHREAD_MUTEX_UNLOCK(sched_mutex); } return 0; }
static int push_task_dummy(struct starpu_task *task) { unsigned sched_ctx_id = task->sched_ctx; struct dummy_sched_data *data = (struct dummy_sched_data*)starpu_sched_ctx_get_policy_data(sched_ctx_id); /* NB: In this simplistic strategy, we assume that the context in which we push task has at least one worker*/ /* lock all workers when pushing tasks on a list where all of them would pop for tasks */ starpu_pthread_mutex_lock(&data->policy_mutex); starpu_task_list_push_front(&data->sched_list, task); starpu_push_task_end(task); starpu_pthread_mutex_unlock(&data->policy_mutex); /*if there are no tasks block */ /* wake people waiting for a task */ unsigned worker = 0; struct starpu_worker_collection *workers = starpu_sched_ctx_get_worker_collection(sched_ctx_id); struct starpu_sched_ctx_iterator it; workers->init_iterator(workers, &it); while(workers->has_next(workers, &it)) { worker = workers->get_next(workers, &it); starpu_pthread_mutex_t *sched_mutex; starpu_pthread_cond_t *sched_cond; starpu_worker_get_sched_condition(worker, &sched_mutex, &sched_cond); starpu_pthread_mutex_lock(sched_mutex); starpu_pthread_cond_signal(sched_cond); starpu_pthread_mutex_unlock(sched_mutex); } return 0; }
int _starpu_push_task_to_workers(struct starpu_task *task) { struct _starpu_sched_ctx *sched_ctx = _starpu_get_sched_ctx_struct(task->sched_ctx); unsigned nworkers = 0; _STARPU_TRACE_JOB_PUSH(task, task->priority > 0); /* if the contexts still does not have workers put the task back to its place in the empty ctx list */ if(!sched_ctx->is_initial_sched) { /*if there are workers in the ctx that are not able to execute tasks we consider the ctx empty */ nworkers = _starpu_nworkers_able_to_execute_task(task, sched_ctx); if (nworkers == 0) { STARPU_PTHREAD_MUTEX_LOCK(&sched_ctx->empty_ctx_mutex); starpu_task_list_push_back(&sched_ctx->empty_ctx_tasks, task); STARPU_PTHREAD_MUTEX_UNLOCK(&sched_ctx->empty_ctx_mutex); #ifdef STARPU_USE_SC_HYPERVISOR if(sched_ctx != NULL && sched_ctx->id != 0 && sched_ctx->perf_counters != NULL && sched_ctx->perf_counters->notify_empty_ctx) { _STARPU_TRACE_HYPERVISOR_BEGIN(); sched_ctx->perf_counters->notify_empty_ctx(sched_ctx->id, task); _STARPU_TRACE_HYPERVISOR_END(); } #endif return -EAGAIN; } } _starpu_profiling_set_task_push_start_time(task); int ret = 0; if (STARPU_UNLIKELY(task->execute_on_a_specific_worker)) { unsigned node = starpu_worker_get_memory_node(task->workerid); if (starpu_get_prefetch_flag()) starpu_prefetch_task_input_on_node(task, node); ret = _starpu_push_task_on_specific_worker(task, task->workerid); } else { struct _starpu_machine_config *config = _starpu_get_machine_config(); /* When a task can only be executed on a given arch and we have * only one memory node for that arch, we can systematically * prefetch before the scheduling decision. */ if (starpu_get_prefetch_flag()) { if (task->cl->where == STARPU_CPU && config->cpus_nodeid >= 0) starpu_prefetch_task_input_on_node(task, config->cpus_nodeid); else if (task->cl->where == STARPU_CUDA && config->cuda_nodeid >= 0) starpu_prefetch_task_input_on_node(task, config->cuda_nodeid); else if (task->cl->where == STARPU_OPENCL && config->opencl_nodeid >= 0) starpu_prefetch_task_input_on_node(task, config->opencl_nodeid); else if (task->cl->where == STARPU_MIC && config->mic_nodeid >= 0) starpu_prefetch_task_input_on_node(task, config->mic_nodeid); else if (task->cl->where == STARPU_SCC && config->scc_nodeid >= 0) starpu_prefetch_task_input_on_node(task, config->scc_nodeid); } if(!sched_ctx->sched_policy) { /* Note: we have to call that early, or else the task may have * disappeared already */ starpu_push_task_end(task); if(!sched_ctx->awake_workers) ret = _starpu_push_task_on_specific_worker(task, sched_ctx->main_master); else { struct starpu_worker_collection *workers = sched_ctx->workers; struct _starpu_job *job = _starpu_get_job_associated_to_task(task); job->task_size = workers->nworkers; job->combined_workerid = -1; // workerid; its a ctx not combined worker job->active_task_alias_count = 0; STARPU_PTHREAD_BARRIER_INIT(&job->before_work_barrier, NULL, workers->nworkers); STARPU_PTHREAD_BARRIER_INIT(&job->after_work_barrier, NULL, workers->nworkers); job->after_work_busy_barrier = workers->nworkers; unsigned workerid; struct starpu_sched_ctx_iterator it; if(workers->init_iterator) workers->init_iterator(workers, &it); while(workers->has_next(workers, &it)) { workerid = workers->get_next(workers, &it); struct starpu_task *alias = starpu_task_dup(task); alias->destroy = 1; ret |= _starpu_push_task_on_specific_worker(alias, workerid); } } } else { STARPU_ASSERT(sched_ctx->sched_policy->push_task); /* check out if there are any workers in the context */ starpu_pthread_rwlock_t *changing_ctx_mutex = _starpu_sched_ctx_get_changing_ctx_mutex(sched_ctx->id); STARPU_PTHREAD_RWLOCK_RDLOCK(changing_ctx_mutex); nworkers = starpu_sched_ctx_get_nworkers(sched_ctx->id); if (nworkers == 0) ret = -1; else { _STARPU_TRACE_WORKER_SCHEDULING_PUSH; ret = sched_ctx->sched_policy->push_task(task); _STARPU_TRACE_WORKER_SCHEDULING_POP; } STARPU_PTHREAD_RWLOCK_UNLOCK(changing_ctx_mutex); } if(ret == -1) { fprintf(stderr, "repush task \n"); _STARPU_TRACE_JOB_POP(task, task->priority > 0); ret = _starpu_push_task_to_workers(task); } } /* Note: from here, the task might have been destroyed already! */ _STARPU_LOG_OUT(); return ret; }
/* Enqueue a task into the list of tasks explicitely attached to a worker. In * case workerid identifies a combined worker, a task will be enqueued into * each worker of the combination. */ static int _starpu_push_task_on_specific_worker(struct starpu_task *task, int workerid) { int nbasic_workers = (int)starpu_worker_get_count(); /* Is this a basic worker or a combined worker ? */ int is_basic_worker = (workerid < nbasic_workers); unsigned memory_node; struct _starpu_worker *worker = NULL; struct _starpu_combined_worker *combined_worker = NULL; if (is_basic_worker) { worker = _starpu_get_worker_struct(workerid); memory_node = worker->memory_node; } else { combined_worker = _starpu_get_combined_worker_struct(workerid); memory_node = combined_worker->memory_node; } if (use_prefetch) starpu_prefetch_task_input_on_node(task, memory_node); if (is_basic_worker) _starpu_push_task_on_specific_worker_notify_sched(task, worker, workerid, workerid); else { /* Notify all workers of the combined worker */ int worker_size = combined_worker->worker_size; int *combined_workerid = combined_worker->combined_workerid; int j; for (j = 0; j < worker_size; j++) { int subworkerid = combined_workerid[j]; _starpu_push_task_on_specific_worker_notify_sched(task, _starpu_get_worker_struct(subworkerid), subworkerid, workerid); } } #ifdef STARPU_USE_SC_HYPERVISOR starpu_sched_ctx_call_pushed_task_cb(workerid, task->sched_ctx); #endif //STARPU_USE_SC_HYPERVISOR unsigned i; if (is_basic_worker) { unsigned node = starpu_worker_get_memory_node(workerid); if (_starpu_task_uses_multiformat_handles(task)) { unsigned nbuffers = STARPU_TASK_GET_NBUFFERS(task); for (i = 0; i < nbuffers; i++) { struct starpu_task *conversion_task; starpu_data_handle_t handle; handle = STARPU_TASK_GET_HANDLE(task, i); if (!_starpu_handle_needs_conversion_task(handle, node)) continue; conversion_task = _starpu_create_conversion_task(handle, node); conversion_task->mf_skip = 1; conversion_task->execute_on_a_specific_worker = 1; conversion_task->workerid = workerid; _starpu_task_submit_conversion_task(conversion_task, workerid); //_STARPU_DEBUG("Pushing a conversion task\n"); } for (i = 0; i < nbuffers; i++) { starpu_data_handle_t handle = STARPU_TASK_GET_HANDLE(task, i); handle->mf_node = node; } } // if(task->sched_ctx != _starpu_get_initial_sched_ctx()->id) if(task->priority > 0) return _starpu_push_local_task(worker, task, 1); else return _starpu_push_local_task(worker, task, 0); } else { /* This is a combined worker so we create task aliases */ int worker_size = combined_worker->worker_size; int *combined_workerid = combined_worker->combined_workerid; int ret = 0; struct _starpu_job *job = _starpu_get_job_associated_to_task(task); job->task_size = worker_size; job->combined_workerid = workerid; job->active_task_alias_count = 0; STARPU_PTHREAD_BARRIER_INIT(&job->before_work_barrier, NULL, worker_size); STARPU_PTHREAD_BARRIER_INIT(&job->after_work_barrier, NULL, worker_size); job->after_work_busy_barrier = worker_size; /* Note: we have to call that early, or else the task may have * disappeared already */ starpu_push_task_end(task); int j; for (j = 0; j < worker_size; j++) { struct starpu_task *alias = starpu_task_dup(task); alias->destroy = 1; worker = _starpu_get_worker_struct(combined_workerid[j]); ret |= _starpu_push_local_task(worker, alias, 0); } return ret; } }