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