// Start the main task.
//
// Warning: this method is not called within a Qthread task context. Do
// not use methods that require task context (e.g., task-local storage).
void chpl_task_callMain(void (*chpl_main)(void))
{
const chpl_bool initial_serial_state = false;
const c_localeid_t initial_locale_id = default_locale_id;
const chapel_wrapper_args_t wrapper_args =
{chpl_main, NULL, NULL, 0,
{initial_serial_state, initial_locale_id, NULL,
chpl_malloc, chpl_calloc, chpl_realloc, chpl_free}};
qthread_debug(CHAPEL_CALLS, "[%d] begin chpl_task_callMain()\n", chpl_localeID);
default_serial_state = initial_serial_state;
#ifdef QTHREAD_MULTINODE
qthread_debug(CHAPEL_BEHAVIOR, "[%d] calling spr_unify\n", chpl_localeID);
int const rc = spr_unify();
assert(SPR_OK == rc);
#endif /* QTHREAD_MULTINODE */
qthread_fork_syncvar(chapel_wrapper, &wrapper_args, &exit_ret);
qthread_syncvar_readFF(NULL, &exit_ret);
qthread_debug(CHAPEL_BEHAVIOR, "[%d] main task finished\n", chpl_localeID);
qthread_debug(CHAPEL_CALLS, "[%d] end chpl_task_callMain()\n", chpl_localeID);
}
// Start the main task.
//
// Warning: this method is not called within a Qthread task context. Do
// not use methods that require task context (e.g., task-local storage).
void chpl_task_callMain(void (*chpl_main)(void))
{
const chpl_qthread_wrapper_args_t wrapper_args =
{chpl_main, NULL, NULL, 0, false, {c_sublocid_any_val, false}};
qthread_debug(CHAPEL_CALLS, "[%d] begin chpl_task_callMain()\n", chpl_localeID);
#ifdef QTHREAD_MULTINODE
qthread_debug(CHAPEL_BEHAVIOR, "[%d] calling spr_unify\n", chpl_localeID);
int const rc = spr_unify();
assert(SPR_OK == rc);
#endif /* QTHREAD_MULTINODE */
qthread_fork_syncvar(chapel_wrapper, &wrapper_args, &exit_ret);
qthread_syncvar_readFF(NULL, &exit_ret);
qthread_debug(CHAPEL_BEHAVIOR, "[%d] main task finished\n", chpl_localeID);
qthread_debug(CHAPEL_CALLS, "[%d] end chpl_task_callMain()\n", chpl_localeID);
}
int main(int argc,
char *argv[])
{
uint64_t total_num_nodes = 0;
qtimer_t timer;
double total_time = 0.0;
CHECK_VERBOSE();
{
unsigned int tmp = (unsigned int)tree_type;
NUMARG(tmp, "UTS_TREE_TYPE");
if (tmp <= BALANCED) {
tree_type = (tree_t)tmp;
} else {
fprintf(stderr, "invalid tree type\n");
return EXIT_FAILURE;
}
tmp = (unsigned int)shape_fn;
NUMARG(tmp, "UTS_SHAPE_FN");
if (tmp <= FIXED) {
shape_fn = (shape_t)tmp;
} else {
fprintf(stderr, "invalid shape function\n");
return EXIT_FAILURE;
}
}
DBLARG(bf_0, "UTS_BF_0");
NUMARG(root_seed, "UTS_ROOT_SEED");
NUMARG(tree_depth, "UTS_TREE_DEPTH");
DBLARG(non_leaf_prob, "UTS_NON_LEAF_PROB");
NUMARG(non_leaf_bf, "UTS_NON_LEAF_NUM");
NUMARG(shift_depth, "UTS_SHIFT_DEPTH");
NUMARG(num_samples, "UTS_NUM_SAMPLES");
// If the operator did not attempt to set a stack size, force
// a reasonable lower bound
if (!getenv("QT_STACK_SIZE") && !getenv("QTHREAD_STACK_SIZE"))
setenv("QT_STACK_SIZE", "32768", 0);
assert(qthread_initialize() == 0);
#ifdef PRINT_STATS
print_stats();
#else
print_banner();
#endif
timer = qtimer_create();
qtimer_start(timer);
node_t root;
root.height = 0;
rng_init(root.state.state, root_seed);
root.num_children = calc_num_children(&root);
aligned_t donecount = 0;
root.dc = &donecount;
qthread_empty(&donecount);
aligned_t tot = 0;
root.acc = &tot;
root.expect = 1;
qthread_fork_syncvar(visit, &root, NULL);
qthread_readFF(NULL, root.dc);
total_num_nodes = tot;
qtimer_stop(timer);
total_time = qtimer_secs(timer);
qtimer_destroy(timer);
#ifdef PRINT_STATS
printf("tree-size %lu\ntree-depth %d\nnum-leaves %llu\nperc-leaves %.2f\n",
(unsigned long)total_num_nodes,
(int)tree_height,
(unsigned long long)num_leaves,
num_leaves / (float)total_num_nodes * 100.0);
printf("exec-time %.3f\ntotal-perf %.0f\npu-perf %.0f\n\n",
total_time,
total_num_nodes / total_time,
total_num_nodes / total_time / qthread_num_workers());
#else
printf("Tree size = %lu, tree depth = %d, num leaves = %llu (%.2f%%)\n",
(unsigned long)total_num_nodes,
(int)tree_height,
(unsigned long long)num_leaves,
num_leaves / (float)total_num_nodes * 100.0);
printf("Wallclock time = %.3f sec, performance = %.0f "
"nodes/sec (%.0f nodes/sec per PE)\n\n",
total_time,
total_num_nodes / total_time,
total_num_nodes / total_time / qthread_num_workers());
#endif /* ifdef PRINT_STATS */
return 0;
}
