/* the generic interface that call the proper underlying implementation */ int _starpu_push_task(starpu_job_t j, unsigned job_is_already_locked) { struct starpu_task *task = j->task; task->status = STARPU_TASK_READY; /* in case there is no codelet associated to the task (that's a control * task), we directly execute its callback and enforce the * corresponding dependencies */ if (task->cl == NULL) { _starpu_handle_job_termination(j, job_is_already_locked); return 0; } if (STARPU_UNLIKELY(task->execute_on_a_specific_worker)) { unsigned workerid = task->workerid; struct starpu_worker_s *worker = _starpu_get_worker_struct(workerid); if (use_prefetch) { uint32_t memory_node = starpu_worker_get_memory_node(workerid); _starpu_prefetch_task_input_on_node(task, memory_node); } return _starpu_push_local_task(worker, j); } else { STARPU_ASSERT(policy.push_task); return policy.push_task(task); } }
/* return true if workerid can execute task, and fill task->predicted and task->predicted_transfer * according to best implementation predictions */ static int find_best_impl(unsigned sched_ctx_id, struct starpu_task * task, int workerid) { double len = DBL_MAX; int best_impl = -1; int impl; for(impl = 0; impl < STARPU_MAXIMPLEMENTATIONS; impl++) { if(starpu_worker_can_execute_task(workerid, task, impl)) { struct starpu_perfmodel_arch* archtype = starpu_worker_get_perf_archtype(workerid, sched_ctx_id); double d = starpu_task_expected_length(task, archtype, impl); if(isnan(d)) { best_impl = impl; len = 0.0; break; } if(d < len) { len = d; best_impl = impl; } } } if(best_impl == -1) return 0; int memory_node = starpu_worker_get_memory_node(workerid); task->predicted = len; task->predicted_transfer = starpu_task_expected_data_transfer_time(memory_node, task); starpu_task_set_implementation(task, best_impl); return 1; }
int main(int argc, char **argv) { starpu_init(NULL); starpu_data_malloc_pinned_if_possible((void **)&v, VECTORSIZE*sizeof(unsigned)); starpu_vector_data_register(&v_handle, 0, (uintptr_t)v, VECTORSIZE, sizeof(unsigned)); unsigned nworker = starpu_worker_get_count(); cnt = nworker*N; unsigned iter, worker; for (iter = 0; iter < N; iter++) { for (worker = 0; worker < nworker; worker++) { /* synchronous prefetch */ unsigned node = starpu_worker_get_memory_node(worker); starpu_data_prefetch_on_node(v_handle, node, 0); /* execute a task */ struct starpu_task *task = starpu_task_create(); task->cl = &cl; task->buffers[0].handle = v_handle; task->buffers[0].mode = select_random_mode(); task->callback_func = callback; task->callback_arg = NULL; task->synchronous = 1; int ret = starpu_task_submit(task, NULL); if (ret == -ENODEV) goto enodev; } } pthread_mutex_lock(&mutex); if (!finished) pthread_cond_wait(&cond, &mutex); pthread_mutex_unlock(&mutex); starpu_shutdown(); return 0; enodev: fprintf(stderr, "WARNING: No one can execute this task\n"); /* yes, we do not perform the computation but we did detect that no one * could perform the kernel, so this is not an error from StarPU */ return 0; }
static void common_data_cpy_func(void *descr[], void *cl_arg) { unsigned interface_id = *(unsigned *)cl_arg; const struct starpu_data_interface_ops *interface_ops = _starpu_data_interface_get_ops(interface_id); const struct starpu_data_copy_methods *copy_methods = interface_ops->copy_methods; int workerid = starpu_worker_get_id(); enum starpu_worker_archtype type = starpu_worker_get_type(workerid); unsigned memory_node = starpu_worker_get_memory_node(workerid); void *dst_interface = descr[0]; void *src_interface = descr[1]; switch (type) { case STARPU_CPU_WORKER: if (copy_methods->ram_to_ram) { copy_methods->ram_to_ram(src_interface, memory_node, dst_interface, memory_node); return; } break; case STARPU_CUDA_WORKER: if (copy_methods->cuda_to_cuda) { copy_methods->cuda_to_cuda(src_interface, memory_node, dst_interface, memory_node); return; } break; case STARPU_OPENCL_WORKER: if (copy_methods->opencl_to_opencl) { copy_methods->opencl_to_opencl(src_interface, memory_node, dst_interface, memory_node); return; } break; default: /* unknown architecture */ STARPU_ABORT(); } STARPU_ASSERT(copy_methods->any_to_any); copy_methods->any_to_any(src_interface, memory_node, dst_interface, memory_node, NULL); }
void starpu_data_unpartition(starpu_data_handle_t root_handle, unsigned gathering_node) { unsigned child; unsigned worker; unsigned nworkers = starpu_worker_get_count(); unsigned node; unsigned sizes[root_handle->nchildren]; _STARPU_TRACE_START_UNPARTITION(root_handle, gathering_node); _starpu_spin_lock(&root_handle->header_lock); STARPU_ASSERT_MSG(root_handle->nchildren != 0, "data %p is not partitioned, can not unpartition it", root_handle); /* first take all the children lock (in order !) */ for (child = 0; child < root_handle->nchildren; child++) { starpu_data_handle_t child_handle = starpu_data_get_child(root_handle, child); /* make sure the intermediate children is unpartitionned as well */ if (child_handle->nchildren > 0) starpu_data_unpartition(child_handle, gathering_node); /* If this is a multiformat handle, we must convert the data now */ #ifdef STARPU_DEVEL #warning TODO: _starpu_fetch_data_on_node should be doing it #endif if (_starpu_data_is_multiformat_handle(child_handle) && starpu_node_get_kind(child_handle->mf_node) != STARPU_CPU_RAM) { struct starpu_codelet cl = { .where = STARPU_CPU, .cpu_funcs = { _starpu_empty_codelet_function }, .modes = { STARPU_RW }, .nbuffers = 1 }; struct starpu_task *task = starpu_task_create(); task->name = "convert_data"; STARPU_TASK_SET_HANDLE(task, child_handle, 0); task->cl = &cl; task->synchronous = 1; if (_starpu_task_submit_internally(task) != 0) _STARPU_ERROR("Could not submit the conversion task while unpartitionning\n"); } int ret; /* for now we pretend that the RAM is almost unlimited and that gathering * data should be possible from the node that does the unpartionning ... we * don't want to have the programming deal with memory shortage at that time, * really */ /* Acquire the child data on the gathering node. This will trigger collapsing any reduction */ ret = starpu_data_acquire_on_node(child_handle, gathering_node, STARPU_RW); STARPU_ASSERT(ret == 0); starpu_data_release_on_node(child_handle, gathering_node); _starpu_spin_lock(&child_handle->header_lock); child_handle->busy_waiting = 1; _starpu_spin_unlock(&child_handle->header_lock); /* Wait for all requests to finish (notably WT requests) */ STARPU_PTHREAD_MUTEX_LOCK(&child_handle->busy_mutex); while (1) { /* Here helgrind would shout that this an unprotected access, * but this is actually fine: all threads who do busy_count-- * are supposed to call _starpu_data_check_not_busy, which will * wake us up through the busy_mutex/busy_cond. */ if (!child_handle->busy_count) break; /* This is woken by _starpu_data_check_not_busy, always called * after decrementing busy_count */ STARPU_PTHREAD_COND_WAIT(&child_handle->busy_cond, &child_handle->busy_mutex); } STARPU_PTHREAD_MUTEX_UNLOCK(&child_handle->busy_mutex); _starpu_spin_lock(&child_handle->header_lock); sizes[child] = _starpu_data_get_size(child_handle); _starpu_data_unregister_ram_pointer(child_handle); for (worker = 0; worker < nworkers; worker++) { struct _starpu_data_replicate *local = &child_handle->per_worker[worker]; STARPU_ASSERT(local->state == STARPU_INVALID); if (local->allocated && local->automatically_allocated) _starpu_request_mem_chunk_removal(child_handle, local, starpu_worker_get_memory_node(worker), sizes[child]); } _starpu_memory_stats_free(child_handle); }
static void _starpu_data_partition(starpu_data_handle_t initial_handle, starpu_data_handle_t *childrenp, unsigned nparts, struct starpu_data_filter *f, int inherit_state) { unsigned i; unsigned node; /* first take care to properly lock the data header */ _starpu_spin_lock(&initial_handle->header_lock); initial_handle->nplans++; STARPU_ASSERT_MSG(nparts > 0, "Partitioning data %p in 0 piece does not make sense", initial_handle); /* allocate the children */ if (inherit_state) { initial_handle->children = (struct _starpu_data_state *) calloc(nparts, sizeof(struct _starpu_data_state)); STARPU_ASSERT(initial_handle->children); /* this handle now has children */ initial_handle->nchildren = nparts; } unsigned nworkers = starpu_worker_get_count(); for (node = 0; node < STARPU_MAXNODES; node++) { if (initial_handle->per_node[node].state != STARPU_INVALID) break; } if (node == STARPU_MAXNODES) { /* This is lazy allocation, allocate it now in main RAM, so as * to have somewhere to gather pieces later */ /* FIXME: mark as unevictable! */ int ret = _starpu_allocate_memory_on_node(initial_handle, &initial_handle->per_node[STARPU_MAIN_RAM], 0); #ifdef STARPU_DEVEL #warning we should reclaim memory if allocation failed #endif STARPU_ASSERT(!ret); } for (i = 0; i < nparts; i++) { starpu_data_handle_t child; if (inherit_state) child = &initial_handle->children[i]; else child = childrenp[i]; STARPU_ASSERT(child); struct starpu_data_interface_ops *ops; /* each child may have his own interface type */ /* what's this child's interface ? */ if (f->get_child_ops) ops = f->get_child_ops(f, i); else ops = initial_handle->ops; _starpu_data_handle_init(child, ops, initial_handle->mf_node); child->nchildren = 0; child->nplans = 0; child->switch_cl = NULL; child->partitioned = 0; child->readonly = 0; child->mpi_data = initial_handle->mpi_data; child->root_handle = initial_handle->root_handle; child->father_handle = initial_handle; child->sibling_index = i; child->depth = initial_handle->depth + 1; child->is_not_important = initial_handle->is_not_important; child->wt_mask = initial_handle->wt_mask; child->home_node = initial_handle->home_node; child->is_readonly = initial_handle->is_readonly; /* initialize the chunk lock */ _starpu_data_requester_list_init(&child->req_list); _starpu_data_requester_list_init(&child->reduction_req_list); child->reduction_tmp_handles = NULL; child->write_invalidation_req = NULL; child->refcnt = 0; child->unlocking_reqs = 0; child->busy_count = 0; child->busy_waiting = 0; STARPU_PTHREAD_MUTEX_INIT(&child->busy_mutex, NULL); STARPU_PTHREAD_COND_INIT(&child->busy_cond, NULL); child->reduction_refcnt = 0; _starpu_spin_init(&child->header_lock); child->sequential_consistency = initial_handle->sequential_consistency; STARPU_PTHREAD_MUTEX_INIT(&child->sequential_consistency_mutex, NULL); child->last_submitted_mode = STARPU_R; child->last_sync_task = NULL; child->last_submitted_accessors.task = NULL; child->last_submitted_accessors.next = &child->last_submitted_accessors; child->last_submitted_accessors.prev = &child->last_submitted_accessors; child->post_sync_tasks = NULL; /* Tell helgrind that the race in _starpu_unlock_post_sync_tasks is fine */ STARPU_HG_DISABLE_CHECKING(child->post_sync_tasks_cnt); child->post_sync_tasks_cnt = 0; /* The methods used for reduction are propagated to the * children. */ child->redux_cl = initial_handle->redux_cl; child->init_cl = initial_handle->init_cl; #ifdef STARPU_USE_FXT child->last_submitted_ghost_sync_id_is_valid = 0; child->last_submitted_ghost_sync_id = 0; child->last_submitted_ghost_accessors_id = NULL; #endif if (_starpu_global_arbiter) /* Just for testing purpose */ starpu_data_assign_arbiter(child, _starpu_global_arbiter); else child->arbiter = NULL; _starpu_data_requester_list_init(&child->arbitered_req_list); for (node = 0; node < STARPU_MAXNODES; node++) { struct _starpu_data_replicate *initial_replicate; struct _starpu_data_replicate *child_replicate; initial_replicate = &initial_handle->per_node[node]; child_replicate = &child->per_node[node]; if (inherit_state) child_replicate->state = initial_replicate->state; else child_replicate->state = STARPU_INVALID; if (inherit_state || !initial_replicate->automatically_allocated) child_replicate->allocated = initial_replicate->allocated; else child_replicate->allocated = 0; /* Do not allow memory reclaiming within the child for parent bits */ child_replicate->automatically_allocated = 0; child_replicate->refcnt = 0; child_replicate->memory_node = node; child_replicate->relaxed_coherency = 0; if (inherit_state) child_replicate->initialized = initial_replicate->initialized; else child_replicate->initialized = 0; /* update the interface */ void *initial_interface = starpu_data_get_interface_on_node(initial_handle, node); void *child_interface = starpu_data_get_interface_on_node(child, node); STARPU_ASSERT_MSG(!(!inherit_state && child_replicate->automatically_allocated && child_replicate->allocated), "partition planning is currently not supported when handle has some automatically allocated buffers"); f->filter_func(initial_interface, child_interface, f, i, nparts); } unsigned worker; for (worker = 0; worker < nworkers; worker++) { struct _starpu_data_replicate *child_replicate; child_replicate = &child->per_worker[worker]; child_replicate->state = STARPU_INVALID; child_replicate->allocated = 0; child_replicate->automatically_allocated = 0; child_replicate->refcnt = 0; child_replicate->memory_node = starpu_worker_get_memory_node(worker); child_replicate->requested = 0; for (node = 0; node < STARPU_MAXNODES; node++) { child_replicate->request[node] = NULL; } child_replicate->relaxed_coherency = 1; child_replicate->initialized = 0; /* duplicate the content of the interface on node 0 */ memcpy(child_replicate->data_interface, child->per_node[0].data_interface, child->ops->interface_size); } /* We compute the size and the footprint of the child once and * store it in the handle */ child->footprint = _starpu_compute_data_footprint(child); void *ptr; ptr = starpu_data_handle_to_pointer(child, STARPU_MAIN_RAM); if (ptr != NULL) _starpu_data_register_ram_pointer(child, ptr); } /* now let the header */ _starpu_spin_unlock(&initial_handle->header_lock); }
int main(int argc, char **argv) { unsigned *foo; starpu_data_handle_t handle; int ret; unsigned n, i, size; ret = starpu_initialize(NULL, &argc, &argv); if (ret == -ENODEV) return STARPU_TEST_SKIPPED; STARPU_CHECK_RETURN_VALUE(ret, "starpu_init"); #ifdef STARPU_USE_OPENCL ret = starpu_opencl_load_opencl_from_file("tests/datawizard/scal_opencl.cl", &opencl_program, NULL); STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file"); #endif n = starpu_worker_get_count(); if (n == 1) { starpu_shutdown(); return STARPU_TEST_SKIPPED; } size = 10 * n; foo = (unsigned *) calloc(size, sizeof(*foo)); for (i = 0; i < size; i++) foo[i] = i; starpu_vector_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)foo, size, sizeof(*foo)); /* Broadcast the data to force in-place partitioning */ for (i = 0; i < n; i++) starpu_data_prefetch_on_node(handle, starpu_worker_get_memory_node(i), 0); struct starpu_data_filter f = { .filter_func = starpu_vector_filter_block, .nchildren = n, }; starpu_data_partition(handle, &f); for (i = 0; i < f.nchildren; i++) { struct starpu_task *task = starpu_task_create(); task->handles[0] = starpu_data_get_sub_data(handle, 1, i); task->cl = &scal_codelet; task->execute_on_a_specific_worker = 1; task->workerid = i; ret = starpu_task_submit(task); if (ret == -ENODEV) goto enodev; STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit"); } ret = starpu_task_wait_for_all(); STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait_for_all"); starpu_data_unpartition(handle, STARPU_MAIN_RAM); starpu_data_unregister(handle); starpu_shutdown(); ret = EXIT_SUCCESS; for (i = 0; i < size; i++) { if (foo[i] != i*2) { FPRINTF(stderr,"value %u is %u instead of %u\n", i, foo[i], 2*i); ret = EXIT_FAILURE; } } return ret; enodev: starpu_data_unregister(handle); fprintf(stderr, "WARNING: No one can execute this task\n"); /* yes, we do not perform the computation but we did detect that no one * could perform the kernel, so this is not an error from StarPU */ starpu_shutdown(); return STARPU_TEST_SKIPPED; }
struct starpu_task *_starpu_pop_task(struct _starpu_worker *worker) { struct starpu_task *task; int worker_id; unsigned node; /* We can't tell in advance which task will be picked up, so we measure * a timestamp, and will attribute it afterwards to the task. */ int profiling = starpu_profiling_status_get(); struct timespec pop_start_time; if (profiling) _starpu_clock_gettime(&pop_start_time); pick: /* perhaps there is some local task to be executed first */ task = _starpu_pop_local_task(worker); /* get tasks from the stacks of the strategy */ if(!task) { struct _starpu_sched_ctx *sched_ctx ; #ifndef STARPU_NON_BLOCKING_DRIVERS int been_here[STARPU_NMAX_SCHED_CTXS]; int i; for(i = 0; i < STARPU_NMAX_SCHED_CTXS; i++) been_here[i] = 0; while(!task) #endif { if(worker->nsched_ctxs == 1) sched_ctx = _starpu_get_initial_sched_ctx(); else { while(1) { sched_ctx = _get_next_sched_ctx_to_pop_into(worker); if(worker->removed_from_ctx[sched_ctx->id] == 1 && worker->shares_tasks_lists[sched_ctx->id] == 1) { _starpu_worker_gets_out_of_ctx(sched_ctx->id, worker); worker->removed_from_ctx[sched_ctx->id] = 0; sched_ctx = NULL; } else break; } } if(sched_ctx && sched_ctx->id != STARPU_NMAX_SCHED_CTXS) { if (sched_ctx->sched_policy && sched_ctx->sched_policy->pop_task) { task = sched_ctx->sched_policy->pop_task(sched_ctx->id); _starpu_pop_task_end(task); } } if(!task) { /* it doesn't matter if it shares tasks list or not in the scheduler, if it does not have any task to pop just get it out of here */ /* however if it shares a task list it will be removed as soon as he finishes this job (in handle_job_termination) */ if(worker->removed_from_ctx[sched_ctx->id]) { _starpu_worker_gets_out_of_ctx(sched_ctx->id, worker); worker->removed_from_ctx[sched_ctx->id] = 0; } #ifdef STARPU_USE_SC_HYPERVISOR if(worker->pop_ctx_priority) { struct starpu_sched_ctx_performance_counters *perf_counters = sched_ctx->perf_counters; if(sched_ctx->id != 0 && perf_counters != NULL && perf_counters->notify_idle_cycle && _starpu_sched_ctx_allow_hypervisor(sched_ctx->id)) { // _STARPU_TRACE_HYPERVISOR_BEGIN(); perf_counters->notify_idle_cycle(sched_ctx->id, worker->workerid, 1.0); // _STARPU_TRACE_HYPERVISOR_END(); } } #endif //STARPU_USE_SC_HYPERVISOR #ifndef STARPU_NON_BLOCKING_DRIVERS if(been_here[sched_ctx->id] || worker->nsched_ctxs == 1) break; been_here[sched_ctx->id] = 1; #endif } } } if (!task) { idle_start[worker->workerid] = starpu_timing_now(); return NULL; } if(idle_start[worker->workerid] != 0.0) { double idle_end = starpu_timing_now(); idle[worker->workerid] += (idle_end - idle_start[worker->workerid]); idle_start[worker->workerid] = 0.0; } #ifdef STARPU_USE_SC_HYPERVISOR struct _starpu_sched_ctx *sched_ctx = _starpu_get_sched_ctx_struct(task->sched_ctx); struct starpu_sched_ctx_performance_counters *perf_counters = sched_ctx->perf_counters; if(sched_ctx->id != 0 && perf_counters != NULL && perf_counters->notify_poped_task && _starpu_sched_ctx_allow_hypervisor(sched_ctx->id)) { // _STARPU_TRACE_HYPERVISOR_BEGIN(); perf_counters->notify_poped_task(task->sched_ctx, worker->workerid); // _STARPU_TRACE_HYPERVISOR_END(); } #endif //STARPU_USE_SC_HYPERVISOR /* Make sure we do not bother with all the multiformat-specific code if * it is not necessary. */ if (!_starpu_task_uses_multiformat_handles(task)) goto profiling; /* This is either a conversion task, or a regular task for which the * conversion tasks have already been created and submitted */ if (task->mf_skip) goto profiling; /* * This worker may not be able to execute this task. In this case, we * should return the task anyway. It will be pushed back almost immediatly. * This way, we avoid computing and executing the conversions tasks. * Here, we do not care about what implementation is used. */ worker_id = starpu_worker_get_id(); if (!starpu_worker_can_execute_task_first_impl(worker_id, task, NULL)) return task; node = starpu_worker_get_memory_node(worker_id); /* * We do have a task that uses multiformat handles. Let's create the * required conversion tasks. */ STARPU_PTHREAD_MUTEX_UNLOCK(&worker->sched_mutex); unsigned i; 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 = worker_id; /* * Next tasks will need to know where these handles have gone. */ handle->mf_node = node; _starpu_task_submit_conversion_task(conversion_task, worker_id); } task->mf_skip = 1; starpu_task_list_push_back(&worker->local_tasks, task); STARPU_PTHREAD_MUTEX_LOCK(&worker->sched_mutex); goto pick; profiling: if (profiling) { struct starpu_profiling_task_info *profiling_info; profiling_info = task->profiling_info; /* The task may have been created before profiling was enabled, * so we check if the profiling_info structure is available * even though we already tested if profiling is enabled. */ if (profiling_info) { memcpy(&profiling_info->pop_start_time, &pop_start_time, sizeof(struct timespec)); _starpu_clock_gettime(&profiling_info->pop_end_time); } } if(task->prologue_callback_pop_func) task->prologue_callback_pop_func(task->prologue_callback_pop_arg); return task; }
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; } }