void __ompc_static_init_4 (int global_id, int sched, int *lower,
int *upper, int *stride, int inc, int chunk)
{
int64_t l = (int64_t) *lower;
int64_t u = (int64_t) *upper;
int64_t s;
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_static_init_4 global_id=%d with "
"sched=%d, lower=%d, upper=%d, inc=%d, chunk=%d\n",
global_id, sched, *lower, *upper, inc, chunk);
#endif
__pmp_sample(PMP_PROFILE_OMPC_STATIC_INIT_4);
__pmp_static_init(global_id, sched, &l, &u, &s, inc, chunk);
*lower = (int) l;
*upper = (int) u;
*stride = (int) s;
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_static_init_4 global_id=%d returns "
"lower=%d, upper=%d, stride=%d\n",
global_id, *lower, *upper, *stride);
#endif
}
static void __pmp_shared_catch_segv (pmp_thread_t *thread)
{
static int32_t installing_segv = 0;
static int32_t installed_segv = 0;
/* For Linuxthreads this only needs to be done once, since sigaction's are
* shared across all of the pthreads. I arrange for it to be set up by the
* first worker thread that is woken up. This tranfers SEGV catching
* responsibility from the serial code in libfoobar to libopenmp as
* soon as parallelism is employed.
*/
if (__pmp_atomic_cmpxchg32(&installing_segv, 0, 1) == 0) {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d "
"is installing the SEGV handler\n", thread->global_id);
__pmp_catch_segv();
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d "
"has installed the SEGV handler\n", thread->global_id);
installed_segv = 1;
}
while (installed_segv == 0) {
/* USER LEVEL SPIN LOCK */
__pmp_yield();
}
}
int __ompc_schedule_next_4 (int global_id, int *lower, int *upper,
int *inc)
{
int result;
int64_t l;
int64_t u;
int64_t i;
__pmp_sample(PMP_PROFILE_OMPC_SCHEDULE_NEXT_4);
result = __pmp_schedule_next(global_id, &l, &u, &i);
*lower = (int) l;
*upper = (int) u;
*inc = (int) i;
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug) {
if (result == 0)
__pmp_debug("CALLS_DEBUG", "__ompc_schedule_next_4 global_id=%d returns "
"result=0\n", global_id);
else
__pmp_debug("CALLS_DEBUG", "__ompc_schedule_next_4 global_id=%d returns "
"lower=%d, upper=%d, inc=%d, result=%d\n",
global_id, *lower, *upper, *inc, result);
}
#endif
return result;
}
void __ompc_copyin_thdprv (int n, ...)
{
pmp_global_id_t global_id;
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_copyin_thdprv: n=%d\n", n);
__pmp_sample(PMP_PROFILE_OMPC_COPYIN_THDPRV);
if (__pmp_get_param()->disabled) {
return;
}
global_id = __pmp_get_current_global_id();
va_list ap;
va_start(ap, n);
while (n > 0) {
void *dst = va_arg(ap, void*);
void *src = va_arg(ap, void*);
int size = va_arg(ap, int);
if (dst != src) {
__pmp_debug(PMP_DEBUG_THREAD, "__ompc_copyin_thdprv: global_id=%d "
"dst: %p, src: %p, size: %d\n", global_id, dst, src, size);
memcpy(dst, src, size);
}
n -= 3;
}
va_end(ap);
}
static inline void __pmp_thread_wake (pmp_thread_t *thread)
{
int32_t sync = __pmp_atomic_cmpxchg32(&thread->sync, PMP_SYNC_IDLE,
PMP_SYNC_UNBLOCKED);
assert(sync != PMP_SYNC_UNBLOCKED);
if (sync == PMP_SYNC_BLOCKED) {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d is being signaled\n",
thread->global_id);
thread->sync = PMP_SYNC_IDLE;
assert(thread->tid != -1);
__pmp_sample(PMP_PROFILE_THREAD_RESCHEDULE);
#if (defined PMP_USE_PTHREAD_SIGNALS)
if (pthread_kill(thread->pthread_id, SIGPMP) != 0) {
__pmp_fatal("unable to wake thread using pthread_kill\n");
}
#elif (defined PMP_NO_NPTL)
if (kill(thread->tid, SIGPMP) != 0) {
__pmp_fatal("unable to wake thread using kill\n");
}
#else
if (tkill(thread->tid, SIGPMP) != 0) {
__pmp_fatal("unable to wake thread using tkill\n");
}
#endif
}
else {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d is woken\n",
thread->global_id);
}
}
static inline void __pmp_thread_wait (pmp_thread_t *thread)
{
int32_t sync;
int thread_spin;
int i;
if (thread->sync == PMP_SYNC_UNBLOCKED) {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d does not block (1)\n",
thread->global_id);
thread->sync = PMP_SYNC_IDLE;
return;
}
thread_spin = __pmp_get_param()->thread_spin;
for (i = 0; i < thread_spin; i++) {
/* USER LEVEL SPIN LOOP */
if (thread->sync == PMP_SYNC_UNBLOCKED) {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d does not block (2)\n",
thread->global_id);
thread->sync = PMP_SYNC_IDLE;
return;
}
__pmp_yield();
}
sync = __pmp_atomic_cmpxchg32(&thread->sync, PMP_SYNC_IDLE,
PMP_SYNC_BLOCKED);
if (sync == PMP_SYNC_IDLE) {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d is waiting\n",
thread->global_id);
__pmp_sample(PMP_PROFILE_THREAD_DESCHEDULE);
#ifdef PMP_USE_PTHREAD_SIGNALS
{
int sig;
do {
sigwait(&__pmp_manager.mask_block_sigpmp, &sig);
} while (sig != SIGPMP);
}
#else
sigsuspend(&__pmp_manager.mask_unblock_sigpmp);
/* NOTE: it is unfortunate that sigsuspend does not tell us which
* signal has been raised. */
#endif
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d is awake\n",
thread->global_id);
}
else {
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d does not block (3)\n",
thread->global_id);
thread->sync = PMP_SYNC_IDLE;
}
}
void __pmp_thread_create_main (void)
{
pmp_thread_t *thread = __pmp_get_main_thread();
int global_id;
global_id = __pmp_idstack_pop(&__pmp_manager.idstack);
assert(global_id == 0);
thread->pthread_id = pthread_self();
#ifdef PMP_NO_NPTL
thread->tid = getpid();
#else
thread->tid = gettid();
#endif
thread->local_id = 0;
#ifndef PMP_NO_TLS
#ifdef PMP_TLS_THREAD
__pmp_tls_current_thread = thread;
#endif
#ifdef PMP_TLS_LOCAL_ID
__pmp_tls_current_local_id = 0;
#endif
#ifdef PMP_TLS_GLOBAL_ID
__pmp_tls_current_global_id = thread->global_id;
#endif
#endif
#ifdef PMP_USE_PTHREAD_SIGNALS
if (pthread_sigmask(SIG_BLOCK, &__pmp_manager.mask_block_sigpmp,
NULL) != 0) {
__pmp_fatal("unable to set thread-specific sigmask\n");
}
#else
if (sigprocmask(SIG_BLOCK, &__pmp_manager.mask_block_sigpmp, NULL) != 0) {
__pmp_fatal("unable to set thread-specific sigmask\n");
}
#endif
if (pthread_setspecific(__pmp_manager.thread_key, (void *) thread) != 0) {
__pmp_fatal("unable to set thread-specific data\n");
}
__pmp_thread_bind(thread); /* early master bind */
__pmp_debug(PMP_DEBUG_THREAD, "created main thread global_id %d\n",
thread->global_id);
__pmp_debug(PMP_DEBUG_THREAD,
"__pmp_thread_create_main: tid=%d, pthread_id=0x%08x "
"global_id=%d, local_id=%d\n",
(int) thread->tid, (int) thread->pthread_id,
(int) thread->global_id, (int) thread->local_id);
}
void __ompc_fork (int nthreads, workfunc_t work, void *fp)
{
pmp_thread_t *master = __pmp_get_current_thread();
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_fork nthreads=%d, work=%p, fp=%p "
" (nesting depth = %d)\n",
nthreads, work, fp, master->nesting_depth);
__pmp_sample(PMP_PROFILE_OMPC_FORK);
__pmp_memory_fence();
__pmp_thread_fork(master, nthreads, work, fp);
__pmp_memory_fence();
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_fork completed"
" (nesting depth = %d)\n",
master->nesting_depth);
}
static inline int __pmp_thread_single (pmp_thread_t *thread)
{
/* NOTE: the compiler optimizes away OMP singles in the serial code
* so there is no need to optimize that case here. The case of a team
* with just one thread is not so common, so don't optimize that path either.
* The most important case is the n-way single where n > 1. However, it
* is possible for a thread with no team to call this (e.g. orphaned
* directive) so in that case one has to check to see if there is a team.
* TODO: allocate a team of 1 thread in the orphaned case to reduce the
* amount of special case code throughout the library looking for NULL
* team pointers. */
pmp_team_t *team = thread->team;
if (__pmp_get_team_size(team) == 1) {
return 1;
}
else {
int32_t thread_count = thread->single_count;
int32_t team_count = __pmp_atomic_cmpxchg32(&team->single_count,
thread_count,
thread_count + 1);
thread->single_count++;
__pmp_debug(PMP_DEBUG_THREAD, "__pmp_thread_single, local_id %d "
"thread_count=%d team_count=%d %s\n",
thread->local_id, thread_count, team_count,
(thread_count == team_count) ? "win" : "lose");
return (thread_count == team_count);
}
}
static inline void __pmp_thread_barrier (pmp_thread_t *thread)
{
/* NOTE: the compiler optimizes away OMP barriers in the serial code
* so there is no need to optimize that case here. The case of a team
* with just one thread is not so common, so ideally don't optimize
* that path either. However, it is currently necessary to check
* (team != NULL) so one might as well check for the 1-thread team too.
* The most important case is the n-way barrier where n > 1. */
pmp_team_t *team = thread->team;
int team_size = __pmp_get_team_size(team);
if (team_size > 1) {
int32_t count = __pmp_atomic_xadd32(&team->barrier_count, -1);
assert(count > 0);
__pmp_debug(PMP_DEBUG_THREAD, "thread hits barrier with count of %d\n",
(int) count);
if (count > 1) {
__pmp_thread_wait(thread);
}
else {
pmp_local_id_t local_id = thread->local_id;
int i;
team->barrier_count = team_size;
for (i = 0; i < team_size; i++) {
pmp_thread_t *t = team->members[i];
if (i != local_id) {
__pmp_thread_wake(t);
}
}
}
}
}
static void __pmp_thread_release (pmp_team_t *team, pmp_thread_t *master)
{
pmp_local_id_t old_local_id;
pmp_idstack_t *idstack = &__pmp_manager.idstack;
int nworkers = team->team_size - 1;
int i;
__pmp_lock(master->global_id, &__pmp_manager.idlock);
for (i = nworkers; i >= 1; i--) {
pmp_thread_t *thread = team->members[i];
assert(thread != master);
old_local_id = thread->local_id;
thread->local_id = -1;
thread->team = NULL;
__pmp_idstack_push(idstack, thread->global_id);
__pmp_debug(PMP_DEBUG_THREAD,
"released thread global_id %d from local_id %d "
"of team at %p\n",
thread->global_id, old_local_id, team);
}
__pmp_unlock(master->global_id, &__pmp_manager.idlock);
__pmp_atomic_add32(&__ompc_cur_numthreads, -nworkers);
__pmp_atomic_xadd32(&__pmp_manager.waiting_threads, nworkers);
}
int __pmp_thread_acquire (int nthreads)
{
int count = 1; /* count from 1 to ignore master thread */
/* NOTE - in the typical case this while construct does not loop */
while (count < nthreads) {
int required = nthreads - count;
int waiting = __pmp_atomic_xadd32(&__pmp_manager.waiting_threads,
-required);
if (waiting >= required) {
count += required;
break;
}
else {
count += waiting;
required -= waiting;
__pmp_atomic_xadd32(&__pmp_manager.waiting_threads, required);
if (__pmp_manager_create_more_threads(required) == 0) {
break;
}
}
}
__pmp_atomic_add32(&__ompc_cur_numthreads, count - 1);
__pmp_debug(PMP_DEBUG_THREAD, "acquired %d out of %d threads\n",
count, nthreads);
return count;
}
static inline void __pmp_thread_master_join (pmp_thread_t *master)
{
pmp_team_t *team = master->team;
int32_t count;
int thread_spin = __pmp_get_param()->thread_spin;
int i;
/* NOTE: insert a small spin loop here to try to arrange for the master
* to arrive just after the last worker thread. If this happens
* then we avoid a much more expensive thread synchronization. */
for (i = 0; i < thread_spin; i++) {
/* USER LEVEL SPIN LOOP */
if (team->working_threads == 1) {
team->working_threads = 0;
return;
}
__pmp_yield();
}
count = __pmp_atomic_xadd32(&team->working_threads, -1);
__pmp_debug(PMP_DEBUG_THREAD, "master thread joins with count of %d\n",
(int) count);
assert(count >= 1);
if (count > 1) {
__pmp_thread_wait(master);
}
}
void __ompc_serialized_parallel (void)
{
pmp_thread_t *thread = __pmp_get_current_thread();
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_serialized_parallel\n");
__pmp_sample(PMP_PROFILE_OMPC_SERIALIZED_PARALLEL);
thread->serialized_parallel++;
}
int __ompc_get_thread_num (void)
{
int global_id = __pmp_get_current_global_id();
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_get_thread_num returns %d\n",global_id);
__pmp_sample(PMP_PROFILE_OMPC_GET_THREAD_NUM);
return global_id;
}
void __ompc_ordered (int global_id)
{
pmp_thread_t *thread = __pmp_get_thread(global_id);
if (__pmp_get_team_size(thread->team) > 1) {
pmp_loop_t *loop = thread->loop;
int64_t ticket_number = thread->ticket_number;
int64_t now_serving;
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_ordered: global_id=%d\n", global_id);
#endif
__pmp_sample(PMP_PROFILE_OMPC_ORDERED);
if (loop == NULL || loop->sched <= PMP_SCHED_ORDERED_OFFSET) {
__pmp_warning("ordered directives must be used inside ordered "
"OpenMP loops\n");
return;
}
assert(loop != NULL);
now_serving = loop->now_serving;
if (now_serving != ticket_number) {
if ((loop->inc >= 0) ? (now_serving > ticket_number) :
(now_serving < ticket_number)) {
__pmp_warning("ordered OpenMP loop may result in program deadlock\n");
__pmp_warning("maybe due to multiple ordered directives "
"in a loop iteration\n");
}
while (loop->now_serving != ticket_number) {
/* USER LEVEL SPIN LOOP */
__pmp_yield();
}
}
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Loop_Debug)
__pmp_debug("LOOPS_DEBUG", "__ompc_ordered: now serving global_id=%d "
" ticket_number=%" PRId64 "\n", global_id, ticket_number);
#endif
}
__pmp_memory_fence();
}
int __ompc_master (int global_id)
{
int master = (__pmp_get_thread(global_id)->local_id == 0);
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_master global_id=%d returns %d\n",
global_id, master);
__pmp_sample(PMP_PROFILE_OMPC_MASTER);
return master;
}
int __ompc_single (int global_id)
{
int result = __pmp_thread_single(__pmp_get_thread(global_id));
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_single global_id=%d returns %d\n",
global_id, result);
__pmp_sample(PMP_PROFILE_OMPC_SINGLE);
return result;
}
void __ompc_barrier (void)
{
pmp_thread_t *thread = __pmp_get_current_thread();
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_barrier\n");
__pmp_sample(PMP_PROFILE_OMPC_BARRIER);
__pmp_memory_fence();
__pmp_thread_barrier(thread);
}
static inline void __pmp_thread_work (pmp_thread_t *thread)
{
pmp_team_t *team = thread->team;
assert(team != NULL);
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d, local_id %d, "
"team at %p has lots of work to do now\n",
thread->global_id, thread->local_id, team);
/* NOTE: the id passed to the work function is the global_id. This is
* passed back to certain library routines as the global_id parameter.
* It can be used to find the thread structure very quickly. */
team->work(thread->global_id, team->fp);
__pmp_debug(PMP_DEBUG_THREAD, "thread global_id %d, local_id %d, "
"team at %p has no more work to do\n",
thread->global_id, thread->local_id, team);
}
void __ompc_static_fini (void)
{
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_static_fini\n");
#endif
__pmp_sample(PMP_PROFILE_OMPC_STATIC_FINI);
/* no work here, NOTE: does not appear to be called by compiler anyway */
}
int __ompc_get_local_thread_num (void)
{
int local_id = __pmp_get_current_local_id();
assert(local_id != -1);
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_get_local_thread_num returns %d\n",
local_id);
__pmp_sample(PMP_PROFILE_OMPC_GET_LOCAL_THREAD_NUM);
return local_id;
}
int __ompc_in_parallel (void)
{
pmp_thread_t *thread = __pmp_get_current_thread();
int in_parallel = (thread->team != NULL) ||
(thread->serialized_parallel > 0);
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_in_parallel returns %d\n", in_parallel);
__pmp_sample(PMP_PROFILE_OMPC_IN_PARALLEL);
return in_parallel;
}
void __ompc_static_init_8 (int global_id, int sched, int64_t *lower,
int64_t *upper, int64_t *stride, int64_t inc,
int64_t chunk)
{
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_static_init_8 global_id=%d with "
"sched=%d, lower=%" PRId64 ", upper=%" PRId64 ", inc=%" PRId64 ", chunk=%" PRId64 "\n",
global_id, sched, *lower, *upper, inc, chunk);
#endif
__pmp_sample(PMP_PROFILE_OMPC_STATIC_INIT_8);
__pmp_static_init(global_id, sched, lower, upper, stride, inc, chunk);
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_static_init_8 global_id=%d returns "
"lower=%" PRId64 ", upper=%" PRId64 ", stride=%" PRId64 "\n",
global_id, *lower, *upper, *stride);
#endif
}
static inline void __pmp_thread_worker_join (pmp_team_t *team)
{
int32_t count = __pmp_atomic_xadd32(&team->working_threads, -1);
__pmp_debug(PMP_DEBUG_THREAD, "worker thread joins with count of %d\n",
(int) count);
assert(count >= 1);
if (count == 1) {
__pmp_thread_wake(team->members[0]);
}
}
int __ompc_can_fork (void)
{
int team_size = __pmp_get_new_team_size();
int has_forked = (__pmp_get_main_thread()->nesting_depth > 0);
pmp_param_t *param = __pmp_get_param();
int serial_outline = param->serial_outline;
int disabled = param->disabled;
int can_fork = (team_size > 1 || has_forked || serial_outline) && !disabled;
__pmp_debug(PMP_DEBUG_CALLS, "__ompc_can_fork returns %d\n", can_fork);
__pmp_sample(PMP_PROFILE_OMPC_CAN_FORK);
return can_fork;
}
void __pmp_thread_assign (pmp_team_t *team, pmp_thread_t *master, int nthreads)
{
pmp_idstack_t *idstack = &__pmp_manager.idstack;
int i;
__pmp_lock(master->global_id, &__pmp_manager.idlock);
for (i = 1; i < nthreads; i++) {
int global_id = __pmp_idstack_pop(idstack);
pmp_thread_t *thread = __pmp_manager.threads + global_id;
__pmp_debug(PMP_DEBUG_THREAD,
"assigning thread global_id %d to local_id %d "
"of team at %p\n",
thread->global_id, i, team);
assert(thread->global_id == global_id);
assert(thread->local_id == -1);
assert(thread->team == NULL);
thread->local_id = i;
thread->team = team;
team->members[i] = thread;
__pmp_debug(PMP_DEBUG_THREAD,
"assigned thread global_id %d to local_id %d "
"of team at %p\n",
thread->global_id, thread->local_id, team);
}
__pmp_unlock(master->global_id, &__pmp_manager.idlock);
for (i = 1; i < nthreads; i++) {
/* This is pulled outside the idlock because waking a thread typically
* involves an inter-CPU synchronization which is relatively expensive.
*/
__pmp_thread_wake(team->members[i]);
}
}
void __ompc_scheduler_init_8 (int global_id, int sched, int64_t lower,
int64_t upper, int64_t inc, int64_t chunk)
{
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_scheduler_init_8 global_id=%d with "
"sched=%d, lower=%" PRId64 ", upper=%" PRId64 ", inc=%" PRId64 ", chunk=%" PRId64 "\n",
global_id, sched, lower, upper, inc, chunk);
#endif
__pmp_sample(PMP_PROFILE_OMPC_SCHEDULER_INIT_8);
__pmp_scheduler_init(global_id, sched, lower, upper, inc, chunk);
}
void __ompc_end_ordered (int global_id)
{
pmp_thread_t *thread = __pmp_get_thread(global_id);
__pmp_memory_fence();
if (__pmp_get_team_size(thread->team) > 1) {
pmp_loop_t *loop = thread->loop;
int64_t ticket_number = thread->ticket_number;
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_end_ordered: global_id=%d\n",
global_id);
#endif
__pmp_sample(PMP_PROFILE_OMPC_END_ORDERED);
if (loop == NULL || loop->sched <= PMP_SCHED_ORDERED_OFFSET) {
if (thread->global_id == 0)
__pmp_warning("ordered directives must be used inside ordered "
"OpenMP loops\n");
return;
}
assert(loop != NULL);
assert(loop->now_serving == ticket_number);
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Loop_Debug)
__pmp_debug("LOOPS_DEBUG", "__ompc_ordered: stop serving global_id=%d "
" ticket_number=%" PRId64 "\n", global_id, ticket_number);
#endif
loop->now_serving += loop->inc;
thread->ticket_number = ticket_number + loop->inc;
}
}
void __ompc_scheduler_init_4 (int global_id, int sched, int lower, int upper,
int inc, int chunk)
{
#ifdef SUPER_DEBUG
if (Enabled_Libomp_Call_Debug)
__pmp_debug("CALLS_DEBUG", "__ompc_scheduler_init_4 global_id=%d with "
"sched=%d, lower=%d, upper=%d, inc=%d, chunk=%d\n",
global_id, sched, lower, upper, inc, chunk);
#endif
__pmp_sample(PMP_PROFILE_OMPC_SCHEDULER_INIT_4);
__pmp_scheduler_init(global_id, sched, (int64_t) lower, (int64_t) upper,
(int64_t) inc, (int64_t) chunk);
}