13 #ifndef KMP_DISPATCH_HIER_H 14 #define KMP_DISPATCH_HIER_H 16 #include "kmp_dispatch.h" 19 enum kmp_hier_layer_e {
30 static inline const char *__kmp_get_hier_str(kmp_hier_layer_e type) {
32 case kmp_hier_layer_e::LAYER_THREAD:
34 case kmp_hier_layer_e::LAYER_L1:
36 case kmp_hier_layer_e::LAYER_L2:
38 case kmp_hier_layer_e::LAYER_L3:
40 case kmp_hier_layer_e::LAYER_NUMA:
42 case kmp_hier_layer_e::LAYER_LOOP:
44 case kmp_hier_layer_e::LAYER_LAST:
53 typedef struct kmp_hier_sched_env_t {
57 kmp_int32 *small_chunks;
58 kmp_int64 *large_chunks;
59 kmp_hier_layer_e *layers;
61 void append(
enum sched_type sched, kmp_int32 chunk, kmp_hier_layer_e layer) {
64 kmp_hier_layer_e::LAYER_LAST);
65 small_chunks = (kmp_int32 *)__kmp_allocate(
sizeof(kmp_int32) *
66 kmp_hier_layer_e::LAYER_LAST);
67 large_chunks = (kmp_int64 *)__kmp_allocate(
sizeof(kmp_int64) *
68 kmp_hier_layer_e::LAYER_LAST);
69 layers = (kmp_hier_layer_e *)__kmp_allocate(
sizeof(kmp_hier_layer_e) *
70 kmp_hier_layer_e::LAYER_LAST);
71 capacity = kmp_hier_layer_e::LAYER_LAST;
73 int current_size = size;
74 KMP_DEBUG_ASSERT(current_size < kmp_hier_layer_e::LAYER_LAST);
75 scheds[current_size] = sched;
76 layers[current_size] = layer;
77 small_chunks[current_size] = chunk;
78 large_chunks[current_size] = (kmp_int64)chunk;
86 for (
int i = 0; i < size; ++i) {
88 for (
int j = i + 1; j < size; ++j) {
89 if (layers[j] < layers[switch_index])
92 if (switch_index != i) {
93 kmp_hier_layer_e temp1 = layers[i];
95 kmp_int32 temp3 = small_chunks[i];
96 kmp_int64 temp4 = large_chunks[i];
97 layers[i] = layers[switch_index];
98 scheds[i] = scheds[switch_index];
99 small_chunks[i] = small_chunks[switch_index];
100 large_chunks[i] = large_chunks[switch_index];
101 layers[switch_index] = temp1;
102 scheds[switch_index] = temp2;
103 small_chunks[switch_index] = temp3;
104 large_chunks[switch_index] = temp4;
113 __kmp_free(small_chunks);
114 __kmp_free(large_chunks);
123 } kmp_hier_sched_env_t;
125 extern int __kmp_dispatch_hand_threading;
126 extern kmp_hier_sched_env_t __kmp_hier_scheds;
129 extern int __kmp_hier_max_units[kmp_hier_layer_e::LAYER_LAST + 1];
130 extern int __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LAST + 1];
132 extern int __kmp_dispatch_get_index(
int tid, kmp_hier_layer_e type);
133 extern int __kmp_dispatch_get_id(
int gtid, kmp_hier_layer_e type);
134 extern int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1,
135 kmp_hier_layer_e t2);
136 extern void __kmp_dispatch_free_hierarchies(kmp_team_t *team);
138 template <
typename T>
struct kmp_hier_shared_bdata_t {
139 typedef typename traits_t<T>::signed_t ST;
140 volatile kmp_uint64 val[2];
145 dispatch_shared_info_template<T> sh[2];
148 status[0] = status[1] = 0;
152 sh[0].u.s.iteration = sh[1].u.s.iteration = 0;
154 void set_next_hand_thread(T nlb, T nub, ST nst, kmp_int32 nstatus,
159 status[1 - index] = nstatus;
161 void set_next(T nlb, T nub, ST nst, kmp_int32 nstatus, kmp_uint64 index) {
165 status[1 - index] = nstatus;
166 sh[1 - index].u.s.iteration = 0;
169 kmp_int32 get_next_status(kmp_uint64 index)
const {
170 return status[1 - index];
172 T get_next_lb(kmp_uint64 index)
const {
return lb[1 - index]; }
173 T get_next_ub(kmp_uint64 index)
const {
return ub[1 - index]; }
174 ST get_next_st(kmp_uint64 index)
const {
return st[1 - index]; }
175 dispatch_shared_info_template<T>
volatile *get_next_sh(kmp_uint64 index) {
176 return &(sh[1 - index]);
179 kmp_int32 get_curr_status(kmp_uint64 index)
const {
return status[index]; }
180 T get_curr_lb(kmp_uint64 index)
const {
return lb[index]; }
181 T get_curr_ub(kmp_uint64 index)
const {
return ub[index]; }
182 ST get_curr_st(kmp_uint64 index)
const {
return st[index]; }
183 dispatch_shared_info_template<T>
volatile *get_curr_sh(kmp_uint64 index) {
209 template <
typename T>
class core_barrier_impl {
210 static inline kmp_uint64 get_wait_val(
int num_active) {
212 switch (num_active) {
217 wait_val = 0x010101LL;
220 wait_val = 0x01010101LL;
223 wait_val = 0x0101010101LL;
226 wait_val = 0x010101010101LL;
229 wait_val = 0x01010101010101LL;
232 wait_val = 0x0101010101010101LL;
242 static void reset_private(kmp_int32 num_active,
243 kmp_hier_private_bdata_t *tdata);
244 static void reset_shared(kmp_int32 num_active,
245 kmp_hier_shared_bdata_t<T> *bdata);
246 static void barrier(kmp_int32
id, kmp_hier_shared_bdata_t<T> *bdata,
247 kmp_hier_private_bdata_t *tdata);
250 template <
typename T>
251 void core_barrier_impl<T>::reset_private(kmp_int32 num_active,
252 kmp_hier_private_bdata_t *tdata) {
253 tdata->num_active = num_active;
255 tdata->wait_val[0] = tdata->wait_val[1] = get_wait_val(num_active);
257 template <
typename T>
258 void core_barrier_impl<T>::reset_shared(kmp_int32 num_active,
259 kmp_hier_shared_bdata_t<T> *bdata) {
260 bdata->val[0] = bdata->val[1] = 0LL;
261 bdata->status[0] = bdata->status[1] = 0LL;
263 template <
typename T>
264 void core_barrier_impl<T>::barrier(kmp_int32
id,
265 kmp_hier_shared_bdata_t<T> *bdata,
266 kmp_hier_private_bdata_t *tdata) {
267 kmp_uint64 current_index = tdata->index;
268 kmp_uint64 next_index = 1 - current_index;
269 kmp_uint64 current_wait_value = tdata->wait_val[current_index];
270 kmp_uint64 next_wait_value =
271 (current_wait_value ? 0 : get_wait_val(tdata->num_active));
272 KD_TRACE(10, (
"core_barrier_impl::barrier(): T#%d current_index:%llu " 273 "next_index:%llu curr_wait:%llu next_wait:%llu\n",
274 __kmp_get_gtid(), current_index, next_index, current_wait_value,
276 char v = (current_wait_value ? 0x1 : 0x0);
277 (RCAST(
volatile char *, &(bdata->val[current_index])))[
id] = v;
278 __kmp_wait<kmp_uint64>(&(bdata->val[current_index]), current_wait_value,
279 __kmp_eq<kmp_uint64> USE_ITT_BUILD_ARG(NULL));
280 tdata->wait_val[current_index] = next_wait_value;
281 tdata->index = next_index;
286 template <
typename T>
class counter_barrier_impl {
288 static void reset_private(kmp_int32 num_active,
289 kmp_hier_private_bdata_t *tdata);
290 static void reset_shared(kmp_int32 num_active,
291 kmp_hier_shared_bdata_t<T> *bdata);
292 static void barrier(kmp_int32
id, kmp_hier_shared_bdata_t<T> *bdata,
293 kmp_hier_private_bdata_t *tdata);
296 template <
typename T>
297 void counter_barrier_impl<T>::reset_private(kmp_int32 num_active,
298 kmp_hier_private_bdata_t *tdata) {
299 tdata->num_active = num_active;
301 tdata->wait_val[0] = tdata->wait_val[1] = (kmp_uint64)num_active;
303 template <
typename T>
304 void counter_barrier_impl<T>::reset_shared(kmp_int32 num_active,
305 kmp_hier_shared_bdata_t<T> *bdata) {
306 bdata->val[0] = bdata->val[1] = 0LL;
307 bdata->status[0] = bdata->status[1] = 0LL;
309 template <
typename T>
310 void counter_barrier_impl<T>::barrier(kmp_int32
id,
311 kmp_hier_shared_bdata_t<T> *bdata,
312 kmp_hier_private_bdata_t *tdata) {
313 volatile kmp_int64 *val;
314 kmp_uint64 current_index = tdata->index;
315 kmp_uint64 next_index = 1 - current_index;
316 kmp_uint64 current_wait_value = tdata->wait_val[current_index];
317 kmp_uint64 next_wait_value = current_wait_value + tdata->num_active;
319 KD_TRACE(10, (
"counter_barrier_impl::barrier(): T#%d current_index:%llu " 320 "next_index:%llu curr_wait:%llu next_wait:%llu\n",
321 __kmp_get_gtid(), current_index, next_index, current_wait_value,
323 val = RCAST(
volatile kmp_int64 *, &(bdata->val[current_index]));
324 KMP_TEST_THEN_INC64(val);
325 __kmp_wait<kmp_uint64>(&(bdata->val[current_index]), current_wait_value,
326 __kmp_ge<kmp_uint64> USE_ITT_BUILD_ARG(NULL));
327 tdata->wait_val[current_index] = next_wait_value;
328 tdata->index = next_index;
333 template <
typename T>
struct kmp_hier_top_unit_t {
334 typedef typename traits_t<T>::signed_t ST;
335 typedef typename traits_t<T>::unsigned_t UT;
338 dispatch_private_info_template<T> hier_pr;
339 kmp_hier_top_unit_t<T> *hier_parent;
340 kmp_hier_shared_bdata_t<T> hier_barrier;
342 kmp_int32 get_hier_id()
const {
return hier_pr.hier_id; }
343 void reset_shared_barrier() {
344 KMP_DEBUG_ASSERT(active > 0);
348 if (active >= 2 && active <= 8) {
349 core_barrier_impl<T>::reset_shared(active, &hier_barrier);
351 counter_barrier_impl<T>::reset_shared(active, &hier_barrier);
354 void reset_private_barrier(kmp_hier_private_bdata_t *tdata) {
355 KMP_DEBUG_ASSERT(tdata);
356 KMP_DEBUG_ASSERT(active > 0);
359 if (active >= 2 && active <= 8) {
360 core_barrier_impl<T>::reset_private(active, tdata);
362 counter_barrier_impl<T>::reset_private(active, tdata);
365 void barrier(kmp_int32
id, kmp_hier_private_bdata_t *tdata) {
366 KMP_DEBUG_ASSERT(tdata);
367 KMP_DEBUG_ASSERT(active > 0);
368 KMP_DEBUG_ASSERT(
id >= 0 &&
id < active);
370 tdata->index = 1 - tdata->index;
373 if (active >= 2 && active <= 8) {
374 core_barrier_impl<T>::barrier(
id, &hier_barrier, tdata);
376 counter_barrier_impl<T>::barrier(
id, &hier_barrier, tdata);
380 kmp_int32 get_next_status(kmp_uint64 index)
const {
381 return hier_barrier.get_next_status(index);
383 T get_next_lb(kmp_uint64 index)
const {
384 return hier_barrier.get_next_lb(index);
386 T get_next_ub(kmp_uint64 index)
const {
387 return hier_barrier.get_next_ub(index);
389 ST get_next_st(kmp_uint64 index)
const {
390 return hier_barrier.get_next_st(index);
392 dispatch_shared_info_template<T>
volatile *get_next_sh(kmp_uint64 index) {
393 return hier_barrier.get_next_sh(index);
396 kmp_int32 get_curr_status(kmp_uint64 index)
const {
397 return hier_barrier.get_curr_status(index);
399 T get_curr_lb(kmp_uint64 index)
const {
400 return hier_barrier.get_curr_lb(index);
402 T get_curr_ub(kmp_uint64 index)
const {
403 return hier_barrier.get_curr_ub(index);
405 ST get_curr_st(kmp_uint64 index)
const {
406 return hier_barrier.get_curr_st(index);
408 dispatch_shared_info_template<T>
volatile *get_curr_sh(kmp_uint64 index) {
409 return hier_barrier.get_curr_sh(index);
412 void set_next_hand_thread(T lb, T ub, ST st, kmp_int32 status,
414 hier_barrier.set_next_hand_thread(lb, ub, st, status, index);
416 void set_next(T lb, T ub, ST st, kmp_int32 status, kmp_uint64 index) {
417 hier_barrier.set_next(lb, ub, st, status, index);
419 dispatch_private_info_template<T> *get_my_pr() {
return &hier_pr; }
420 kmp_hier_top_unit_t<T> *get_parent() {
return hier_parent; }
421 dispatch_private_info_template<T> *get_parent_pr() {
422 return &(hier_parent->hier_pr);
425 kmp_int32 is_active()
const {
return active; }
426 kmp_int32 get_num_active()
const {
return active; }
430 (
" kmp_hier_top_unit_t: active:%d pr:%p lb:%d ub:%d st:%d tc:%d\n",
431 active, &hier_pr, hier_pr.u.p.lb, hier_pr.u.p.ub, hier_pr.u.p.st,
437 template <
typename T>
struct kmp_hier_layer_info_t {
439 kmp_hier_layer_e type;
441 typename traits_t<T>::signed_t chunk;
446 const char *t = __kmp_get_hier_str(type);
449 (
" kmp_hier_layer_info_t: num_active:%d type:%s sched:%d chunk:%d " 451 num_active, t, sched, chunk, length));
468 template <
typename T>
struct kmp_hier_t {
470 typedef typename traits_t<T>::unsigned_t UT;
471 typedef typename traits_t<T>::signed_t ST;
474 int next_recurse(
ident_t *loc,
int gtid, kmp_hier_top_unit_t<T> *current,
475 kmp_int32 *p_last, T *p_lb, T *p_ub, ST *p_st,
476 kmp_int32 previous_id,
int hier_level) {
478 kmp_info_t *th = __kmp_threads[gtid];
479 auto parent = current->get_parent();
480 bool last_layer = (hier_level == get_num_layers() - 1);
481 KMP_DEBUG_ASSERT(th);
482 kmp_hier_private_bdata_t *tdata = &(th->th.th_hier_bar_data[hier_level]);
483 KMP_DEBUG_ASSERT(current);
484 KMP_DEBUG_ASSERT(hier_level >= 0);
485 KMP_DEBUG_ASSERT(hier_level < get_num_layers());
486 KMP_DEBUG_ASSERT(tdata);
487 KMP_DEBUG_ASSERT(parent || last_layer);
490 1, (
"kmp_hier_t.next_recurse(): T#%d (%d) called\n", gtid, hier_level));
492 T hier_id = (T)current->get_hier_id();
494 if (previous_id == 0) {
495 KD_TRACE(1, (
"kmp_hier_t.next_recurse(): T#%d (%d) is master of unit\n",
497 kmp_int32 contains_last;
501 dispatch_shared_info_template<T>
volatile *my_sh;
502 dispatch_private_info_template<T> *my_pr;
507 (
"kmp_hier_t.next_recurse(): T#%d (%d) using top level sh\n",
509 my_sh =
reinterpret_cast<dispatch_shared_info_template<T>
volatile *
>(
510 th->th.th_dispatch->th_dispatch_sh_current);
511 nproc = (T)get_top_level_nproc();
515 KD_TRACE(10, (
"kmp_hier_t.next_recurse(): T#%d (%d) using hier sh\n",
518 parent->get_curr_sh(th->th.th_hier_bar_data[hier_level + 1].index);
519 nproc = (T)parent->get_num_active();
521 my_pr = current->get_my_pr();
522 KMP_DEBUG_ASSERT(my_sh);
523 KMP_DEBUG_ASSERT(my_pr);
524 enum sched_type schedule = get_sched(hier_level);
525 ST chunk = (ST)get_chunk(hier_level);
526 status = __kmp_dispatch_next_algorithm<T>(gtid, my_pr, my_sh,
527 &contains_last, &my_lb, &my_ub,
528 &my_st, nproc, hier_id);
531 (
"kmp_hier_t.next_recurse(): T#%d (%d) next_pr_sh() returned %d\n",
532 gtid, hier_level, status));
535 if (status == 0 && !last_layer) {
536 status = next_recurse(loc, gtid, parent, &contains_last, &my_lb, &my_ub,
537 &my_st, hier_id, hier_level + 1);
540 (
"kmp_hier_t.next_recurse(): T#%d (%d) hier_next() returned %d\n",
541 gtid, hier_level, status));
543 kmp_hier_private_bdata_t *upper_tdata =
544 &(th->th.th_hier_bar_data[hier_level + 1]);
545 my_sh = parent->get_curr_sh(upper_tdata->index);
546 KD_TRACE(10, (
"kmp_hier_t.next_recurse(): T#%d (%d) about to init\n",
548 __kmp_dispatch_init_algorithm(loc, gtid, my_pr, schedule,
549 parent->get_curr_lb(upper_tdata->index),
550 parent->get_curr_ub(upper_tdata->index),
551 parent->get_curr_st(upper_tdata->index),
555 chunk, nproc, hier_id);
556 status = __kmp_dispatch_next_algorithm<T>(
557 gtid, my_pr, my_sh, &contains_last, &my_lb, &my_ub, &my_st, nproc,
560 KD_TRACE(10, (
"kmp_hier_t.next_recurse(): T#%d (%d) status not 1 " 567 current->set_next(my_lb, my_ub, my_st, status, tdata->index);
576 if (last_layer || parent->hier_pr.flags.contains_last) {
577 KD_TRACE(10, (
"kmp_hier_t.next_recurse(): T#%d (%d) Setting this pr " 578 "to contain last.\n",
580 current->hier_pr.flags.contains_last = contains_last;
582 if (!current->hier_pr.flags.contains_last)
583 contains_last = FALSE;
586 *p_last = contains_last;
588 if (hier_level > 0 || !__kmp_dispatch_hand_threading) {
590 (
"kmp_hier_t.next_recurse(): T#%d (%d) going into barrier.\n",
592 current->barrier(previous_id, tdata);
594 (
"kmp_hier_t.next_recurse(): T#%d (%d) released and exit %d\n",
595 gtid, hier_level, current->get_curr_status(tdata->index)));
597 KMP_DEBUG_ASSERT(previous_id == 0);
600 return current->get_curr_status(tdata->index);
608 kmp_hier_layer_info_t<T> *info;
609 kmp_hier_top_unit_t<T> **layers;
612 for (
int i = 0; i < num_layers; ++i)
613 if (layers[i] != NULL) {
614 __kmp_free(layers[i]);
616 if (layers != NULL) {
628 bool need_to_reallocate(
int n,
const kmp_hier_layer_e *new_layers,
630 const ST *new_chunks)
const {
631 if (!valid || layers == NULL || info == NULL ||
632 traits_t<T>::type_size != type_size || n != num_layers)
634 for (
int i = 0; i < n; ++i) {
635 if (info[i].type != new_layers[i])
637 if (info[i].sched != new_scheds[i])
639 if (info[i].chunk != new_chunks[i])
649 void allocate_hier(
int n,
const kmp_hier_layer_e *new_layers,
650 const enum sched_type *new_scheds,
const ST *new_chunks) {
652 if (!need_to_reallocate(n, new_layers, new_scheds, new_chunks)) {
655 (
"kmp_hier_t<T>::allocate_hier: T#0 do not need to reallocate\n"));
656 for (
int i = 0; i < n; ++i) {
657 info[i].num_active = 0;
658 for (
int j = 0; j < get_length(i); ++j)
659 layers[i][j].active = 0;
663 KD_TRACE(10, (
"kmp_hier_t<T>::allocate_hier: T#0 full alloc\n"));
665 type_size = traits_t<T>::type_size;
667 info = (kmp_hier_layer_info_t<T> *)__kmp_allocate(
668 sizeof(kmp_hier_layer_info_t<T>) * n);
669 layers = (kmp_hier_top_unit_t<T> **)__kmp_allocate(
670 sizeof(kmp_hier_top_unit_t<T> *) * n);
671 for (
int i = 0; i < n; ++i) {
673 kmp_hier_layer_e layer = new_layers[i];
674 info[i].num_active = 0;
675 info[i].type = layer;
676 info[i].sched = new_scheds[i];
677 info[i].chunk = new_chunks[i];
678 max = __kmp_hier_max_units[layer + 1];
681 KMP_WARNING(HierSchedInvalid, __kmp_get_hier_str(layer));
685 info[i].length = max;
686 layers[i] = (kmp_hier_top_unit_t<T> *)__kmp_allocate(
687 sizeof(kmp_hier_top_unit_t<T>) * max);
688 for (
int j = 0; j < max; ++j) {
689 layers[i][j].active = 0;
705 int next(
ident_t *loc,
int gtid, dispatch_private_info_template<T> *pr,
706 kmp_int32 *p_last, T *p_lb, T *p_ub, ST *p_st) {
708 kmp_int32 contains_last = 0;
709 kmp_info_t *th = __kmp_threads[gtid];
710 kmp_hier_private_bdata_t *tdata = &(th->th.th_hier_bar_data[0]);
711 auto parent = pr->get_parent();
712 KMP_DEBUG_ASSERT(parent);
713 KMP_DEBUG_ASSERT(th);
714 KMP_DEBUG_ASSERT(tdata);
715 KMP_DEBUG_ASSERT(parent);
716 T nproc = (T)parent->get_num_active();
717 T unit_id = (T)pr->get_hier_id();
720 (
"kmp_hier_t.next(): T#%d THREAD LEVEL nproc:%d unit_id:%d called\n",
721 gtid, nproc, unit_id));
730 if (__kmp_dispatch_hand_threading) {
732 (
"kmp_hier_t.next(): T#%d THREAD LEVEL using hand threading\n",
738 auto sh = &(parent->hier_barrier.sh[0]);
739 KMP_DEBUG_ASSERT(sh);
740 status = __kmp_dispatch_next_algorithm<T>(
741 gtid, pr, sh, &contains_last, p_lb, p_ub, p_st, nproc, unit_id);
746 status = next_recurse(loc, gtid, parent, &contains_last, p_lb, p_ub,
749 __kmp_dispatch_init_algorithm(loc, gtid, pr, pr->schedule,
750 parent->get_next_lb(tdata->index),
751 parent->get_next_ub(tdata->index),
752 parent->get_next_st(tdata->index),
756 pr->u.p.parm1, nproc, unit_id);
757 sh->u.s.iteration = 0;
758 status = __kmp_dispatch_next_algorithm<T>(
759 gtid, pr, sh, &contains_last, p_lb, p_ub, p_st, nproc,
763 (
"kmp_hier_t.next(): T#%d THREAD LEVEL status == 0 " 769 }
else if (status == 2) {
770 KD_TRACE(10, (
"kmp_hier_t.next(): T#%d THREAD LEVEL status == 2 " 777 parent->set_next_hand_thread(*p_lb, *p_ub, *p_st, status, tdata->index);
779 parent->barrier(pr->get_hier_id(), tdata);
781 *p_lb = parent->get_curr_lb(tdata->index);
782 *p_ub = parent->get_curr_ub(tdata->index);
783 *p_st = parent->get_curr_st(tdata->index);
784 status = parent->get_curr_status(tdata->index);
789 auto sh = parent->get_curr_sh(tdata->index);
790 KMP_DEBUG_ASSERT(sh);
791 status = __kmp_dispatch_next_algorithm<T>(
792 gtid, pr, sh, &contains_last, p_lb, p_ub, p_st, nproc, unit_id);
794 (
"kmp_hier_t.next(): T#%d THREAD LEVEL next_algorithm status:%d " 795 "contains_last:%d p_lb:%d p_ub:%d p_st:%d\n",
796 gtid, status, contains_last, *p_lb, *p_ub, *p_st));
801 status = next_recurse(loc, gtid, parent, &contains_last, p_lb, p_ub,
804 sh = parent->get_curr_sh(tdata->index);
805 __kmp_dispatch_init_algorithm(loc, gtid, pr, pr->schedule,
806 parent->get_curr_lb(tdata->index),
807 parent->get_curr_ub(tdata->index),
808 parent->get_curr_st(tdata->index),
812 pr->u.p.parm1, nproc, unit_id);
813 status = __kmp_dispatch_next_algorithm<T>(
814 gtid, pr, sh, &contains_last, p_lb, p_ub, p_st, nproc, unit_id);
816 KD_TRACE(10, (
"kmp_hier_t.next(): T#%d THREAD LEVEL status == 0 " 822 }
else if (status == 2) {
823 KD_TRACE(10, (
"kmp_hier_t.next(): T#%d THREAD LEVEL status == 2 " 831 if (contains_last && !parent->hier_pr.flags.contains_last) {
832 KD_TRACE(10, (
"kmp_hier_t.next(): T#%d THREAD LEVEL resetting " 833 "contains_last to FALSE\n",
835 contains_last = FALSE;
838 *p_last = contains_last;
839 KD_TRACE(10, (
"kmp_hier_t.next(): T#%d THREAD LEVEL exit status %d\n", gtid,
845 kmp_hier_layer_e get_type(
int level)
const {
846 KMP_DEBUG_ASSERT(level >= 0);
847 KMP_DEBUG_ASSERT(level < num_layers);
848 return info[level].type;
852 KMP_DEBUG_ASSERT(level >= 0);
853 KMP_DEBUG_ASSERT(level < num_layers);
854 return info[level].sched;
857 ST get_chunk(
int level)
const {
858 KMP_DEBUG_ASSERT(level >= 0);
859 KMP_DEBUG_ASSERT(level < num_layers);
860 return info[level].chunk;
863 int get_num_active(
int level)
const {
864 KMP_DEBUG_ASSERT(level >= 0);
865 KMP_DEBUG_ASSERT(level < num_layers);
866 return info[level].num_active;
869 int get_length(
int level)
const {
870 KMP_DEBUG_ASSERT(level >= 0);
871 KMP_DEBUG_ASSERT(level < num_layers);
872 return info[level].length;
875 kmp_hier_top_unit_t<T> *get_unit(
int level,
int index) {
876 KMP_DEBUG_ASSERT(level >= 0);
877 KMP_DEBUG_ASSERT(level < num_layers);
878 KMP_DEBUG_ASSERT(index >= 0);
879 KMP_DEBUG_ASSERT(index < get_length(level));
880 return &(layers[level][index]);
883 int get_num_layers()
const {
return num_layers; }
887 int get_top_level_nproc()
const {
return top_level_nproc; }
889 bool is_valid()
const {
return valid; }
892 KD_TRACE(10, (
"kmp_hier_t:\n"));
893 for (
int i = num_layers - 1; i >= 0; --i) {
894 KD_TRACE(10, (
"Info[%d] = ", i));
897 for (
int i = num_layers - 1; i >= 0; --i) {
898 KD_TRACE(10, (
"Layer[%d] =\n", i));
899 for (
int j = 0; j < info[i].length; ++j) {
900 layers[i][j].print();
906 template <
typename T>
907 void __kmp_dispatch_init_hierarchy(
ident_t *loc,
int n,
908 kmp_hier_layer_e *new_layers,
910 typename traits_t<T>::signed_t *new_chunks,
912 typename traits_t<T>::signed_t st) {
913 typedef typename traits_t<T>::signed_t ST;
914 typedef typename traits_t<T>::unsigned_t UT;
915 int tid, gtid, num_hw_threads, num_threads_per_layer1, active;
919 dispatch_private_info_template<T> *pr;
920 dispatch_shared_info_template<T>
volatile *sh;
921 gtid = __kmp_entry_gtid();
922 tid = __kmp_tid_from_gtid(gtid);
924 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d called: %d layer(s)\n",
926 for (
int i = 0; i < n; ++i) {
927 const char *layer = __kmp_get_hier_str(new_layers[i]);
928 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d: new_layers[%d] = %s, " 929 "new_scheds[%d] = %d, new_chunks[%d] = %u\n",
930 gtid, i, layer, i, (
int)new_scheds[i], i, new_chunks[i]));
933 KMP_DEBUG_ASSERT(n > 0);
934 KMP_DEBUG_ASSERT(new_layers);
935 KMP_DEBUG_ASSERT(new_scheds);
936 KMP_DEBUG_ASSERT(new_chunks);
937 if (!TCR_4(__kmp_init_parallel))
938 __kmp_parallel_initialize();
940 __kmp_resume_if_soft_paused();
943 th = __kmp_threads[gtid];
944 team = th->th.th_team;
945 active = !team->t.t_serialized;
946 th->th.th_ident = loc;
947 num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
949 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d not active parallel. " 950 "Using normal dispatch functions.\n",
952 pr =
reinterpret_cast<dispatch_private_info_template<T> *
>(
953 th->th.th_dispatch->th_disp_buffer);
954 KMP_DEBUG_ASSERT(pr);
955 pr->flags.use_hier = FALSE;
956 pr->flags.contains_last = FALSE;
959 KMP_DEBUG_ASSERT(th->th.th_dispatch ==
960 &th->th.th_team->t.t_dispatch[th->th.th_info.ds.ds_tid]);
962 my_buffer_index = th->th.th_dispatch->th_disp_index;
963 pr =
reinterpret_cast<dispatch_private_info_template<T> *
>(
965 ->th_disp_buffer[my_buffer_index % __kmp_dispatch_num_buffers]);
966 sh =
reinterpret_cast<dispatch_shared_info_template<T>
volatile *
>(
967 &team->t.t_disp_buffer[my_buffer_index % __kmp_dispatch_num_buffers]);
968 KMP_DEBUG_ASSERT(pr);
969 KMP_DEBUG_ASSERT(sh);
970 pr->flags.use_hier = TRUE;
973 if (__kmp_tid_from_gtid(gtid) == 0) {
974 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d pr:%p sh:%p allocating " 977 if (sh->hier == NULL) {
978 sh->hier = (kmp_hier_t<T> *)__kmp_allocate(
sizeof(kmp_hier_t<T>));
980 sh->hier->allocate_hier(n, new_layers, new_scheds, new_chunks);
981 sh->u.s.iteration = 0;
983 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
985 kmp_hier_t<T> *hier = sh->hier;
986 if (!sh->hier->is_valid()) {
987 pr->flags.use_hier = FALSE;
992 if (th->th.th_hier_bar_data == NULL) {
993 th->th.th_hier_bar_data = (kmp_hier_private_bdata_t *)__kmp_allocate(
994 sizeof(kmp_hier_private_bdata_t) * kmp_hier_layer_e::LAYER_LAST);
999 for (
int i = 0; i < n; ++i) {
1000 int index = __kmp_dispatch_get_index(tid, hier->get_type(i));
1001 kmp_hier_top_unit_t<T> *my_unit = hier->get_unit(i, index);
1004 pr->hier_parent = my_unit;
1006 if (my_unit->is_active()) {
1007 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d my_unit (%p) " 1008 "is already active (%d)\n",
1009 gtid, my_unit, my_unit->active));
1010 KMP_TEST_THEN_INC32(&(my_unit->active));
1014 if (KMP_COMPARE_AND_STORE_ACQ32(&(my_unit->active), 0, 1)) {
1018 my_unit->get_my_pr()->hier_id =
1019 index % __kmp_dispatch_get_t1_per_t2(hier->get_type(i),
1020 hier->get_type(i + 1));
1021 int parent_index = __kmp_dispatch_get_index(tid, hier->get_type(i + 1));
1022 my_unit->hier_parent = hier->get_unit(i + 1, parent_index);
1025 my_unit->get_my_pr()->hier_id =
1026 index % __kmp_dispatch_get_t1_per_t2(hier->get_type(i),
1027 kmp_hier_layer_e::LAYER_LOOP);
1028 KMP_TEST_THEN_INC32(&(hier->top_level_nproc));
1029 my_unit->hier_parent =
nullptr;
1033 my_unit->get_my_pr()->u.p.tc = 0;
1035 KMP_TEST_THEN_INC32(&(hier->info[i].num_active));
1036 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d my_unit (%p) " 1037 "incrementing num_active\n",
1040 KMP_TEST_THEN_INC32(&(my_unit->active));
1045 num_threads_per_layer1 = __kmp_dispatch_get_t1_per_t2(
1046 kmp_hier_layer_e::LAYER_THREAD, hier->get_type(0));
1047 pr->hier_id = tid % num_threads_per_layer1;
1050 if (tid >= num_hw_threads)
1051 pr->hier_id += ((tid / num_hw_threads) * num_threads_per_layer1);
1053 10, (
"__kmp_dispatch_init_hierarchy: T#%d setting lowest hier_id to %d\n",
1054 gtid, pr->hier_id));
1056 pr->flags.contains_last = FALSE;
1057 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
1062 int prev_id = pr->get_hier_id();
1063 for (
int i = 0; i < n; ++i) {
1066 int index = __kmp_dispatch_get_index(tid, hier->get_type(i));
1067 kmp_hier_top_unit_t<T> *my_unit = hier->get_unit(i, index);
1069 KD_TRACE(10, (
"__kmp_dispatch_init_hierarchy: T#%d (%d) prev_id is 0\n",
1071 my_unit->reset_shared_barrier();
1072 my_unit->hier_pr.flags.contains_last = FALSE;
1077 __kmp_dispatch_init_algorithm<T>(
1078 loc, gtid, my_unit->get_my_pr(), hier->get_sched(i), lb, ub, st,
1082 hier->get_chunk(i), hier->get_num_active(i), my_unit->get_hier_id());
1084 prev_id = my_unit->get_hier_id();
1087 kmp_hier_top_unit_t<T> *unit = pr->hier_parent;
1088 for (
int i = 0; i < n && unit; ++i, unit = unit->get_parent()) {
1089 kmp_hier_private_bdata_t *tdata = &(th->th.th_hier_bar_data[i]);
1090 unit->reset_private_barrier(tdata);
1092 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
1095 if (__kmp_tid_from_gtid(gtid) == 0) {
1096 for (
int i = 0; i < n; ++i) {
1098 (
"__kmp_dispatch_init_hierarchy: T#%d active count[%d] = %d\n",
1099 gtid, i, hier->get_num_active(i)));
1103 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);