1	/*
2	* kmp_affinity.cpp -- affinity management
3	*/
4
5	//===----------------------------------------------------------------------===//
6	//
7	// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8	// See https://llvm.org/LICENSE.txt for license information.
9	// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10	//
11	//===----------------------------------------------------------------------===//
12
13	#include "kmp.h"
14	#include "kmp_affinity.h"
15	#include "kmp_i18n.h"
16	#include "kmp_io.h"
17	#include "kmp_str.h"
18	#include "kmp_wrapper_getpid.h"
19	#if KMP_USE_HIER_SCHED
20	#include "kmp_dispatch_hier.h"
21	#endif
22	#if KMP_USE_HWLOC
23	// Copied from hwloc
24	#define HWLOC_GROUP_KIND_INTEL_MODULE 102
25	#define HWLOC_GROUP_KIND_INTEL_TILE 103
26	#define HWLOC_GROUP_KIND_INTEL_DIE 104
27	#define HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP 220
28	#endif
29	#include <ctype.h>
30
31	// The machine topology
32	kmp_topology_t *__kmp_topology = nullptr;
33	// KMP_HW_SUBSET environment variable
34	kmp_hw_subset_t *__kmp_hw_subset = nullptr;
35
36	// Store the real or imagined machine hierarchy here
37	static hierarchy_info machine_hierarchy;
38
39	void __kmp_cleanup_hierarchy() { machine_hierarchy.fini(); }
40
41	#if KMP_AFFINITY_SUPPORTED
42	// Helper class to see if place lists further restrict the fullMask
43	class kmp_full_mask_modifier_t {
44	kmp_affin_mask_t *mask;
45
46	public:
47	kmp_full_mask_modifier_t() {
48	KMP_CPU_ALLOC(mask);
49	KMP_CPU_ZERO(mask);
50	}
51	~kmp_full_mask_modifier_t() {
52	KMP_CPU_FREE(mask);
53	mask = nullptr;
54	}
55	void include(const kmp_affin_mask_t *other) { KMP_CPU_UNION(mask, other); }
56	// If the new full mask is different from the current full mask,
57	// then switch them. Returns true if full mask was affected, false otherwise.
58	bool restrict_to_mask() {
59	// See if the new mask further restricts or changes the full mask
60	if (KMP_CPU_EQUAL(__kmp_affin_fullMask, mask) || KMP_CPU_ISEMPTY(mask))
61	return false;
62	return __kmp_topology->restrict_to_mask(mask);
63	}
64	};
65
66	static inline const char *
67	__kmp_get_affinity_env_var(const kmp_affinity_t &affinity,
68	bool for_binding = false) {
69	if (affinity.flags.omp_places) {
70	if (for_binding)
71	return "OMP_PROC_BIND";
72	return "OMP_PLACES";
73	}
74	return affinity.env_var;
75	}
76	#endif // KMP_AFFINITY_SUPPORTED
77
78	void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
79	kmp_uint32 depth;
80	// The test below is true if affinity is available, but set to "none". Need to
81	// init on first use of hierarchical barrier.
82	if (TCR_1(machine_hierarchy.uninitialized))
83	machine_hierarchy.init(num_addrs: nproc);
84
85	// Adjust the hierarchy in case num threads exceeds original
86	if (nproc > machine_hierarchy.base_num_threads)
87	machine_hierarchy.resize(nproc);
88
89	depth = machine_hierarchy.depth;
90	KMP_DEBUG_ASSERT(depth > 0);
91
92	thr_bar->depth = depth;
93	__kmp_type_convert(src: machine_hierarchy.numPerLevel[0] - 1,
94	dest: &(thr_bar->base_leaf_kids));
95	thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
96	}
97
98	static int nCoresPerPkg, nPackages;
99	static int __kmp_nThreadsPerCore;
100	#ifndef KMP_DFLT_NTH_CORES
101	static int __kmp_ncores;
102	#endif
103
104	const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural) {
105	switch (type) {
106	case KMP_HW_SOCKET:
107	return ((plural) ? KMP_I18N_STR(Sockets) : KMP_I18N_STR(Socket));
108	case KMP_HW_DIE:
109	return ((plural) ? KMP_I18N_STR(Dice) : KMP_I18N_STR(Die));
110	case KMP_HW_MODULE:
111	return ((plural) ? KMP_I18N_STR(Modules) : KMP_I18N_STR(Module));
112	case KMP_HW_TILE:
113	return ((plural) ? KMP_I18N_STR(Tiles) : KMP_I18N_STR(Tile));
114	case KMP_HW_NUMA:
115	return ((plural) ? KMP_I18N_STR(NumaDomains) : KMP_I18N_STR(NumaDomain));
116	case KMP_HW_L3:
117	return ((plural) ? KMP_I18N_STR(L3Caches) : KMP_I18N_STR(L3Cache));
118	case KMP_HW_L2:
119	return ((plural) ? KMP_I18N_STR(L2Caches) : KMP_I18N_STR(L2Cache));
120	case KMP_HW_L1:
121	return ((plural) ? KMP_I18N_STR(L1Caches) : KMP_I18N_STR(L1Cache));
122	case KMP_HW_LLC:
123	return ((plural) ? KMP_I18N_STR(LLCaches) : KMP_I18N_STR(LLCache));
124	case KMP_HW_CORE:
125	return ((plural) ? KMP_I18N_STR(Cores) : KMP_I18N_STR(Core));
126	case KMP_HW_THREAD:
127	return ((plural) ? KMP_I18N_STR(Threads) : KMP_I18N_STR(Thread));
128	case KMP_HW_PROC_GROUP:
129	return ((plural) ? KMP_I18N_STR(ProcGroups) : KMP_I18N_STR(ProcGroup));
130	case KMP_HW_UNKNOWN:
131	case KMP_HW_LAST:
132	return KMP_I18N_STR(Unknown);
133	}
134	KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
135	KMP_BUILTIN_UNREACHABLE;
136	}
137
138	const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural) {
139	switch (type) {
140	case KMP_HW_SOCKET:
141	return ((plural) ? "sockets" : "socket");
142	case KMP_HW_DIE:
143	return ((plural) ? "dice" : "die");
144	case KMP_HW_MODULE:
145	return ((plural) ? "modules" : "module");
146	case KMP_HW_TILE:
147	return ((plural) ? "tiles" : "tile");
148	case KMP_HW_NUMA:
149	return ((plural) ? "numa_domains" : "numa_domain");
150	case KMP_HW_L3:
151	return ((plural) ? "l3_caches" : "l3_cache");
152	case KMP_HW_L2:
153	return ((plural) ? "l2_caches" : "l2_cache");
154	case KMP_HW_L1:
155	return ((plural) ? "l1_caches" : "l1_cache");
156	case KMP_HW_LLC:
157	return ((plural) ? "ll_caches" : "ll_cache");
158	case KMP_HW_CORE:
159	return ((plural) ? "cores" : "core");
160	case KMP_HW_THREAD:
161	return ((plural) ? "threads" : "thread");
162	case KMP_HW_PROC_GROUP:
163	return ((plural) ? "proc_groups" : "proc_group");
164	case KMP_HW_UNKNOWN:
165	case KMP_HW_LAST:
166	return ((plural) ? "unknowns" : "unknown");
167	}
168	KMP_ASSERT2(false, "Unhandled kmp_hw_t enumeration");
169	KMP_BUILTIN_UNREACHABLE;
170	}
171
172	const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type) {
173	switch (type) {
174	case KMP_HW_CORE_TYPE_UNKNOWN:
175	case KMP_HW_MAX_NUM_CORE_TYPES:
176	return "unknown";
177	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
178	case KMP_HW_CORE_TYPE_ATOM:
179	return "Intel Atom(R) processor";
180	case KMP_HW_CORE_TYPE_CORE:
181	return "Intel(R) Core(TM) processor";
182	#endif
183	}
184	KMP_ASSERT2(false, "Unhandled kmp_hw_core_type_t enumeration");
185	KMP_BUILTIN_UNREACHABLE;
186	}
187
188	#if KMP_AFFINITY_SUPPORTED
189	// If affinity is supported, check the affinity
190	// verbose and warning flags before printing warning
191	#define KMP_AFF_WARNING(s, ...) \
192	if (s.flags.verbose || (s.flags.warnings && (s.type != affinity_none))) { \
193	KMP_WARNING(__VA_ARGS__); \
194	}
195	#else
196	#define KMP_AFF_WARNING(s, ...) KMP_WARNING(__VA_ARGS__)
197	#endif
198
199	////////////////////////////////////////////////////////////////////////////////
200	// kmp_hw_thread_t methods
201	int kmp_hw_thread_t::compare_ids(const void *a, const void *b) {
202	const kmp_hw_thread_t *ahwthread = (const kmp_hw_thread_t *)a;
203	const kmp_hw_thread_t *bhwthread = (const kmp_hw_thread_t *)b;
204	int depth = __kmp_topology->get_depth();
205	for (int level = 0; level < depth; ++level) {
206	if (ahwthread->ids[level] < bhwthread->ids[level])
207	return -1;
208	else if (ahwthread->ids[level] > bhwthread->ids[level])
209	return 1;
210	}
211	if (ahwthread->os_id < bhwthread->os_id)
212	return -1;
213	else if (ahwthread->os_id > bhwthread->os_id)
214	return 1;
215	return 0;
216	}
217
218	#if KMP_AFFINITY_SUPPORTED
219	int kmp_hw_thread_t::compare_compact(const void *a, const void *b) {
220	int i;
221	const kmp_hw_thread_t *aa = (const kmp_hw_thread_t *)a;
222	const kmp_hw_thread_t *bb = (const kmp_hw_thread_t *)b;
223	int depth = __kmp_topology->get_depth();
224	int compact = __kmp_topology->compact;
225	KMP_DEBUG_ASSERT(compact >= 0);
226	KMP_DEBUG_ASSERT(compact <= depth);
227	for (i = 0; i < compact; i++) {
228	int j = depth - i - 1;
229	if (aa->sub_ids[j] < bb->sub_ids[j])
230	return -1;
231	if (aa->sub_ids[j] > bb->sub_ids[j])
232	return 1;
233	}
234	for (; i < depth; i++) {
235	int j = i - compact;
236	if (aa->sub_ids[j] < bb->sub_ids[j])
237	return -1;
238	if (aa->sub_ids[j] > bb->sub_ids[j])
239	return 1;
240	}
241	return 0;
242	}
243	#endif
244
245	void kmp_hw_thread_t::print() const {
246	int depth = __kmp_topology->get_depth();
247	printf(format: "%4d ", os_id);
248	for (int i = 0; i < depth; ++i) {
249	printf(format: "%4d ", ids[i]);
250	}
251	if (attrs) {
252	if (attrs.is_core_type_valid())
253	printf(format: " (%s)", __kmp_hw_get_core_type_string(type: attrs.get_core_type()));
254	if (attrs.is_core_eff_valid())
255	printf(format: " (eff=%d)", attrs.get_core_eff());
256	}
257	if (leader)
258	printf(format: " (leader)");
259	printf(format: "\n");
260	}
261
262	////////////////////////////////////////////////////////////////////////////////
263	// kmp_topology_t methods
264
265	// Add a layer to the topology based on the ids. Assume the topology
266	// is perfectly nested (i.e., so no object has more than one parent)
267	void kmp_topology_t::_insert_layer(kmp_hw_t type, const int *ids) {
268	// Figure out where the layer should go by comparing the ids of the current
269	// layers with the new ids
270	int target_layer;
271	int previous_id = kmp_hw_thread_t::UNKNOWN_ID;
272	int previous_new_id = kmp_hw_thread_t::UNKNOWN_ID;
273
274	// Start from the highest layer and work down to find target layer
275	// If new layer is equal to another layer then put the new layer above
276	for (target_layer = 0; target_layer < depth; ++target_layer) {
277	bool layers_equal = true;
278	bool strictly_above_target_layer = false;
279	for (int i = 0; i < num_hw_threads; ++i) {
280	int id = hw_threads[i].ids[target_layer];
281	int new_id = ids[i];
282	if (id != previous_id && new_id == previous_new_id) {
283	// Found the layer we are strictly above
284	strictly_above_target_layer = true;
285	layers_equal = false;
286	break;
287	} else if (id == previous_id && new_id != previous_new_id) {
288	// Found a layer we are below. Move to next layer and check.
289	layers_equal = false;
290	break;
291	}
292	previous_id = id;
293	previous_new_id = new_id;
294	}
295	if (strictly_above_target_layer || layers_equal)
296	break;
297	}
298
299	// Found the layer we are above. Now move everything to accommodate the new
300	// layer. And put the new ids and type into the topology.
301	for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
302	types[j] = types[i];
303	types[target_layer] = type;
304	for (int k = 0; k < num_hw_threads; ++k) {
305	for (int i = depth - 1, j = depth; i >= target_layer; --i, --j)
306	hw_threads[k].ids[j] = hw_threads[k].ids[i];
307	hw_threads[k].ids[target_layer] = ids[k];
308	}
309	equivalent[type] = type;
310	depth++;
311	}
312
313	#if KMP_GROUP_AFFINITY
314	// Insert the Windows Processor Group structure into the topology
315	void kmp_topology_t::_insert_windows_proc_groups() {
316	// Do not insert the processor group structure for a single group
317	if (__kmp_num_proc_groups == 1)
318	return;
319	kmp_affin_mask_t *mask;
320	int *ids = (int *)__kmp_allocate(sizeof(int) * num_hw_threads);
321	KMP_CPU_ALLOC(mask);
322	for (int i = 0; i < num_hw_threads; ++i) {
323	KMP_CPU_ZERO(mask);
324	KMP_CPU_SET(hw_threads[i].os_id, mask);
325	ids[i] = __kmp_get_proc_group(mask);
326	}
327	KMP_CPU_FREE(mask);
328	_insert_layer(KMP_HW_PROC_GROUP, ids);
329	__kmp_free(ids);
330
331	// sort topology after adding proc groups
332	__kmp_topology->sort_ids();
333	}
334	#endif
335
336	// Remove layers that don't add information to the topology.
337	// This is done by having the layer take on the id = UNKNOWN_ID (-1)
338	void kmp_topology_t::_remove_radix1_layers() {
339	int preference[KMP_HW_LAST];
340	int top_index1, top_index2;
341	// Set up preference associative array
342	preference[KMP_HW_SOCKET] = 110;
343	preference[KMP_HW_PROC_GROUP] = 100;
344	preference[KMP_HW_CORE] = 95;
345	preference[KMP_HW_THREAD] = 90;
346	preference[KMP_HW_NUMA] = 85;
347	preference[KMP_HW_DIE] = 80;
348	preference[KMP_HW_TILE] = 75;
349	preference[KMP_HW_MODULE] = 73;
350	preference[KMP_HW_L3] = 70;
351	preference[KMP_HW_L2] = 65;
352	preference[KMP_HW_L1] = 60;
353	preference[KMP_HW_LLC] = 5;
354	top_index1 = 0;
355	top_index2 = 1;
356	while (top_index1 < depth - 1 && top_index2 < depth) {
357	kmp_hw_t type1 = types[top_index1];
358	kmp_hw_t type2 = types[top_index2];
359	KMP_ASSERT_VALID_HW_TYPE(type1);
360	KMP_ASSERT_VALID_HW_TYPE(type2);
361	// Do not allow the three main topology levels (sockets, cores, threads) to
362	// be compacted down
363	if ((type1 == KMP_HW_THREAD || type1 == KMP_HW_CORE ||
364	type1 == KMP_HW_SOCKET) &&
365	(type2 == KMP_HW_THREAD || type2 == KMP_HW_CORE ||
366	type2 == KMP_HW_SOCKET)) {
367	top_index1 = top_index2++;
368	continue;
369	}
370	bool radix1 = true;
371	bool all_same = true;
372	int id1 = hw_threads[0].ids[top_index1];
373	int id2 = hw_threads[0].ids[top_index2];
374	int pref1 = preference[type1];
375	int pref2 = preference[type2];
376	for (int hwidx = 1; hwidx < num_hw_threads; ++hwidx) {
377	if (hw_threads[hwidx].ids[top_index1] == id1 &&
378	hw_threads[hwidx].ids[top_index2] != id2) {
379	radix1 = false;
380	break;
381	}
382	if (hw_threads[hwidx].ids[top_index2] != id2)
383	all_same = false;
384	id1 = hw_threads[hwidx].ids[top_index1];
385	id2 = hw_threads[hwidx].ids[top_index2];
386	}
387	if (radix1) {
388	// Select the layer to remove based on preference
389	kmp_hw_t remove_type, keep_type;
390	int remove_layer, remove_layer_ids;
391	if (pref1 > pref2) {
392	remove_type = type2;
393	remove_layer = remove_layer_ids = top_index2;
394	keep_type = type1;
395	} else {
396	remove_type = type1;
397	remove_layer = remove_layer_ids = top_index1;
398	keep_type = type2;
399	}
400	// If all the indexes for the second (deeper) layer are the same.
401	// e.g., all are zero, then make sure to keep the first layer's ids
402	if (all_same)
403	remove_layer_ids = top_index2;
404	// Remove radix one type by setting the equivalence, removing the id from
405	// the hw threads and removing the layer from types and depth
406	set_equivalent_type(type1: remove_type, type2: keep_type);
407	for (int idx = 0; idx < num_hw_threads; ++idx) {
408	kmp_hw_thread_t &hw_thread = hw_threads[idx];
409	for (int d = remove_layer_ids; d < depth - 1; ++d)
410	hw_thread.ids[d] = hw_thread.ids[d + 1];
411	}
412	for (int idx = remove_layer; idx < depth - 1; ++idx)
413	types[idx] = types[idx + 1];
414	depth--;
415	} else {
416	top_index1 = top_index2++;
417	}
418	}
419	KMP_ASSERT(depth > 0);
420	}
421
422	void kmp_topology_t::_set_last_level_cache() {
423	if (get_equivalent_type(type: KMP_HW_L3) != KMP_HW_UNKNOWN)
424	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_L3);
425	else if (get_equivalent_type(type: KMP_HW_L2) != KMP_HW_UNKNOWN)
426	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_L2);
427	#if KMP_MIC_SUPPORTED
428	else if (__kmp_mic_type == mic3) {
429	if (get_equivalent_type(type: KMP_HW_L2) != KMP_HW_UNKNOWN)
430	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_L2);
431	else if (get_equivalent_type(type: KMP_HW_TILE) != KMP_HW_UNKNOWN)
432	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_TILE);
433	// L2/Tile wasn't detected so just say L1
434	else
435	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_L1);
436	}
437	#endif
438	else if (get_equivalent_type(type: KMP_HW_L1) != KMP_HW_UNKNOWN)
439	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_L1);
440	// Fallback is to set last level cache to socket or core
441	if (get_equivalent_type(type: KMP_HW_LLC) == KMP_HW_UNKNOWN) {
442	if (get_equivalent_type(type: KMP_HW_SOCKET) != KMP_HW_UNKNOWN)
443	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_SOCKET);
444	else if (get_equivalent_type(type: KMP_HW_CORE) != KMP_HW_UNKNOWN)
445	set_equivalent_type(type1: KMP_HW_LLC, type2: KMP_HW_CORE);
446	}
447	KMP_ASSERT(get_equivalent_type(KMP_HW_LLC) != KMP_HW_UNKNOWN);
448	}
449
450	// Gather the count of each topology layer and the ratio
451	void kmp_topology_t::_gather_enumeration_information() {
452	int previous_id[KMP_HW_LAST];
453	int max[KMP_HW_LAST];
454
455	for (int i = 0; i < depth; ++i) {
456	previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
457	max[i] = 0;
458	count[i] = 0;
459	ratio[i] = 0;
460	}
461	int core_level = get_level(type: KMP_HW_CORE);
462	for (int i = 0; i < num_hw_threads; ++i) {
463	kmp_hw_thread_t &hw_thread = hw_threads[i];
464	for (int layer = 0; layer < depth; ++layer) {
465	int id = hw_thread.ids[layer];
466	if (id != previous_id[layer]) {
467	// Add an additional increment to each count
468	for (int l = layer; l < depth; ++l)
469	count[l]++;
470	// Keep track of topology layer ratio statistics
471	max[layer]++;
472	for (int l = layer + 1; l < depth; ++l) {
473	if (max[l] > ratio[l])
474	ratio[l] = max[l];
475	max[l] = 1;
476	}
477	// Figure out the number of different core types
478	// and efficiencies for hybrid CPUs
479	if (__kmp_is_hybrid_cpu() && core_level >= 0 && layer <= core_level) {
480	if (hw_thread.attrs.is_core_eff_valid() &&
481	hw_thread.attrs.core_eff >= num_core_efficiencies) {
482	// Because efficiencies can range from 0 to max efficiency - 1,
483	// the number of efficiencies is max efficiency + 1
484	num_core_efficiencies = hw_thread.attrs.core_eff + 1;
485	}
486	if (hw_thread.attrs.is_core_type_valid()) {
487	bool found = false;
488	for (int j = 0; j < num_core_types; ++j) {
489	if (hw_thread.attrs.get_core_type() == core_types[j]) {
490	found = true;
491	break;
492	}
493	}
494	if (!found) {
495	KMP_ASSERT(num_core_types < KMP_HW_MAX_NUM_CORE_TYPES);
496	core_types[num_core_types++] = hw_thread.attrs.get_core_type();
497	}
498	}
499	}
500	break;
501	}
502	}
503	for (int layer = 0; layer < depth; ++layer) {
504	previous_id[layer] = hw_thread.ids[layer];
505	}
506	}
507	for (int layer = 0; layer < depth; ++layer) {
508	if (max[layer] > ratio[layer])
509	ratio[layer] = max[layer];
510	}
511	}
512
513	int kmp_topology_t::_get_ncores_with_attr(const kmp_hw_attr_t &attr,
514	int above_level,
515	bool find_all) const {
516	int current, current_max;
517	int previous_id[KMP_HW_LAST];
518	for (int i = 0; i < depth; ++i)
519	previous_id[i] = kmp_hw_thread_t::UNKNOWN_ID;
520	int core_level = get_level(type: KMP_HW_CORE);
521	if (find_all)
522	above_level = -1;
523	KMP_ASSERT(above_level < core_level);
524	current_max = 0;
525	current = 0;
526	for (int i = 0; i < num_hw_threads; ++i) {
527	kmp_hw_thread_t &hw_thread = hw_threads[i];
528	if (!find_all && hw_thread.ids[above_level] != previous_id[above_level]) {
529	if (current > current_max)
530	current_max = current;
531	current = hw_thread.attrs.contains(other: attr);
532	} else {
533	for (int level = above_level + 1; level <= core_level; ++level) {
534	if (hw_thread.ids[level] != previous_id[level]) {
535	if (hw_thread.attrs.contains(other: attr))
536	current++;
537	break;
538	}
539	}
540	}
541	for (int level = 0; level < depth; ++level)
542	previous_id[level] = hw_thread.ids[level];
543	}
544	if (current > current_max)
545	current_max = current;
546	return current_max;
547	}
548
549	// Find out if the topology is uniform
550	void kmp_topology_t::_discover_uniformity() {
551	int num = 1;
552	for (int level = 0; level < depth; ++level)
553	num *= ratio[level];
554	flags.uniform = (num == count[depth - 1]);
555	}
556
557	// Set all the sub_ids for each hardware thread
558	void kmp_topology_t::_set_sub_ids() {
559	int previous_id[KMP_HW_LAST];
560	int sub_id[KMP_HW_LAST];
561
562	for (int i = 0; i < depth; ++i) {
563	previous_id[i] = -1;
564	sub_id[i] = -1;
565	}
566	for (int i = 0; i < num_hw_threads; ++i) {
567	kmp_hw_thread_t &hw_thread = hw_threads[i];
568	// Setup the sub_id
569	for (int j = 0; j < depth; ++j) {
570	if (hw_thread.ids[j] != previous_id[j]) {
571	sub_id[j]++;
572	for (int k = j + 1; k < depth; ++k) {
573	sub_id[k] = 0;
574	}
575	break;
576	}
577	}
578	// Set previous_id
579	for (int j = 0; j < depth; ++j) {
580	previous_id[j] = hw_thread.ids[j];
581	}
582	// Set the sub_ids field
583	for (int j = 0; j < depth; ++j) {
584	hw_thread.sub_ids[j] = sub_id[j];
585	}
586	}
587	}
588
589	void kmp_topology_t::_set_globals() {
590	// Set nCoresPerPkg, nPackages, __kmp_nThreadsPerCore, __kmp_ncores
591	int core_level, thread_level, package_level;
592	package_level = get_level(type: KMP_HW_SOCKET);
593	#if KMP_GROUP_AFFINITY
594	if (package_level == -1)
595	package_level = get_level(KMP_HW_PROC_GROUP);
596	#endif
597	core_level = get_level(type: KMP_HW_CORE);
598	thread_level = get_level(type: KMP_HW_THREAD);
599
600	KMP_ASSERT(core_level != -1);
601	KMP_ASSERT(thread_level != -1);
602
603	__kmp_nThreadsPerCore = calculate_ratio(level1: thread_level, level2: core_level);
604	if (package_level != -1) {
605	nCoresPerPkg = calculate_ratio(level1: core_level, level2: package_level);
606	nPackages = get_count(level: package_level);
607	} else {
608	// assume one socket
609	nCoresPerPkg = get_count(level: core_level);
610	nPackages = 1;
611	}
612	#ifndef KMP_DFLT_NTH_CORES
613	__kmp_ncores = get_count(level: core_level);
614	#endif
615	}
616
617	kmp_topology_t *kmp_topology_t::allocate(int nproc, int ndepth,
618	const kmp_hw_t *types) {
619	kmp_topology_t *retval;
620	// Allocate all data in one large allocation
621	size_t size = sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc +
622	sizeof(int) * (size_t)KMP_HW_LAST * 3;
623	char *bytes = (char *)__kmp_allocate(size);
624	retval = (kmp_topology_t *)bytes;
625	if (nproc > 0) {
626	retval->hw_threads = (kmp_hw_thread_t *)(bytes + sizeof(kmp_topology_t));
627	} else {
628	retval->hw_threads = nullptr;
629	}
630	retval->num_hw_threads = nproc;
631	retval->depth = ndepth;
632	int *arr =
633	(int *)(bytes + sizeof(kmp_topology_t) + sizeof(kmp_hw_thread_t) * nproc);
634	retval->types = (kmp_hw_t *)arr;
635	retval->ratio = arr + (size_t)KMP_HW_LAST;
636	retval->count = arr + 2 * (size_t)KMP_HW_LAST;
637	retval->num_core_efficiencies = 0;
638	retval->num_core_types = 0;
639	retval->compact = 0;
640	for (int i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
641	retval->core_types[i] = KMP_HW_CORE_TYPE_UNKNOWN;
642	KMP_FOREACH_HW_TYPE(type) { retval->equivalent[type] = KMP_HW_UNKNOWN; }
643	for (int i = 0; i < ndepth; ++i) {
644	retval->types[i] = types[i];
645	retval->equivalent[types[i]] = types[i];
646	}
647	return retval;
648	}
649
650	void kmp_topology_t::deallocate(kmp_topology_t *topology) {
651	if (topology)
652	__kmp_free(topology);
653	}
654
655	bool kmp_topology_t::check_ids() const {
656	// Assume ids have been sorted
657	if (num_hw_threads == 0)
658	return true;
659	for (int i = 1; i < num_hw_threads; ++i) {
660	kmp_hw_thread_t &current_thread = hw_threads[i];
661	kmp_hw_thread_t &previous_thread = hw_threads[i - 1];
662	bool unique = false;
663	for (int j = 0; j < depth; ++j) {
664	if (previous_thread.ids[j] != current_thread.ids[j]) {
665	unique = true;
666	break;
667	}
668	}
669	if (unique)
670	continue;
671	return false;
672	}
673	return true;
674	}
675
676	void kmp_topology_t::dump() const {
677	printf(format: "***********************\n");
678	printf(format: "*** __kmp_topology: ***\n");
679	printf(format: "***********************\n");
680	printf(format: "* depth: %d\n", depth);
681
682	printf(format: "* types: ");
683	for (int i = 0; i < depth; ++i)
684	printf(format: "%15s ", __kmp_hw_get_keyword(type: types[i]));
685	printf(format: "\n");
686
687	printf(format: "* ratio: ");
688	for (int i = 0; i < depth; ++i) {
689	printf(format: "%15d ", ratio[i]);
690	}
691	printf(format: "\n");
692
693	printf(format: "* count: ");
694	for (int i = 0; i < depth; ++i) {
695	printf(format: "%15d ", count[i]);
696	}
697	printf(format: "\n");
698
699	printf(format: "* num_core_eff: %d\n", num_core_efficiencies);
700	printf(format: "* num_core_types: %d\n", num_core_types);
701	printf(format: "* core_types: ");
702	for (int i = 0; i < num_core_types; ++i)
703	printf(format: "%3d ", core_types[i]);
704	printf(format: "\n");
705
706	printf(format: "* equivalent map:\n");
707	KMP_FOREACH_HW_TYPE(i) {
708	const char *key = __kmp_hw_get_keyword(type: i);
709	const char *value = __kmp_hw_get_keyword(type: equivalent[i]);
710	printf(format: "%-15s -> %-15s\n", key, value);
711	}
712
713	printf(format: "* uniform: %s\n", (is_uniform() ? "Yes" : "No"));
714
715	printf(format: "* num_hw_threads: %d\n", num_hw_threads);
716	printf(format: "* hw_threads:\n");
717	for (int i = 0; i < num_hw_threads; ++i) {
718	hw_threads[i].print();
719	}
720	printf(format: "***********************\n");
721	}
722
723	void kmp_topology_t::print(const char *env_var) const {
724	kmp_str_buf_t buf;
725	int print_types_depth;
726	__kmp_str_buf_init(&buf);
727	kmp_hw_t print_types[KMP_HW_LAST + 2];
728
729	// Num Available Threads
730	if (num_hw_threads) {
731	KMP_INFORM(AvailableOSProc, env_var, num_hw_threads);
732	} else {
733	KMP_INFORM(AvailableOSProc, env_var, __kmp_xproc);
734	}
735
736	// Uniform or not
737	if (is_uniform()) {
738	KMP_INFORM(Uniform, env_var);
739	} else {
740	KMP_INFORM(NonUniform, env_var);
741	}
742
743	// Equivalent types
744	KMP_FOREACH_HW_TYPE(type) {
745	kmp_hw_t eq_type = equivalent[type];
746	if (eq_type != KMP_HW_UNKNOWN && eq_type != type) {
747	KMP_INFORM(AffEqualTopologyTypes, env_var,
748	__kmp_hw_get_catalog_string(type),
749	__kmp_hw_get_catalog_string(eq_type));
750	}
751	}
752
753	// Quick topology
754	KMP_ASSERT(depth > 0 && depth <= (int)KMP_HW_LAST);
755	// Create a print types array that always guarantees printing
756	// the core and thread level
757	print_types_depth = 0;
758	for (int level = 0; level < depth; ++level)
759	print_types[print_types_depth++] = types[level];
760	if (equivalent[KMP_HW_CORE] != KMP_HW_CORE) {
761	// Force in the core level for quick topology
762	if (print_types[print_types_depth - 1] == KMP_HW_THREAD) {
763	// Force core before thread e.g., 1 socket X 2 threads/socket
764	// becomes 1 socket X 1 core/socket X 2 threads/socket
765	print_types[print_types_depth - 1] = KMP_HW_CORE;
766	print_types[print_types_depth++] = KMP_HW_THREAD;
767	} else {
768	print_types[print_types_depth++] = KMP_HW_CORE;
769	}
770	}
771	// Always put threads at very end of quick topology
772	if (equivalent[KMP_HW_THREAD] != KMP_HW_THREAD)
773	print_types[print_types_depth++] = KMP_HW_THREAD;
774
775	__kmp_str_buf_clear(buffer: &buf);
776	kmp_hw_t numerator_type;
777	kmp_hw_t denominator_type = KMP_HW_UNKNOWN;
778	int core_level = get_level(type: KMP_HW_CORE);
779	int ncores = get_count(level: core_level);
780
781	for (int plevel = 0, level = 0; plevel < print_types_depth; ++plevel) {
782	int c;
783	bool plural;
784	numerator_type = print_types[plevel];
785	KMP_ASSERT_VALID_HW_TYPE(numerator_type);
786	if (equivalent[numerator_type] != numerator_type)
787	c = 1;
788	else
789	c = get_ratio(level: level++);
790	plural = (c > 1);
791	if (plevel == 0) {
792	__kmp_str_buf_print(buffer: &buf, format: "%d %s", c,
793	__kmp_hw_get_catalog_string(type: numerator_type, plural));
794	} else {
795	__kmp_str_buf_print(buffer: &buf, format: " x %d %s/%s", c,
796	__kmp_hw_get_catalog_string(type: numerator_type, plural),
797	__kmp_hw_get_catalog_string(type: denominator_type));
798	}
799	denominator_type = numerator_type;
800	}
801	KMP_INFORM(TopologyGeneric, env_var, buf.str, ncores);
802
803	// Hybrid topology information
804	if (__kmp_is_hybrid_cpu()) {
805	for (int i = 0; i < num_core_types; ++i) {
806	kmp_hw_core_type_t core_type = core_types[i];
807	kmp_hw_attr_t attr;
808	attr.clear();
809	attr.set_core_type(core_type);
810	int ncores = get_ncores_with_attr(attr);
811	if (ncores > 0) {
812	KMP_INFORM(TopologyHybrid, env_var, ncores,
813	__kmp_hw_get_core_type_string(core_type));
814	KMP_ASSERT(num_core_efficiencies <= KMP_HW_MAX_NUM_CORE_EFFS)
815	for (int eff = 0; eff < num_core_efficiencies; ++eff) {
816	attr.set_core_eff(eff);
817	int ncores_with_eff = get_ncores_with_attr(attr);
818	if (ncores_with_eff > 0) {
819	KMP_INFORM(TopologyHybridCoreEff, env_var, ncores_with_eff, eff);
820	}
821	}
822	}
823	}
824	}
825
826	if (num_hw_threads <= 0) {
827	__kmp_str_buf_free(buffer: &buf);
828	return;
829	}
830
831	// Full OS proc to hardware thread map
832	KMP_INFORM(OSProcToPhysicalThreadMap, env_var);
833	for (int i = 0; i < num_hw_threads; i++) {
834	__kmp_str_buf_clear(buffer: &buf);
835	for (int level = 0; level < depth; ++level) {
836	kmp_hw_t type = types[level];
837	__kmp_str_buf_print(buffer: &buf, format: "%s ", __kmp_hw_get_catalog_string(type));
838	__kmp_str_buf_print(buffer: &buf, format: "%d ", hw_threads[i].ids[level]);
839	}
840	if (__kmp_is_hybrid_cpu())
841	__kmp_str_buf_print(
842	buffer: &buf, format: "(%s)",
843	__kmp_hw_get_core_type_string(type: hw_threads[i].attrs.get_core_type()));
844	KMP_INFORM(OSProcMapToPack, env_var, hw_threads[i].os_id, buf.str);
845	}
846
847	__kmp_str_buf_free(buffer: &buf);
848	}
849
850	#if KMP_AFFINITY_SUPPORTED
851	void kmp_topology_t::set_granularity(kmp_affinity_t &affinity) const {
852	const char *env_var = __kmp_get_affinity_env_var(affinity);
853	// If requested hybrid CPU attributes for granularity (either OMP_PLACES or
854	// KMP_AFFINITY), but none exist, then reset granularity and have below method
855	// select a granularity and warn user.
856	if (!__kmp_is_hybrid_cpu()) {
857	if (affinity.core_attr_gran.valid) {
858	// OMP_PLACES with cores:<attribute> but non-hybrid arch, use cores
859	// instead
860	KMP_AFF_WARNING(
861	affinity, AffIgnoringNonHybrid, env_var,
862	__kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
863	affinity.gran = KMP_HW_CORE;
864	affinity.gran_levels = -1;
865	affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
866	affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
867	} else if (affinity.flags.core_types_gran ||
868	affinity.flags.core_effs_gran) {
869	// OMP_PLACES=core_types|core_effs but non-hybrid, use cores instead
870	if (affinity.flags.omp_places) {
871	KMP_AFF_WARNING(
872	affinity, AffIgnoringNonHybrid, env_var,
873	__kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true));
874	} else {
875	// KMP_AFFINITY=granularity=core_type|core_eff,...
876	KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
877	"Intel(R) Hybrid Technology core attribute",
878	__kmp_hw_get_catalog_string(KMP_HW_CORE));
879	}
880	affinity.gran = KMP_HW_CORE;
881	affinity.gran_levels = -1;
882	affinity.core_attr_gran = KMP_AFFINITY_ATTRS_UNKNOWN;
883	affinity.flags.core_types_gran = affinity.flags.core_effs_gran = 0;
884	}
885	}
886	// Set the number of affinity granularity levels
887	if (affinity.gran_levels < 0) {
888	kmp_hw_t gran_type = get_equivalent_type(type: affinity.gran);
889	// Check if user's granularity request is valid
890	if (gran_type == KMP_HW_UNKNOWN) {
891	// First try core, then thread, then package
892	kmp_hw_t gran_types[3] = {KMP_HW_CORE, KMP_HW_THREAD, KMP_HW_SOCKET};
893	for (auto g : gran_types) {
894	if (get_equivalent_type(type: g) != KMP_HW_UNKNOWN) {
895	gran_type = g;
896	break;
897	}
898	}
899	KMP_ASSERT(gran_type != KMP_HW_UNKNOWN);
900	// Warn user what granularity setting will be used instead
901	KMP_AFF_WARNING(affinity, AffGranularityBad, env_var,
902	__kmp_hw_get_catalog_string(affinity.gran),
903	__kmp_hw_get_catalog_string(gran_type));
904	affinity.gran = gran_type;
905	}
906	#if KMP_GROUP_AFFINITY
907	// If more than one processor group exists, and the level of
908	// granularity specified by the user is too coarse, then the
909	// granularity must be adjusted "down" to processor group affinity
910	// because threads can only exist within one processor group.
911	// For example, if a user sets granularity=socket and there are two
912	// processor groups that cover a socket, then the runtime must
913	// restrict the granularity down to the processor group level.
914	if (__kmp_num_proc_groups > 1) {
915	int gran_depth = get_level(gran_type);
916	int proc_group_depth = get_level(KMP_HW_PROC_GROUP);
917	if (gran_depth >= 0 && proc_group_depth >= 0 &&
918	gran_depth < proc_group_depth) {
919	KMP_AFF_WARNING(affinity, AffGranTooCoarseProcGroup, env_var,
920	__kmp_hw_get_catalog_string(affinity.gran));
921	affinity.gran = gran_type = KMP_HW_PROC_GROUP;
922	}
923	}
924	#endif
925	affinity.gran_levels = 0;
926	for (int i = depth - 1; i >= 0 && get_type(level: i) != gran_type; --i)
927	affinity.gran_levels++;
928	}
929	}
930	#endif
931
932	void kmp_topology_t::canonicalize() {
933	#if KMP_GROUP_AFFINITY
934	_insert_windows_proc_groups();
935	#endif
936	_remove_radix1_layers();
937	_gather_enumeration_information();
938	_discover_uniformity();
939	_set_sub_ids();
940	_set_globals();
941	_set_last_level_cache();
942
943	#if KMP_MIC_SUPPORTED
944	// Manually Add L2 = Tile equivalence
945	if (__kmp_mic_type == mic3) {
946	if (get_level(type: KMP_HW_L2) != -1)
947	set_equivalent_type(type1: KMP_HW_TILE, type2: KMP_HW_L2);
948	else if (get_level(type: KMP_HW_TILE) != -1)
949	set_equivalent_type(type1: KMP_HW_L2, type2: KMP_HW_TILE);
950	}
951	#endif
952
953	// Perform post canonicalization checking
954	KMP_ASSERT(depth > 0);
955	for (int level = 0; level < depth; ++level) {
956	// All counts, ratios, and types must be valid
957	KMP_ASSERT(count[level] > 0 && ratio[level] > 0);
958	KMP_ASSERT_VALID_HW_TYPE(types[level]);
959	// Detected types must point to themselves
960	KMP_ASSERT(equivalent[types[level]] == types[level]);
961	}
962	}
963
964	// Canonicalize an explicit packages X cores/pkg X threads/core topology
965	void kmp_topology_t::canonicalize(int npackages, int ncores_per_pkg,
966	int nthreads_per_core, int ncores) {
967	int ndepth = 3;
968	depth = ndepth;
969	KMP_FOREACH_HW_TYPE(i) { equivalent[i] = KMP_HW_UNKNOWN; }
970	for (int level = 0; level < depth; ++level) {
971	count[level] = 0;
972	ratio[level] = 0;
973	}
974	count[0] = npackages;
975	count[1] = ncores;
976	count[2] = __kmp_xproc;
977	ratio[0] = npackages;
978	ratio[1] = ncores_per_pkg;
979	ratio[2] = nthreads_per_core;
980	equivalent[KMP_HW_SOCKET] = KMP_HW_SOCKET;
981	equivalent[KMP_HW_CORE] = KMP_HW_CORE;
982	equivalent[KMP_HW_THREAD] = KMP_HW_THREAD;
983	types[0] = KMP_HW_SOCKET;
984	types[1] = KMP_HW_CORE;
985	types[2] = KMP_HW_THREAD;
986	//__kmp_avail_proc = __kmp_xproc;
987	_discover_uniformity();
988	}
989
990	#if KMP_AFFINITY_SUPPORTED
991	static kmp_str_buf_t *
992	__kmp_hw_get_catalog_core_string(const kmp_hw_attr_t &attr, kmp_str_buf_t *buf,
993	bool plural) {
994	__kmp_str_buf_init(buf);
995	if (attr.is_core_type_valid())
996	__kmp_str_buf_print(buffer: buf, format: "%s %s",
997	__kmp_hw_get_core_type_string(type: attr.get_core_type()),
998	__kmp_hw_get_catalog_string(type: KMP_HW_CORE, plural));
999	else
1000	__kmp_str_buf_print(buffer: buf, format: "%s eff=%d",
1001	__kmp_hw_get_catalog_string(type: KMP_HW_CORE, plural),
1002	attr.get_core_eff());
1003	return buf;
1004	}
1005
1006	bool kmp_topology_t::restrict_to_mask(const kmp_affin_mask_t *mask) {
1007	// Apply the filter
1008	bool affected;
1009	int new_index = 0;
1010	for (int i = 0; i < num_hw_threads; ++i) {
1011	int os_id = hw_threads[i].os_id;
1012	if (KMP_CPU_ISSET(os_id, mask)) {
1013	if (i != new_index)
1014	hw_threads[new_index] = hw_threads[i];
1015	new_index++;
1016	} else {
1017	KMP_CPU_CLR(os_id, __kmp_affin_fullMask);
1018	__kmp_avail_proc--;
1019	}
1020	}
1021
1022	KMP_DEBUG_ASSERT(new_index <= num_hw_threads);
1023	affected = (num_hw_threads != new_index);
1024	num_hw_threads = new_index;
1025
1026	// Post hardware subset canonicalization
1027	if (affected) {
1028	_gather_enumeration_information();
1029	_discover_uniformity();
1030	_set_globals();
1031	_set_last_level_cache();
1032	#if KMP_OS_WINDOWS
1033	// Copy filtered full mask if topology has single processor group
1034	if (__kmp_num_proc_groups <= 1)
1035	#endif
1036	__kmp_affin_origMask->copy(src: __kmp_affin_fullMask);
1037	}
1038	return affected;
1039	}
1040
1041	// Apply the KMP_HW_SUBSET envirable to the topology
1042	// Returns true if KMP_HW_SUBSET filtered any processors
1043	// otherwise, returns false
1044	bool kmp_topology_t::filter_hw_subset() {
1045	// If KMP_HW_SUBSET wasn't requested, then do nothing.
1046	if (!__kmp_hw_subset)
1047	return false;
1048
1049	// First, sort the KMP_HW_SUBSET items by the machine topology
1050	__kmp_hw_subset->sort();
1051
1052	__kmp_hw_subset->canonicalize(top: __kmp_topology);
1053
1054	// Check to see if KMP_HW_SUBSET is a valid subset of the detected topology
1055	bool using_core_types = false;
1056	bool using_core_effs = false;
1057	bool is_absolute = __kmp_hw_subset->is_absolute();
1058	int hw_subset_depth = __kmp_hw_subset->get_depth();
1059	kmp_hw_t specified[KMP_HW_LAST];
1060	int *topology_levels = (int *)KMP_ALLOCA(sizeof(int) * hw_subset_depth);
1061	KMP_ASSERT(hw_subset_depth > 0);
1062	KMP_FOREACH_HW_TYPE(i) { specified[i] = KMP_HW_UNKNOWN; }
1063	int core_level = get_level(type: KMP_HW_CORE);
1064	for (int i = 0; i < hw_subset_depth; ++i) {
1065	int max_count;
1066	const kmp_hw_subset_t::item_t &item = __kmp_hw_subset->at(index: i);
1067	int num = item.num[0];
1068	int offset = item.offset[0];
1069	kmp_hw_t type = item.type;
1070	kmp_hw_t equivalent_type = equivalent[type];
1071	int level = get_level(type);
1072	topology_levels[i] = level;
1073
1074	// Check to see if current layer is in detected machine topology
1075	if (equivalent_type != KMP_HW_UNKNOWN) {
1076	__kmp_hw_subset->at(index: i).type = equivalent_type;
1077	} else {
1078	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetNotExistGeneric,
1079	__kmp_hw_get_catalog_string(type));
1080	return false;
1081	}
1082
1083	// Check to see if current layer has already been
1084	// specified either directly or through an equivalent type
1085	if (specified[equivalent_type] != KMP_HW_UNKNOWN) {
1086	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetEqvLayers,
1087	__kmp_hw_get_catalog_string(type),
1088	__kmp_hw_get_catalog_string(specified[equivalent_type]));
1089	return false;
1090	}
1091	specified[equivalent_type] = type;
1092
1093	// Check to see if each layer's num & offset parameters are valid
1094	max_count = get_ratio(level);
1095	if (!is_absolute) {
1096	if (max_count < 0 ||
1097	(num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1098	bool plural = (num > 1);
1099	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric,
1100	__kmp_hw_get_catalog_string(type, plural));
1101	return false;
1102	}
1103	}
1104
1105	// Check to see if core attributes are consistent
1106	if (core_level == level) {
1107	// Determine which core attributes are specified
1108	for (int j = 0; j < item.num_attrs; ++j) {
1109	if (item.attr[j].is_core_type_valid())
1110	using_core_types = true;
1111	if (item.attr[j].is_core_eff_valid())
1112	using_core_effs = true;
1113	}
1114
1115	// Check if using a single core attribute on non-hybrid arch.
1116	// Do not ignore all of KMP_HW_SUBSET, just ignore the attribute.
1117	//
1118	// Check if using multiple core attributes on non-hyrbid arch.
1119	// Ignore all of KMP_HW_SUBSET if this is the case.
1120	if ((using_core_effs || using_core_types) && !__kmp_is_hybrid_cpu()) {
1121	if (item.num_attrs == 1) {
1122	if (using_core_effs) {
1123	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1124	"efficiency");
1125	} else {
1126	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIgnoringAttr,
1127	"core_type");
1128	}
1129	using_core_effs = false;
1130	using_core_types = false;
1131	} else {
1132	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrsNonHybrid);
1133	return false;
1134	}
1135	}
1136
1137	// Check if using both core types and core efficiencies together
1138	if (using_core_types && using_core_effs) {
1139	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat, "core_type",
1140	"efficiency");
1141	return false;
1142	}
1143
1144	// Check that core efficiency values are valid
1145	if (using_core_effs) {
1146	for (int j = 0; j < item.num_attrs; ++j) {
1147	if (item.attr[j].is_core_eff_valid()) {
1148	int core_eff = item.attr[j].get_core_eff();
1149	if (core_eff < 0 || core_eff >= num_core_efficiencies) {
1150	kmp_str_buf_t buf;
1151	__kmp_str_buf_init(&buf);
1152	__kmp_str_buf_print(buffer: &buf, format: "%d", item.attr[j].get_core_eff());
1153	__kmp_msg(kmp_ms_warning,
1154	KMP_MSG(AffHWSubsetAttrInvalid, "efficiency", buf.str),
1155	KMP_HNT(ValidValuesRange, 0, num_core_efficiencies - 1),
1156	__kmp_msg_null);
1157	__kmp_str_buf_free(buffer: &buf);
1158	return false;
1159	}
1160	}
1161	}
1162	}
1163
1164	// Check that the number of requested cores with attributes is valid
1165	if ((using_core_types || using_core_effs) && !is_absolute) {
1166	for (int j = 0; j < item.num_attrs; ++j) {
1167	int num = item.num[j];
1168	int offset = item.offset[j];
1169	int level_above = core_level - 1;
1170	if (level_above >= 0) {
1171	max_count = get_ncores_with_attr_per(attr: item.attr[j], above: level_above);
1172	if (max_count <= 0 ||
1173	(num != kmp_hw_subset_t::USE_ALL && num + offset > max_count)) {
1174	kmp_str_buf_t buf;
1175	__kmp_hw_get_catalog_core_string(attr: item.attr[j], buf: &buf, plural: num > 0);
1176	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetManyGeneric, buf.str);
1177	__kmp_str_buf_free(buffer: &buf);
1178	return false;
1179	}
1180	}
1181	}
1182	}
1183
1184	if ((using_core_types || using_core_effs) && item.num_attrs > 1) {
1185	for (int j = 0; j < item.num_attrs; ++j) {
1186	// Ambiguous use of specific core attribute + generic core
1187	// e.g., 4c & 3c:intel_core or 4c & 3c:eff1
1188	if (!item.attr[j]) {
1189	kmp_hw_attr_t other_attr;
1190	for (int k = 0; k < item.num_attrs; ++k) {
1191	if (item.attr[k] != item.attr[j]) {
1192	other_attr = item.attr[k];
1193	break;
1194	}
1195	}
1196	kmp_str_buf_t buf;
1197	__kmp_hw_get_catalog_core_string(attr: other_attr, buf: &buf, plural: item.num[j] > 0);
1198	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetIncompat,
1199	__kmp_hw_get_catalog_string(KMP_HW_CORE), buf.str);
1200	__kmp_str_buf_free(buffer: &buf);
1201	return false;
1202	}
1203	// Allow specifying a specific core type or core eff exactly once
1204	for (int k = 0; k < j; ++k) {
1205	if (!item.attr[j] || !item.attr[k])
1206	continue;
1207	if (item.attr[k] == item.attr[j]) {
1208	kmp_str_buf_t buf;
1209	__kmp_hw_get_catalog_core_string(attr: item.attr[j], buf: &buf,
1210	plural: item.num[j] > 0);
1211	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAttrRepeat, buf.str);
1212	__kmp_str_buf_free(buffer: &buf);
1213	return false;
1214	}
1215	}
1216	}
1217	}
1218	}
1219	}
1220
1221	// For keeping track of sub_ids for an absolute KMP_HW_SUBSET
1222	// or core attributes (core type or efficiency)
1223	int prev_sub_ids[KMP_HW_LAST];
1224	int abs_sub_ids[KMP_HW_LAST];
1225	int core_eff_sub_ids[KMP_HW_MAX_NUM_CORE_EFFS];
1226	int core_type_sub_ids[KMP_HW_MAX_NUM_CORE_TYPES];
1227	for (size_t i = 0; i < KMP_HW_LAST; ++i) {
1228	abs_sub_ids[i] = -1;
1229	prev_sub_ids[i] = -1;
1230	}
1231	for (size_t i = 0; i < KMP_HW_MAX_NUM_CORE_EFFS; ++i)
1232	core_eff_sub_ids[i] = -1;
1233	for (size_t i = 0; i < KMP_HW_MAX_NUM_CORE_TYPES; ++i)
1234	core_type_sub_ids[i] = -1;
1235
1236	// Determine which hardware threads should be filtered.
1237
1238	// Helpful to determine if a topology layer is targeted by an absolute subset
1239	auto is_targeted = [&](int level) {
1240	if (is_absolute) {
1241	for (int i = 0; i < hw_subset_depth; ++i)
1242	if (topology_levels[i] == level)
1243	return true;
1244	return false;
1245	}
1246	// If not absolute KMP_HW_SUBSET, then every layer is seen as targeted
1247	return true;
1248	};
1249
1250	// Helpful to index into core type sub Ids array
1251	auto get_core_type_index = [](const kmp_hw_thread_t &t) {
1252	switch (t.attrs.get_core_type()) {
1253	case KMP_HW_CORE_TYPE_UNKNOWN:
1254	case KMP_HW_MAX_NUM_CORE_TYPES:
1255	return 0;
1256	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1257	case KMP_HW_CORE_TYPE_ATOM:
1258	return 1;
1259	case KMP_HW_CORE_TYPE_CORE:
1260	return 2;
1261	#endif
1262	}
1263	KMP_ASSERT2(false, "Unhandled kmp_hw_thread_t enumeration");
1264	KMP_BUILTIN_UNREACHABLE;
1265	};
1266
1267	// Helpful to index into core efficiencies sub Ids array
1268	auto get_core_eff_index = [](const kmp_hw_thread_t &t) {
1269	return t.attrs.get_core_eff();
1270	};
1271
1272	int num_filtered = 0;
1273	kmp_affin_mask_t *filtered_mask;
1274	KMP_CPU_ALLOC(filtered_mask);
1275	KMP_CPU_COPY(filtered_mask, __kmp_affin_fullMask);
1276	for (int i = 0; i < num_hw_threads; ++i) {
1277	kmp_hw_thread_t &hw_thread = hw_threads[i];
1278
1279	// Figure out the absolute sub ids and core eff/type sub ids
1280	if (is_absolute || using_core_effs || using_core_types) {
1281	for (int level = 0; level < get_depth(); ++level) {
1282	if (hw_thread.sub_ids[level] != prev_sub_ids[level]) {
1283	bool found_targeted = false;
1284	for (int j = level; j < get_depth(); ++j) {
1285	bool targeted = is_targeted(j);
1286	if (!found_targeted && targeted) {
1287	found_targeted = true;
1288	abs_sub_ids[j]++;
1289	if (j == core_level && using_core_effs)
1290	core_eff_sub_ids[get_core_eff_index(hw_thread)]++;
1291	if (j == core_level && using_core_types)
1292	core_type_sub_ids[get_core_type_index(hw_thread)]++;
1293	} else if (targeted) {
1294	abs_sub_ids[j] = 0;
1295	if (j == core_level && using_core_effs)
1296	core_eff_sub_ids[get_core_eff_index(hw_thread)] = 0;
1297	if (j == core_level && using_core_types)
1298	core_type_sub_ids[get_core_type_index(hw_thread)] = 0;
1299	}
1300	}
1301	break;
1302	}
1303	}
1304	for (int level = 0; level < get_depth(); ++level)
1305	prev_sub_ids[level] = hw_thread.sub_ids[level];
1306	}
1307
1308	// Check to see if this hardware thread should be filtered
1309	bool should_be_filtered = false;
1310	for (int hw_subset_index = 0; hw_subset_index < hw_subset_depth;
1311	++hw_subset_index) {
1312	const auto &hw_subset_item = __kmp_hw_subset->at(index: hw_subset_index);
1313	int level = topology_levels[hw_subset_index];
1314	if (level == -1)
1315	continue;
1316	if ((using_core_effs || using_core_types) && level == core_level) {
1317	// Look for the core attribute in KMP_HW_SUBSET which corresponds
1318	// to this hardware thread's core attribute. Use this num,offset plus
1319	// the running sub_id for the particular core attribute of this hardware
1320	// thread to determine if the hardware thread should be filtered or not.
1321	int attr_idx;
1322	kmp_hw_core_type_t core_type = hw_thread.attrs.get_core_type();
1323	int core_eff = hw_thread.attrs.get_core_eff();
1324	for (attr_idx = 0; attr_idx < hw_subset_item.num_attrs; ++attr_idx) {
1325	if (using_core_types &&
1326	hw_subset_item.attr[attr_idx].get_core_type() == core_type)
1327	break;
1328	if (using_core_effs &&
1329	hw_subset_item.attr[attr_idx].get_core_eff() == core_eff)
1330	break;
1331	}
1332	// This core attribute isn't in the KMP_HW_SUBSET so always filter it.
1333	if (attr_idx == hw_subset_item.num_attrs) {
1334	should_be_filtered = true;
1335	break;
1336	}
1337	int sub_id;
1338	int num = hw_subset_item.num[attr_idx];
1339	int offset = hw_subset_item.offset[attr_idx];
1340	if (using_core_types)
1341	sub_id = core_type_sub_ids[get_core_type_index(hw_thread)];
1342	else
1343	sub_id = core_eff_sub_ids[get_core_eff_index(hw_thread)];
1344	if (sub_id < offset ||
1345	(num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1346	should_be_filtered = true;
1347	break;
1348	}
1349	} else {
1350	int sub_id;
1351	int num = hw_subset_item.num[0];
1352	int offset = hw_subset_item.offset[0];
1353	if (is_absolute)
1354	sub_id = abs_sub_ids[level];
1355	else
1356	sub_id = hw_thread.sub_ids[level];
1357	if (sub_id < offset ||
1358	(num != kmp_hw_subset_t::USE_ALL && sub_id >= offset + num)) {
1359	should_be_filtered = true;
1360	break;
1361	}
1362	}
1363	}
1364	// Collect filtering information
1365	if (should_be_filtered) {
1366	KMP_CPU_CLR(hw_thread.os_id, filtered_mask);
1367	num_filtered++;
1368	}
1369	}
1370
1371	// One last check that we shouldn't allow filtering entire machine
1372	if (num_filtered == num_hw_threads) {
1373	KMP_AFF_WARNING(__kmp_affinity, AffHWSubsetAllFiltered);
1374	return false;
1375	}
1376
1377	// Apply the filter
1378	restrict_to_mask(mask: filtered_mask);
1379	return true;
1380	}
1381
1382	bool kmp_topology_t::is_close(int hwt1, int hwt2,
1383	const kmp_affinity_t &stgs) const {
1384	int hw_level = stgs.gran_levels;
1385	if (hw_level >= depth)
1386	return true;
1387	bool retval = true;
1388	const kmp_hw_thread_t &t1 = hw_threads[hwt1];
1389	const kmp_hw_thread_t &t2 = hw_threads[hwt2];
1390	if (stgs.flags.core_types_gran)
1391	return t1.attrs.get_core_type() == t2.attrs.get_core_type();
1392	if (stgs.flags.core_effs_gran)
1393	return t1.attrs.get_core_eff() == t2.attrs.get_core_eff();
1394	for (int i = 0; i < (depth - hw_level); ++i) {
1395	if (t1.ids[i] != t2.ids[i])
1396	return false;
1397	}
1398	return retval;
1399	}
1400
1401	////////////////////////////////////////////////////////////////////////////////
1402
1403	bool KMPAffinity::picked_api = false;
1404
1405	void *KMPAffinity::Mask::operator new(size_t n) { return __kmp_allocate(n); }
1406	void *KMPAffinity::Mask::operator new[](size_t n) { return __kmp_allocate(n); }
1407	void KMPAffinity::Mask::operator delete(void *p) { __kmp_free(p); }
1408	void KMPAffinity::Mask::operator delete[](void *p) { __kmp_free(p); }
1409	void *KMPAffinity::operator new(size_t n) { return __kmp_allocate(n); }
1410	void KMPAffinity::operator delete(void *p) { __kmp_free(p); }
1411
1412	void KMPAffinity::pick_api() {
1413	KMPAffinity *affinity_dispatch;
1414	if (picked_api)
1415	return;
1416	#if KMP_USE_HWLOC
1417	// Only use Hwloc if affinity isn't explicitly disabled and
1418	// user requests Hwloc topology method
1419	if (__kmp_affinity_top_method == affinity_top_method_hwloc &&
1420	__kmp_affinity.type != affinity_disabled) {
1421	affinity_dispatch = new KMPHwlocAffinity();
1422	} else
1423	#endif
1424	{
1425	affinity_dispatch = new KMPNativeAffinity();
1426	}
1427	__kmp_affinity_dispatch = affinity_dispatch;
1428	picked_api = true;
1429	}
1430
1431	void KMPAffinity::destroy_api() {
1432	if (__kmp_affinity_dispatch != NULL) {
1433	delete __kmp_affinity_dispatch;
1434	__kmp_affinity_dispatch = NULL;
1435	picked_api = false;
1436	}
1437	}
1438
1439	#define KMP_ADVANCE_SCAN(scan) \
1440	while (*scan != '\0') { \
1441	scan++; \
1442	}
1443
1444	// Print the affinity mask to the character array in a pretty format.
1445	// The format is a comma separated list of non-negative integers or integer
1446	// ranges: e.g., 1,2,3-5,7,9-15
1447	// The format can also be the string "{<empty>}" if no bits are set in mask
1448	char *__kmp_affinity_print_mask(char *buf, int buf_len,
1449	kmp_affin_mask_t *mask) {
1450	int start = 0, finish = 0, previous = 0;
1451	bool first_range;
1452	KMP_ASSERT(buf);
1453	KMP_ASSERT(buf_len >= 40);
1454	KMP_ASSERT(mask);
1455	char *scan = buf;
1456	char *end = buf + buf_len - 1;
1457
1458	// Check for empty set.
1459	if (mask->begin() == mask->end()) {
1460	KMP_SNPRINTF(s: scan, maxlen: end - scan + 1, format: "{<empty>}");
1461	KMP_ADVANCE_SCAN(scan);
1462	KMP_ASSERT(scan <= end);
1463	return buf;
1464	}
1465
1466	first_range = true;
1467	start = mask->begin();
1468	while (1) {
1469	// Find next range
1470	// [start, previous] is inclusive range of contiguous bits in mask
1471	for (finish = mask->next(previous: start), previous = start;
1472	finish == previous + 1 && finish != mask->end();
1473	finish = mask->next(previous: finish)) {
1474	previous = finish;
1475	}
1476
1477	// The first range does not need a comma printed before it, but the rest
1478	// of the ranges do need a comma beforehand
1479	if (!first_range) {
1480	KMP_SNPRINTF(s: scan, maxlen: end - scan + 1, format: "%s", ",");
1481	KMP_ADVANCE_SCAN(scan);
1482	} else {
1483	first_range = false;
1484	}
1485	// Range with three or more contiguous bits in the affinity mask
1486	if (previous - start > 1) {
1487	KMP_SNPRINTF(s: scan, maxlen: end - scan + 1, format: "%u-%u", start, previous);
1488	} else {
1489	// Range with one or two contiguous bits in the affinity mask
1490	KMP_SNPRINTF(s: scan, maxlen: end - scan + 1, format: "%u", start);
1491	KMP_ADVANCE_SCAN(scan);
1492	if (previous - start > 0) {
1493	KMP_SNPRINTF(s: scan, maxlen: end - scan + 1, format: ",%u", previous);
1494	}
1495	}
1496	KMP_ADVANCE_SCAN(scan);
1497	// Start over with new start point
1498	start = finish;
1499	if (start == mask->end())
1500	break;
1501	// Check for overflow
1502	if (end - scan < 2)
1503	break;
1504	}
1505
1506	// Check for overflow
1507	KMP_ASSERT(scan <= end);
1508	return buf;
1509	}
1510	#undef KMP_ADVANCE_SCAN
1511
1512	// Print the affinity mask to the string buffer object in a pretty format
1513	// The format is a comma separated list of non-negative integers or integer
1514	// ranges: e.g., 1,2,3-5,7,9-15
1515	// The format can also be the string "{<empty>}" if no bits are set in mask
1516	kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
1517	kmp_affin_mask_t *mask) {
1518	int start = 0, finish = 0, previous = 0;
1519	bool first_range;
1520	KMP_ASSERT(buf);
1521	KMP_ASSERT(mask);
1522
1523	__kmp_str_buf_clear(buffer: buf);
1524
1525	// Check for empty set.
1526	if (mask->begin() == mask->end()) {
1527	__kmp_str_buf_print(buffer: buf, format: "%s", "{<empty>}");
1528	return buf;
1529	}
1530
1531	first_range = true;
1532	start = mask->begin();
1533	while (1) {
1534	// Find next range
1535	// [start, previous] is inclusive range of contiguous bits in mask
1536	for (finish = mask->next(previous: start), previous = start;
1537	finish == previous + 1 && finish != mask->end();
1538	finish = mask->next(previous: finish)) {
1539	previous = finish;
1540	}
1541
1542	// The first range does not need a comma printed before it, but the rest
1543	// of the ranges do need a comma beforehand
1544	if (!first_range) {
1545	__kmp_str_buf_print(buffer: buf, format: "%s", ",");
1546	} else {
1547	first_range = false;
1548	}
1549	// Range with three or more contiguous bits in the affinity mask
1550	if (previous - start > 1) {
1551	__kmp_str_buf_print(buffer: buf, format: "%u-%u", start, previous);
1552	} else {
1553	// Range with one or two contiguous bits in the affinity mask
1554	__kmp_str_buf_print(buffer: buf, format: "%u", start);
1555	if (previous - start > 0) {
1556	__kmp_str_buf_print(buffer: buf, format: ",%u", previous);
1557	}
1558	}
1559	// Start over with new start point
1560	start = finish;
1561	if (start == mask->end())
1562	break;
1563	}
1564	return buf;
1565	}
1566
1567	// Return (possibly empty) affinity mask representing the offline CPUs
1568	// Caller must free the mask
1569	kmp_affin_mask_t *__kmp_affinity_get_offline_cpus() {
1570	kmp_affin_mask_t *offline;
1571	KMP_CPU_ALLOC(offline);
1572	KMP_CPU_ZERO(offline);
1573	#if KMP_OS_LINUX
1574	int n, begin_cpu, end_cpu;
1575	kmp_safe_raii_file_t offline_file;
1576	auto skip_ws = [](FILE *f) {
1577	int c;
1578	do {
1579	c = fgetc(stream: f);
1580	} while (isspace(c));
1581	if (c != EOF)
1582	ungetc(c: c, stream: f);
1583	};
1584	// File contains CSV of integer ranges representing the offline CPUs
1585	// e.g., 1,2,4-7,9,11-15
1586	int status = offline_file.try_open(filename: "/sys/devices/system/cpu/offline", mode: "r");
1587	if (status != 0)
1588	return offline;
1589	while (!feof(stream: offline_file)) {
1590	skip_ws(offline_file);
1591	n = fscanf(stream: offline_file, format: "%d", &begin_cpu);
1592	if (n != 1)
1593	break;
1594	skip_ws(offline_file);
1595	int c = fgetc(stream: offline_file);
1596	if (c == EOF || c == ',') {
1597	// Just single CPU
1598	end_cpu = begin_cpu;
1599	} else if (c == '-') {
1600	// Range of CPUs
1601	skip_ws(offline_file);
1602	n = fscanf(stream: offline_file, format: "%d", &end_cpu);
1603	if (n != 1)
1604	break;
1605	skip_ws(offline_file);
1606	c = fgetc(stream: offline_file); // skip ','
1607	} else {
1608	// Syntax problem
1609	break;
1610	}
1611	// Ensure a valid range of CPUs
1612	if (begin_cpu < 0 || begin_cpu >= __kmp_xproc || end_cpu < 0 ||
1613	end_cpu >= __kmp_xproc || begin_cpu > end_cpu) {
1614	continue;
1615	}
1616	// Insert [begin_cpu, end_cpu] into offline mask
1617	for (int cpu = begin_cpu; cpu <= end_cpu; ++cpu) {
1618	KMP_CPU_SET(cpu, offline);
1619	}
1620	}
1621	#endif
1622	return offline;
1623	}
1624
1625	// Return the number of available procs
1626	int __kmp_affinity_entire_machine_mask(kmp_affin_mask_t *mask) {
1627	int avail_proc = 0;
1628	KMP_CPU_ZERO(mask);
1629
1630	#if KMP_GROUP_AFFINITY
1631
1632	if (__kmp_num_proc_groups > 1) {
1633	int group;
1634	KMP_DEBUG_ASSERT(__kmp_GetActiveProcessorCount != NULL);
1635	for (group = 0; group < __kmp_num_proc_groups; group++) {
1636	int i;
1637	int num = __kmp_GetActiveProcessorCount(group);
1638	for (i = 0; i < num; i++) {
1639	KMP_CPU_SET(i + group * (CHAR_BIT * sizeof(DWORD_PTR)), mask);
1640	avail_proc++;
1641	}
1642	}
1643	} else
1644
1645	#endif /* KMP_GROUP_AFFINITY */
1646
1647	{
1648	int proc;
1649	kmp_affin_mask_t *offline_cpus = __kmp_affinity_get_offline_cpus();
1650	for (proc = 0; proc < __kmp_xproc; proc++) {
1651	// Skip offline CPUs
1652	if (KMP_CPU_ISSET(proc, offline_cpus))
1653	continue;
1654	KMP_CPU_SET(proc, mask);
1655	avail_proc++;
1656	}
1657	KMP_CPU_FREE(offline_cpus);
1658	}
1659
1660	return avail_proc;
1661	}
1662
1663	// All of the __kmp_affinity_create_*_map() routines should allocate the
1664	// internal topology object and set the layer ids for it. Each routine
1665	// returns a boolean on whether it was successful at doing so.
1666	kmp_affin_mask_t *__kmp_affin_fullMask = NULL;
1667	// Original mask is a subset of full mask in multiple processor groups topology
1668	kmp_affin_mask_t *__kmp_affin_origMask = NULL;
1669
1670	#if KMP_USE_HWLOC
1671	static inline bool __kmp_hwloc_is_cache_type(hwloc_obj_t obj) {
1672	#if HWLOC_API_VERSION >= 0x00020000
1673	return hwloc_obj_type_is_cache(obj->type);
1674	#else
1675	return obj->type == HWLOC_OBJ_CACHE;
1676	#endif
1677	}
1678
1679	// Returns KMP_HW_* type derived from HWLOC_* type
1680	static inline kmp_hw_t __kmp_hwloc_type_2_topology_type(hwloc_obj_t obj) {
1681
1682	if (__kmp_hwloc_is_cache_type(obj)) {
1683	if (obj->attr->cache.type == HWLOC_OBJ_CACHE_INSTRUCTION)
1684	return KMP_HW_UNKNOWN;
1685	switch (obj->attr->cache.depth) {
1686	case 1:
1687	return KMP_HW_L1;
1688	case 2:
1689	#if KMP_MIC_SUPPORTED
1690	if (__kmp_mic_type == mic3) {
1691	return KMP_HW_TILE;
1692	}
1693	#endif
1694	return KMP_HW_L2;
1695	case 3:
1696	return KMP_HW_L3;
1697	}
1698	return KMP_HW_UNKNOWN;
1699	}
1700
1701	switch (obj->type) {
1702	case HWLOC_OBJ_PACKAGE:
1703	return KMP_HW_SOCKET;
1704	case HWLOC_OBJ_NUMANODE:
1705	return KMP_HW_NUMA;
1706	case HWLOC_OBJ_CORE:
1707	return KMP_HW_CORE;
1708	case HWLOC_OBJ_PU:
1709	return KMP_HW_THREAD;
1710	case HWLOC_OBJ_GROUP:
1711	#if HWLOC_API_VERSION >= 0x00020000
1712	if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_DIE)
1713	return KMP_HW_DIE;
1714	else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_TILE)
1715	return KMP_HW_TILE;
1716	else if (obj->attr->group.kind == HWLOC_GROUP_KIND_INTEL_MODULE)
1717	return KMP_HW_MODULE;
1718	else if (obj->attr->group.kind == HWLOC_GROUP_KIND_WINDOWS_PROCESSOR_GROUP)
1719	return KMP_HW_PROC_GROUP;
1720	#endif
1721	return KMP_HW_UNKNOWN;
1722	#if HWLOC_API_VERSION >= 0x00020100
1723	case HWLOC_OBJ_DIE:
1724	return KMP_HW_DIE;
1725	#endif
1726	}
1727	return KMP_HW_UNKNOWN;
1728	}
1729
1730	// Returns the number of objects of type 'type' below 'obj' within the topology
1731	// tree structure. e.g., if obj is a HWLOC_OBJ_PACKAGE object, and type is
1732	// HWLOC_OBJ_PU, then this will return the number of PU's under the SOCKET
1733	// object.
1734	static int __kmp_hwloc_get_nobjs_under_obj(hwloc_obj_t obj,
1735	hwloc_obj_type_t type) {
1736	int retval = 0;
1737	hwloc_obj_t first;
1738	for (first = hwloc_get_obj_below_by_type(__kmp_hwloc_topology, obj->type,
1739	obj->logical_index, type, 0);
1740	first != NULL && hwloc_get_ancestor_obj_by_type(__kmp_hwloc_topology,
1741	obj->type, first) == obj;
1742	first = hwloc_get_next_obj_by_type(__kmp_hwloc_topology, first->type,
1743	first)) {
1744	++retval;
1745	}
1746	return retval;
1747	}
1748
1749	// This gets the sub_id for a lower object under a higher object in the
1750	// topology tree
1751	static int __kmp_hwloc_get_sub_id(hwloc_topology_t t, hwloc_obj_t higher,
1752	hwloc_obj_t lower) {
1753	hwloc_obj_t obj;
1754	hwloc_obj_type_t ltype = lower->type;
1755	int lindex = lower->logical_index - 1;
1756	int sub_id = 0;
1757	// Get the previous lower object
1758	obj = hwloc_get_obj_by_type(t, ltype, lindex);
1759	while (obj && lindex >= 0 &&
1760	hwloc_bitmap_isincluded(obj->cpuset, higher->cpuset)) {
1761	if (obj->userdata) {
1762	sub_id = (int)(RCAST(kmp_intptr_t, obj->userdata));
1763	break;
1764	}
1765	sub_id++;
1766	lindex--;
1767	obj = hwloc_get_obj_by_type(t, ltype, lindex);
1768	}
1769	// store sub_id + 1 so that 0 is differed from NULL
1770	lower->userdata = RCAST(void *, sub_id + 1);
1771	return sub_id;
1772	}
1773
1774	static bool __kmp_affinity_create_hwloc_map(kmp_i18n_id_t *const msg_id) {
1775	kmp_hw_t type;
1776	int hw_thread_index, sub_id;
1777	int depth;
1778	hwloc_obj_t pu, obj, root, prev;
1779	kmp_hw_t types[KMP_HW_LAST];
1780	hwloc_obj_type_t hwloc_types[KMP_HW_LAST];
1781
1782	hwloc_topology_t tp = __kmp_hwloc_topology;
1783	*msg_id = kmp_i18n_null;
1784	if (__kmp_affinity.flags.verbose) {
1785	KMP_INFORM(AffUsingHwloc, "KMP_AFFINITY");
1786	}
1787
1788	if (!KMP_AFFINITY_CAPABLE()) {
1789	// Hack to try and infer the machine topology using only the data
1790	// available from hwloc on the current thread, and __kmp_xproc.
1791	KMP_ASSERT(__kmp_affinity.type == affinity_none);
1792	// hwloc only guarantees existance of PU object, so check PACKAGE and CORE
1793	hwloc_obj_t o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_PACKAGE, 0);
1794	if (o != NULL)
1795	nCoresPerPkg = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_CORE);
1796	else
1797	nCoresPerPkg = 1; // no PACKAGE found
1798	o = hwloc_get_obj_by_type(tp, HWLOC_OBJ_CORE, 0);
1799	if (o != NULL)
1800	__kmp_nThreadsPerCore = __kmp_hwloc_get_nobjs_under_obj(o, HWLOC_OBJ_PU);
1801	else
1802	__kmp_nThreadsPerCore = 1; // no CORE found
1803	if (__kmp_nThreadsPerCore == 0)
1804	__kmp_nThreadsPerCore = 1;
1805	__kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
1806	if (nCoresPerPkg == 0)
1807	nCoresPerPkg = 1; // to prevent possible division by 0
1808	nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
1809	return true;
1810	}
1811
1812	#if HWLOC_API_VERSION >= 0x00020400
1813	// Handle multiple types of cores if they exist on the system
1814	int nr_cpu_kinds = hwloc_cpukinds_get_nr(tp, 0);
1815
1816	typedef struct kmp_hwloc_cpukinds_info_t {
1817	int efficiency;
1818	kmp_hw_core_type_t core_type;
1819	hwloc_bitmap_t mask;
1820	} kmp_hwloc_cpukinds_info_t;
1821	kmp_hwloc_cpukinds_info_t *cpukinds = nullptr;
1822
1823	if (nr_cpu_kinds > 0) {
1824	unsigned nr_infos;
1825	struct hwloc_info_s *infos;
1826	cpukinds = (kmp_hwloc_cpukinds_info_t *)__kmp_allocate(
1827	sizeof(kmp_hwloc_cpukinds_info_t) * nr_cpu_kinds);
1828	for (unsigned idx = 0; idx < (unsigned)nr_cpu_kinds; ++idx) {
1829	cpukinds[idx].efficiency = -1;
1830	cpukinds[idx].core_type = KMP_HW_CORE_TYPE_UNKNOWN;
1831	cpukinds[idx].mask = hwloc_bitmap_alloc();
1832	if (hwloc_cpukinds_get_info(tp, idx, cpukinds[idx].mask,
1833	&cpukinds[idx].efficiency, &nr_infos, &infos,
1834	0) == 0) {
1835	for (unsigned i = 0; i < nr_infos; ++i) {
1836	if (__kmp_str_match("CoreType", 8, infos[i].name)) {
1837	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1838	if (__kmp_str_match("IntelAtom", 9, infos[i].value)) {
1839	cpukinds[idx].core_type = KMP_HW_CORE_TYPE_ATOM;
1840	break;
1841	} else if (__kmp_str_match("IntelCore", 9, infos[i].value)) {
1842	cpukinds[idx].core_type = KMP_HW_CORE_TYPE_CORE;
1843	break;
1844	}
1845	#endif
1846	}
1847	}
1848	}
1849	}
1850	}
1851	#endif
1852
1853	root = hwloc_get_root_obj(tp);
1854
1855	// Figure out the depth and types in the topology
1856	depth = 0;
1857	obj = hwloc_get_pu_obj_by_os_index(tp, __kmp_affin_fullMask->begin());
1858	while (obj && obj != root) {
1859	#if HWLOC_API_VERSION >= 0x00020000
1860	if (obj->memory_arity) {
1861	hwloc_obj_t memory;
1862	for (memory = obj->memory_first_child; memory;
1863	memory = hwloc_get_next_child(tp, obj, memory)) {
1864	if (memory->type == HWLOC_OBJ_NUMANODE)
1865	break;
1866	}
1867	if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1868	types[depth] = KMP_HW_NUMA;
1869	hwloc_types[depth] = memory->type;
1870	depth++;
1871	}
1872	}
1873	#endif
1874	type = __kmp_hwloc_type_2_topology_type(obj);
1875	if (type != KMP_HW_UNKNOWN) {
1876	types[depth] = type;
1877	hwloc_types[depth] = obj->type;
1878	depth++;
1879	}
1880	obj = obj->parent;
1881	}
1882	KMP_ASSERT(depth > 0);
1883
1884	// Get the order for the types correct
1885	for (int i = 0, j = depth - 1; i < j; ++i, --j) {
1886	hwloc_obj_type_t hwloc_temp = hwloc_types[i];
1887	kmp_hw_t temp = types[i];
1888	types[i] = types[j];
1889	types[j] = temp;
1890	hwloc_types[i] = hwloc_types[j];
1891	hwloc_types[j] = hwloc_temp;
1892	}
1893
1894	// Allocate the data structure to be returned.
1895	__kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
1896
1897	hw_thread_index = 0;
1898	pu = NULL;
1899	while ((pu = hwloc_get_next_obj_by_type(tp, HWLOC_OBJ_PU, pu))) {
1900	int index = depth - 1;
1901	bool included = KMP_CPU_ISSET(pu->os_index, __kmp_affin_fullMask);
1902	kmp_hw_thread_t &hw_thread = __kmp_topology->at(hw_thread_index);
1903	if (included) {
1904	hw_thread.clear();
1905	hw_thread.ids[index] = pu->logical_index;
1906	hw_thread.os_id = pu->os_index;
1907	// If multiple core types, then set that attribute for the hardware thread
1908	#if HWLOC_API_VERSION >= 0x00020400
1909	if (cpukinds) {
1910	int cpukind_index = -1;
1911	for (int i = 0; i < nr_cpu_kinds; ++i) {
1912	if (hwloc_bitmap_isset(cpukinds[i].mask, hw_thread.os_id)) {
1913	cpukind_index = i;
1914	break;
1915	}
1916	}
1917	if (cpukind_index >= 0) {
1918	hw_thread.attrs.set_core_type(cpukinds[cpukind_index].core_type);
1919	hw_thread.attrs.set_core_eff(cpukinds[cpukind_index].efficiency);
1920	}
1921	}
1922	#endif
1923	index--;
1924	}
1925	obj = pu;
1926	prev = obj;
1927	while (obj != root && obj != NULL) {
1928	obj = obj->parent;
1929	#if HWLOC_API_VERSION >= 0x00020000
1930	// NUMA Nodes are handled differently since they are not within the
1931	// parent/child structure anymore. They are separate children
1932	// of obj (memory_first_child points to first memory child)
1933	if (obj->memory_arity) {
1934	hwloc_obj_t memory;
1935	for (memory = obj->memory_first_child; memory;
1936	memory = hwloc_get_next_child(tp, obj, memory)) {
1937	if (memory->type == HWLOC_OBJ_NUMANODE)
1938	break;
1939	}
1940	if (memory && memory->type == HWLOC_OBJ_NUMANODE) {
1941	sub_id = __kmp_hwloc_get_sub_id(tp, memory, prev);
1942	if (included) {
1943	hw_thread.ids[index] = memory->logical_index;
1944	hw_thread.ids[index + 1] = sub_id;
1945	index--;
1946	}
1947	prev = memory;
1948	}
1949	prev = obj;
1950	}
1951	#endif
1952	type = __kmp_hwloc_type_2_topology_type(obj);
1953	if (type != KMP_HW_UNKNOWN) {
1954	sub_id = __kmp_hwloc_get_sub_id(tp, obj, prev);
1955	if (included) {
1956	hw_thread.ids[index] = obj->logical_index;
1957	hw_thread.ids[index + 1] = sub_id;
1958	index--;
1959	}
1960	prev = obj;
1961	}
1962	}
1963	if (included)
1964	hw_thread_index++;
1965	}
1966
1967	#if HWLOC_API_VERSION >= 0x00020400
1968	// Free the core types information
1969	if (cpukinds) {
1970	for (int idx = 0; idx < nr_cpu_kinds; ++idx)
1971	hwloc_bitmap_free(cpukinds[idx].mask);
1972	__kmp_free(cpukinds);
1973	}
1974	#endif
1975	__kmp_topology->sort_ids();
1976	return true;
1977	}
1978	#endif // KMP_USE_HWLOC
1979
1980	// If we don't know how to retrieve the machine's processor topology, or
1981	// encounter an error in doing so, this routine is called to form a "flat"
1982	// mapping of os thread id's <-> processor id's.
1983	static bool __kmp_affinity_create_flat_map(kmp_i18n_id_t *const msg_id) {
1984	*msg_id = kmp_i18n_null;
1985	int depth = 3;
1986	kmp_hw_t types[] = {KMP_HW_SOCKET, KMP_HW_CORE, KMP_HW_THREAD};
1987
1988	if (__kmp_affinity.flags.verbose) {
1989	KMP_INFORM(UsingFlatOS, "KMP_AFFINITY");
1990	}
1991
1992	// Even if __kmp_affinity.type == affinity_none, this routine might still
1993	// be called to set __kmp_ncores, as well as
1994	// __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
1995	if (!KMP_AFFINITY_CAPABLE()) {
1996	KMP_ASSERT(__kmp_affinity.type == affinity_none);
1997	__kmp_ncores = nPackages = __kmp_xproc;
1998	__kmp_nThreadsPerCore = nCoresPerPkg = 1;
1999	return true;
2000	}
2001
2002	// When affinity is off, this routine will still be called to set
2003	// __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2004	// Make sure all these vars are set correctly, and return now if affinity is
2005	// not enabled.
2006	__kmp_ncores = nPackages = __kmp_avail_proc;
2007	__kmp_nThreadsPerCore = nCoresPerPkg = 1;
2008
2009	// Construct the data structure to be returned.
2010	__kmp_topology = kmp_topology_t::allocate(nproc: __kmp_avail_proc, ndepth: depth, types);
2011	int avail_ct = 0;
2012	int i;
2013	KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2014	// Skip this proc if it is not included in the machine model.
2015	if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2016	continue;
2017	}
2018	kmp_hw_thread_t &hw_thread = __kmp_topology->at(index: avail_ct);
2019	hw_thread.clear();
2020	hw_thread.os_id = i;
2021	hw_thread.ids[0] = i;
2022	hw_thread.ids[1] = 0;
2023	hw_thread.ids[2] = 0;
2024	avail_ct++;
2025	}
2026	if (__kmp_affinity.flags.verbose) {
2027	KMP_INFORM(OSProcToPackage, "KMP_AFFINITY");
2028	}
2029	return true;
2030	}
2031
2032	#if KMP_GROUP_AFFINITY
2033	// If multiple Windows* OS processor groups exist, we can create a 2-level
2034	// topology map with the groups at level 0 and the individual procs at level 1.
2035	// This facilitates letting the threads float among all procs in a group,
2036	// if granularity=group (the default when there are multiple groups).
2037	static bool __kmp_affinity_create_proc_group_map(kmp_i18n_id_t *const msg_id) {
2038	*msg_id = kmp_i18n_null;
2039	int depth = 3;
2040	kmp_hw_t types[] = {KMP_HW_PROC_GROUP, KMP_HW_CORE, KMP_HW_THREAD};
2041	const static size_t BITS_PER_GROUP = CHAR_BIT * sizeof(DWORD_PTR);
2042
2043	if (__kmp_affinity.flags.verbose) {
2044	KMP_INFORM(AffWindowsProcGroupMap, "KMP_AFFINITY");
2045	}
2046
2047	// If we aren't affinity capable, then use flat topology
2048	if (!KMP_AFFINITY_CAPABLE()) {
2049	KMP_ASSERT(__kmp_affinity.type == affinity_none);
2050	nPackages = __kmp_num_proc_groups;
2051	__kmp_nThreadsPerCore = 1;
2052	__kmp_ncores = __kmp_xproc;
2053	nCoresPerPkg = nPackages / __kmp_ncores;
2054	return true;
2055	}
2056
2057	// Construct the data structure to be returned.
2058	__kmp_topology = kmp_topology_t::allocate(__kmp_avail_proc, depth, types);
2059	int avail_ct = 0;
2060	int i;
2061	KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2062	// Skip this proc if it is not included in the machine model.
2063	if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2064	continue;
2065	}
2066	kmp_hw_thread_t &hw_thread = __kmp_topology->at(avail_ct++);
2067	hw_thread.clear();
2068	hw_thread.os_id = i;
2069	hw_thread.ids[0] = i / BITS_PER_GROUP;
2070	hw_thread.ids[1] = hw_thread.ids[2] = i % BITS_PER_GROUP;
2071	}
2072	return true;
2073	}
2074	#endif /* KMP_GROUP_AFFINITY */
2075
2076	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
2077
2078	template <kmp_uint32 LSB, kmp_uint32 MSB>
2079	static inline unsigned __kmp_extract_bits(kmp_uint32 v) {
2080	const kmp_uint32 SHIFT_LEFT = sizeof(kmp_uint32) * 8 - 1 - MSB;
2081	const kmp_uint32 SHIFT_RIGHT = LSB;
2082	kmp_uint32 retval = v;
2083	retval <<= SHIFT_LEFT;
2084	retval >>= (SHIFT_LEFT + SHIFT_RIGHT);
2085	return retval;
2086	}
2087
2088	static int __kmp_cpuid_mask_width(int count) {
2089	int r = 0;
2090
2091	while ((1 << r) < count)
2092	++r;
2093	return r;
2094	}
2095
2096	class apicThreadInfo {
2097	public:
2098	unsigned osId; // param to __kmp_affinity_bind_thread
2099	unsigned apicId; // from cpuid after binding
2100	unsigned maxCoresPerPkg; // ""
2101	unsigned maxThreadsPerPkg; // ""
2102	unsigned pkgId; // inferred from above values
2103	unsigned coreId; // ""
2104	unsigned threadId; // ""
2105	};
2106
2107	static int __kmp_affinity_cmp_apicThreadInfo_phys_id(const void *a,
2108	const void *b) {
2109	const apicThreadInfo *aa = (const apicThreadInfo *)a;
2110	const apicThreadInfo *bb = (const apicThreadInfo *)b;
2111	if (aa->pkgId < bb->pkgId)
2112	return -1;
2113	if (aa->pkgId > bb->pkgId)
2114	return 1;
2115	if (aa->coreId < bb->coreId)
2116	return -1;
2117	if (aa->coreId > bb->coreId)
2118	return 1;
2119	if (aa->threadId < bb->threadId)
2120	return -1;
2121	if (aa->threadId > bb->threadId)
2122	return 1;
2123	return 0;
2124	}
2125
2126	class kmp_cache_info_t {
2127	public:
2128	struct info_t {
2129	unsigned level, mask;
2130	};
2131	kmp_cache_info_t() : depth(0) { get_leaf4_levels(); }
2132	size_t get_depth() const { return depth; }
2133	info_t &operator[](size_t index) { return table[index]; }
2134	const info_t &operator[](size_t index) const { return table[index]; }
2135
2136	static kmp_hw_t get_topology_type(unsigned level) {
2137	KMP_DEBUG_ASSERT(level >= 1 && level <= MAX_CACHE_LEVEL);
2138	switch (level) {
2139	case 1:
2140	return KMP_HW_L1;
2141	case 2:
2142	return KMP_HW_L2;
2143	case 3:
2144	return KMP_HW_L3;
2145	}
2146	return KMP_HW_UNKNOWN;
2147	}
2148
2149	private:
2150	static const int MAX_CACHE_LEVEL = 3;
2151
2152	size_t depth;
2153	info_t table[MAX_CACHE_LEVEL];
2154
2155	void get_leaf4_levels() {
2156	unsigned level = 0;
2157	while (depth < MAX_CACHE_LEVEL) {
2158	unsigned cache_type, max_threads_sharing;
2159	unsigned cache_level, cache_mask_width;
2160	kmp_cpuid buf2;
2161	__kmp_x86_cpuid(leaf: 4, subleaf: level, p: &buf2);
2162	cache_type = __kmp_extract_bits<0, 4>(v: buf2.eax);
2163	if (!cache_type)
2164	break;
2165	// Skip instruction caches
2166	if (cache_type == 2) {
2167	level++;
2168	continue;
2169	}
2170	max_threads_sharing = __kmp_extract_bits<14, 25>(v: buf2.eax) + 1;
2171	cache_mask_width = __kmp_cpuid_mask_width(count: max_threads_sharing);
2172	cache_level = __kmp_extract_bits<5, 7>(v: buf2.eax);
2173	table[depth].level = cache_level;
2174	table[depth].mask = ((-1) << cache_mask_width);
2175	depth++;
2176	level++;
2177	}
2178	}
2179	};
2180
2181	// On IA-32 architecture and Intel(R) 64 architecture, we attempt to use
2182	// an algorithm which cycles through the available os threads, setting
2183	// the current thread's affinity mask to that thread, and then retrieves
2184	// the Apic Id for each thread context using the cpuid instruction.
2185	static bool __kmp_affinity_create_apicid_map(kmp_i18n_id_t *const msg_id) {
2186	kmp_cpuid buf;
2187	*msg_id = kmp_i18n_null;
2188
2189	if (__kmp_affinity.flags.verbose) {
2190	KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(DecodingLegacyAPIC));
2191	}
2192
2193	// Check if cpuid leaf 4 is supported.
2194	__kmp_x86_cpuid(leaf: 0, subleaf: 0, p: &buf);
2195	if (buf.eax < 4) {
2196	*msg_id = kmp_i18n_str_NoLeaf4Support;
2197	return false;
2198	}
2199
2200	// The algorithm used starts by setting the affinity to each available thread
2201	// and retrieving info from the cpuid instruction, so if we are not capable of
2202	// calling __kmp_get_system_affinity() and _kmp_get_system_affinity(), then we
2203	// need to do something else - use the defaults that we calculated from
2204	// issuing cpuid without binding to each proc.
2205	if (!KMP_AFFINITY_CAPABLE()) {
2206	// Hack to try and infer the machine topology using only the data
2207	// available from cpuid on the current thread, and __kmp_xproc.
2208	KMP_ASSERT(__kmp_affinity.type == affinity_none);
2209
2210	// Get an upper bound on the number of threads per package using cpuid(1).
2211	// On some OS/chps combinations where HT is supported by the chip but is
2212	// disabled, this value will be 2 on a single core chip. Usually, it will be
2213	// 2 if HT is enabled and 1 if HT is disabled.
2214	__kmp_x86_cpuid(leaf: 1, subleaf: 0, p: &buf);
2215	int maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2216	if (maxThreadsPerPkg == 0) {
2217	maxThreadsPerPkg = 1;
2218	}
2219
2220	// The num cores per pkg comes from cpuid(4). 1 must be added to the encoded
2221	// value.
2222	//
2223	// The author of cpu_count.cpp treated this only an upper bound on the
2224	// number of cores, but I haven't seen any cases where it was greater than
2225	// the actual number of cores, so we will treat it as exact in this block of
2226	// code.
2227	//
2228	// First, we need to check if cpuid(4) is supported on this chip. To see if
2229	// cpuid(n) is supported, issue cpuid(0) and check if eax has the value n or
2230	// greater.
2231	__kmp_x86_cpuid(leaf: 0, subleaf: 0, p: &buf);
2232	if (buf.eax >= 4) {
2233	__kmp_x86_cpuid(leaf: 4, subleaf: 0, p: &buf);
2234	nCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2235	} else {
2236	nCoresPerPkg = 1;
2237	}
2238
2239	// There is no way to reliably tell if HT is enabled without issuing the
2240	// cpuid instruction from every thread, can correlating the cpuid info, so
2241	// if the machine is not affinity capable, we assume that HT is off. We have
2242	// seen quite a few machines where maxThreadsPerPkg is 2, yet the machine
2243	// does not support HT.
2244	//
2245	// - Older OSes are usually found on machines with older chips, which do not
2246	// support HT.
2247	// - The performance penalty for mistakenly identifying a machine as HT when
2248	// it isn't (which results in blocktime being incorrectly set to 0) is
2249	// greater than the penalty when for mistakenly identifying a machine as
2250	// being 1 thread/core when it is really HT enabled (which results in
2251	// blocktime being incorrectly set to a positive value).
2252	__kmp_ncores = __kmp_xproc;
2253	nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2254	__kmp_nThreadsPerCore = 1;
2255	return true;
2256	}
2257
2258	// From here on, we can assume that it is safe to call
2259	// __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2260	// __kmp_affinity.type = affinity_none.
2261
2262	// Save the affinity mask for the current thread.
2263	kmp_affinity_raii_t previous_affinity;
2264
2265	// Run through each of the available contexts, binding the current thread
2266	// to it, and obtaining the pertinent information using the cpuid instr.
2267	//
2268	// The relevant information is:
2269	// - Apic Id: Bits 24:31 of ebx after issuing cpuid(1) - each thread context
2270	// has a uniqie Apic Id, which is of the form pkg# : core# : thread#.
2271	// - Max Threads Per Pkg: Bits 16:23 of ebx after issuing cpuid(1). The value
2272	// of this field determines the width of the core# + thread# fields in the
2273	// Apic Id. It is also an upper bound on the number of threads per
2274	// package, but it has been verified that situations happen were it is not
2275	// exact. In particular, on certain OS/chip combinations where Intel(R)
2276	// Hyper-Threading Technology is supported by the chip but has been
2277	// disabled, the value of this field will be 2 (for a single core chip).
2278	// On other OS/chip combinations supporting Intel(R) Hyper-Threading
2279	// Technology, the value of this field will be 1 when Intel(R)
2280	// Hyper-Threading Technology is disabled and 2 when it is enabled.
2281	// - Max Cores Per Pkg: Bits 26:31 of eax after issuing cpuid(4). The value
2282	// of this field (+1) determines the width of the core# field in the Apic
2283	// Id. The comments in "cpucount.cpp" say that this value is an upper
2284	// bound, but the IA-32 architecture manual says that it is exactly the
2285	// number of cores per package, and I haven't seen any case where it
2286	// wasn't.
2287	//
2288	// From this information, deduce the package Id, core Id, and thread Id,
2289	// and set the corresponding fields in the apicThreadInfo struct.
2290	unsigned i;
2291	apicThreadInfo *threadInfo = (apicThreadInfo *)__kmp_allocate(
2292	__kmp_avail_proc * sizeof(apicThreadInfo));
2293	unsigned nApics = 0;
2294	KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
2295	// Skip this proc if it is not included in the machine model.
2296	if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
2297	continue;
2298	}
2299	KMP_DEBUG_ASSERT((int)nApics < __kmp_avail_proc);
2300
2301	__kmp_affinity_dispatch->bind_thread(proc: i);
2302	threadInfo[nApics].osId = i;
2303
2304	// The apic id and max threads per pkg come from cpuid(1).
2305	__kmp_x86_cpuid(leaf: 1, subleaf: 0, p: &buf);
2306	if (((buf.edx >> 9) & 1) == 0) {
2307	__kmp_free(threadInfo);
2308	*msg_id = kmp_i18n_str_ApicNotPresent;
2309	return false;
2310	}
2311	threadInfo[nApics].apicId = (buf.ebx >> 24) & 0xff;
2312	threadInfo[nApics].maxThreadsPerPkg = (buf.ebx >> 16) & 0xff;
2313	if (threadInfo[nApics].maxThreadsPerPkg == 0) {
2314	threadInfo[nApics].maxThreadsPerPkg = 1;
2315	}
2316
2317	// Max cores per pkg comes from cpuid(4). 1 must be added to the encoded
2318	// value.
2319	//
2320	// First, we need to check if cpuid(4) is supported on this chip. To see if
2321	// cpuid(n) is supported, issue cpuid(0) and check if eax has the value n
2322	// or greater.
2323	__kmp_x86_cpuid(leaf: 0, subleaf: 0, p: &buf);
2324	if (buf.eax >= 4) {
2325	__kmp_x86_cpuid(leaf: 4, subleaf: 0, p: &buf);
2326	threadInfo[nApics].maxCoresPerPkg = ((buf.eax >> 26) & 0x3f) + 1;
2327	} else {
2328	threadInfo[nApics].maxCoresPerPkg = 1;
2329	}
2330
2331	// Infer the pkgId / coreId / threadId using only the info obtained locally.
2332	int widthCT = __kmp_cpuid_mask_width(count: threadInfo[nApics].maxThreadsPerPkg);
2333	threadInfo[nApics].pkgId = threadInfo[nApics].apicId >> widthCT;
2334
2335	int widthC = __kmp_cpuid_mask_width(count: threadInfo[nApics].maxCoresPerPkg);
2336	int widthT = widthCT - widthC;
2337	if (widthT < 0) {
2338	// I've never seen this one happen, but I suppose it could, if the cpuid
2339	// instruction on a chip was really screwed up. Make sure to restore the
2340	// affinity mask before the tail call.
2341	__kmp_free(threadInfo);
2342	*msg_id = kmp_i18n_str_InvalidCpuidInfo;
2343	return false;
2344	}
2345
2346	int maskC = (1 << widthC) - 1;
2347	threadInfo[nApics].coreId = (threadInfo[nApics].apicId >> widthT) & maskC;
2348
2349	int maskT = (1 << widthT) - 1;
2350	threadInfo[nApics].threadId = threadInfo[nApics].apicId & maskT;
2351
2352	nApics++;
2353	}
2354
2355	// We've collected all the info we need.
2356	// Restore the old affinity mask for this thread.
2357	previous_affinity.restore();
2358
2359	// Sort the threadInfo table by physical Id.
2360	qsort(base: threadInfo, nmemb: nApics, size: sizeof(*threadInfo),
2361	compar: __kmp_affinity_cmp_apicThreadInfo_phys_id);
2362
2363	// The table is now sorted by pkgId / coreId / threadId, but we really don't
2364	// know the radix of any of the fields. pkgId's may be sparsely assigned among
2365	// the chips on a system. Although coreId's are usually assigned
2366	// [0 .. coresPerPkg-1] and threadId's are usually assigned
2367	// [0..threadsPerCore-1], we don't want to make any such assumptions.
2368	//
2369	// For that matter, we don't know what coresPerPkg and threadsPerCore (or the
2370	// total # packages) are at this point - we want to determine that now. We
2371	// only have an upper bound on the first two figures.
2372	//
2373	// We also perform a consistency check at this point: the values returned by
2374	// the cpuid instruction for any thread bound to a given package had better
2375	// return the same info for maxThreadsPerPkg and maxCoresPerPkg.
2376	nPackages = 1;
2377	nCoresPerPkg = 1;
2378	__kmp_nThreadsPerCore = 1;
2379	unsigned nCores = 1;
2380
2381	unsigned pkgCt = 1; // to determine radii
2382	unsigned lastPkgId = threadInfo[0].pkgId;
2383	unsigned coreCt = 1;
2384	unsigned lastCoreId = threadInfo[0].coreId;
2385	unsigned threadCt = 1;
2386	unsigned lastThreadId = threadInfo[0].threadId;
2387
2388	// intra-pkg consist checks
2389	unsigned prevMaxCoresPerPkg = threadInfo[0].maxCoresPerPkg;
2390	unsigned prevMaxThreadsPerPkg = threadInfo[0].maxThreadsPerPkg;
2391
2392	for (i = 1; i < nApics; i++) {
2393	if (threadInfo[i].pkgId != lastPkgId) {
2394	nCores++;
2395	pkgCt++;
2396	lastPkgId = threadInfo[i].pkgId;
2397	if ((int)coreCt > nCoresPerPkg)
2398	nCoresPerPkg = coreCt;
2399	coreCt = 1;
2400	lastCoreId = threadInfo[i].coreId;
2401	if ((int)threadCt > __kmp_nThreadsPerCore)
2402	__kmp_nThreadsPerCore = threadCt;
2403	threadCt = 1;
2404	lastThreadId = threadInfo[i].threadId;
2405
2406	// This is a different package, so go on to the next iteration without
2407	// doing any consistency checks. Reset the consistency check vars, though.
2408	prevMaxCoresPerPkg = threadInfo[i].maxCoresPerPkg;
2409	prevMaxThreadsPerPkg = threadInfo[i].maxThreadsPerPkg;
2410	continue;
2411	}
2412
2413	if (threadInfo[i].coreId != lastCoreId) {
2414	nCores++;
2415	coreCt++;
2416	lastCoreId = threadInfo[i].coreId;
2417	if ((int)threadCt > __kmp_nThreadsPerCore)
2418	__kmp_nThreadsPerCore = threadCt;
2419	threadCt = 1;
2420	lastThreadId = threadInfo[i].threadId;
2421	} else if (threadInfo[i].threadId != lastThreadId) {
2422	threadCt++;
2423	lastThreadId = threadInfo[i].threadId;
2424	} else {
2425	__kmp_free(threadInfo);
2426	*msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2427	return false;
2428	}
2429
2430	// Check to make certain that the maxCoresPerPkg and maxThreadsPerPkg
2431	// fields agree between all the threads bounds to a given package.
2432	if ((prevMaxCoresPerPkg != threadInfo[i].maxCoresPerPkg) ||
2433	(prevMaxThreadsPerPkg != threadInfo[i].maxThreadsPerPkg)) {
2434	__kmp_free(threadInfo);
2435	*msg_id = kmp_i18n_str_InconsistentCpuidInfo;
2436	return false;
2437	}
2438	}
2439	// When affinity is off, this routine will still be called to set
2440	// __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
2441	// Make sure all these vars are set correctly
2442	nPackages = pkgCt;
2443	if ((int)coreCt > nCoresPerPkg)
2444	nCoresPerPkg = coreCt;
2445	if ((int)threadCt > __kmp_nThreadsPerCore)
2446	__kmp_nThreadsPerCore = threadCt;
2447	__kmp_ncores = nCores;
2448	KMP_DEBUG_ASSERT(nApics == (unsigned)__kmp_avail_proc);
2449
2450	// Now that we've determined the number of packages, the number of cores per
2451	// package, and the number of threads per core, we can construct the data
2452	// structure that is to be returned.
2453	int idx = 0;
2454	int pkgLevel = 0;
2455	int coreLevel = 1;
2456	int threadLevel = 2;
2457	//(__kmp_nThreadsPerCore <= 1) ? -1 : ((coreLevel >= 0) ? 2 : 1);
2458	int depth = (pkgLevel >= 0) + (coreLevel >= 0) + (threadLevel >= 0);
2459	kmp_hw_t types[3];
2460	if (pkgLevel >= 0)
2461	types[idx++] = KMP_HW_SOCKET;
2462	if (coreLevel >= 0)
2463	types[idx++] = KMP_HW_CORE;
2464	if (threadLevel >= 0)
2465	types[idx++] = KMP_HW_THREAD;
2466
2467	KMP_ASSERT(depth > 0);
2468	__kmp_topology = kmp_topology_t::allocate(nproc: nApics, ndepth: depth, types);
2469
2470	for (i = 0; i < nApics; ++i) {
2471	idx = 0;
2472	unsigned os = threadInfo[i].osId;
2473	kmp_hw_thread_t &hw_thread = __kmp_topology->at(index: i);
2474	hw_thread.clear();
2475
2476	if (pkgLevel >= 0) {
2477	hw_thread.ids[idx++] = threadInfo[i].pkgId;
2478	}
2479	if (coreLevel >= 0) {
2480	hw_thread.ids[idx++] = threadInfo[i].coreId;
2481	}
2482	if (threadLevel >= 0) {
2483	hw_thread.ids[idx++] = threadInfo[i].threadId;
2484	}
2485	hw_thread.os_id = os;
2486	}
2487
2488	__kmp_free(threadInfo);
2489	__kmp_topology->sort_ids();
2490	if (!__kmp_topology->check_ids()) {
2491	kmp_topology_t::deallocate(topology: __kmp_topology);
2492	__kmp_topology = nullptr;
2493	*msg_id = kmp_i18n_str_LegacyApicIDsNotUnique;
2494	return false;
2495	}
2496	return true;
2497	}
2498
2499	// Hybrid cpu detection using CPUID.1A
2500	// Thread should be pinned to processor already
2501	static void __kmp_get_hybrid_info(kmp_hw_core_type_t *type, int *efficiency,
2502	unsigned *native_model_id) {
2503	kmp_cpuid buf;
2504	__kmp_x86_cpuid(leaf: 0x1a, subleaf: 0, p: &buf);
2505	*type = (kmp_hw_core_type_t)__kmp_extract_bits<24, 31>(v: buf.eax);
2506	switch (*type) {
2507	case KMP_HW_CORE_TYPE_ATOM:
2508	*efficiency = 0;
2509	break;
2510	case KMP_HW_CORE_TYPE_CORE:
2511	*efficiency = 1;
2512	break;
2513	default:
2514	*efficiency = 0;
2515	}
2516	*native_model_id = __kmp_extract_bits<0, 23>(v: buf.eax);
2517	}
2518
2519	// Intel(R) microarchitecture code name Nehalem, Dunnington and later
2520	// architectures support a newer interface for specifying the x2APIC Ids,
2521	// based on CPUID.B or CPUID.1F
2522	/*
2523	* CPUID.B or 1F, Input ECX (sub leaf # aka level number)
2524	Bits Bits Bits Bits
2525	31-16 15-8 7-4 4-0
2526	---+-----------+--------------+-------------+-----------------+
2527	EAX| reserved | reserved | reserved | Bits to Shift |
2528	---+-----------|--------------+-------------+-----------------|
2529	EBX| reserved | Num logical processors at level (16 bits) |
2530	---+-----------|--------------+-------------------------------|
2531	ECX| reserved | Level Type | Level Number (8 bits) |
2532	---+-----------+--------------+-------------------------------|
2533	EDX| X2APIC ID (32 bits) |
2534	---+----------------------------------------------------------+
2535	*/
2536
2537	enum {
2538	INTEL_LEVEL_TYPE_INVALID = 0, // Package level
2539	INTEL_LEVEL_TYPE_SMT = 1,
2540	INTEL_LEVEL_TYPE_CORE = 2,
2541	INTEL_LEVEL_TYPE_MODULE = 3,
2542	INTEL_LEVEL_TYPE_TILE = 4,
2543	INTEL_LEVEL_TYPE_DIE = 5,
2544	INTEL_LEVEL_TYPE_LAST = 6,
2545	};
2546
2547	struct cpuid_level_info_t {
2548	unsigned level_type, mask, mask_width, nitems, cache_mask;
2549	};
2550
2551	static kmp_hw_t __kmp_intel_type_2_topology_type(int intel_type) {
2552	switch (intel_type) {
2553	case INTEL_LEVEL_TYPE_INVALID:
2554	return KMP_HW_SOCKET;
2555	case INTEL_LEVEL_TYPE_SMT:
2556	return KMP_HW_THREAD;
2557	case INTEL_LEVEL_TYPE_CORE:
2558	return KMP_HW_CORE;
2559	case INTEL_LEVEL_TYPE_TILE:
2560	return KMP_HW_TILE;
2561	case INTEL_LEVEL_TYPE_MODULE:
2562	return KMP_HW_MODULE;
2563	case INTEL_LEVEL_TYPE_DIE:
2564	return KMP_HW_DIE;
2565	}
2566	return KMP_HW_UNKNOWN;
2567	}
2568
2569	// This function takes the topology leaf, a levels array to store the levels
2570	// detected and a bitmap of the known levels.
2571	// Returns the number of levels in the topology
2572	static unsigned
2573	__kmp_x2apicid_get_levels(int leaf,
2574	cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST],
2575	kmp_uint64 known_levels) {
2576	unsigned level, levels_index;
2577	unsigned level_type, mask_width, nitems;
2578	kmp_cpuid buf;
2579
2580	// New algorithm has known topology layers act as highest unknown topology
2581	// layers when unknown topology layers exist.
2582	// e.g., Suppose layers were SMT <X> CORE <Y> <Z> PACKAGE, where <X> <Y> <Z>
2583	// are unknown topology layers, Then SMT will take the characteristics of
2584	// (SMT x <X>) and CORE will take the characteristics of (CORE x <Y> x <Z>).
2585	// This eliminates unknown portions of the topology while still keeping the
2586	// correct structure.
2587	level = levels_index = 0;
2588	do {
2589	__kmp_x86_cpuid(leaf, subleaf: level, p: &buf);
2590	level_type = __kmp_extract_bits<8, 15>(v: buf.ecx);
2591	mask_width = __kmp_extract_bits<0, 4>(v: buf.eax);
2592	nitems = __kmp_extract_bits<0, 15>(v: buf.ebx);
2593	if (level_type != INTEL_LEVEL_TYPE_INVALID && nitems == 0)
2594	return 0;
2595
2596	if (known_levels & (1ull << level_type)) {
2597	// Add a new level to the topology
2598	KMP_ASSERT(levels_index < INTEL_LEVEL_TYPE_LAST);
2599	levels[levels_index].level_type = level_type;
2600	levels[levels_index].mask_width = mask_width;
2601	levels[levels_index].nitems = nitems;
2602	levels_index++;
2603	} else {
2604	// If it is an unknown level, then logically move the previous layer up
2605	if (levels_index > 0) {
2606	levels[levels_index - 1].mask_width = mask_width;
2607	levels[levels_index - 1].nitems = nitems;
2608	}
2609	}
2610	level++;
2611	} while (level_type != INTEL_LEVEL_TYPE_INVALID);
2612
2613	// Ensure the INTEL_LEVEL_TYPE_INVALID (Socket) layer isn't first
2614	if (levels_index == 0 || levels[0].level_type == INTEL_LEVEL_TYPE_INVALID)
2615	return 0;
2616
2617	// Set the masks to & with apicid
2618	for (unsigned i = 0; i < levels_index; ++i) {
2619	if (levels[i].level_type != INTEL_LEVEL_TYPE_INVALID) {
2620	levels[i].mask = ~((-1) << levels[i].mask_width);
2621	levels[i].cache_mask = (-1) << levels[i].mask_width;
2622	for (unsigned j = 0; j < i; ++j)
2623	levels[i].mask ^= levels[j].mask;
2624	} else {
2625	KMP_DEBUG_ASSERT(i > 0);
2626	levels[i].mask = (-1) << levels[i - 1].mask_width;
2627	levels[i].cache_mask = 0;
2628	}
2629	}
2630	return levels_index;
2631	}
2632
2633	static bool __kmp_affinity_create_x2apicid_map(kmp_i18n_id_t *const msg_id) {
2634
2635	cpuid_level_info_t levels[INTEL_LEVEL_TYPE_LAST];
2636	kmp_hw_t types[INTEL_LEVEL_TYPE_LAST];
2637	unsigned levels_index;
2638	kmp_cpuid buf;
2639	kmp_uint64 known_levels;
2640	int topology_leaf, highest_leaf, apic_id;
2641	int num_leaves;
2642	static int leaves[] = {0, 0};
2643
2644	kmp_i18n_id_t leaf_message_id;
2645
2646	KMP_BUILD_ASSERT(sizeof(known_levels) * CHAR_BIT > KMP_HW_LAST);
2647
2648	*msg_id = kmp_i18n_null;
2649	if (__kmp_affinity.flags.verbose) {
2650	KMP_INFORM(AffInfoStr, "KMP_AFFINITY", KMP_I18N_STR(Decodingx2APIC));
2651	}
2652
2653	// Figure out the known topology levels
2654	known_levels = 0ull;
2655	for (int i = 0; i < INTEL_LEVEL_TYPE_LAST; ++i) {
2656	if (__kmp_intel_type_2_topology_type(intel_type: i) != KMP_HW_UNKNOWN) {
2657	known_levels |= (1ull << i);
2658	}
2659	}
2660
2661	// Get the highest cpuid leaf supported
2662	__kmp_x86_cpuid(leaf: 0, subleaf: 0, p: &buf);
2663	highest_leaf = buf.eax;
2664
2665	// If a specific topology method was requested, only allow that specific leaf
2666	// otherwise, try both leaves 31 and 11 in that order
2667	num_leaves = 0;
2668	if (__kmp_affinity_top_method == affinity_top_method_x2apicid) {
2669	num_leaves = 1;
2670	leaves[0] = 11;
2671	leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2672	} else if (__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
2673	num_leaves = 1;
2674	leaves[0] = 31;
2675	leaf_message_id = kmp_i18n_str_NoLeaf31Support;
2676	} else {
2677	num_leaves = 2;
2678	leaves[0] = 31;
2679	leaves[1] = 11;
2680	leaf_message_id = kmp_i18n_str_NoLeaf11Support;
2681	}
2682
2683	// Check to see if cpuid leaf 31 or 11 is supported.
2684	__kmp_nThreadsPerCore = nCoresPerPkg = nPackages = 1;
2685	topology_leaf = -1;
2686	for (int i = 0; i < num_leaves; ++i) {
2687	int leaf = leaves[i];
2688	if (highest_leaf < leaf)
2689	continue;
2690	__kmp_x86_cpuid(leaf, subleaf: 0, p: &buf);
2691	if (buf.ebx == 0)
2692	continue;
2693	topology_leaf = leaf;
2694	levels_index = __kmp_x2apicid_get_levels(leaf, levels, known_levels);
2695	if (levels_index == 0)
2696	continue;
2697	break;
2698	}
2699	if (topology_leaf == -1 || levels_index == 0) {
2700	*msg_id = leaf_message_id;
2701	return false;
2702	}
2703	KMP_ASSERT(levels_index <= INTEL_LEVEL_TYPE_LAST);
2704
2705	// The algorithm used starts by setting the affinity to each available thread
2706	// and retrieving info from the cpuid instruction, so if we are not capable of
2707	// calling __kmp_get_system_affinity() and __kmp_get_system_affinity(), then
2708	// we need to do something else - use the defaults that we calculated from
2709	// issuing cpuid without binding to each proc.
2710	if (!KMP_AFFINITY_CAPABLE()) {
2711	// Hack to try and infer the machine topology using only the data
2712	// available from cpuid on the current thread, and __kmp_xproc.
2713	KMP_ASSERT(__kmp_affinity.type == affinity_none);
2714	for (unsigned i = 0; i < levels_index; ++i) {
2715	if (levels[i].level_type == INTEL_LEVEL_TYPE_SMT) {
2716	__kmp_nThreadsPerCore = levels[i].nitems;
2717	} else if (levels[i].level_type == INTEL_LEVEL_TYPE_CORE) {
2718	nCoresPerPkg = levels[i].nitems;
2719	}
2720	}
2721	__kmp_ncores = __kmp_xproc / __kmp_nThreadsPerCore;
2722	nPackages = (__kmp_xproc + nCoresPerPkg - 1) / nCoresPerPkg;
2723	return true;
2724	}
2725
2726	// Allocate the data structure to be returned.
2727	int depth = levels_index;
2728	for (int i = depth - 1, j = 0; i >= 0; --i, ++j)
2729	types[j] = __kmp_intel_type_2_topology_type(intel_type: levels[i].level_type);
2730	__kmp_topology =
2731	kmp_topology_t::allocate(nproc: __kmp_avail_proc, ndepth: levels_index, types);
2732
2733	// Insert equivalent cache types if they exist
2734	kmp_cache_info_t cache_info;
2735	for (size_t i = 0; i < cache_info.get_depth(); ++i) {
2736	const kmp_cache_info_t::info_t &info = cache_info[i];
2737	unsigned cache_mask = info.mask;
2738	unsigned cache_level = info.level;
2739	for (unsigned j = 0; j < levels_index; ++j) {
2740	unsigned hw_cache_mask = levels[j].cache_mask;
2741	kmp_hw_t cache_type = kmp_cache_info_t::get_topology_type(level: cache_level);
2742	if (hw_cache_mask == cache_mask && j < levels_index - 1) {
2743	kmp_hw_t type =
2744	__kmp_intel_type_2_topology_type(intel_type: levels[j + 1].level_type);
2745	__kmp_topology->set_equivalent_type(type1: cache_type, type2: type);
2746	}
2747	}
2748	}
2749
2750	// From here on, we can assume that it is safe to call
2751	// __kmp_get_system_affinity() and __kmp_set_system_affinity(), even if
2752	// __kmp_affinity.type = affinity_none.
2753
2754	// Save the affinity mask for the current thread.
2755	kmp_affinity_raii_t previous_affinity;
2756
2757	// Run through each of the available contexts, binding the current thread
2758	// to it, and obtaining the pertinent information using the cpuid instr.
2759	unsigned int proc;
2760	int hw_thread_index = 0;
2761	KMP_CPU_SET_ITERATE(proc, __kmp_affin_fullMask) {
2762	cpuid_level_info_t my_levels[INTEL_LEVEL_TYPE_LAST];
2763	unsigned my_levels_index;
2764
2765	// Skip this proc if it is not included in the machine model.
2766	if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
2767	continue;
2768	}
2769	KMP_DEBUG_ASSERT(hw_thread_index < __kmp_avail_proc);
2770
2771	__kmp_affinity_dispatch->bind_thread(proc);
2772
2773	// New algorithm
2774	__kmp_x86_cpuid(leaf: topology_leaf, subleaf: 0, p: &buf);
2775	apic_id = buf.edx;
2776	kmp_hw_thread_t &hw_thread = __kmp_topology->at(index: hw_thread_index);
2777	my_levels_index =
2778	__kmp_x2apicid_get_levels(leaf: topology_leaf, levels: my_levels, known_levels);
2779	if (my_levels_index == 0 || my_levels_index != levels_index) {
2780	*msg_id = kmp_i18n_str_InvalidCpuidInfo;
2781	return false;
2782	}
2783	hw_thread.clear();
2784	hw_thread.os_id = proc;
2785	// Put in topology information
2786	for (unsigned j = 0, idx = depth - 1; j < my_levels_index; ++j, --idx) {
2787	hw_thread.ids[idx] = apic_id & my_levels[j].mask;
2788	if (j > 0) {
2789	hw_thread.ids[idx] >>= my_levels[j - 1].mask_width;
2790	}
2791	}
2792	// Hybrid information
2793	if (__kmp_is_hybrid_cpu() && highest_leaf >= 0x1a) {
2794	kmp_hw_core_type_t type;
2795	unsigned native_model_id;
2796	int efficiency;
2797	__kmp_get_hybrid_info(type: &type, efficiency: &efficiency, native_model_id: &native_model_id);
2798	hw_thread.attrs.set_core_type(type);
2799	hw_thread.attrs.set_core_eff(efficiency);
2800	}
2801	hw_thread_index++;
2802	}
2803	KMP_ASSERT(hw_thread_index > 0);
2804	__kmp_topology->sort_ids();
2805	if (!__kmp_topology->check_ids()) {
2806	kmp_topology_t::deallocate(topology: __kmp_topology);
2807	__kmp_topology = nullptr;
2808	*msg_id = kmp_i18n_str_x2ApicIDsNotUnique;
2809	return false;
2810	}
2811	return true;
2812	}
2813	#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2814
2815	#define osIdIndex 0
2816	#define threadIdIndex 1
2817	#define coreIdIndex 2
2818	#define pkgIdIndex 3
2819	#define nodeIdIndex 4
2820
2821	typedef unsigned *ProcCpuInfo;
2822	static unsigned maxIndex = pkgIdIndex;
2823
2824	static int __kmp_affinity_cmp_ProcCpuInfo_phys_id(const void *a,
2825	const void *b) {
2826	unsigned i;
2827	const unsigned *aa = *(unsigned *const *)a;
2828	const unsigned *bb = *(unsigned *const *)b;
2829	for (i = maxIndex;; i--) {
2830	if (aa[i] < bb[i])
2831	return -1;
2832	if (aa[i] > bb[i])
2833	return 1;
2834	if (i == osIdIndex)
2835	break;
2836	}
2837	return 0;
2838	}
2839
2840	#if KMP_USE_HIER_SCHED
2841	// Set the array sizes for the hierarchy layers
2842	static void __kmp_dispatch_set_hierarchy_values() {
2843	// Set the maximum number of L1's to number of cores
2844	// Set the maximum number of L2's to either number of cores / 2 for
2845	// Intel(R) Xeon Phi(TM) coprocessor formally codenamed Knights Landing
2846	// Or the number of cores for Intel(R) Xeon(R) processors
2847	// Set the maximum number of NUMA nodes and L3's to number of packages
2848	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
2849	nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2850	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
2851	#if KMP_ARCH_X86_64 && \
2852	(KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
2853	KMP_OS_WINDOWS) && \
2854	KMP_MIC_SUPPORTED
2855	if (__kmp_mic_type >= mic3)
2856	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
2857	else
2858	#endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2859	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores;
2860	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_L3 + 1] = nPackages;
2861	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_NUMA + 1] = nPackages;
2862	__kmp_hier_max_units[kmp_hier_layer_e::LAYER_LOOP + 1] = 1;
2863	// Set the number of threads per unit
2864	// Number of hardware threads per L1/L2/L3/NUMA/LOOP
2865	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
2866	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
2867	__kmp_nThreadsPerCore;
2868	#if KMP_ARCH_X86_64 && \
2869	(KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
2870	KMP_OS_WINDOWS) && \
2871	KMP_MIC_SUPPORTED
2872	if (__kmp_mic_type >= mic3)
2873	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2874	2 * __kmp_nThreadsPerCore;
2875	else
2876	#endif // KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_WINDOWS)
2877	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
2878	__kmp_nThreadsPerCore;
2879	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L3 + 1] =
2880	nCoresPerPkg * __kmp_nThreadsPerCore;
2881	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_NUMA + 1] =
2882	nCoresPerPkg * __kmp_nThreadsPerCore;
2883	__kmp_hier_threads_per[kmp_hier_layer_e::LAYER_LOOP + 1] =
2884	nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
2885	}
2886
2887	// Return the index into the hierarchy for this tid and layer type (L1, L2, etc)
2888	// i.e., this thread's L1 or this thread's L2, etc.
2889	int __kmp_dispatch_get_index(int tid, kmp_hier_layer_e type) {
2890	int index = type + 1;
2891	int num_hw_threads = __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1];
2892	KMP_DEBUG_ASSERT(type != kmp_hier_layer_e::LAYER_LAST);
2893	if (type == kmp_hier_layer_e::LAYER_THREAD)
2894	return tid;
2895	else if (type == kmp_hier_layer_e::LAYER_LOOP)
2896	return 0;
2897	KMP_DEBUG_ASSERT(__kmp_hier_max_units[index] != 0);
2898	if (tid >= num_hw_threads)
2899	tid = tid % num_hw_threads;
2900	return (tid / __kmp_hier_threads_per[index]) % __kmp_hier_max_units[index];
2901	}
2902
2903	// Return the number of t1's per t2
2904	int __kmp_dispatch_get_t1_per_t2(kmp_hier_layer_e t1, kmp_hier_layer_e t2) {
2905	int i1 = t1 + 1;
2906	int i2 = t2 + 1;
2907	KMP_DEBUG_ASSERT(i1 <= i2);
2908	KMP_DEBUG_ASSERT(t1 != kmp_hier_layer_e::LAYER_LAST);
2909	KMP_DEBUG_ASSERT(t2 != kmp_hier_layer_e::LAYER_LAST);
2910	KMP_DEBUG_ASSERT(__kmp_hier_threads_per[i1] != 0);
2911	// (nthreads/t2) / (nthreads/t1) = t1 / t2
2912	return __kmp_hier_threads_per[i2] / __kmp_hier_threads_per[i1];
2913	}
2914	#endif // KMP_USE_HIER_SCHED
2915
2916	static inline const char *__kmp_cpuinfo_get_filename() {
2917	const char *filename;
2918	if (__kmp_cpuinfo_file != nullptr)
2919	filename = __kmp_cpuinfo_file;
2920	else
2921	filename = "/proc/cpuinfo";
2922	return filename;
2923	}
2924
2925	static inline const char *__kmp_cpuinfo_get_envvar() {
2926	const char *envvar = nullptr;
2927	if (__kmp_cpuinfo_file != nullptr)
2928	envvar = "KMP_CPUINFO_FILE";
2929	return envvar;
2930	}
2931
2932	// Parse /proc/cpuinfo (or an alternate file in the same format) to obtain the
2933	// affinity map. On AIX, the map is obtained through system SRAD (Scheduler
2934	// Resource Allocation Domain).
2935	static bool __kmp_affinity_create_cpuinfo_map(int *line,
2936	kmp_i18n_id_t *const msg_id) {
2937	*msg_id = kmp_i18n_null;
2938
2939	#if KMP_OS_AIX
2940	unsigned num_records = __kmp_xproc;
2941	#else
2942	const char *filename = __kmp_cpuinfo_get_filename();
2943	const char *envvar = __kmp_cpuinfo_get_envvar();
2944
2945	if (__kmp_affinity.flags.verbose) {
2946	KMP_INFORM(AffParseFilename, "KMP_AFFINITY", filename);
2947	}
2948
2949	kmp_safe_raii_file_t f(filename, "r", envvar);
2950
2951	// Scan of the file, and count the number of "processor" (osId) fields,
2952	// and find the highest value of <n> for a node_<n> field.
2953	char buf[256];
2954	unsigned num_records = 0;
2955	while (!feof(stream: f)) {
2956	buf[sizeof(buf) - 1] = 1;
2957	if (!fgets(s: buf, n: sizeof(buf), stream: f)) {
2958	// Read errors presumably because of EOF
2959	break;
2960	}
2961
2962	char s1[] = "processor";
2963	if (strncmp(s1: buf, s2: s1, n: sizeof(s1) - 1) == 0) {
2964	num_records++;
2965	continue;
2966	}
2967
2968	// FIXME - this will match "node_<n> <garbage>"
2969	unsigned level;
2970	if (KMP_SSCANF(s: buf, format: "node_%u id", &level) == 1) {
2971	// validate the input fisrt:
2972	if (level > (unsigned)__kmp_xproc) { // level is too big
2973	level = __kmp_xproc;
2974	}
2975	if (nodeIdIndex + level >= maxIndex) {
2976	maxIndex = nodeIdIndex + level;
2977	}
2978	continue;
2979	}
2980	}
2981
2982	// Check for empty file / no valid processor records, or too many. The number
2983	// of records can't exceed the number of valid bits in the affinity mask.
2984	if (num_records == 0) {
2985	*msg_id = kmp_i18n_str_NoProcRecords;
2986	return false;
2987	}
2988	if (num_records > (unsigned)__kmp_xproc) {
2989	*msg_id = kmp_i18n_str_TooManyProcRecords;
2990	return false;
2991	}
2992
2993	// Set the file pointer back to the beginning, so that we can scan the file
2994	// again, this time performing a full parse of the data. Allocate a vector of
2995	// ProcCpuInfo object, where we will place the data. Adding an extra element
2996	// at the end allows us to remove a lot of extra checks for termination
2997	// conditions.
2998	if (fseek(stream: f, off: 0, SEEK_SET) != 0) {
2999	*msg_id = kmp_i18n_str_CantRewindCpuinfo;
3000	return false;
3001	}
3002	#endif // KMP_OS_AIX
3003
3004	// Allocate the array of records to store the proc info in. The dummy
3005	// element at the end makes the logic in filling them out easier to code.
3006	unsigned **threadInfo =
3007	(unsigned **)__kmp_allocate((num_records + 1) * sizeof(unsigned *));
3008	unsigned i;
3009	for (i = 0; i <= num_records; i++) {
3010	threadInfo[i] =
3011	(unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3012	}
3013
3014	#define CLEANUP_THREAD_INFO \
3015	for (i = 0; i <= num_records; i++) { \
3016	__kmp_free(threadInfo[i]); \
3017	} \
3018	__kmp_free(threadInfo);
3019
3020	// A value of UINT_MAX means that we didn't find the field
3021	unsigned __index;
3022
3023	#define INIT_PROC_INFO(p) \
3024	for (__index = 0; __index <= maxIndex; __index++) { \
3025	(p)[__index] = UINT_MAX; \
3026	}
3027
3028	for (i = 0; i <= num_records; i++) {
3029	INIT_PROC_INFO(threadInfo[i]);
3030	}
3031
3032	#if KMP_OS_AIX
3033	int smt_threads;
3034	lpar_info_format1_t cpuinfo;
3035	unsigned num_avail = __kmp_xproc;
3036
3037	if (__kmp_affinity.flags.verbose)
3038	KMP_INFORM(AffParseFilename, "KMP_AFFINITY", "system info for topology");
3039
3040	// Get the number of SMT threads per core.
3041	int retval =
3042	lpar_get_info(LPAR_INFO_FORMAT1, &cpuinfo, sizeof(lpar_info_format1_t));
3043	if (!retval)
3044	smt_threads = cpuinfo.smt_threads;
3045	else {
3046	CLEANUP_THREAD_INFO;
3047	*msg_id = kmp_i18n_str_UnknownTopology;
3048	return false;
3049	}
3050
3051	// Allocate a resource set containing available system resourses.
3052	rsethandle_t sys_rset = rs_alloc(RS_SYSTEM);
3053	if (sys_rset == NULL) {
3054	CLEANUP_THREAD_INFO;
3055	*msg_id = kmp_i18n_str_UnknownTopology;
3056	return false;
3057	}
3058	// Allocate a resource set for the SRAD info.
3059	rsethandle_t srad = rs_alloc(RS_EMPTY);
3060	if (srad == NULL) {
3061	rs_free(sys_rset);
3062	CLEANUP_THREAD_INFO;
3063	*msg_id = kmp_i18n_str_UnknownTopology;
3064	return false;
3065	}
3066
3067	// Get the SRAD system detail level.
3068	int sradsdl = rs_getinfo(NULL, R_SRADSDL, 0);
3069	if (sradsdl < 0) {
3070	rs_free(sys_rset);
3071	rs_free(srad);
3072	CLEANUP_THREAD_INFO;
3073	*msg_id = kmp_i18n_str_UnknownTopology;
3074	return false;
3075	}
3076	// Get the number of RADs at that SRAD SDL.
3077	int num_rads = rs_numrads(sys_rset, sradsdl, 0);
3078	if (num_rads < 0) {
3079	rs_free(sys_rset);
3080	rs_free(srad);
3081	CLEANUP_THREAD_INFO;
3082	*msg_id = kmp_i18n_str_UnknownTopology;
3083	return false;
3084	}
3085
3086	// Get the maximum number of procs that may be contained in a resource set.
3087	int max_procs = rs_getinfo(NULL, R_MAXPROCS, 0);
3088	if (max_procs < 0) {
3089	rs_free(sys_rset);
3090	rs_free(srad);
3091	CLEANUP_THREAD_INFO;
3092	*msg_id = kmp_i18n_str_UnknownTopology;
3093	return false;
3094	}
3095
3096	int cur_rad = 0;
3097	int num_set = 0;
3098	for (int srad_idx = 0; cur_rad < num_rads && srad_idx < VMI_MAXRADS;
3099	++srad_idx) {
3100	// Check if the SRAD is available in the RSET.
3101	if (rs_getrad(sys_rset, srad, sradsdl, srad_idx, 0) < 0)
3102	continue;
3103
3104	for (int cpu = 0; cpu < max_procs; cpu++) {
3105	// Set the info for the cpu if it is in the SRAD.
3106	if (rs_op(RS_TESTRESOURCE, srad, NULL, R_PROCS, cpu)) {
3107	threadInfo[cpu][osIdIndex] = cpu;
3108	threadInfo[cpu][pkgIdIndex] = cur_rad;
3109	threadInfo[cpu][coreIdIndex] = cpu / smt_threads;
3110	++num_set;
3111	if (num_set >= num_avail) {
3112	// Done if all available CPUs have been set.
3113	break;
3114	}
3115	}
3116	}
3117	++cur_rad;
3118	}
3119	rs_free(sys_rset);
3120	rs_free(srad);
3121
3122	// The topology is already sorted.
3123
3124	#else // !KMP_OS_AIX
3125	unsigned num_avail = 0;
3126	*line = 0;
3127	#if KMP_ARCH_S390X
3128	bool reading_s390x_sys_info = true;
3129	#endif
3130	while (!feof(stream: f)) {
3131	// Create an inner scoping level, so that all the goto targets at the end of
3132	// the loop appear in an outer scoping level. This avoids warnings about
3133	// jumping past an initialization to a target in the same block.
3134	{
3135	buf[sizeof(buf) - 1] = 1;
3136	bool long_line = false;
3137	if (!fgets(s: buf, n: sizeof(buf), stream: f)) {
3138	// Read errors presumably because of EOF
3139	// If there is valid data in threadInfo[num_avail], then fake
3140	// a blank line in ensure that the last address gets parsed.
3141	bool valid = false;
3142	for (i = 0; i <= maxIndex; i++) {
3143	if (threadInfo[num_avail][i] != UINT_MAX) {
3144	valid = true;
3145	}
3146	}
3147	if (!valid) {
3148	break;
3149	}
3150	buf[0] = 0;
3151	} else if (!buf[sizeof(buf) - 1]) {
3152	// The line is longer than the buffer. Set a flag and don't
3153	// emit an error if we were going to ignore the line, anyway.
3154	long_line = true;
3155
3156	#define CHECK_LINE \
3157	if (long_line) { \
3158	CLEANUP_THREAD_INFO; \
3159	*msg_id = kmp_i18n_str_LongLineCpuinfo; \
3160	return false; \
3161	}
3162	}
3163	(*line)++;
3164
3165	#if KMP_ARCH_LOONGARCH64
3166	// The parsing logic of /proc/cpuinfo in this function highly depends on
3167	// the blank lines between each processor info block. But on LoongArch a
3168	// blank line exists before the first processor info block (i.e. after the
3169	// "system type" line). This blank line was added because the "system
3170	// type" line is unrelated to any of the CPUs. We must skip this line so
3171	// that the original logic works on LoongArch.
3172	if (*buf == '\n' && *line == 2)
3173	continue;
3174	#endif
3175	#if KMP_ARCH_S390X
3176	// s390x /proc/cpuinfo starts with a variable number of lines containing
3177	// the overall system information. Skip them.
3178	if (reading_s390x_sys_info) {
3179	if (*buf == '\n')
3180	reading_s390x_sys_info = false;
3181	continue;
3182	}
3183	#endif
3184
3185	#if KMP_ARCH_S390X
3186	char s1[] = "cpu number";
3187	#else
3188	char s1[] = "processor";
3189	#endif
3190	if (strncmp(s1: buf, s2: s1, n: sizeof(s1) - 1) == 0) {
3191	CHECK_LINE;
3192	char *p = strchr(s: buf + sizeof(s1) - 1, c: ':');
3193	unsigned val;
3194	if ((p == NULL) || (KMP_SSCANF(s: p + 1, format: "%u\n", &val) != 1))
3195	goto no_val;
3196	if (threadInfo[num_avail][osIdIndex] != UINT_MAX)
3197	#if KMP_ARCH_AARCH64
3198	// Handle the old AArch64 /proc/cpuinfo layout differently,
3199	// it contains all of the 'processor' entries listed in a
3200	// single 'Processor' section, therefore the normal looking
3201	// for duplicates in that section will always fail.
3202	num_avail++;
3203	#else
3204	goto dup_field;
3205	#endif
3206	threadInfo[num_avail][osIdIndex] = val;
3207	#if KMP_OS_LINUX && !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
3208	char path[256];
3209	KMP_SNPRINTF(
3210	path, sizeof(path),
3211	"/sys/devices/system/cpu/cpu%u/topology/physical_package_id",
3212	threadInfo[num_avail][osIdIndex]);
3213	__kmp_read_from_file(path, "%u", &threadInfo[num_avail][pkgIdIndex]);
3214
3215	#if KMP_ARCH_S390X
3216	// Disambiguate physical_package_id.
3217	unsigned book_id;
3218	KMP_SNPRINTF(path, sizeof(path),
3219	"/sys/devices/system/cpu/cpu%u/topology/book_id",
3220	threadInfo[num_avail][osIdIndex]);
3221	__kmp_read_from_file(path, "%u", &book_id);
3222	threadInfo[num_avail][pkgIdIndex] |= (book_id << 8);
3223
3224	unsigned drawer_id;
3225	KMP_SNPRINTF(path, sizeof(path),
3226	"/sys/devices/system/cpu/cpu%u/topology/drawer_id",
3227	threadInfo[num_avail][osIdIndex]);
3228	__kmp_read_from_file(path, "%u", &drawer_id);
3229	threadInfo[num_avail][pkgIdIndex] |= (drawer_id << 16);
3230	#endif
3231
3232	KMP_SNPRINTF(path, sizeof(path),
3233	"/sys/devices/system/cpu/cpu%u/topology/core_id",
3234	threadInfo[num_avail][osIdIndex]);
3235	__kmp_read_from_file(path, "%u", &threadInfo[num_avail][coreIdIndex]);
3236	continue;
3237	#else
3238	}
3239	char s2[] = "physical id";
3240	if (strncmp(s1: buf, s2: s2, n: sizeof(s2) - 1) == 0) {
3241	CHECK_LINE;
3242	char *p = strchr(s: buf + sizeof(s2) - 1, c: ':');
3243	unsigned val;
3244	if ((p == NULL) || (KMP_SSCANF(s: p + 1, format: "%u\n", &val) != 1))
3245	goto no_val;
3246	if (threadInfo[num_avail][pkgIdIndex] != UINT_MAX)
3247	goto dup_field;
3248	threadInfo[num_avail][pkgIdIndex] = val;
3249	continue;
3250	}
3251	char s3[] = "core id";
3252	if (strncmp(s1: buf, s2: s3, n: sizeof(s3) - 1) == 0) {
3253	CHECK_LINE;
3254	char *p = strchr(s: buf + sizeof(s3) - 1, c: ':');
3255	unsigned val;
3256	if ((p == NULL) || (KMP_SSCANF(s: p + 1, format: "%u\n", &val) != 1))
3257	goto no_val;
3258	if (threadInfo[num_avail][coreIdIndex] != UINT_MAX)
3259	goto dup_field;
3260	threadInfo[num_avail][coreIdIndex] = val;
3261	continue;
3262	#endif // KMP_OS_LINUX && USE_SYSFS_INFO
3263	}
3264	char s4[] = "thread id";
3265	if (strncmp(s1: buf, s2: s4, n: sizeof(s4) - 1) == 0) {
3266	CHECK_LINE;
3267	char *p = strchr(s: buf + sizeof(s4) - 1, c: ':');
3268	unsigned val;
3269	if ((p == NULL) || (KMP_SSCANF(s: p + 1, format: "%u\n", &val) != 1))
3270	goto no_val;
3271	if (threadInfo[num_avail][threadIdIndex] != UINT_MAX)
3272	goto dup_field;
3273	threadInfo[num_avail][threadIdIndex] = val;
3274	continue;
3275	}
3276	unsigned level;
3277	if (KMP_SSCANF(s: buf, format: "node_%u id", &level) == 1) {
3278	CHECK_LINE;
3279	char *p = strchr(s: buf + sizeof(s4) - 1, c: ':');
3280	unsigned val;
3281	if ((p == NULL) || (KMP_SSCANF(s: p + 1, format: "%u\n", &val) != 1))
3282	goto no_val;
3283	// validate the input before using level:
3284	if (level > (unsigned)__kmp_xproc) { // level is too big
3285	level = __kmp_xproc;
3286	}
3287	if (threadInfo[num_avail][nodeIdIndex + level] != UINT_MAX)
3288	goto dup_field;
3289	threadInfo[num_avail][nodeIdIndex + level] = val;
3290	continue;
3291	}
3292
3293	// We didn't recognize the leading token on the line. There are lots of
3294	// leading tokens that we don't recognize - if the line isn't empty, go on
3295	// to the next line.
3296	if ((*buf != 0) && (*buf != '\n')) {
3297	// If the line is longer than the buffer, read characters
3298	// until we find a newline.
3299	if (long_line) {
3300	int ch;
3301	while (((ch = fgetc(stream: f)) != EOF) && (ch != '\n'))
3302	;
3303	}
3304	continue;
3305	}
3306
3307	// A newline has signalled the end of the processor record.
3308	// Check that there aren't too many procs specified.
3309	if ((int)num_avail == __kmp_xproc) {
3310	CLEANUP_THREAD_INFO;
3311	*msg_id = kmp_i18n_str_TooManyEntries;
3312	return false;
3313	}
3314
3315	// Check for missing fields. The osId field must be there, and we
3316	// currently require that the physical id field is specified, also.
3317	if (threadInfo[num_avail][osIdIndex] == UINT_MAX) {
3318	CLEANUP_THREAD_INFO;
3319	*msg_id = kmp_i18n_str_MissingProcField;
3320	return false;
3321	}
3322	if (threadInfo[0][pkgIdIndex] == UINT_MAX) {
3323	CLEANUP_THREAD_INFO;
3324	*msg_id = kmp_i18n_str_MissingPhysicalIDField;
3325	return false;
3326	}
3327
3328	// Skip this proc if it is not included in the machine model.
3329	if (KMP_AFFINITY_CAPABLE() &&
3330	!KMP_CPU_ISSET(threadInfo[num_avail][osIdIndex],
3331	__kmp_affin_fullMask)) {
3332	INIT_PROC_INFO(threadInfo[num_avail]);
3333	continue;
3334	}
3335
3336	// We have a successful parse of this proc's info.
3337	// Increment the counter, and prepare for the next proc.
3338	num_avail++;
3339	KMP_ASSERT(num_avail <= num_records);
3340	INIT_PROC_INFO(threadInfo[num_avail]);
3341	}
3342	continue;
3343
3344	no_val:
3345	CLEANUP_THREAD_INFO;
3346	*msg_id = kmp_i18n_str_MissingValCpuinfo;
3347	return false;
3348
3349	dup_field:
3350	CLEANUP_THREAD_INFO;
3351	*msg_id = kmp_i18n_str_DuplicateFieldCpuinfo;
3352	return false;
3353	}
3354	*line = 0;
3355
3356	#if KMP_MIC && REDUCE_TEAM_SIZE
3357	unsigned teamSize = 0;
3358	#endif // KMP_MIC && REDUCE_TEAM_SIZE
3359
3360	// check for num_records == __kmp_xproc ???
3361
3362	// If it is configured to omit the package level when there is only a single
3363	// package, the logic at the end of this routine won't work if there is only a
3364	// single thread
3365	KMP_ASSERT(num_avail > 0);
3366	KMP_ASSERT(num_avail <= num_records);
3367
3368	// Sort the threadInfo table by physical Id.
3369	qsort(base: threadInfo, nmemb: num_avail, size: sizeof(*threadInfo),
3370	compar: __kmp_affinity_cmp_ProcCpuInfo_phys_id);
3371
3372	#endif // KMP_OS_AIX
3373
3374	// The table is now sorted by pkgId / coreId / threadId, but we really don't
3375	// know the radix of any of the fields. pkgId's may be sparsely assigned among
3376	// the chips on a system. Although coreId's are usually assigned
3377	// [0 .. coresPerPkg-1] and threadId's are usually assigned
3378	// [0..threadsPerCore-1], we don't want to make any such assumptions.
3379	//
3380	// For that matter, we don't know what coresPerPkg and threadsPerCore (or the
3381	// total # packages) are at this point - we want to determine that now. We
3382	// only have an upper bound on the first two figures.
3383	unsigned *counts =
3384	(unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3385	unsigned *maxCt =
3386	(unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3387	unsigned *totals =
3388	(unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3389	unsigned *lastId =
3390	(unsigned *)__kmp_allocate((maxIndex + 1) * sizeof(unsigned));
3391
3392	bool assign_thread_ids = false;
3393	unsigned threadIdCt;
3394	unsigned index;
3395
3396	restart_radix_check:
3397	threadIdCt = 0;
3398
3399	// Initialize the counter arrays with data from threadInfo[0].
3400	if (assign_thread_ids) {
3401	if (threadInfo[0][threadIdIndex] == UINT_MAX) {
3402	threadInfo[0][threadIdIndex] = threadIdCt++;
3403	} else if (threadIdCt <= threadInfo[0][threadIdIndex]) {
3404	threadIdCt = threadInfo[0][threadIdIndex] + 1;
3405	}
3406	}
3407	for (index = 0; index <= maxIndex; index++) {
3408	counts[index] = 1;
3409	maxCt[index] = 1;
3410	totals[index] = 1;
3411	lastId[index] = threadInfo[0][index];
3412	;
3413	}
3414
3415	// Run through the rest of the OS procs.
3416	for (i = 1; i < num_avail; i++) {
3417	// Find the most significant index whose id differs from the id for the
3418	// previous OS proc.
3419	for (index = maxIndex; index >= threadIdIndex; index--) {
3420	if (assign_thread_ids && (index == threadIdIndex)) {
3421	// Auto-assign the thread id field if it wasn't specified.
3422	if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3423	threadInfo[i][threadIdIndex] = threadIdCt++;
3424	}
3425	// Apparently the thread id field was specified for some entries and not
3426	// others. Start the thread id counter off at the next higher thread id.
3427	else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3428	threadIdCt = threadInfo[i][threadIdIndex] + 1;
3429	}
3430	}
3431	if (threadInfo[i][index] != lastId[index]) {
3432	// Run through all indices which are less significant, and reset the
3433	// counts to 1. At all levels up to and including index, we need to
3434	// increment the totals and record the last id.
3435	unsigned index2;
3436	for (index2 = threadIdIndex; index2 < index; index2++) {
3437	totals[index2]++;
3438	if (counts[index2] > maxCt[index2]) {
3439	maxCt[index2] = counts[index2];
3440	}
3441	counts[index2] = 1;
3442	lastId[index2] = threadInfo[i][index2];
3443	}
3444	counts[index]++;
3445	totals[index]++;
3446	lastId[index] = threadInfo[i][index];
3447
3448	if (assign_thread_ids && (index > threadIdIndex)) {
3449
3450	#if KMP_MIC && REDUCE_TEAM_SIZE
3451	// The default team size is the total #threads in the machine
3452	// minus 1 thread for every core that has 3 or more threads.
3453	teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3454	#endif // KMP_MIC && REDUCE_TEAM_SIZE
3455
3456	// Restart the thread counter, as we are on a new core.
3457	threadIdCt = 0;
3458
3459	// Auto-assign the thread id field if it wasn't specified.
3460	if (threadInfo[i][threadIdIndex] == UINT_MAX) {
3461	threadInfo[i][threadIdIndex] = threadIdCt++;
3462	}
3463
3464	// Apparently the thread id field was specified for some entries and
3465	// not others. Start the thread id counter off at the next higher
3466	// thread id.
3467	else if (threadIdCt <= threadInfo[i][threadIdIndex]) {
3468	threadIdCt = threadInfo[i][threadIdIndex] + 1;
3469	}
3470	}
3471	break;
3472	}
3473	}
3474	if (index < threadIdIndex) {
3475	// If thread ids were specified, it is an error if they are not unique.
3476	// Also, check that we waven't already restarted the loop (to be safe -
3477	// shouldn't need to).
3478	if ((threadInfo[i][threadIdIndex] != UINT_MAX) || assign_thread_ids) {
3479	__kmp_free(lastId);
3480	__kmp_free(totals);
3481	__kmp_free(maxCt);
3482	__kmp_free(counts);
3483	CLEANUP_THREAD_INFO;
3484	*msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3485	return false;
3486	}
3487
3488	// If the thread ids were not specified and we see entries that
3489	// are duplicates, start the loop over and assign the thread ids manually.
3490	assign_thread_ids = true;
3491	goto restart_radix_check;
3492	}
3493	}
3494
3495	#if KMP_MIC && REDUCE_TEAM_SIZE
3496	// The default team size is the total #threads in the machine
3497	// minus 1 thread for every core that has 3 or more threads.
3498	teamSize += (threadIdCt <= 2) ? (threadIdCt) : (threadIdCt - 1);
3499	#endif // KMP_MIC && REDUCE_TEAM_SIZE
3500
3501	for (index = threadIdIndex; index <= maxIndex; index++) {
3502	if (counts[index] > maxCt[index]) {
3503	maxCt[index] = counts[index];
3504	}
3505	}
3506
3507	__kmp_nThreadsPerCore = maxCt[threadIdIndex];
3508	nCoresPerPkg = maxCt[coreIdIndex];
3509	nPackages = totals[pkgIdIndex];
3510
3511	// When affinity is off, this routine will still be called to set
3512	// __kmp_ncores, as well as __kmp_nThreadsPerCore, nCoresPerPkg, & nPackages.
3513	// Make sure all these vars are set correctly, and return now if affinity is
3514	// not enabled.
3515	__kmp_ncores = totals[coreIdIndex];
3516	if (!KMP_AFFINITY_CAPABLE()) {
3517	KMP_ASSERT(__kmp_affinity.type == affinity_none);
3518	return true;
3519	}
3520
3521	#if KMP_MIC && REDUCE_TEAM_SIZE
3522	// Set the default team size.
3523	if ((__kmp_dflt_team_nth == 0) && (teamSize > 0)) {
3524	__kmp_dflt_team_nth = teamSize;
3525	KA_TRACE(20, ("__kmp_affinity_create_cpuinfo_map: setting "
3526	"__kmp_dflt_team_nth = %d\n",
3527	__kmp_dflt_team_nth));
3528	}
3529	#endif // KMP_MIC && REDUCE_TEAM_SIZE
3530
3531	KMP_DEBUG_ASSERT(num_avail == (unsigned)__kmp_avail_proc);
3532
3533	// Count the number of levels which have more nodes at that level than at the
3534	// parent's level (with there being an implicit root node of the top level).
3535	// This is equivalent to saying that there is at least one node at this level
3536	// which has a sibling. These levels are in the map, and the package level is
3537	// always in the map.
3538	bool *inMap = (bool *)__kmp_allocate((maxIndex + 1) * sizeof(bool));
3539	for (index = threadIdIndex; index < maxIndex; index++) {
3540	KMP_ASSERT(totals[index] >= totals[index + 1]);
3541	inMap[index] = (totals[index] > totals[index + 1]);
3542	}
3543	inMap[maxIndex] = (totals[maxIndex] > 1);
3544	inMap[pkgIdIndex] = true;
3545	inMap[coreIdIndex] = true;
3546	inMap[threadIdIndex] = true;
3547
3548	int depth = 0;
3549	int idx = 0;
3550	kmp_hw_t types[KMP_HW_LAST];
3551	int pkgLevel = -1;
3552	int coreLevel = -1;
3553	int threadLevel = -1;
3554	for (index = threadIdIndex; index <= maxIndex; index++) {
3555	if (inMap[index]) {
3556	depth++;
3557	}
3558	}
3559	if (inMap[pkgIdIndex]) {
3560	pkgLevel = idx;
3561	types[idx++] = KMP_HW_SOCKET;
3562	}
3563	if (inMap[coreIdIndex]) {
3564	coreLevel = idx;
3565	types[idx++] = KMP_HW_CORE;
3566	}
3567	if (inMap[threadIdIndex]) {
3568	threadLevel = idx;
3569	types[idx++] = KMP_HW_THREAD;
3570	}
3571	KMP_ASSERT(depth > 0);
3572
3573	// Construct the data structure that is to be returned.
3574	__kmp_topology = kmp_topology_t::allocate(nproc: num_avail, ndepth: depth, types);
3575
3576	for (i = 0; i < num_avail; ++i) {
3577	unsigned os = threadInfo[i][osIdIndex];
3578	int src_index;
3579	kmp_hw_thread_t &hw_thread = __kmp_topology->at(index: i);
3580	hw_thread.clear();
3581	hw_thread.os_id = os;
3582
3583	idx = 0;
3584	for (src_index = maxIndex; src_index >= threadIdIndex; src_index--) {
3585	if (!inMap[src_index]) {
3586	continue;
3587	}
3588	if (src_index == pkgIdIndex) {
3589	hw_thread.ids[pkgLevel] = threadInfo[i][src_index];
3590	} else if (src_index == coreIdIndex) {
3591	hw_thread.ids[coreLevel] = threadInfo[i][src_index];
3592	} else if (src_index == threadIdIndex) {
3593	hw_thread.ids[threadLevel] = threadInfo[i][src_index];
3594	}
3595	}
3596	}
3597
3598	__kmp_free(inMap);
3599	__kmp_free(lastId);
3600	__kmp_free(totals);
3601	__kmp_free(maxCt);
3602	__kmp_free(counts);
3603	CLEANUP_THREAD_INFO;
3604	__kmp_topology->sort_ids();
3605	if (!__kmp_topology->check_ids()) {
3606	kmp_topology_t::deallocate(topology: __kmp_topology);
3607	__kmp_topology = nullptr;
3608	*msg_id = kmp_i18n_str_PhysicalIDsNotUnique;
3609	return false;
3610	}
3611	return true;
3612	}
3613
3614	// Create and return a table of affinity masks, indexed by OS thread ID.
3615	// This routine handles OR'ing together all the affinity masks of threads
3616	// that are sufficiently close, if granularity > fine.
3617	template <typename FindNextFunctionType>
3618	static void __kmp_create_os_id_masks(unsigned *numUnique,
3619	kmp_affinity_t &affinity,
3620	FindNextFunctionType find_next) {
3621	// First form a table of affinity masks in order of OS thread id.
3622	int maxOsId;
3623	int i;
3624	int numAddrs = __kmp_topology->get_num_hw_threads();
3625	int depth = __kmp_topology->get_depth();
3626	const char *env_var = __kmp_get_affinity_env_var(affinity);
3627	KMP_ASSERT(numAddrs);
3628	KMP_ASSERT(depth);
3629
3630	i = find_next(-1);
3631	// If could not find HW thread location with attributes, then return and
3632	// fallback to increment find_next and disregard core attributes.
3633	if (i >= numAddrs)
3634	return;
3635
3636	maxOsId = 0;
3637	for (i = numAddrs - 1;; --i) {
3638	int osId = __kmp_topology->at(index: i).os_id;
3639	if (osId > maxOsId) {
3640	maxOsId = osId;
3641	}
3642	if (i == 0)
3643	break;
3644	}
3645	affinity.num_os_id_masks = maxOsId + 1;
3646	KMP_CPU_ALLOC_ARRAY(affinity.os_id_masks, affinity.num_os_id_masks);
3647	KMP_ASSERT(affinity.gran_levels >= 0);
3648	if (affinity.flags.verbose && (affinity.gran_levels > 0)) {
3649	KMP_INFORM(ThreadsMigrate, env_var, affinity.gran_levels);
3650	}
3651	if (affinity.gran_levels >= (int)depth) {
3652	KMP_AFF_WARNING(affinity, AffThreadsMayMigrate);
3653	}
3654
3655	// Run through the table, forming the masks for all threads on each core.
3656	// Threads on the same core will have identical kmp_hw_thread_t objects, not
3657	// considering the last level, which must be the thread id. All threads on a
3658	// core will appear consecutively.
3659	int unique = 0;
3660	int j = 0; // index of 1st thread on core
3661	int leader = 0;
3662	kmp_affin_mask_t *sum;
3663	KMP_CPU_ALLOC_ON_STACK(sum);
3664	KMP_CPU_ZERO(sum);
3665
3666	i = j = leader = find_next(-1);
3667	KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3668	kmp_full_mask_modifier_t full_mask;
3669	for (i = find_next(i); i < numAddrs; i = find_next(i)) {
3670	// If this thread is sufficiently close to the leader (within the
3671	// granularity setting), then set the bit for this os thread in the
3672	// affinity mask for this group, and go on to the next thread.
3673	if (__kmp_topology->is_close(hwt1: leader, hwt2: i, stgs: affinity)) {
3674	KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3675	continue;
3676	}
3677
3678	// For every thread in this group, copy the mask to the thread's entry in
3679	// the OS Id mask table. Mark the first address as a leader.
3680	for (; j < i; j = find_next(j)) {
3681	int osId = __kmp_topology->at(index: j).os_id;
3682	KMP_DEBUG_ASSERT(osId <= maxOsId);
3683	kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3684	KMP_CPU_COPY(mask, sum);
3685	__kmp_topology->at(index: j).leader = (j == leader);
3686	}
3687	unique++;
3688
3689	// Start a new mask.
3690	leader = i;
3691	full_mask.include(other: sum);
3692	KMP_CPU_ZERO(sum);
3693	KMP_CPU_SET(__kmp_topology->at(i).os_id, sum);
3694	}
3695
3696	// For every thread in last group, copy the mask to the thread's
3697	// entry in the OS Id mask table.
3698	for (; j < i; j = find_next(j)) {
3699	int osId = __kmp_topology->at(index: j).os_id;
3700	KMP_DEBUG_ASSERT(osId <= maxOsId);
3701	kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.os_id_masks, osId);
3702	KMP_CPU_COPY(mask, sum);
3703	__kmp_topology->at(index: j).leader = (j == leader);
3704	}
3705	full_mask.include(other: sum);
3706	unique++;
3707	KMP_CPU_FREE_FROM_STACK(sum);
3708
3709	// See if the OS Id mask table further restricts or changes the full mask
3710	if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
3711	__kmp_topology->print(env_var);
3712	}
3713
3714	*numUnique = unique;
3715	}
3716
3717	// Stuff for the affinity proclist parsers. It's easier to declare these vars
3718	// as file-static than to try and pass them through the calling sequence of
3719	// the recursive-descent OMP_PLACES parser.
3720	static kmp_affin_mask_t *newMasks;
3721	static int numNewMasks;
3722	static int nextNewMask;
3723
3724	#define ADD_MASK(_mask) \
3725	{ \
3726	if (nextNewMask >= numNewMasks) { \
3727	int i; \
3728	numNewMasks *= 2; \
3729	kmp_affin_mask_t *temp; \
3730	KMP_CPU_INTERNAL_ALLOC_ARRAY(temp, numNewMasks); \
3731	for (i = 0; i < numNewMasks / 2; i++) { \
3732	kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i); \
3733	kmp_affin_mask_t *dest = KMP_CPU_INDEX(temp, i); \
3734	KMP_CPU_COPY(dest, src); \
3735	} \
3736	KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks / 2); \
3737	newMasks = temp; \
3738	} \
3739	KMP_CPU_COPY(KMP_CPU_INDEX(newMasks, nextNewMask), (_mask)); \
3740	nextNewMask++; \
3741	}
3742
3743	#define ADD_MASK_OSID(_osId, _osId2Mask, _maxOsId) \
3744	{ \
3745	if (((_osId) > _maxOsId) || \
3746	(!KMP_CPU_ISSET((_osId), KMP_CPU_INDEX((_osId2Mask), (_osId))))) { \
3747	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, _osId); \
3748	} else { \
3749	ADD_MASK(KMP_CPU_INDEX(_osId2Mask, (_osId))); \
3750	} \
3751	}
3752
3753	// Re-parse the proclist (for the explicit affinity type), and form the list
3754	// of affinity newMasks indexed by gtid.
3755	static void __kmp_affinity_process_proclist(kmp_affinity_t &affinity) {
3756	int i;
3757	kmp_affin_mask_t **out_masks = &affinity.masks;
3758	unsigned *out_numMasks = &affinity.num_masks;
3759	const char *proclist = affinity.proclist;
3760	kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3761	int maxOsId = affinity.num_os_id_masks - 1;
3762	const char *scan = proclist;
3763	const char *next = proclist;
3764
3765	// We use malloc() for the temporary mask vector, so that we can use
3766	// realloc() to extend it.
3767	numNewMasks = 2;
3768	KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
3769	nextNewMask = 0;
3770	kmp_affin_mask_t *sumMask;
3771	KMP_CPU_ALLOC(sumMask);
3772	int setSize = 0;
3773
3774	for (;;) {
3775	int start, end, stride;
3776
3777	SKIP_WS(scan);
3778	next = scan;
3779	if (*next == '\0') {
3780	break;
3781	}
3782
3783	if (*next == '{') {
3784	int num;
3785	setSize = 0;
3786	next++; // skip '{'
3787	SKIP_WS(next);
3788	scan = next;
3789
3790	// Read the first integer in the set.
3791	KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad proclist");
3792	SKIP_DIGITS(next);
3793	num = __kmp_str_to_int(str: scan, sentinel: *next);
3794	KMP_ASSERT2(num >= 0, "bad explicit proc list");
3795
3796	// Copy the mask for that osId to the sum (union) mask.
3797	if ((num > maxOsId) ||
3798	(!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3799	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
3800	KMP_CPU_ZERO(sumMask);
3801	} else {
3802	KMP_CPU_COPY(sumMask, KMP_CPU_INDEX(osId2Mask, num));
3803	setSize = 1;
3804	}
3805
3806	for (;;) {
3807	// Check for end of set.
3808	SKIP_WS(next);
3809	if (*next == '}') {
3810	next++; // skip '}'
3811	break;
3812	}
3813
3814	// Skip optional comma.
3815	if (*next == ',') {
3816	next++;
3817	}
3818	SKIP_WS(next);
3819
3820	// Read the next integer in the set.
3821	scan = next;
3822	KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3823
3824	SKIP_DIGITS(next);
3825	num = __kmp_str_to_int(str: scan, sentinel: *next);
3826	KMP_ASSERT2(num >= 0, "bad explicit proc list");
3827
3828	// Add the mask for that osId to the sum mask.
3829	if ((num > maxOsId) ||
3830	(!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
3831	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
3832	} else {
3833	KMP_CPU_UNION(sumMask, KMP_CPU_INDEX(osId2Mask, num));
3834	setSize++;
3835	}
3836	}
3837	if (setSize > 0) {
3838	ADD_MASK(sumMask);
3839	}
3840
3841	SKIP_WS(next);
3842	if (*next == ',') {
3843	next++;
3844	}
3845	scan = next;
3846	continue;
3847	}
3848
3849	// Read the first integer.
3850	KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3851	SKIP_DIGITS(next);
3852	start = __kmp_str_to_int(str: scan, sentinel: *next);
3853	KMP_ASSERT2(start >= 0, "bad explicit proc list");
3854	SKIP_WS(next);
3855
3856	// If this isn't a range, then add a mask to the list and go on.
3857	if (*next != '-') {
3858	ADD_MASK_OSID(start, osId2Mask, maxOsId);
3859
3860	// Skip optional comma.
3861	if (*next == ',') {
3862	next++;
3863	}
3864	scan = next;
3865	continue;
3866	}
3867
3868	// This is a range. Skip over the '-' and read in the 2nd int.
3869	next++; // skip '-'
3870	SKIP_WS(next);
3871	scan = next;
3872	KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3873	SKIP_DIGITS(next);
3874	end = __kmp_str_to_int(str: scan, sentinel: *next);
3875	KMP_ASSERT2(end >= 0, "bad explicit proc list");
3876
3877	// Check for a stride parameter
3878	stride = 1;
3879	SKIP_WS(next);
3880	if (*next == ':') {
3881	// A stride is specified. Skip over the ':" and read the 3rd int.
3882	int sign = +1;
3883	next++; // skip ':'
3884	SKIP_WS(next);
3885	scan = next;
3886	if (*next == '-') {
3887	sign = -1;
3888	next++;
3889	SKIP_WS(next);
3890	scan = next;
3891	}
3892	KMP_ASSERT2((*next >= '0') && (*next <= '9'), "bad explicit proc list");
3893	SKIP_DIGITS(next);
3894	stride = __kmp_str_to_int(str: scan, sentinel: *next);
3895	KMP_ASSERT2(stride >= 0, "bad explicit proc list");
3896	stride *= sign;
3897	}
3898
3899	// Do some range checks.
3900	KMP_ASSERT2(stride != 0, "bad explicit proc list");
3901	if (stride > 0) {
3902	KMP_ASSERT2(start <= end, "bad explicit proc list");
3903	} else {
3904	KMP_ASSERT2(start >= end, "bad explicit proc list");
3905	}
3906	KMP_ASSERT2((end - start) / stride <= 65536, "bad explicit proc list");
3907
3908	// Add the mask for each OS proc # to the list.
3909	if (stride > 0) {
3910	do {
3911	ADD_MASK_OSID(start, osId2Mask, maxOsId);
3912	start += stride;
3913	} while (start <= end);
3914	} else {
3915	do {
3916	ADD_MASK_OSID(start, osId2Mask, maxOsId);
3917	start += stride;
3918	} while (start >= end);
3919	}
3920
3921	// Skip optional comma.
3922	SKIP_WS(next);
3923	if (*next == ',') {
3924	next++;
3925	}
3926	scan = next;
3927	}
3928
3929	*out_numMasks = nextNewMask;
3930	if (nextNewMask == 0) {
3931	*out_masks = NULL;
3932	KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3933	return;
3934	}
3935	KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
3936	for (i = 0; i < nextNewMask; i++) {
3937	kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
3938	kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
3939	KMP_CPU_COPY(dest, src);
3940	}
3941	KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
3942	KMP_CPU_FREE(sumMask);
3943	}
3944
3945	/*-----------------------------------------------------------------------------
3946	Re-parse the OMP_PLACES proc id list, forming the newMasks for the different
3947	places. Again, Here is the grammar:
3948
3949	place_list := place
3950	place_list := place , place_list
3951	place := num
3952	place := place : num
3953	place := place : num : signed
3954	place := { subplacelist }
3955	place := ! place // (lowest priority)
3956	subplace_list := subplace
3957	subplace_list := subplace , subplace_list
3958	subplace := num
3959	subplace := num : num
3960	subplace := num : num : signed
3961	signed := num
3962	signed := + signed
3963	signed := - signed
3964	-----------------------------------------------------------------------------*/
3965	static void __kmp_process_subplace_list(const char **scan,
3966	kmp_affinity_t &affinity, int maxOsId,
3967	kmp_affin_mask_t *tempMask,
3968	int *setSize) {
3969	const char *next;
3970	kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
3971
3972	for (;;) {
3973	int start, count, stride, i;
3974
3975	// Read in the starting proc id
3976	SKIP_WS(*scan);
3977	KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
3978	next = *scan;
3979	SKIP_DIGITS(next);
3980	start = __kmp_str_to_int(str: *scan, sentinel: *next);
3981	KMP_ASSERT(start >= 0);
3982	*scan = next;
3983
3984	// valid follow sets are ',' ':' and '}'
3985	SKIP_WS(*scan);
3986	if (**scan == '}' || **scan == ',') {
3987	if ((start > maxOsId) ||
3988	(!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
3989	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
3990	} else {
3991	KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
3992	(*setSize)++;
3993	}
3994	if (**scan == '}') {
3995	break;
3996	}
3997	(*scan)++; // skip ','
3998	continue;
3999	}
4000	KMP_ASSERT2(**scan == ':', "bad explicit places list");
4001	(*scan)++; // skip ':'
4002
4003	// Read count parameter
4004	SKIP_WS(*scan);
4005	KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4006	next = *scan;
4007	SKIP_DIGITS(next);
4008	count = __kmp_str_to_int(str: *scan, sentinel: *next);
4009	KMP_ASSERT(count >= 0);
4010	*scan = next;
4011
4012	// valid follow sets are ',' ':' and '}'
4013	SKIP_WS(*scan);
4014	if (**scan == '}' || **scan == ',') {
4015	for (i = 0; i < count; i++) {
4016	if ((start > maxOsId) ||
4017	(!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4018	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4019	break; // don't proliferate warnings for large count
4020	} else {
4021	KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4022	start++;
4023	(*setSize)++;
4024	}
4025	}
4026	if (**scan == '}') {
4027	break;
4028	}
4029	(*scan)++; // skip ','
4030	continue;
4031	}
4032	KMP_ASSERT2(**scan == ':', "bad explicit places list");
4033	(*scan)++; // skip ':'
4034
4035	// Read stride parameter
4036	int sign = +1;
4037	for (;;) {
4038	SKIP_WS(*scan);
4039	if (**scan == '+') {
4040	(*scan)++; // skip '+'
4041	continue;
4042	}
4043	if (**scan == '-') {
4044	sign *= -1;
4045	(*scan)++; // skip '-'
4046	continue;
4047	}
4048	break;
4049	}
4050	SKIP_WS(*scan);
4051	KMP_ASSERT2((**scan >= '0') && (**scan <= '9'), "bad explicit places list");
4052	next = *scan;
4053	SKIP_DIGITS(next);
4054	stride = __kmp_str_to_int(str: *scan, sentinel: *next);
4055	KMP_ASSERT(stride >= 0);
4056	*scan = next;
4057	stride *= sign;
4058
4059	// valid follow sets are ',' and '}'
4060	SKIP_WS(*scan);
4061	if (**scan == '}' || **scan == ',') {
4062	for (i = 0; i < count; i++) {
4063	if ((start > maxOsId) ||
4064	(!KMP_CPU_ISSET(start, KMP_CPU_INDEX(osId2Mask, start)))) {
4065	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, start);
4066	break; // don't proliferate warnings for large count
4067	} else {
4068	KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, start));
4069	start += stride;
4070	(*setSize)++;
4071	}
4072	}
4073	if (**scan == '}') {
4074	break;
4075	}
4076	(*scan)++; // skip ','
4077	continue;
4078	}
4079
4080	KMP_ASSERT2(0, "bad explicit places list");
4081	}
4082	}
4083
4084	static void __kmp_process_place(const char **scan, kmp_affinity_t &affinity,
4085	int maxOsId, kmp_affin_mask_t *tempMask,
4086	int *setSize) {
4087	const char *next;
4088	kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4089
4090	// valid follow sets are '{' '!' and num
4091	SKIP_WS(*scan);
4092	if (**scan == '{') {
4093	(*scan)++; // skip '{'
4094	__kmp_process_subplace_list(scan, affinity, maxOsId, tempMask, setSize);
4095	KMP_ASSERT2(**scan == '}', "bad explicit places list");
4096	(*scan)++; // skip '}'
4097	} else if (**scan == '!') {
4098	(*scan)++; // skip '!'
4099	__kmp_process_place(scan, affinity, maxOsId, tempMask, setSize);
4100	KMP_CPU_COMPLEMENT(maxOsId, tempMask);
4101	} else if ((**scan >= '0') && (**scan <= '9')) {
4102	next = *scan;
4103	SKIP_DIGITS(next);
4104	int num = __kmp_str_to_int(str: *scan, sentinel: *next);
4105	KMP_ASSERT(num >= 0);
4106	if ((num > maxOsId) ||
4107	(!KMP_CPU_ISSET(num, KMP_CPU_INDEX(osId2Mask, num)))) {
4108	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, num);
4109	} else {
4110	KMP_CPU_UNION(tempMask, KMP_CPU_INDEX(osId2Mask, num));
4111	(*setSize)++;
4112	}
4113	*scan = next; // skip num
4114	} else {
4115	KMP_ASSERT2(0, "bad explicit places list");
4116	}
4117	}
4118
4119	// static void
4120	void __kmp_affinity_process_placelist(kmp_affinity_t &affinity) {
4121	int i, j, count, stride, sign;
4122	kmp_affin_mask_t **out_masks = &affinity.masks;
4123	unsigned *out_numMasks = &affinity.num_masks;
4124	const char *placelist = affinity.proclist;
4125	kmp_affin_mask_t *osId2Mask = affinity.os_id_masks;
4126	int maxOsId = affinity.num_os_id_masks - 1;
4127	const char *scan = placelist;
4128	const char *next = placelist;
4129
4130	numNewMasks = 2;
4131	KMP_CPU_INTERNAL_ALLOC_ARRAY(newMasks, numNewMasks);
4132	nextNewMask = 0;
4133
4134	// tempMask is modified based on the previous or initial
4135	// place to form the current place
4136	// previousMask contains the previous place
4137	kmp_affin_mask_t *tempMask;
4138	kmp_affin_mask_t *previousMask;
4139	KMP_CPU_ALLOC(tempMask);
4140	KMP_CPU_ZERO(tempMask);
4141	KMP_CPU_ALLOC(previousMask);
4142	KMP_CPU_ZERO(previousMask);
4143	int setSize = 0;
4144
4145	for (;;) {
4146	__kmp_process_place(scan: &scan, affinity, maxOsId, tempMask, setSize: &setSize);
4147
4148	// valid follow sets are ',' ':' and EOL
4149	SKIP_WS(scan);
4150	if (*scan == '\0' || *scan == ',') {
4151	if (setSize > 0) {
4152	ADD_MASK(tempMask);
4153	}
4154	KMP_CPU_ZERO(tempMask);
4155	setSize = 0;
4156	if (*scan == '\0') {
4157	break;
4158	}
4159	scan++; // skip ','
4160	continue;
4161	}
4162
4163	KMP_ASSERT2(*scan == ':', "bad explicit places list");
4164	scan++; // skip ':'
4165
4166	// Read count parameter
4167	SKIP_WS(scan);
4168	KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4169	next = scan;
4170	SKIP_DIGITS(next);
4171	count = __kmp_str_to_int(str: scan, sentinel: *next);
4172	KMP_ASSERT(count >= 0);
4173	scan = next;
4174
4175	// valid follow sets are ',' ':' and EOL
4176	SKIP_WS(scan);
4177	if (*scan == '\0' || *scan == ',') {
4178	stride = +1;
4179	} else {
4180	KMP_ASSERT2(*scan == ':', "bad explicit places list");
4181	scan++; // skip ':'
4182
4183	// Read stride parameter
4184	sign = +1;
4185	for (;;) {
4186	SKIP_WS(scan);
4187	if (*scan == '+') {
4188	scan++; // skip '+'
4189	continue;
4190	}
4191	if (*scan == '-') {
4192	sign *= -1;
4193	scan++; // skip '-'
4194	continue;
4195	}
4196	break;
4197	}
4198	SKIP_WS(scan);
4199	KMP_ASSERT2((*scan >= '0') && (*scan <= '9'), "bad explicit places list");
4200	next = scan;
4201	SKIP_DIGITS(next);
4202	stride = __kmp_str_to_int(str: scan, sentinel: *next);
4203	KMP_DEBUG_ASSERT(stride >= 0);
4204	scan = next;
4205	stride *= sign;
4206	}
4207
4208	// Add places determined by initial_place : count : stride
4209	for (i = 0; i < count; i++) {
4210	if (setSize == 0) {
4211	break;
4212	}
4213	// Add the current place, then build the next place (tempMask) from that
4214	KMP_CPU_COPY(previousMask, tempMask);
4215	ADD_MASK(previousMask);
4216	KMP_CPU_ZERO(tempMask);
4217	setSize = 0;
4218	KMP_CPU_SET_ITERATE(j, previousMask) {
4219	if (!KMP_CPU_ISSET(j, previousMask)) {
4220	continue;
4221	}
4222	if ((j + stride > maxOsId) || (j + stride < 0) ||
4223	(!KMP_CPU_ISSET(j, __kmp_affin_fullMask)) ||
4224	(!KMP_CPU_ISSET(j + stride,
4225	KMP_CPU_INDEX(osId2Mask, j + stride)))) {
4226	if (i < count - 1) {
4227	KMP_AFF_WARNING(affinity, AffIgnoreInvalidProcID, j + stride);
4228	}
4229	continue;
4230	}
4231	KMP_CPU_SET(j + stride, tempMask);
4232	setSize++;
4233	}
4234	}
4235	KMP_CPU_ZERO(tempMask);
4236	setSize = 0;
4237
4238	// valid follow sets are ',' and EOL
4239	SKIP_WS(scan);
4240	if (*scan == '\0') {
4241	break;
4242	}
4243	if (*scan == ',') {
4244	scan++; // skip ','
4245	continue;
4246	}
4247
4248	KMP_ASSERT2(0, "bad explicit places list");
4249	}
4250
4251	*out_numMasks = nextNewMask;
4252	if (nextNewMask == 0) {
4253	*out_masks = NULL;
4254	KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4255	return;
4256	}
4257	KMP_CPU_ALLOC_ARRAY((*out_masks), nextNewMask);
4258	KMP_CPU_FREE(tempMask);
4259	KMP_CPU_FREE(previousMask);
4260	for (i = 0; i < nextNewMask; i++) {
4261	kmp_affin_mask_t *src = KMP_CPU_INDEX(newMasks, i);
4262	kmp_affin_mask_t *dest = KMP_CPU_INDEX((*out_masks), i);
4263	KMP_CPU_COPY(dest, src);
4264	}
4265	KMP_CPU_INTERNAL_FREE_ARRAY(newMasks, numNewMasks);
4266	}
4267
4268	#undef ADD_MASK
4269	#undef ADD_MASK_OSID
4270
4271	// This function figures out the deepest level at which there is at least one
4272	// cluster/core with more than one processing unit bound to it.
4273	static int __kmp_affinity_find_core_level(int nprocs, int bottom_level) {
4274	int core_level = 0;
4275
4276	for (int i = 0; i < nprocs; i++) {
4277	const kmp_hw_thread_t &hw_thread = __kmp_topology->at(index: i);
4278	for (int j = bottom_level; j > 0; j--) {
4279	if (hw_thread.ids[j] > 0) {
4280	if (core_level < (j - 1)) {
4281	core_level = j - 1;
4282	}
4283	}
4284	}
4285	}
4286	return core_level;
4287	}
4288
4289	// This function counts number of clusters/cores at given level.
4290	static int __kmp_affinity_compute_ncores(int nprocs, int bottom_level,
4291	int core_level) {
4292	return __kmp_topology->get_count(level: core_level);
4293	}
4294	// This function finds to which cluster/core given processing unit is bound.
4295	static int __kmp_affinity_find_core(int proc, int bottom_level,
4296	int core_level) {
4297	int core = 0;
4298	KMP_DEBUG_ASSERT(proc >= 0 && proc < __kmp_topology->get_num_hw_threads());
4299	for (int i = 0; i <= proc; ++i) {
4300	if (i + 1 <= proc) {
4301	for (int j = 0; j <= core_level; ++j) {
4302	if (__kmp_topology->at(index: i + 1).sub_ids[j] !=
4303	__kmp_topology->at(index: i).sub_ids[j]) {
4304	core++;
4305	break;
4306	}
4307	}
4308	}
4309	}
4310	return core;
4311	}
4312
4313	// This function finds maximal number of processing units bound to a
4314	// cluster/core at given level.
4315	static int __kmp_affinity_max_proc_per_core(int nprocs, int bottom_level,
4316	int core_level) {
4317	if (core_level >= bottom_level)
4318	return 1;
4319	int thread_level = __kmp_topology->get_level(type: KMP_HW_THREAD);
4320	return __kmp_topology->calculate_ratio(level1: thread_level, level2: core_level);
4321	}
4322
4323	static int *procarr = NULL;
4324	static int __kmp_aff_depth = 0;
4325	static int *__kmp_osid_to_hwthread_map = NULL;
4326
4327	static void __kmp_affinity_get_mask_topology_info(const kmp_affin_mask_t *mask,
4328	kmp_affinity_ids_t &ids,
4329	kmp_affinity_attrs_t &attrs) {
4330	if (!KMP_AFFINITY_CAPABLE())
4331	return;
4332
4333	// Initiailze ids and attrs thread data
4334	for (int i = 0; i < KMP_HW_LAST; ++i)
4335	ids.ids[i] = kmp_hw_thread_t::UNKNOWN_ID;
4336	attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
4337
4338	// Iterate through each os id within the mask and determine
4339	// the topology id and attribute information
4340	int cpu;
4341	int depth = __kmp_topology->get_depth();
4342	KMP_CPU_SET_ITERATE(cpu, mask) {
4343	int osid_idx = __kmp_osid_to_hwthread_map[cpu];
4344	ids.os_id = cpu;
4345	const kmp_hw_thread_t &hw_thread = __kmp_topology->at(index: osid_idx);
4346	for (int level = 0; level < depth; ++level) {
4347	kmp_hw_t type = __kmp_topology->get_type(level);
4348	int id = hw_thread.sub_ids[level];
4349	if (ids.ids[type] == kmp_hw_thread_t::UNKNOWN_ID || ids.ids[type] == id) {
4350	ids.ids[type] = id;
4351	} else {
4352	// This mask spans across multiple topology units, set it as such
4353	// and mark every level below as such as well.
4354	ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4355	for (; level < depth; ++level) {
4356	kmp_hw_t type = __kmp_topology->get_type(level);
4357	ids.ids[type] = kmp_hw_thread_t::MULTIPLE_ID;
4358	}
4359	}
4360	}
4361	if (!attrs.valid) {
4362	attrs.core_type = hw_thread.attrs.get_core_type();
4363	attrs.core_eff = hw_thread.attrs.get_core_eff();
4364	attrs.valid = 1;
4365	} else {
4366	// This mask spans across multiple attributes, set it as such
4367	if (attrs.core_type != hw_thread.attrs.get_core_type())
4368	attrs.core_type = KMP_HW_CORE_TYPE_UNKNOWN;
4369	if (attrs.core_eff != hw_thread.attrs.get_core_eff())
4370	attrs.core_eff = kmp_hw_attr_t::UNKNOWN_CORE_EFF;
4371	}
4372	}
4373	}
4374
4375	static void __kmp_affinity_get_thread_topology_info(kmp_info_t *th) {
4376	if (!KMP_AFFINITY_CAPABLE())
4377	return;
4378	const kmp_affin_mask_t *mask = th->th.th_affin_mask;
4379	kmp_affinity_ids_t &ids = th->th.th_topology_ids;
4380	kmp_affinity_attrs_t &attrs = th->th.th_topology_attrs;
4381	__kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4382	}
4383
4384	// Assign the topology information to each place in the place list
4385	// A thread can then grab not only its affinity mask, but the topology
4386	// information associated with that mask. e.g., Which socket is a thread on
4387	static void __kmp_affinity_get_topology_info(kmp_affinity_t &affinity) {
4388	if (!KMP_AFFINITY_CAPABLE())
4389	return;
4390	if (affinity.type != affinity_none) {
4391	KMP_ASSERT(affinity.num_os_id_masks);
4392	KMP_ASSERT(affinity.os_id_masks);
4393	}
4394	KMP_ASSERT(affinity.num_masks);
4395	KMP_ASSERT(affinity.masks);
4396	KMP_ASSERT(__kmp_affin_fullMask);
4397
4398	int max_cpu = __kmp_affin_fullMask->get_max_cpu();
4399	int num_hw_threads = __kmp_topology->get_num_hw_threads();
4400
4401	// Allocate thread topology information
4402	if (!affinity.ids) {
4403	affinity.ids = (kmp_affinity_ids_t *)__kmp_allocate(
4404	sizeof(kmp_affinity_ids_t) * affinity.num_masks);
4405	}
4406	if (!affinity.attrs) {
4407	affinity.attrs = (kmp_affinity_attrs_t *)__kmp_allocate(
4408	sizeof(kmp_affinity_attrs_t) * affinity.num_masks);
4409	}
4410	if (!__kmp_osid_to_hwthread_map) {
4411	// Want the +1 because max_cpu should be valid index into map
4412	__kmp_osid_to_hwthread_map =
4413	(int *)__kmp_allocate(sizeof(int) * (max_cpu + 1));
4414	}
4415
4416	// Create the OS proc to hardware thread map
4417	for (int hw_thread = 0; hw_thread < num_hw_threads; ++hw_thread) {
4418	int os_id = __kmp_topology->at(index: hw_thread).os_id;
4419	if (KMP_CPU_ISSET(os_id, __kmp_affin_fullMask))
4420	__kmp_osid_to_hwthread_map[os_id] = hw_thread;
4421	}
4422
4423	for (unsigned i = 0; i < affinity.num_masks; ++i) {
4424	kmp_affinity_ids_t &ids = affinity.ids[i];
4425	kmp_affinity_attrs_t &attrs = affinity.attrs[i];
4426	kmp_affin_mask_t *mask = KMP_CPU_INDEX(affinity.masks, i);
4427	__kmp_affinity_get_mask_topology_info(mask, ids, attrs);
4428	}
4429	}
4430
4431	// Called when __kmp_topology is ready
4432	static void __kmp_aux_affinity_initialize_other_data(kmp_affinity_t &affinity) {
4433	// Initialize other data structures which depend on the topology
4434	if (__kmp_topology && __kmp_topology->get_num_hw_threads()) {
4435	machine_hierarchy.init(num_addrs: __kmp_topology->get_num_hw_threads());
4436	__kmp_affinity_get_topology_info(affinity);
4437	#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
4438	__kmp_first_osid_with_ecore = __kmp_get_first_osid_with_ecore();
4439	#endif
4440	}
4441	}
4442
4443	// Create a one element mask array (set of places) which only contains the
4444	// initial process's affinity mask
4445	static void __kmp_create_affinity_none_places(kmp_affinity_t &affinity) {
4446	KMP_ASSERT(__kmp_affin_fullMask != NULL);
4447	KMP_ASSERT(affinity.type == affinity_none);
4448	KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4449	affinity.num_masks = 1;
4450	KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4451	kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, 0);
4452	KMP_CPU_COPY(dest, __kmp_affin_fullMask);
4453	__kmp_aux_affinity_initialize_other_data(affinity);
4454	}
4455
4456	static void __kmp_aux_affinity_initialize_masks(kmp_affinity_t &affinity) {
4457	// Create the "full" mask - this defines all of the processors that we
4458	// consider to be in the machine model. If respect is set, then it is the
4459	// initialization thread's affinity mask. Otherwise, it is all processors that
4460	// we know about on the machine.
4461	int verbose = affinity.flags.verbose;
4462	const char *env_var = affinity.env_var;
4463
4464	// Already initialized
4465	if (__kmp_affin_fullMask && __kmp_affin_origMask)
4466	return;
4467
4468	if (__kmp_affin_fullMask == NULL) {
4469	KMP_CPU_ALLOC(__kmp_affin_fullMask);
4470	}
4471	if (__kmp_affin_origMask == NULL) {
4472	KMP_CPU_ALLOC(__kmp_affin_origMask);
4473	}
4474	if (KMP_AFFINITY_CAPABLE()) {
4475	__kmp_get_system_affinity(__kmp_affin_fullMask, TRUE);
4476	// Make a copy before possible expanding to the entire machine mask
4477	__kmp_affin_origMask->copy(src: __kmp_affin_fullMask);
4478	if (affinity.flags.respect) {
4479	// Count the number of available processors.
4480	unsigned i;
4481	__kmp_avail_proc = 0;
4482	KMP_CPU_SET_ITERATE(i, __kmp_affin_fullMask) {
4483	if (!KMP_CPU_ISSET(i, __kmp_affin_fullMask)) {
4484	continue;
4485	}
4486	__kmp_avail_proc++;
4487	}
4488	if (__kmp_avail_proc > __kmp_xproc) {
4489	KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4490	affinity.type = affinity_none;
4491	KMP_AFFINITY_DISABLE();
4492	return;
4493	}
4494
4495	if (verbose) {
4496	char buf[KMP_AFFIN_MASK_PRINT_LEN];
4497	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4498	mask: __kmp_affin_fullMask);
4499	KMP_INFORM(InitOSProcSetRespect, env_var, buf);
4500	}
4501	} else {
4502	if (verbose) {
4503	char buf[KMP_AFFIN_MASK_PRINT_LEN];
4504	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
4505	mask: __kmp_affin_fullMask);
4506	KMP_INFORM(InitOSProcSetNotRespect, env_var, buf);
4507	}
4508	__kmp_avail_proc =
4509	__kmp_affinity_entire_machine_mask(mask: __kmp_affin_fullMask);
4510	#if KMP_OS_WINDOWS
4511	if (__kmp_num_proc_groups <= 1) {
4512	// Copy expanded full mask if topology has single processor group
4513	__kmp_affin_origMask->copy(__kmp_affin_fullMask);
4514	}
4515	// Set the process affinity mask since threads' affinity
4516	// masks must be subset of process mask in Windows* OS
4517	__kmp_affin_fullMask->set_process_affinity(true);
4518	#endif
4519	}
4520	}
4521	}
4522
4523	static bool __kmp_aux_affinity_initialize_topology(kmp_affinity_t &affinity) {
4524	bool success = false;
4525	const char *env_var = affinity.env_var;
4526	kmp_i18n_id_t msg_id = kmp_i18n_null;
4527	int verbose = affinity.flags.verbose;
4528
4529	// For backward compatibility, setting KMP_CPUINFO_FILE =>
4530	// KMP_TOPOLOGY_METHOD=cpuinfo
4531	if ((__kmp_cpuinfo_file != NULL) &&
4532	(__kmp_affinity_top_method == affinity_top_method_all)) {
4533	__kmp_affinity_top_method = affinity_top_method_cpuinfo;
4534	}
4535
4536	if (__kmp_affinity_top_method == affinity_top_method_all) {
4537	// In the default code path, errors are not fatal - we just try using
4538	// another method. We only emit a warning message if affinity is on, or the
4539	// verbose flag is set, an the nowarnings flag was not set.
4540	#if KMP_USE_HWLOC
4541	if (!success &&
4542	__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC) {
4543	if (!__kmp_hwloc_error) {
4544	success = __kmp_affinity_create_hwloc_map(&msg_id);
4545	if (!success && verbose) {
4546	KMP_INFORM(AffIgnoringHwloc, env_var);
4547	}
4548	} else if (verbose) {
4549	KMP_INFORM(AffIgnoringHwloc, env_var);
4550	}
4551	}
4552	#endif
4553
4554	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
4555	if (!success) {
4556	success = __kmp_affinity_create_x2apicid_map(&msg_id);
4557	if (!success && verbose && msg_id != kmp_i18n_null) {
4558	KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4559	}
4560	}
4561	if (!success) {
4562	success = __kmp_affinity_create_apicid_map(&msg_id);
4563	if (!success && verbose && msg_id != kmp_i18n_null) {
4564	KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4565	}
4566	}
4567	#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4568
4569	#if KMP_OS_LINUX || KMP_OS_AIX
4570	if (!success) {
4571	int line = 0;
4572	success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4573	if (!success && verbose && msg_id != kmp_i18n_null) {
4574	KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4575	}
4576	}
4577	#endif /* KMP_OS_LINUX */
4578
4579	#if KMP_GROUP_AFFINITY
4580	if (!success && (__kmp_num_proc_groups > 1)) {
4581	success = __kmp_affinity_create_proc_group_map(&msg_id);
4582	if (!success && verbose && msg_id != kmp_i18n_null) {
4583	KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4584	}
4585	}
4586	#endif /* KMP_GROUP_AFFINITY */
4587
4588	if (!success) {
4589	success = __kmp_affinity_create_flat_map(&msg_id);
4590	if (!success && verbose && msg_id != kmp_i18n_null) {
4591	KMP_INFORM(AffInfoStr, env_var, __kmp_i18n_catgets(msg_id));
4592	}
4593	KMP_ASSERT(success);
4594	}
4595	}
4596
4597	// If the user has specified that a paricular topology discovery method is to be
4598	// used, then we abort if that method fails. The exception is group affinity,
4599	// which might have been implicitly set.
4600	#if KMP_USE_HWLOC
4601	else if (__kmp_affinity_top_method == affinity_top_method_hwloc) {
4602	KMP_ASSERT(__kmp_affinity_dispatch->get_api_type() == KMPAffinity::HWLOC);
4603	success = __kmp_affinity_create_hwloc_map(&msg_id);
4604	if (!success) {
4605	KMP_ASSERT(msg_id != kmp_i18n_null);
4606	KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4607	}
4608	}
4609	#endif // KMP_USE_HWLOC
4610
4611	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
4612	else if (__kmp_affinity_top_method == affinity_top_method_x2apicid ||
4613	__kmp_affinity_top_method == affinity_top_method_x2apicid_1f) {
4614	success = __kmp_affinity_create_x2apicid_map(&msg_id);
4615	if (!success) {
4616	KMP_ASSERT(msg_id != kmp_i18n_null);
4617	KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4618	}
4619	} else if (__kmp_affinity_top_method == affinity_top_method_apicid) {
4620	success = __kmp_affinity_create_apicid_map(&msg_id);
4621	if (!success) {
4622	KMP_ASSERT(msg_id != kmp_i18n_null);
4623	KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4624	}
4625	}
4626	#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4627
4628	else if (__kmp_affinity_top_method == affinity_top_method_cpuinfo) {
4629	int line = 0;
4630	success = __kmp_affinity_create_cpuinfo_map(&line, &msg_id);
4631	if (!success) {
4632	KMP_ASSERT(msg_id != kmp_i18n_null);
4633	const char *filename = __kmp_cpuinfo_get_filename();
4634	if (line > 0) {
4635	KMP_FATAL(FileLineMsgExiting, filename, line,
4636	__kmp_i18n_catgets(msg_id));
4637	} else {
4638	KMP_FATAL(FileMsgExiting, filename, __kmp_i18n_catgets(msg_id));
4639	}
4640	}
4641	}
4642
4643	#if KMP_GROUP_AFFINITY
4644	else if (__kmp_affinity_top_method == affinity_top_method_group) {
4645	success = __kmp_affinity_create_proc_group_map(&msg_id);
4646	KMP_ASSERT(success);
4647	if (!success) {
4648	KMP_ASSERT(msg_id != kmp_i18n_null);
4649	KMP_FATAL(MsgExiting, __kmp_i18n_catgets(msg_id));
4650	}
4651	}
4652	#endif /* KMP_GROUP_AFFINITY */
4653
4654	else if (__kmp_affinity_top_method == affinity_top_method_flat) {
4655	success = __kmp_affinity_create_flat_map(&msg_id);
4656	// should not fail
4657	KMP_ASSERT(success);
4658	}
4659
4660	// Early exit if topology could not be created
4661	if (!__kmp_topology) {
4662	if (KMP_AFFINITY_CAPABLE()) {
4663	KMP_AFF_WARNING(affinity, ErrorInitializeAffinity);
4664	}
4665	if (nPackages > 0 && nCoresPerPkg > 0 && __kmp_nThreadsPerCore > 0 &&
4666	__kmp_ncores > 0) {
4667	__kmp_topology = kmp_topology_t::allocate(nproc: 0, ndepth: 0, NULL);
4668	__kmp_topology->canonicalize(npackages: nPackages, ncores_per_pkg: nCoresPerPkg,
4669	nthreads_per_core: __kmp_nThreadsPerCore, ncores: __kmp_ncores);
4670	if (verbose) {
4671	__kmp_topology->print(env_var);
4672	}
4673	}
4674	return false;
4675	}
4676
4677	// Canonicalize, print (if requested), apply KMP_HW_SUBSET
4678	__kmp_topology->canonicalize();
4679	if (verbose)
4680	__kmp_topology->print(env_var);
4681	bool filtered = __kmp_topology->filter_hw_subset();
4682	if (filtered && verbose)
4683	__kmp_topology->print(env_var: "KMP_HW_SUBSET");
4684	return success;
4685	}
4686
4687	static void __kmp_aux_affinity_initialize(kmp_affinity_t &affinity) {
4688	bool is_regular_affinity = (&affinity == &__kmp_affinity);
4689	bool is_hidden_helper_affinity = (&affinity == &__kmp_hh_affinity);
4690	const char *env_var = __kmp_get_affinity_env_var(affinity);
4691
4692	if (affinity.flags.initialized) {
4693	KMP_ASSERT(__kmp_affin_fullMask != NULL);
4694	return;
4695	}
4696
4697	if (is_regular_affinity && (!__kmp_affin_fullMask || !__kmp_affin_origMask))
4698	__kmp_aux_affinity_initialize_masks(affinity);
4699
4700	if (is_regular_affinity && !__kmp_topology) {
4701	bool success = __kmp_aux_affinity_initialize_topology(affinity);
4702	if (success) {
4703	KMP_ASSERT(__kmp_avail_proc == __kmp_topology->get_num_hw_threads());
4704	} else {
4705	affinity.type = affinity_none;
4706	KMP_AFFINITY_DISABLE();
4707	}
4708	}
4709
4710	// If KMP_AFFINITY=none, then only create the single "none" place
4711	// which is the process's initial affinity mask or the number of
4712	// hardware threads depending on respect,norespect
4713	if (affinity.type == affinity_none) {
4714	__kmp_create_affinity_none_places(affinity);
4715	#if KMP_USE_HIER_SCHED
4716	__kmp_dispatch_set_hierarchy_values();
4717	#endif
4718	affinity.flags.initialized = TRUE;
4719	return;
4720	}
4721
4722	__kmp_topology->set_granularity(affinity);
4723	int depth = __kmp_topology->get_depth();
4724
4725	// Create the table of masks, indexed by thread Id.
4726	unsigned numUnique;
4727	int numAddrs = __kmp_topology->get_num_hw_threads();
4728	// If OMP_PLACES=cores:<attribute> specified, then attempt
4729	// to make OS Id mask table using those attributes
4730	if (affinity.core_attr_gran.valid) {
4731	__kmp_create_os_id_masks(numUnique: &numUnique, affinity, find_next: [&](int idx) {
4732	KMP_ASSERT(idx >= -1);
4733	for (int i = idx + 1; i < numAddrs; ++i)
4734	if (__kmp_topology->at(index: i).attrs.contains(attr: affinity.core_attr_gran))
4735	return i;
4736	return numAddrs;
4737	});
4738	if (!affinity.os_id_masks) {
4739	const char *core_attribute;
4740	if (affinity.core_attr_gran.core_eff != kmp_hw_attr_t::UNKNOWN_CORE_EFF)
4741	core_attribute = "core_efficiency";
4742	else
4743	core_attribute = "core_type";
4744	KMP_AFF_WARNING(affinity, AffIgnoringNotAvailable, env_var,
4745	core_attribute,
4746	__kmp_hw_get_catalog_string(KMP_HW_CORE, /*plural=*/true))
4747	}
4748	}
4749	// If core attributes did not work, or none were specified,
4750	// then make OS Id mask table using typical incremental way.
4751	if (!affinity.os_id_masks) {
4752	__kmp_create_os_id_masks(numUnique: &numUnique, affinity, find_next: [](int idx) {
4753	KMP_ASSERT(idx >= -1);
4754	return idx + 1;
4755	});
4756	}
4757	if (affinity.gran_levels == 0) {
4758	KMP_DEBUG_ASSERT((int)numUnique == __kmp_avail_proc);
4759	}
4760
4761	switch (affinity.type) {
4762
4763	case affinity_explicit:
4764	KMP_DEBUG_ASSERT(affinity.proclist != NULL);
4765	if (is_hidden_helper_affinity ||
4766	__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) {
4767	__kmp_affinity_process_proclist(affinity);
4768	} else {
4769	__kmp_affinity_process_placelist(affinity);
4770	}
4771	if (affinity.num_masks == 0) {
4772	KMP_AFF_WARNING(affinity, AffNoValidProcID);
4773	affinity.type = affinity_none;
4774	__kmp_create_affinity_none_places(affinity);
4775	affinity.flags.initialized = TRUE;
4776	return;
4777	}
4778	break;
4779
4780	// The other affinity types rely on sorting the hardware threads according to
4781	// some permutation of the machine topology tree. Set affinity.compact
4782	// and affinity.offset appropriately, then jump to a common code
4783	// fragment to do the sort and create the array of affinity masks.
4784	case affinity_logical:
4785	affinity.compact = 0;
4786	if (affinity.offset) {
4787	affinity.offset =
4788	__kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
4789	}
4790	goto sortTopology;
4791
4792	case affinity_physical:
4793	if (__kmp_nThreadsPerCore > 1) {
4794	affinity.compact = 1;
4795	if (affinity.compact >= depth) {
4796	affinity.compact = 0;
4797	}
4798	} else {
4799	affinity.compact = 0;
4800	}
4801	if (affinity.offset) {
4802	affinity.offset =
4803	__kmp_nThreadsPerCore * affinity.offset % __kmp_avail_proc;
4804	}
4805	goto sortTopology;
4806
4807	case affinity_scatter:
4808	if (affinity.compact >= depth) {
4809	affinity.compact = 0;
4810	} else {
4811	affinity.compact = depth - 1 - affinity.compact;
4812	}
4813	goto sortTopology;
4814
4815	case affinity_compact:
4816	if (affinity.compact >= depth) {
4817	affinity.compact = depth - 1;
4818	}
4819	goto sortTopology;
4820
4821	case affinity_balanced:
4822	if (depth <= 1 || is_hidden_helper_affinity) {
4823	KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
4824	affinity.type = affinity_none;
4825	__kmp_create_affinity_none_places(affinity);
4826	affinity.flags.initialized = TRUE;
4827	return;
4828	} else if (!__kmp_topology->is_uniform()) {
4829	// Save the depth for further usage
4830	__kmp_aff_depth = depth;
4831
4832	int core_level =
4833	__kmp_affinity_find_core_level(nprocs: __kmp_avail_proc, bottom_level: depth - 1);
4834	int ncores = __kmp_affinity_compute_ncores(nprocs: __kmp_avail_proc, bottom_level: depth - 1,
4835	core_level);
4836	int maxprocpercore = __kmp_affinity_max_proc_per_core(
4837	nprocs: __kmp_avail_proc, bottom_level: depth - 1, core_level);
4838
4839	int nproc = ncores * maxprocpercore;
4840	if ((nproc < 2) || (nproc < __kmp_avail_proc)) {
4841	KMP_AFF_WARNING(affinity, AffBalancedNotAvail, env_var);
4842	affinity.type = affinity_none;
4843	__kmp_create_affinity_none_places(affinity);
4844	affinity.flags.initialized = TRUE;
4845	return;
4846	}
4847
4848	procarr = (int *)__kmp_allocate(sizeof(int) * nproc);
4849	for (int i = 0; i < nproc; i++) {
4850	procarr[i] = -1;
4851	}
4852
4853	int lastcore = -1;
4854	int inlastcore = 0;
4855	for (int i = 0; i < __kmp_avail_proc; i++) {
4856	int proc = __kmp_topology->at(index: i).os_id;
4857	int core = __kmp_affinity_find_core(proc: i, bottom_level: depth - 1, core_level);
4858
4859	if (core == lastcore) {
4860	inlastcore++;
4861	} else {
4862	inlastcore = 0;
4863	}
4864	lastcore = core;
4865
4866	procarr[core * maxprocpercore + inlastcore] = proc;
4867	}
4868	}
4869	if (affinity.compact >= depth) {
4870	affinity.compact = depth - 1;
4871	}
4872
4873	sortTopology:
4874	// Allocate the gtid->affinity mask table.
4875	if (affinity.flags.dups) {
4876	affinity.num_masks = __kmp_avail_proc;
4877	} else {
4878	affinity.num_masks = numUnique;
4879	}
4880
4881	if ((__kmp_nested_proc_bind.bind_types[0] != proc_bind_intel) &&
4882	(__kmp_affinity_num_places > 0) &&
4883	((unsigned)__kmp_affinity_num_places < affinity.num_masks) &&
4884	!is_hidden_helper_affinity) {
4885	affinity.num_masks = __kmp_affinity_num_places;
4886	}
4887
4888	KMP_CPU_ALLOC_ARRAY(affinity.masks, affinity.num_masks);
4889
4890	// Sort the topology table according to the current setting of
4891	// affinity.compact, then fill out affinity.masks.
4892	__kmp_topology->sort_compact(affinity);
4893	{
4894	int i;
4895	unsigned j;
4896	int num_hw_threads = __kmp_topology->get_num_hw_threads();
4897	kmp_full_mask_modifier_t full_mask;
4898	for (i = 0, j = 0; i < num_hw_threads; i++) {
4899	if ((!affinity.flags.dups) && (!__kmp_topology->at(index: i).leader)) {
4900	continue;
4901	}
4902	int osId = __kmp_topology->at(index: i).os_id;
4903
4904	kmp_affin_mask_t *src = KMP_CPU_INDEX(affinity.os_id_masks, osId);
4905	kmp_affin_mask_t *dest = KMP_CPU_INDEX(affinity.masks, j);
4906	KMP_ASSERT(KMP_CPU_ISSET(osId, src));
4907	KMP_CPU_COPY(dest, src);
4908	full_mask.include(other: src);
4909	if (++j >= affinity.num_masks) {
4910	break;
4911	}
4912	}
4913	KMP_DEBUG_ASSERT(j == affinity.num_masks);
4914	// See if the places list further restricts or changes the full mask
4915	if (full_mask.restrict_to_mask() && affinity.flags.verbose) {
4916	__kmp_topology->print(env_var);
4917	}
4918	}
4919	// Sort the topology back using ids
4920	__kmp_topology->sort_ids();
4921	break;
4922
4923	default:
4924	KMP_ASSERT2(0, "Unexpected affinity setting");
4925	}
4926	__kmp_aux_affinity_initialize_other_data(affinity);
4927	affinity.flags.initialized = TRUE;
4928	}
4929
4930	void __kmp_affinity_initialize(kmp_affinity_t &affinity) {
4931	// Much of the code above was written assuming that if a machine was not
4932	// affinity capable, then affinity type == affinity_none.
4933	// We now explicitly represent this as affinity type == affinity_disabled.
4934	// There are too many checks for affinity type == affinity_none in this code.
4935	// Instead of trying to change them all, check if
4936	// affinity type == affinity_disabled, and if so, slam it with affinity_none,
4937	// call the real initialization routine, then restore affinity type to
4938	// affinity_disabled.
4939	int disabled = (affinity.type == affinity_disabled);
4940	if (!KMP_AFFINITY_CAPABLE())
4941	KMP_ASSERT(disabled);
4942	if (disabled)
4943	affinity.type = affinity_none;
4944	__kmp_aux_affinity_initialize(affinity);
4945	if (disabled)
4946	affinity.type = affinity_disabled;
4947	}
4948
4949	void __kmp_affinity_uninitialize(void) {
4950	for (kmp_affinity_t *affinity : __kmp_affinities) {
4951	if (affinity->masks != NULL)
4952	KMP_CPU_FREE_ARRAY(affinity->masks, affinity->num_masks);
4953	if (affinity->os_id_masks != NULL)
4954	KMP_CPU_FREE_ARRAY(affinity->os_id_masks, affinity->num_os_id_masks);
4955	if (affinity->proclist != NULL)
4956	__kmp_free(affinity->proclist);
4957	if (affinity->ids != NULL)
4958	__kmp_free(affinity->ids);
4959	if (affinity->attrs != NULL)
4960	__kmp_free(affinity->attrs);
4961	*affinity = KMP_AFFINITY_INIT(affinity->env_var);
4962	}
4963	if (__kmp_affin_origMask != NULL) {
4964	if (KMP_AFFINITY_CAPABLE()) {
4965	#if KMP_OS_AIX
4966	// Uninitialize by unbinding the thread.
4967	bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
4968	#else
4969	__kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
4970	#endif
4971	}
4972	KMP_CPU_FREE(__kmp_affin_origMask);
4973	__kmp_affin_origMask = NULL;
4974	}
4975	__kmp_affinity_num_places = 0;
4976	if (procarr != NULL) {
4977	__kmp_free(procarr);
4978	procarr = NULL;
4979	}
4980	if (__kmp_osid_to_hwthread_map) {
4981	__kmp_free(__kmp_osid_to_hwthread_map);
4982	__kmp_osid_to_hwthread_map = NULL;
4983	}
4984	#if KMP_USE_HWLOC
4985	if (__kmp_hwloc_topology != NULL) {
4986	hwloc_topology_destroy(__kmp_hwloc_topology);
4987	__kmp_hwloc_topology = NULL;
4988	}
4989	#endif
4990	if (__kmp_hw_subset) {
4991	kmp_hw_subset_t::deallocate(subset: __kmp_hw_subset);
4992	__kmp_hw_subset = nullptr;
4993	}
4994	if (__kmp_topology) {
4995	kmp_topology_t::deallocate(topology: __kmp_topology);
4996	__kmp_topology = nullptr;
4997	}
4998	KMPAffinity::destroy_api();
4999	}
5000
5001	static void __kmp_select_mask_by_gtid(int gtid, const kmp_affinity_t *affinity,
5002	int *place, kmp_affin_mask_t **mask) {
5003	int mask_idx;
5004	bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5005	if (is_hidden_helper)
5006	// The first gtid is the regular primary thread, the second gtid is the main
5007	// thread of hidden team which does not participate in task execution.
5008	mask_idx = gtid - 2;
5009	else
5010	mask_idx = __kmp_adjust_gtid_for_hidden_helpers(gtid);
5011	KMP_DEBUG_ASSERT(affinity->num_masks > 0);
5012	*place = (mask_idx + affinity->offset) % affinity->num_masks;
5013	*mask = KMP_CPU_INDEX(affinity->masks, *place);
5014	}
5015
5016	// This function initializes the per-thread data concerning affinity including
5017	// the mask and topology information
5018	void __kmp_affinity_set_init_mask(int gtid, int isa_root) {
5019
5020	kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5021
5022	// Set the thread topology information to default of unknown
5023	for (int id = 0; id < KMP_HW_LAST; ++id)
5024	th->th.th_topology_ids.ids[id] = kmp_hw_thread_t::UNKNOWN_ID;
5025	th->th.th_topology_attrs = KMP_AFFINITY_ATTRS_UNKNOWN;
5026
5027	if (!KMP_AFFINITY_CAPABLE()) {
5028	return;
5029	}
5030
5031	if (th->th.th_affin_mask == NULL) {
5032	KMP_CPU_ALLOC(th->th.th_affin_mask);
5033	} else {
5034	KMP_CPU_ZERO(th->th.th_affin_mask);
5035	}
5036
5037	// Copy the thread mask to the kmp_info_t structure. If
5038	// __kmp_affinity.type == affinity_none, copy the "full" mask, i.e.
5039	// one that has all of the OS proc ids set, or if
5040	// __kmp_affinity.flags.respect is set, then the full mask is the
5041	// same as the mask of the initialization thread.
5042	kmp_affin_mask_t *mask;
5043	int i;
5044	const kmp_affinity_t *affinity;
5045	bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5046
5047	if (is_hidden_helper)
5048	affinity = &__kmp_hh_affinity;
5049	else
5050	affinity = &__kmp_affinity;
5051
5052	if (KMP_AFFINITY_NON_PROC_BIND || is_hidden_helper) {
5053	if ((affinity->type == affinity_none) ||
5054	(affinity->type == affinity_balanced) ||
5055	KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5056	#if KMP_GROUP_AFFINITY
5057	if (__kmp_num_proc_groups > 1) {
5058	return;
5059	}
5060	#endif
5061	KMP_ASSERT(__kmp_affin_fullMask != NULL);
5062	i = 0;
5063	mask = __kmp_affin_fullMask;
5064	} else {
5065	__kmp_select_mask_by_gtid(gtid, affinity, place: &i, mask: &mask);
5066	}
5067	} else {
5068	if (!isa_root || __kmp_nested_proc_bind.bind_types[0] == proc_bind_false) {
5069	#if KMP_GROUP_AFFINITY
5070	if (__kmp_num_proc_groups > 1) {
5071	return;
5072	}
5073	#endif
5074	KMP_ASSERT(__kmp_affin_fullMask != NULL);
5075	i = KMP_PLACE_ALL;
5076	mask = __kmp_affin_fullMask;
5077	} else {
5078	__kmp_select_mask_by_gtid(gtid, affinity, place: &i, mask: &mask);
5079	}
5080	}
5081
5082	th->th.th_current_place = i;
5083	if (isa_root && !is_hidden_helper) {
5084	th->th.th_new_place = i;
5085	th->th.th_first_place = 0;
5086	th->th.th_last_place = affinity->num_masks - 1;
5087	} else if (KMP_AFFINITY_NON_PROC_BIND) {
5088	// When using a Non-OMP_PROC_BIND affinity method,
5089	// set all threads' place-partition-var to the entire place list
5090	th->th.th_first_place = 0;
5091	th->th.th_last_place = affinity->num_masks - 1;
5092	}
5093	// Copy topology information associated with the place
5094	if (i >= 0) {
5095	th->th.th_topology_ids = __kmp_affinity.ids[i];
5096	th->th.th_topology_attrs = __kmp_affinity.attrs[i];
5097	}
5098
5099	if (i == KMP_PLACE_ALL) {
5100	KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to all places\n",
5101	gtid));
5102	} else {
5103	KA_TRACE(100, ("__kmp_affinity_set_init_mask: setting T#%d to place %d\n",
5104	gtid, i));
5105	}
5106
5107	KMP_CPU_COPY(th->th.th_affin_mask, mask);
5108	}
5109
5110	void __kmp_affinity_bind_init_mask(int gtid) {
5111	if (!KMP_AFFINITY_CAPABLE()) {
5112	return;
5113	}
5114	kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5115	const kmp_affinity_t *affinity;
5116	const char *env_var;
5117	bool is_hidden_helper = KMP_HIDDEN_HELPER_THREAD(gtid);
5118
5119	if (is_hidden_helper)
5120	affinity = &__kmp_hh_affinity;
5121	else
5122	affinity = &__kmp_affinity;
5123	env_var = __kmp_get_affinity_env_var(affinity: *affinity, /*for_binding=*/true);
5124	/* to avoid duplicate printing (will be correctly printed on barrier) */
5125	if (affinity->flags.verbose && (affinity->type == affinity_none ||
5126	(th->th.th_current_place != KMP_PLACE_ALL &&
5127	affinity->type != affinity_balanced)) &&
5128	!KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)) {
5129	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5130	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5131	mask: th->th.th_affin_mask);
5132	KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5133	gtid, buf);
5134	}
5135
5136	#if KMP_OS_WINDOWS
5137	// On Windows* OS, the process affinity mask might have changed. If the user
5138	// didn't request affinity and this call fails, just continue silently.
5139	// See CQ171393.
5140	if (affinity->type == affinity_none) {
5141	__kmp_set_system_affinity(th->th.th_affin_mask, FALSE);
5142	} else
5143	#endif
5144	#ifndef KMP_OS_AIX
5145	// Do not set the full mask as the init mask on AIX.
5146	__kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5147	#endif
5148	}
5149
5150	void __kmp_affinity_bind_place(int gtid) {
5151	// Hidden helper threads should not be affected by OMP_PLACES/OMP_PROC_BIND
5152	if (!KMP_AFFINITY_CAPABLE() || KMP_HIDDEN_HELPER_THREAD(gtid)) {
5153	return;
5154	}
5155
5156	kmp_info_t *th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[gtid]);
5157
5158	KA_TRACE(100, ("__kmp_affinity_bind_place: binding T#%d to place %d (current "
5159	"place = %d)\n",
5160	gtid, th->th.th_new_place, th->th.th_current_place));
5161
5162	// Check that the new place is within this thread's partition.
5163	KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5164	KMP_ASSERT(th->th.th_new_place >= 0);
5165	KMP_ASSERT((unsigned)th->th.th_new_place <= __kmp_affinity.num_masks);
5166	if (th->th.th_first_place <= th->th.th_last_place) {
5167	KMP_ASSERT((th->th.th_new_place >= th->th.th_first_place) &&
5168	(th->th.th_new_place <= th->th.th_last_place));
5169	} else {
5170	KMP_ASSERT((th->th.th_new_place <= th->th.th_first_place) ||
5171	(th->th.th_new_place >= th->th.th_last_place));
5172	}
5173
5174	// Copy the thread mask to the kmp_info_t structure,
5175	// and set this thread's affinity.
5176	kmp_affin_mask_t *mask =
5177	KMP_CPU_INDEX(__kmp_affinity.masks, th->th.th_new_place);
5178	KMP_CPU_COPY(th->th.th_affin_mask, mask);
5179	th->th.th_current_place = th->th.th_new_place;
5180
5181	if (__kmp_affinity.flags.verbose) {
5182	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5183	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5184	mask: th->th.th_affin_mask);
5185	KMP_INFORM(BoundToOSProcSet, "OMP_PROC_BIND", (kmp_int32)getpid(),
5186	__kmp_gettid(), gtid, buf);
5187	}
5188	__kmp_set_system_affinity(th->th.th_affin_mask, TRUE);
5189	}
5190
5191	int __kmp_aux_set_affinity(void **mask) {
5192	int gtid;
5193	kmp_info_t *th;
5194	int retval;
5195
5196	if (!KMP_AFFINITY_CAPABLE()) {
5197	return -1;
5198	}
5199
5200	gtid = __kmp_entry_gtid();
5201	KA_TRACE(
5202	1000, (""); {
5203	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5204	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5205	(kmp_affin_mask_t *)(*mask));
5206	__kmp_debug_printf(
5207	"kmp_set_affinity: setting affinity mask for thread %d = %s\n",
5208	gtid, buf);
5209	});
5210
5211	if (__kmp_env_consistency_check) {
5212	if ((mask == NULL) || (*mask == NULL)) {
5213	KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5214	} else {
5215	unsigned proc;
5216	int num_procs = 0;
5217
5218	KMP_CPU_SET_ITERATE(proc, ((kmp_affin_mask_t *)(*mask))) {
5219	if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5220	KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5221	}
5222	if (!KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask))) {
5223	continue;
5224	}
5225	num_procs++;
5226	}
5227	if (num_procs == 0) {
5228	KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5229	}
5230
5231	#if KMP_GROUP_AFFINITY
5232	if (__kmp_get_proc_group((kmp_affin_mask_t *)(*mask)) < 0) {
5233	KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity");
5234	}
5235	#endif /* KMP_GROUP_AFFINITY */
5236	}
5237	}
5238
5239	th = __kmp_threads[gtid];
5240	KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5241	retval = __kmp_set_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5242	if (retval == 0) {
5243	KMP_CPU_COPY(th->th.th_affin_mask, (kmp_affin_mask_t *)(*mask));
5244	}
5245
5246	th->th.th_current_place = KMP_PLACE_UNDEFINED;
5247	th->th.th_new_place = KMP_PLACE_UNDEFINED;
5248	th->th.th_first_place = 0;
5249	th->th.th_last_place = __kmp_affinity.num_masks - 1;
5250
5251	// Turn off 4.0 affinity for the current tread at this parallel level.
5252	th->th.th_current_task->td_icvs.proc_bind = proc_bind_false;
5253
5254	return retval;
5255	}
5256
5257	int __kmp_aux_get_affinity(void **mask) {
5258	int gtid;
5259	int retval;
5260	#if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5261	kmp_info_t *th;
5262	#endif
5263	if (!KMP_AFFINITY_CAPABLE()) {
5264	return -1;
5265	}
5266
5267	gtid = __kmp_entry_gtid();
5268	#if KMP_OS_WINDOWS || KMP_OS_AIX || KMP_DEBUG
5269	th = __kmp_threads[gtid];
5270	#else
5271	(void)gtid; // unused variable
5272	#endif
5273	KMP_DEBUG_ASSERT(th->th.th_affin_mask != NULL);
5274
5275	KA_TRACE(
5276	1000, (""); {
5277	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5278	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5279	th->th.th_affin_mask);
5280	__kmp_printf(
5281	"kmp_get_affinity: stored affinity mask for thread %d = %s\n", gtid,
5282	buf);
5283	});
5284
5285	if (__kmp_env_consistency_check) {
5286	if ((mask == NULL) || (*mask == NULL)) {
5287	KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity");
5288	}
5289	}
5290
5291	#if !KMP_OS_WINDOWS && !KMP_OS_AIX
5292
5293	retval = __kmp_get_system_affinity((kmp_affin_mask_t *)(*mask), FALSE);
5294	KA_TRACE(
5295	1000, (""); {
5296	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5297	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5298	(kmp_affin_mask_t *)(*mask));
5299	__kmp_printf(
5300	"kmp_get_affinity: system affinity mask for thread %d = %s\n", gtid,
5301	buf);
5302	});
5303	return retval;
5304
5305	#else
5306	(void)retval;
5307
5308	KMP_CPU_COPY((kmp_affin_mask_t *)(*mask), th->th.th_affin_mask);
5309	return 0;
5310
5311	#endif /* !KMP_OS_WINDOWS && !KMP_OS_AIX */
5312	}
5313
5314	int __kmp_aux_get_affinity_max_proc() {
5315	if (!KMP_AFFINITY_CAPABLE()) {
5316	return 0;
5317	}
5318	#if KMP_GROUP_AFFINITY
5319	if (__kmp_num_proc_groups > 1) {
5320	return (int)(__kmp_num_proc_groups * sizeof(DWORD_PTR) * CHAR_BIT);
5321	}
5322	#endif
5323	return __kmp_xproc;
5324	}
5325
5326	int __kmp_aux_set_affinity_mask_proc(int proc, void **mask) {
5327	if (!KMP_AFFINITY_CAPABLE()) {
5328	return -1;
5329	}
5330
5331	KA_TRACE(
5332	1000, (""); {
5333	int gtid = __kmp_entry_gtid();
5334	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5335	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5336	(kmp_affin_mask_t *)(*mask));
5337	__kmp_debug_printf("kmp_set_affinity_mask_proc: setting proc %d in "
5338	"affinity mask for thread %d = %s\n",
5339	proc, gtid, buf);
5340	});
5341
5342	if (__kmp_env_consistency_check) {
5343	if ((mask == NULL) || (*mask == NULL)) {
5344	KMP_FATAL(AffinityInvalidMask, "kmp_set_affinity_mask_proc");
5345	}
5346	}
5347
5348	if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5349	return -1;
5350	}
5351	if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5352	return -2;
5353	}
5354
5355	KMP_CPU_SET(proc, (kmp_affin_mask_t *)(*mask));
5356	return 0;
5357	}
5358
5359	int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask) {
5360	if (!KMP_AFFINITY_CAPABLE()) {
5361	return -1;
5362	}
5363
5364	KA_TRACE(
5365	1000, (""); {
5366	int gtid = __kmp_entry_gtid();
5367	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5368	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5369	(kmp_affin_mask_t *)(*mask));
5370	__kmp_debug_printf("kmp_unset_affinity_mask_proc: unsetting proc %d in "
5371	"affinity mask for thread %d = %s\n",
5372	proc, gtid, buf);
5373	});
5374
5375	if (__kmp_env_consistency_check) {
5376	if ((mask == NULL) || (*mask == NULL)) {
5377	KMP_FATAL(AffinityInvalidMask, "kmp_unset_affinity_mask_proc");
5378	}
5379	}
5380
5381	if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5382	return -1;
5383	}
5384	if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5385	return -2;
5386	}
5387
5388	KMP_CPU_CLR(proc, (kmp_affin_mask_t *)(*mask));
5389	return 0;
5390	}
5391
5392	int __kmp_aux_get_affinity_mask_proc(int proc, void **mask) {
5393	if (!KMP_AFFINITY_CAPABLE()) {
5394	return -1;
5395	}
5396
5397	KA_TRACE(
5398	1000, (""); {
5399	int gtid = __kmp_entry_gtid();
5400	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5401	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN,
5402	(kmp_affin_mask_t *)(*mask));
5403	__kmp_debug_printf("kmp_get_affinity_mask_proc: getting proc %d in "
5404	"affinity mask for thread %d = %s\n",
5405	proc, gtid, buf);
5406	});
5407
5408	if (__kmp_env_consistency_check) {
5409	if ((mask == NULL) || (*mask == NULL)) {
5410	KMP_FATAL(AffinityInvalidMask, "kmp_get_affinity_mask_proc");
5411	}
5412	}
5413
5414	if ((proc < 0) || (proc >= __kmp_aux_get_affinity_max_proc())) {
5415	return -1;
5416	}
5417	if (!KMP_CPU_ISSET(proc, __kmp_affin_fullMask)) {
5418	return 0;
5419	}
5420
5421	return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
5422	}
5423
5424	#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
5425	// Returns first os proc id with ATOM core
5426	int __kmp_get_first_osid_with_ecore(void) {
5427	int low = 0;
5428	int high = __kmp_topology->get_num_hw_threads() - 1;
5429	int mid = 0;
5430	while (high - low > 1) {
5431	mid = (high + low) / 2;
5432	if (__kmp_topology->at(index: mid).attrs.get_core_type() ==
5433	KMP_HW_CORE_TYPE_CORE) {
5434	low = mid + 1;
5435	} else {
5436	high = mid;
5437	}
5438	}
5439	if (__kmp_topology->at(index: mid).attrs.get_core_type() == KMP_HW_CORE_TYPE_ATOM) {
5440	return mid;
5441	}
5442	return -1;
5443	}
5444	#endif
5445
5446	// Dynamic affinity settings - Affinity balanced
5447	void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
5448	KMP_DEBUG_ASSERT(th);
5449	bool fine_gran = true;
5450	int tid = th->th.th_info.ds.ds_tid;
5451	const char *env_var = "KMP_AFFINITY";
5452
5453	// Do not perform balanced affinity for the hidden helper threads
5454	if (KMP_HIDDEN_HELPER_THREAD(__kmp_gtid_from_thread(th)))
5455	return;
5456
5457	switch (__kmp_affinity.gran) {
5458	case KMP_HW_THREAD:
5459	break;
5460	case KMP_HW_CORE:
5461	if (__kmp_nThreadsPerCore > 1) {
5462	fine_gran = false;
5463	}
5464	break;
5465	case KMP_HW_SOCKET:
5466	if (nCoresPerPkg > 1) {
5467	fine_gran = false;
5468	}
5469	break;
5470	default:
5471	fine_gran = false;
5472	}
5473
5474	if (__kmp_topology->is_uniform()) {
5475	int coreID;
5476	int threadID;
5477	// Number of hyper threads per core in HT machine
5478	int __kmp_nth_per_core = __kmp_avail_proc / __kmp_ncores;
5479	// Number of cores
5480	int ncores = __kmp_ncores;
5481	if ((nPackages > 1) && (__kmp_nth_per_core <= 1)) {
5482	__kmp_nth_per_core = __kmp_avail_proc / nPackages;
5483	ncores = nPackages;
5484	}
5485	// How many threads will be bound to each core
5486	int chunk = nthreads / ncores;
5487	// How many cores will have an additional thread bound to it - "big cores"
5488	int big_cores = nthreads % ncores;
5489	// Number of threads on the big cores
5490	int big_nth = (chunk + 1) * big_cores;
5491	if (tid < big_nth) {
5492	coreID = tid / (chunk + 1);
5493	threadID = (tid % (chunk + 1)) % __kmp_nth_per_core;
5494	} else { // tid >= big_nth
5495	coreID = (tid - big_cores) / chunk;
5496	threadID = ((tid - big_cores) % chunk) % __kmp_nth_per_core;
5497	}
5498	KMP_DEBUG_ASSERT2(KMP_AFFINITY_CAPABLE(),
5499	"Illegal set affinity operation when not capable");
5500
5501	kmp_affin_mask_t *mask = th->th.th_affin_mask;
5502	KMP_CPU_ZERO(mask);
5503
5504	if (fine_gran) {
5505	int osID =
5506	__kmp_topology->at(index: coreID * __kmp_nth_per_core + threadID).os_id;
5507	KMP_CPU_SET(osID, mask);
5508	} else {
5509	for (int i = 0; i < __kmp_nth_per_core; i++) {
5510	int osID;
5511	osID = __kmp_topology->at(index: coreID * __kmp_nth_per_core + i).os_id;
5512	KMP_CPU_SET(osID, mask);
5513	}
5514	}
5515	if (__kmp_affinity.flags.verbose) {
5516	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5517	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5518	KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5519	tid, buf);
5520	}
5521	__kmp_affinity_get_thread_topology_info(th);
5522	__kmp_set_system_affinity(mask, TRUE);
5523	} else { // Non-uniform topology
5524
5525	kmp_affin_mask_t *mask = th->th.th_affin_mask;
5526	KMP_CPU_ZERO(mask);
5527
5528	int core_level =
5529	__kmp_affinity_find_core_level(nprocs: __kmp_avail_proc, bottom_level: __kmp_aff_depth - 1);
5530	int ncores = __kmp_affinity_compute_ncores(nprocs: __kmp_avail_proc,
5531	bottom_level: __kmp_aff_depth - 1, core_level);
5532	int nth_per_core = __kmp_affinity_max_proc_per_core(
5533	nprocs: __kmp_avail_proc, bottom_level: __kmp_aff_depth - 1, core_level);
5534
5535	// For performance gain consider the special case nthreads ==
5536	// __kmp_avail_proc
5537	if (nthreads == __kmp_avail_proc) {
5538	if (fine_gran) {
5539	int osID = __kmp_topology->at(index: tid).os_id;
5540	KMP_CPU_SET(osID, mask);
5541	} else {
5542	int core =
5543	__kmp_affinity_find_core(proc: tid, bottom_level: __kmp_aff_depth - 1, core_level);
5544	for (int i = 0; i < __kmp_avail_proc; i++) {
5545	int osID = __kmp_topology->at(index: i).os_id;
5546	if (__kmp_affinity_find_core(proc: i, bottom_level: __kmp_aff_depth - 1, core_level) ==
5547	core) {
5548	KMP_CPU_SET(osID, mask);
5549	}
5550	}
5551	}
5552	} else if (nthreads <= ncores) {
5553
5554	int core = 0;
5555	for (int i = 0; i < ncores; i++) {
5556	// Check if this core from procarr[] is in the mask
5557	int in_mask = 0;
5558	for (int j = 0; j < nth_per_core; j++) {
5559	if (procarr[i * nth_per_core + j] != -1) {
5560	in_mask = 1;
5561	break;
5562	}
5563	}
5564	if (in_mask) {
5565	if (tid == core) {
5566	for (int j = 0; j < nth_per_core; j++) {
5567	int osID = procarr[i * nth_per_core + j];
5568	if (osID != -1) {
5569	KMP_CPU_SET(osID, mask);
5570	// For fine granularity it is enough to set the first available
5571	// osID for this core
5572	if (fine_gran) {
5573	break;
5574	}
5575	}
5576	}
5577	break;
5578	} else {
5579	core++;
5580	}
5581	}
5582	}
5583	} else { // nthreads > ncores
5584	// Array to save the number of processors at each core
5585	int *nproc_at_core = (int *)KMP_ALLOCA(sizeof(int) * ncores);
5586	// Array to save the number of cores with "x" available processors;
5587	int *ncores_with_x_procs =
5588	(int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5589	// Array to save the number of cores with # procs from x to nth_per_core
5590	int *ncores_with_x_to_max_procs =
5591	(int *)KMP_ALLOCA(sizeof(int) * (nth_per_core + 1));
5592
5593	for (int i = 0; i <= nth_per_core; i++) {
5594	ncores_with_x_procs[i] = 0;
5595	ncores_with_x_to_max_procs[i] = 0;
5596	}
5597
5598	for (int i = 0; i < ncores; i++) {
5599	int cnt = 0;
5600	for (int j = 0; j < nth_per_core; j++) {
5601	if (procarr[i * nth_per_core + j] != -1) {
5602	cnt++;
5603	}
5604	}
5605	nproc_at_core[i] = cnt;
5606	ncores_with_x_procs[cnt]++;
5607	}
5608
5609	for (int i = 0; i <= nth_per_core; i++) {
5610	for (int j = i; j <= nth_per_core; j++) {
5611	ncores_with_x_to_max_procs[i] += ncores_with_x_procs[j];
5612	}
5613	}
5614
5615	// Max number of processors
5616	int nproc = nth_per_core * ncores;
5617	// An array to keep number of threads per each context
5618	int *newarr = (int *)__kmp_allocate(sizeof(int) * nproc);
5619	for (int i = 0; i < nproc; i++) {
5620	newarr[i] = 0;
5621	}
5622
5623	int nth = nthreads;
5624	int flag = 0;
5625	while (nth > 0) {
5626	for (int j = 1; j <= nth_per_core; j++) {
5627	int cnt = ncores_with_x_to_max_procs[j];
5628	for (int i = 0; i < ncores; i++) {
5629	// Skip the core with 0 processors
5630	if (nproc_at_core[i] == 0) {
5631	continue;
5632	}
5633	for (int k = 0; k < nth_per_core; k++) {
5634	if (procarr[i * nth_per_core + k] != -1) {
5635	if (newarr[i * nth_per_core + k] == 0) {
5636	newarr[i * nth_per_core + k] = 1;
5637	cnt--;
5638	nth--;
5639	break;
5640	} else {
5641	if (flag != 0) {
5642	newarr[i * nth_per_core + k]++;
5643	cnt--;
5644	nth--;
5645	break;
5646	}
5647	}
5648	}
5649	}
5650	if (cnt == 0 || nth == 0) {
5651	break;
5652	}
5653	}
5654	if (nth == 0) {
5655	break;
5656	}
5657	}
5658	flag = 1;
5659	}
5660	int sum = 0;
5661	for (int i = 0; i < nproc; i++) {
5662	sum += newarr[i];
5663	if (sum > tid) {
5664	if (fine_gran) {
5665	int osID = procarr[i];
5666	KMP_CPU_SET(osID, mask);
5667	} else {
5668	int coreID = i / nth_per_core;
5669	for (int ii = 0; ii < nth_per_core; ii++) {
5670	int osID = procarr[coreID * nth_per_core + ii];
5671	if (osID != -1) {
5672	KMP_CPU_SET(osID, mask);
5673	}
5674	}
5675	}
5676	break;
5677	}
5678	}
5679	__kmp_free(newarr);
5680	}
5681
5682	if (__kmp_affinity.flags.verbose) {
5683	char buf[KMP_AFFIN_MASK_PRINT_LEN];
5684	__kmp_affinity_print_mask(buf, KMP_AFFIN_MASK_PRINT_LEN, mask);
5685	KMP_INFORM(BoundToOSProcSet, env_var, (kmp_int32)getpid(), __kmp_gettid(),
5686	tid, buf);
5687	}
5688	__kmp_affinity_get_thread_topology_info(th);
5689	__kmp_set_system_affinity(mask, TRUE);
5690	}
5691	}
5692
5693	#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
5694	KMP_OS_AIX
5695	// We don't need this entry for Windows because
5696	// there is GetProcessAffinityMask() api
5697	//
5698	// The intended usage is indicated by these steps:
5699	// 1) The user gets the current affinity mask
5700	// 2) Then sets the affinity by calling this function
5701	// 3) Error check the return value
5702	// 4) Use non-OpenMP parallelization
5703	// 5) Reset the affinity to what was stored in step 1)
5704	#ifdef __cplusplus
5705	extern "C"
5706	#endif
5707	int
5708	kmp_set_thread_affinity_mask_initial()
5709	// the function returns 0 on success,
5710	// -1 if we cannot bind thread
5711	// >0 (errno) if an error happened during binding
5712	{
5713	int gtid = __kmp_get_gtid();
5714	if (gtid < 0) {
5715	// Do not touch non-omp threads
5716	KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5717	"non-omp thread, returning\n"));
5718	return -1;
5719	}
5720	if (!KMP_AFFINITY_CAPABLE() || !__kmp_init_middle) {
5721	KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5722	"affinity not initialized, returning\n"));
5723	return -1;
5724	}
5725	KA_TRACE(30, ("kmp_set_thread_affinity_mask_initial: "
5726	"set full mask for thread %d\n",
5727	gtid));
5728	KMP_DEBUG_ASSERT(__kmp_affin_fullMask != NULL);
5729	#if KMP_OS_AIX
5730	return bindprocessor(BINDTHREAD, thread_self(), PROCESSOR_CLASS_ANY);
5731	#else
5732	return __kmp_set_system_affinity(__kmp_affin_fullMask, FALSE);
5733	#endif
5734	}
5735	#endif
5736
5737	#endif // KMP_AFFINITY_SUPPORTED
5738